Design of the Adaptive Cruise Control Systems: An Optimal Control Approach

by

Sanggyum Kim

A dissertation submitted in partial satisfaction of the

requirements for the degree of
Doctor of Philosophy

in

Mechanical Engineering

in the

Graduate Division

of the

University of California, Berkeley

Committee in charge:

Professor Masayoshi Tomizuka, Chair

Professor J. Karl Hedrick
Professor John Strain

Spring 2012
Design of the Adaptive Cruise Control Systems: An Optimal Control Approach

Copyright 2012
by
Sanggyum Kim
 1

Abstract

Design of the Adaptive Cruise Control Systems: An Optimal Control Approach

by
Sanggyum Kim
Doctor of Philosophy in Mechanical Engineering
University of California, Berkeley
Professor Masayoshi Tomizuka, Chair

Modern automobiles are equipped with various driver assistance functions which enhance
safety and relieve driver fatigue. With the recent development of sensor technology, the
Adaptive Cruise Control (ACC) system has been put into practice. This thesis investigates
several aspects for the ACC system including (1) smooth reaction of the host vehicle to the
cutting in and out of lead vehicles, (2) real-time optimal profile generation for stop-and-go
motions, (3) optimal feedback controller design, and (4) extension to Cooperative Adaptive
Cruise Control (CACC) systems.
The ACC system should maintain an appropriate relative distance to the lead vehicle and
should also maintain the desired speed set by the driver if there is no lead vehicle or if the
speed of the lead vehicle is faster than the desired speed. Also, it should react smoothly when
the lead vehicle cuts out or if a new lead vehicle cuts in from a side lane. This thesis introduces
the virtual lead vehicle scheme to prevent the switching between the distance control and the
speed control. By controlling the motion of the virtual lead vehicle to be smooth, the scheme
could provide smooth reaction of the host vehicle to the cutting in and out of lead vehicles.
Linear Quadratic (LQ) optimal control scheme is utilized to find the control gains for the
virtual lead vehicle and the host vehicle. Variable weights are utilized in LQ for the virtual
lead vehicle. With the variable weights, the motion of the virtual lead vehicle is controlled
to be smooth when there is no safety threat while ensuring that the virtual lead vehicel is
still responsive and fast when a dangerous situation occurs. ACC with Stop-and-Go and
the Cooperative Adaptive Cruise Control (CACC) system are extensions of the conventional
ACC system. Stop-and-Go system is targeted to be used in urban driving situation where
the lead vehicle can stop completely. In that case, the Stop-and-Go system should have a
capability to stop the host vehicle completely. The constant time-headway policy used to
find the appropriate relative distance causes undesirable motion for a complete stop. In this
thesis a sliding controller is utilized to control the complete stopping motion. To find the
optimal stopping trajectory, a constrained Quadratic Programming (QP) problem is solved.
A constrained QP is also used to find the optimal velocity profile when the stopped vehicle
is to resume motion. Multi-resolution formulations and the Lemke algorithm are utilized to
 2

find the optimal trajectories in real time. The CACC system utilizes wireless communication
so that the vehicles in the network can share information with other vehicles. In this thesis,
a centralized controller is designed by LQ optimal control scheme and potential benefits
and problems are addressed. A Kalman filter with variable measurement noise covariance
is introduced to compensate the lost data through the wireless network associated with the
CACC system. The proposed control schemes have been verified through simulations.
 i

To my family
 ii

Contents

Contents ii

List of Figures v

List of Tables ix

1 Introduction 1
1.1 Adaptive Cruise Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Performance Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Control System Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Thesis Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Vehicle Model and Baseline Controller 7

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Vehicle Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.1 Longitudinal Vehicle Body Dynamics . . . . . . . . . . . . . . . . . . 9
2.2.2 Powertrain and Engine Model . . . . . . . . . . . . . . . . . . . . . . 11
2.2.3 Brake Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3 Model Based Disturbance Compensator . . . . . . . . . . . . . . . . . . . . . 15
2.3.1 Recursive Least Square Method . . . . . . . . . . . . . . . . . . . . . 15
2.3.2 Disturbance Compensator . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3.3 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4 Feedback Controller for Vehicle Following . . . . . . . . . . . . . . . . . . . . 23
2.4.1 Desired Distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.4.2 Tuning of Feedback Gains . . . . . . . . . . . . . . . . . . . . . . . . 24
2.4.3 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3 Virtual Lead Vehicle Scheme 29

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2 Concept of the Virtual Lead Vehicle Scheme . . . . . . . . . . . . . . . . . . 30
3.2.1 Mode Decision of the ACC system . . . . . . . . . . . . . . . . . . . 31
 iii

3.2.2 Mode Switching Scheme . . . . . . . . . . . . . . . . . . . . . . . . . 32

3.2.3 Virtual Lead Vehicle Scheme . . . . . . . . . . . . . . . . . . . . . . . 32
3.3 Control of the Virtual Lead Vehicle in Transient . . . . . . . . . . . . . . . . 32
3.3.1 Double Integrator Model for the Virtual Lead Vehicle . . . . . . . . . 34
3.3.2 Linear Quadratic Control . . . . . . . . . . . . . . . . . . . . . . . . 35
3.3.3 Comparison with the Model Predictive Control . . . . . . . . . . . . 36
3.4 Linear Quadratic Control With Variable Weights . . . . . . . . . . . . . . . 36
3.4.1 Tuning of the LQ Weights . . . . . . . . . . . . . . . . . . . . . . . . 37
3.4.2 Simulation Results with Constant Weights . . . . . . . . . . . . . . . 37
3.4.3 LQ with Variable Weights . . . . . . . . . . . . . . . . . . . . . . . . 38
3.4.4 Simulation Results with Variable Weights . . . . . . . . . . . . . . . 39
3.5 Stability Analysis of the Linear Quadratic Controller with Variable Weights . 40
3.6 Disturbance Observer for the Virtual Lead Vehicle . . . . . . . . . . . . . . . 46
3.6.1 Model of the Virtual Lead Vehicle . . . . . . . . . . . . . . . . . . . . 47
3.6.2 Disturbance Observer . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.6.3 Utilizing the Lead Vehicle Acceleration . . . . . . . . . . . . . . . . . 55
3.6.4 Extended State DOB . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

4 Optimal Profile Generation for Stop-and-Go 61

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.2 Optimal Relative Distance Profile during a Complete Stop . . . . . . . . . . 63
4.2.1 Optimization Problem Formulation . . . . . . . . . . . . . . . . . . . 63
4.2.2 Velocity Profile of a Starting Motion . . . . . . . . . . . . . . . . . . 67
4.3 Sliding Control and Virtual Lead Vehicle Scheme . . . . . . . . . . . . . . . 68
4.3.1 Sliding Controller Design for a Complete Stop . . . . . . . . . . . . . 68
4.3.2 Simulation Results for a Complete stop . . . . . . . . . . . . . . . . . 69
4.3.3 Virtual Lead Vehicle for a Starting Motion . . . . . . . . . . . . . . . 71
4.4 Real Time Calculation of Quadratic Programming Problem . . . . . . . . . . 74
4.4.1 Algorithms to Solve the Constrained QP . . . . . . . . . . . . . . . . 76
4.4.2 Multi-resolution Optimization Problem Formulation . . . . . . . . . . 81
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

5 Cooperative Adaptive Cruise Control 87

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.2 Linear Quadratic Optimal Controller for a Vehicle Platoon . . . . . . . . . . 88
5.2.1 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.2.2 Linear Quadratic Optimal Control . . . . . . . . . . . . . . . . . . . 89
5.2.3 Simulation Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.3 Cut In/Out of a Vehicle in a Platoon . . . . . . . . . . . . . . . . . . . . . . 91
5.3.1 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.3.2 Virtual Lead Vehicle Scheme . . . . . . . . . . . . . . . . . . . . . . 94
 iv

5.3.3 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

5.4 CACC in Lossy Network Condition . . . . . . . . . . . . . . . . . . . . . . . 100
5.4.1 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.4.2 Control Schemes for Dropped Data . . . . . . . . . . . . . . . . . . . 101
5.4.3 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.4.4 Estimation with Kalman Filter . . . . . . . . . . . . . . . . . . . . . 105
5.4.5 Simulation Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.5 Limitation of the Centralized Controller . . . . . . . . . . . . . . . . . . . . 112
5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

6 Concluding Remarks and Open Issues 114

6.1 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6.1.1 Smooth Reaction to Cutting In and Out . . . . . . . . . . . . . . . . 114
6.1.2 Real-time Optimal Profile Generation for Stop-and-Go . . . . . . . . 114
6.1.3 Optimal Feedback Controller Design . . . . . . . . . . . . . . . . . . 115
6.1.4 Cooperative Adaptive Cruise Control . . . . . . . . . . . . . . . . . . 115
6.2 Open Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.2.1 Lower Level Control and Brake/Engine Switching Logic . . . . . . . . 115
6.2.2 Mixed Traffic CACC . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.2.3 Implementation and Verification . . . . . . . . . . . . . . . . . . . . . 116

Bibliography 117
 v

List of Figures

1.1 Overall ACC system structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.1 Free body diagram of the longitudinal vehicle model . . . . . . . . . . . . . . . 9

2.2 Stopping a vehicle with a constant brake torque: Experimental results vs sim-
ulation results based on vehicle model with DDT tire model (a) longitudinal
acceleration; (b) pitch rate [22] . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Powertrain model block diagram . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4 Braking System [13] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.5 Simulation results with fuzzy logic: brake torque . . . . . . . . . . . . . . . . . . 14
2.6 Simulation result with g1 = g2 = g3 = 1 . . . . . . . . . . . . . . . . . . . . . . . 20
2.7 Simulation result with g1 = g2 = 0.1, g3 = 1 . . . . . . . . . . . . . . . . . . . . 20
2.8 Simulation result where there is not enough excitations: with a vector-type for-
getting factor without the constant trace method . . . . . . . . . . . . . . . . . 21
2.9 Simulation result where there is not enough excitations: with a vector-type for-
getting factor and a constant trace method . . . . . . . . . . . . . . . . . . . . . 21
2.10 Simulation results with and without the disturbance compensator . . . . . . . . 22
2.11 Simulation results with the PID and PD controllers . . . . . . . . . . . . . . . . 27
2.12 Simulation results of a vehicle platoon with the PID controller . . . . . . . . . . 27

3.1 Mode switching logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3.2 Block diagram of the mode switching scheme . . . . . . . . . . . . . . . . . . . . 33
3.3 Concept of the virtual lead vehicle scheme . . . . . . . . . . . . . . . . . . . . . 33
3.4 Block diagram of the virtual lead vehicle scheme . . . . . . . . . . . . . . . . . . 34
3.5 Simulation results with MPC and LQ control: Position and velocity of the virtual
lead vehicle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.6 The symmetric root locus - Arrows indicate decreasing a . . . . . . . . . . . . 38
3.7 Simulation results with constant weights position and velocity of the virtual lead
vehicle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.8 Simulation results when a lead vehicle cuts out position and velocity of the
virtual lead vehicle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.9 Simulation results when a lead vehicle cuts out variable weights . . . . . . . . . 41
 vi

3.10 Simulation results when a lead vehicle cuts out position and velocity of the host
vehicle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.11 Simulation results when a lead vehicle cuts in position and velocity of the virtual
lead vehicle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.12 Simulation results when a lead vehicle cuts in variable weights . . . . . . . . . . 43
3.13 Simulation results when a lead vehicle cuts out position and velocity of the host
vehicle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.14 Phase plane plot of the closed loop system . . . . . . . . . . . . . . . . . . . . . 44
3.15 Phase plane plot of the closed loop system at each quadrant comparing the worst
case gains and the original gains . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.16 Phase plane plot of the closed loop system with the worst case gains . . . . . . . 46
3.17 Block diagram of the speed and the position errors between the virtual and the
actual lead vehicles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.18 Block diagram of the DOB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.19 Simulation results with and without the DOB when the lead vehicle accelerates:
velocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.20 Simulation results with and without the DOB when the lead vehicle accelerates:
position error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.21 Simulation results with and without the DOB when the lead vehicle decelerates:
velocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.22 Simulation results with and without the DOB when the lead vehicle decelerates:
position error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.23 Block diagram of the DOB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.24 Simulation with the measurement noise . . . . . . . . . . . . . . . . . . . . . . . 51
3.25 Simulation results with measurement noise using the velocity measurement: velocity 52
3.26 Simulation results with measurement noise using the velocity measurement: po-
sition error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.27 Simulation results with measurement noise using the velocity measurement: ac-
celeration of the virtual lead vehicle . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.28 Simulation results with a Q-filter: velocity . . . . . . . . . . . . . . . . . . . . . 54
3.29 Simulation results with a Q-filter: acceleration of the virtual lead vehicle . . . . 54
3.30 Simulation results showing the vehicle behavior with and without the feedforward
control when the lead vehicle decelerates: the lead and virtual vehicle‘s velocity 55
3.31 Simulation results showing the vehicle behavior with and without the feedforward
control when the lead vehicle decelerates: the host vehicle‘s position . . . . . . . 56
3.32 Simulation results showing the vehicle behavior with and without the feedforward
control when the lead vehicle accelerates: the lead and virtual vehicle‘s velocity 56
3.33 Simulation results showing the vehicle behavior with and without the feedforward
control when the lead vehicle accelerates: the host vehicle‘s position . . . . . . . 57

4.1 Velocity of the lead and the host vehicle in case of a complete stop . . . . . . . 62
4.2 Sketch of the complete stop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
 vii

4.3 Solution of the QP comparing l 1 and l 2 norms, x = 1, a = 1, j = 0.5 . . 66

4.4 Solution of the QP comparing x = 1 and x = 100, a = 1, j = 0.5 . . . . . . . 67
4.5 Optimal profile generation result for starting motion . . . . . . . . . . . . . . . 68
4.6 Simulation results with the sliding mode controller: relative distance . . . . . . 70
4.7 Simulation results with the sliding mode controller: acceleration and control com-
mand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.8 Simulation results with the smooth sliding mode controller: relative distance . . 71
4.9 Simulation results with the smooth sliding mode controller: acceleration and
control command . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.10 Simulation result for a starting motion with a sliding mode controller . . . . . . 73
4.11 Simulation result with a virtual lead vehicle utilizing the optimal velocity profile 74
4.12 Simulation result with a virtual lead vehicle and a slow lead vehicle: velocity . . 75
4.13 Simulation result with a virtual lead vehicle and a slow lead vehicle: distance error 75
4.14 Intermediate and final solutions of the Lemke algorithm . . . . . . . . . . . . . . 80
4.15 Intermediate and final solutions of the interior point algorithm . . . . . . . . . . 80
4.16 Sketch of the multi-resolution problem concept . . . . . . . . . . . . . . . . . . . 82
4.17 Solutions of the optimization problems with uniform time step and multi-resolution
time step . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.18 Concept of Multi-resolution formulation to calculate the optimal trajectory for
the sliding control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.19 Solutions of the optimization problems with uniform time step and multi-resolution
time step . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.20 Simulation result with a sliding controller in case of a complete stop: velocity
[m/sec] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.21 Simulation result with a sliding controller in case of a complete stop: position [m] 85

5.1 Simulation results with LQ controllers: velocity (solid lines: with CACC, dashed
lines: with ACC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.2 Simulation results with LQ controllers: error (solid lines: with CACC, dashed
lines: with ACC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.3 Simulation results with LQ controllers: product of the error and the lead vehicle
acceleration (solid lines: with CACC, dashed lines: with ACC) . . . . . . . . . . 92
5.4 Simulation results when a vehicle cuts out from the platoon . . . . . . . . . . . 93
5.5 Sketch of the cutting out scenario . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.6 Simulation results when a vehicle cuts out from the platoon (Solid line: Actual
vehicle, Dashed line: Virtual lead vehicle . . . . . . . . . . . . . . . . . . . . . . 97
5.7 Simulation results when two vehicles cut out from the platoon (Solid line: Actual
vehicle, Dashed line: Virtual lead vehicle . . . . . . . . . . . . . . . . . . . . . . 97
5.8 Simulation results when two vehicles cut out from the platoon (Solid line: Actual
vehicle, Dashed line: Virtual lead vehicle . . . . . . . . . . . . . . . . . . . . . . 98
5.9 Sketch of the cutting in scenario . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
 viii

5.10 Simulation results without the virtual lead vehicle when a vehicle cuts into the
platoon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.11 Simulation results when a vehicle cuts into the platoon (Solid line: Actual vehicle,
Dashed line: Virtual lead vehicle . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.12 Simulation results when a vehicle cuts into the platoon (Solid line: Actual vehicle,
Dashed line: Virtual lead vehicle . . . . . . . . . . . . . . . . . . . . . . . . . . 100
5.13 Simulation results over lossy network: error for vehicle #1 . . . . . . . . . . . . 102
5.14 Simulation results over lossy network: error for vehicle #2 . . . . . . . . . . . . 103
5.15 Simulation results over lossy network: error for vehicle #3 . . . . . . . . . . . . 103
5.16 Simulation results over lossy network: error for vehicle #4 . . . . . . . . . . . . 104
5.17 Simulation results over lossy network: error for vehicle #5 . . . . . . . . . . . . 104
5.18 Simulation results over lossy network with ¯ = 0.50: error for vehicle #3 . . . . 105
5.19 Simulation results with Kalman Filter over lossy network: error for vehicle #1 . 109
5.20 Simulation results with Kalman Filter over lossy network: error for vehicle #2 . 109
5.21 Simulation results with Kalman Filter over lossy network: error for vehicle #3 . 110
5.22 Simulation results with Kalman Filter over lossy network: error for vehicle #4 . 110
5.23 Simulation results with Kalman Filter over lossy network: error for vehicle #5 . 111
5.24 Simulation results with Kalman Filter over lossy network with ¯ = 0.50: error
for vehicle #3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
 ix

List of Tables

2.1 Typical gear ratios of an automatic transmission . . . . . . . . . . . . . . . . . . 12

3.1 Desired performance and weights . . . . . . . . . . . . . . . . . . . . . . . . . . 40

3.2 Simulation results comparing the DOB and the Extended State DOB(ESDOB) . 59

4.1 Calculation time [sec] of the Lemke algorithm / Interior point algorithm (Initial
Condition : d0 [m],v0 [m/s],a0 [m/s2 ]) . . . . . . . . . . . . . . . . . . . . . . . . . 79
 x

Acknowledgments
Writing the acknowledgements is the very last and one of the most difficult part for writing
my dissertation. This dissertation may have not been done without the support and encour-
agement of many people. I would never be able to list all the people, however I want to
express my warmest thanks to everyone.
I would like to thank my advisor, Professor Masayoshi Tomizuka, for providing me ex-
cellent research opportunities. Without his insightful advice, persistent inspiration, and
considerate support, this dissertation would not exist. His inexhaustible academic passion
and energy always impress and motivate myself. I would also want to express my thanks to
Professor Miwako Tomizuka for her warm care to the entire lab and our families.
I want to acknowledge the superb instruction and support I have received from all of
my professors in Berkeley. Special thanks to Professor Karl Hedrick, Professor John Strain,
Professor Roberto Horowitz, and Professor Avideh Zakhor for their kind support as my
dissertation and/or qualifying exam committee. Their kind advices and valuable comments
really helped me improve the quality of my work.
To all my colleagues and friends in Mechanical Systems Control Laboratory, I appreciate
all of your support, discussion, conversation, advice, and friendship over the past years:
Wenjie, Pedro, Joonbum, Kyoungchul, Xu, Chi-shen, Raechel, Oak, Sandipan, Soo, Chun-
Chih, Emma, Hoday, Ben, Ahmed, Wenlong, Cong, Haifei, Nancy, Kiyo, Nora, Omar, Chen-
yu, Chung-yen, Junkai, Minghui, Hiroshi, and all others. Special thanks to Mike for his hard
work as my English editor. Thank you Evan, Kay, and Kenta for your kind love to my family
and being my neighbor. I also want to thank Dr. Shiang-Lung Ku, Dr. Takashi Nagata, Mr.
Yasuyuki Matzuda, and Mr. Sumio Sugita for their support and advice on my research.
I would like to thank the Industrial Technology Research Institute in Taiwan for their
kind financial support on my research project. Special thanks to Dr. Kuo-Hsiang Cheng for
his advice and discussions about the project for past years. Big thanks to my Korean friends
for sharing precious moments at Berkeley.
I really want to express my deepest love and respect to my brother Soonkyum. You
are the best friend, mentor, and brother in my life. I would like to especially thank to my
parents Youngki Kim and Kwangsook Baik for their endless love, support, and advice during
my entire life. Also, I appreciate the kind support of my parents-in-law Yoonhyun Park and
Samsoon Hong.
Finally, I want to say to my wife, Bo Ae Park that I love you and you are the best gift
in my life. Without out your love and dedication, I could not do it. Also, I want to express
my love to my children Jae Hee (Amy) and Jae Yoon (Kevin). You are the meaning of my
life. Lastly, to my family, I love you all.
 1

Chapter 1

Introduction

1.1 Adaptive Cruise Control

Since the very first modern automobile was invented by Karl Benz in 19th century, automo-
biles have became a↵ordable, largely due to improvements in mass production technologies,
and have been widely used all over the world. In 2009, more than 681 millions of cars
were registered in the world [10]. With automobiles, longer and faster travels can be easily
achieved and the quality of human life has been greatly enhanced.
Since the first automobile was invented many driver assistance features were introduced
to enhance the safety and to reduce the drivers’ fatigue. Many of these features are related
to the automation of the driving. One of the most popular automation features is the
cruise control system. The cruise control system takes over the throttle of the vehicle and
maintains the speed of the vehicle at the driver’s desired value. This feature is usually
utilized in highway driving conditions where speeds remain constant for prolonged periods
of time.
Various research and works have been done in both academia and industry to make
driving be more automatic. The Automated Highway System (AHS) was one of the most
popular works regarding the automation of vehicles. The California Partners for Advanced
Transit and Highways (PATH) program has one of the successful research projects on the
AHS. The project was focused on the highway driving of specially equipped vehicles in a
separate lane from other manually driven vehicles. Various research on the longitudinal and
lateral control of the vehicle body were performed in the project. The PATH AHS was
successfully demonstrated in tests done in San Diego, California in 1997 [54].
With the recent development in sensor technology, the Adaptive Cruise Control(ACC)
system has been introduced. The ACC system is an advanced version of the conventional
cruise control system. While the conventional cruise control system maintains the desired
speed set by the driver, the ACC system also maintains an appropriate relative distance
to the lead vehicle. If the lead vehicle is detected to be driven at a speed lower than the
desired speed, the system slows down the host vehicle and maintains an appropriate relative
CHAPTER 1. INTRODUCTION 2

distance, which usually depends on the states of the host vehicle and the environment.
The ACC system has extensions like stop-and-go and Cooperative Adaptive Cruise Con-
trol (CACC). Stop-and-go system is targeted to be used in urban driving conditions where
the ACC cannot be easily utilized. The stop-and-go should have the capability to stop the
host vehicle completely when there is a stopped lead vehicle [38]. Also in urban driving, the
environments are more complex and the requirements for the sensor information are more
strict [56]. Since it is expected to be utilizable at low speed in urban driving situations, the
stop-and-go should be able to properly react to frequent cutting in and out of lead vehicles.
Cooperative Adaptive Cruise Control(CACC) is another enhanced version of the ACC.
CACC utilizes communication between the vehicles and/or between the vehicle and the
road structure. This communication allows the control system on a single vehicle to get
information about other vehicles in the platoon such as the acceleration, the velocity, the
throttle control command, and the brake control command. Various research have been
done about the CACC in the area of traffic control and communication which show that the
CACC has the potential to improve the traffic flow [3] and the string stability [33]. For a
vehicle platoon, the string stability signifies whether spacing errors are increased upstream.
A sudden deceleration or acceleration of the first vehicle of a platoon introduces errors for
following vehicles. If the platoon is string unstable, the magnitudes of the errors are increased
for later vehicles.

1.2 Performance Requirements

Stability
Stability is the most important feature for all control systems. Improper feedback controllers
can make stable systems become unstable. In case of ACC systems, stability means that
the speed of the host vehicle and the relative distance to the lead vehicle converge to their
desired values. In addition to the stability of a single vehicle, string stability should be
achieved for a platoon of vehicles. If not, it can cause a problem for the case of a platoon
of vehicles equipped with identical ACC systems. If the feedback controller is not properly
designed, when the very first vehicle of the platoon accelerates or decelerates, the distance
errors increases along the vehicle stream. To guarantee string stability, the desired relative
distance should be properly determined [42] or vehicle-to-vehicle communication is required
[47].

Distance Control & Speed Control

ACC systems should be able to maintain an appropriate relative distance to the lead vehicle.
However, if the lead vehicle is moving faster than the desired speed or no lead vehicle is
detected by sensors, ACC systems should maintain the desired speed set by the driver like
the conventional cruise control system. It is important that not only both distance controller
and speed controller should operate properly and stably, but also the transient performance
CHAPTER 1. INTRODUCTION 3

should be smooth and safe when the control scheme is switched from one to the other. The
transient motion is important for urban driving where there are frequent transitions. For
example, when a lead vehicle cuts out, the relative distance to the new lead vehicle may be
very large so that the speed control is activated without an appropriate transient controller.
In that case, the host vehicle accelerates unnecessarily, which makes the driver uncomfortable
and is not fuel efficient.

Passenger Comfort
Accelerations and decelerations are also important parameters in ACC systems, because
they are closely related to the comfort of passengers. Accelerations and decelerations are
unavoidable during the operation of the ACC system and excessive accelerations or deceler-
ations make passengers feel uncomfortable and uneasy [51]. Therefore, an ACC system must
perform so that acceleration and deceleration remain within a certain range. In addition to
acceleration and deceleration, jerk, the time derivative of the acceleration, can be used as
a parameter to measure the variation of acceleration and deceleration and should be main-
tained as small as possible to ensure the comfort of passengers. Also, the acceleration or
deceleration of the host vehicle should not be the opposite from the drivers’ expectation. For
example, if the vehicle accelerates when the driver expect it to decelerate the driver would
feel very uncomfortable with the ACC system.

Safety
While the comfort of passengers is important, it should not disrupt safety. In the case of
an emergency stop, safety is the highest priority. So, the optimal profile of acceleration and
deceleration should be determined according to the situation. For example, in the case of an
emergency stop, the host vehicle should decelerate quickly to minimize the stopping distance
whereas in non-emergency stopping situations, the host vehicle should decelerate slowly so
that passengers do not feel discomfort.

1.3 Control System Structure

Figure 1.1 depicts the overall ACC system structure, which consists of the driver, the upper
level controller, the lower level controllers, the host vehicle dynamics, and the environment.
The driver decides whether the ACC system is turned on or o↵. The desired velocity and the
desired time headway are determined by the preference of the driver. The desired velocity
refers to the target velocity of the host vehicle when there is no vehicle or obstacle in front.
The time headway represents the amount of time after which the lead vehicle and the host
vehicle will collide when the lead vehicle suddenly stops and the host vehicle maintains its
original velocity. The time headway determines the desired relative distance to the lead
vehicle, and will be examined in Chapter 2.
CHAPTER 1. INTRODUCTION 4

Observer&
Angular%Velocity,%AcceleraAon%
EsAmate%
unmeasurable%
parameters% Environment&

EsAmated% Road%Slope%
%Parameters% Surface%CondiAon%
Aerodynamic%Resistance%

Upper&Level& Lower&Level&
Driver& Host&Vehicle&
Controller& Controllers&
On/Off% Distance%Control% Engine% Vehicle%Dynamics%
Desired%Velocity% Speed%Control% Brake% Tire%Dynamics%
Time%Headway% Stop%&%Go% Transmission%

Angular%Velocity,%AcceleraAon%
Control%Command%
Force% Lead&
Sensors&
InformaAon% RelaAve%Velocity% Vehicle&
%RelaAve%Distance%

Figure 1.1: Overall ACC system structure

The control system is composed of an observer, an upper level controller, and lower level
controllers. The observer directly utilizes measured feedback signals, such as angular veloci-
ties of each wheel and acceleration, to estimate unmeasurable parameters such as the vehicle
mass, the aerodynamic coefficient, and the road slope. Information about the lead vehicle
are also measured by sensors such as the radar and vision sensors. With measured feedback
signals, all this information are utilized in the upper controller. The upper level controller
analyses the information and decides the desired states of the host vehicle and produces the
desired acceleration of the host vehicle accordingly. The desired acceleration commands from
the upper level controller manage the lower level controllers such as the engine control unit,
the transmission controller and the brake controller. However, the vehicle may not behave
as predicted by the controller because there are uncertain environmental parameters such as
the road slope, the rolling resistance, and the aerodynamic resistance as well as uncertainties
in the vehicle dynamics. To overcome such uncertainties the performance of the vehicle is
measured and provided to the control system as feedback signals.
CHAPTER 1. INTRODUCTION 5

1.4 Thesis Overview

To address the requirements of the ACC system, several control schemes are suggested and
studied in this thesis. This thesis is organized as follows.

Chapter 2: Vehicle Model and Baseline Controller

The main purpose of this thesis is designing the upper level controller of the ACC system.
However, other elements of the system should be reliable to e↵ectively verify the performance
of the designed upper level controller. In this thesis, the performance of the suggested con-
troller is evaluated and verified via simulations. To do that, reliable and not too complicated
models of vehicles are necessary.
In Chapter 2, the dynamics models used in simulations are introduced. A vehicle model
including the longitudinal vehicle body dynamics, an engine model, a tire model, and a
brake model are introduced. An observer which utilizes a recursive least square method is
introduced to estimate the vehicle parameters and external disturbances. A model based
disturbance compensator is used to compensate the external disturbances. As a basic upper
level controller for distance control, a Linear Quadratic (LQ) optimal controller is studied
to find optimal feedback control gains.

Chapter 3: Virtual Lead Vehicle Scheme

The main task of the ACC system is to maintain an appropriate relative distance to the lead
vehicle. However when no lead vehicle is detected or the lead vehicle is moving faster than
the desired speed, the host vehicle should maintain the desired speed set by the driver. The
switching between the distance control and the speed control should be stable and smooth.
To achieve that, the virtual lead vehicle scheme is introduced.
In Chapter 3, the modes of the ACC system and the concept of the virtual lead vehicle
scheme are introduced. An LQ controller is designed to control the virtual lead vehicle.
Variable weights are utilized for the LQ controller to better control the virtual lead vehicle
during transient motion. The stability of the LQ controller with variable weights is analyzed.
A Disturbance Observer (DOB) is utilized to estimate and compensate the acceleration of
the lead vehicle.

Chapter 4: Optimal Profile Generation for Stop-and-Go

While the ACC is intended to assist the driver at relatively higher speeds, the Stop-and-
Go is for low speeds and should have a capability to stop the vehicle completely if there is
a stopped lead vehicle [38]. When the lead vehicle is stopped, the constant time-headway
scheme, which determines the desired relative distance to the lead vehicle based on the speed
of the host vehicle, may cause very slow stopping motion. To prevent it, a separate control
scheme should be introduced when the lead vehicle stops completely.
CHAPTER 1. INTRODUCTION 6

In Chapter 4, optimization problems are formulated to find the optimal acceleration

profile during a complete stop and a starting motion. Once the optimal profiles are achieved,
a sliding mode controller is designed for a complete stop and the virtual lead vehicle scheme
is utilized for the starting motion. Algorithms solving the optimization problem are studied
and multi-resolution formulation is introduced to solve the problem in real time.

Chapter 5: Cooperative Adaptive Cruise Control

Cooperative Adaptive Cruise Control(CACC) is an enhanced version of the ACC. The main
di↵erence between CACC and ACC is that CACC utilizes communication between the vehi-
cles and/or between the vehicle and the road structure. This communication allows the ACC
system of a vehicle to get information about other vehicles in the platoon. Many previous
research show that the CACC has the potential to improve both the traffic flow [3] and the
string stability [33].
In Chapter 5, an LQ optimal control problem is solved for a vehicle platoon to find the
feedback control gains. The virtual lead vehicle scheme is also utilized in CACC to control
the platoon such that it will react smoothly when vehicles cut in or out from the platoon.
Since wireless networks used in the CACC are usually lossy, packet loss during wireless
communications is considered and various control scheme are compared by simulations.

Chapter 6: Concluding Remarks and Open Issues

In Chapter 6, concluding remarks of this thesis are summarized and some remaining issues
are discussed as future works.
 7

Chapter 2

Vehicle Model and Baseline Controller

2.1 Introduction
As described in Section 1.3, the ACC system consists of two control layers. The upper
level controller determines the desired acceleration of the host vehicle while the lower level
controllers determine the control commands to the engine and the brake based on the desired
acceleration and current states of the engine and the brake. The main purpose of this thesis
is to design the upper level controller. However, a reliable vehicle model and lower level
controllers are necessary to evaluate and verify the performance of the upper level controller
by simulations. In this chapter, a longitudinal vehicle model, an engine model, and a brake
model are introduced and studied. These models are implemented in MATLAB/Simulink
for simulations.
The longitudinal dynamic model of a vehicle body is based on the force equilibrium of
the external forces which include the tire forces, the rolling resistances, the aerodynamic
drag force, and the gravity force due to the road slope [42]. Among the external forces, the
tire force can be controller by the lower level controllers through the powertrain and the
brake. Due to the compliance of the tire, the tire force and the input torque to the driving
shaft have a dynamic relationship and it can be described by a tire model. In general, the
tire force depends on the slip ratio and many tire models relating the slip ratio and the tire
force have been studied. Very well known examples are the linear tire model and the magic
formula tire model. Both models give the relationships between the longitudinal tire force
and the slip ratio based on the empirical data, and describe the steady state behavior of the
tire [60, 39]. However, for the Stop-and-Go feature which may introduce complete stops, a
smooth motion control is important for the comfort of the driver and passengers. To predict
the oscillation during a complete stop, a dynamic deflection tire model [22] is used in the
simulation for this thesis. The engine and the brake models should be able to reproduce the
behaviors of the actual engine and the brake, however do not need to be very complicated.
In this thesis, an engine model and a lower level controller in [30, 31] are used. A brake
model studied in [13] are used for simulations in this thesis.
CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 8

Other external forces except the tire force are from environment. Considering the vehicle
body as a double integrator whose input is the acceleration and the outputs are the longi-
tudinal velocity and the longitudinal position, these external forces are disturbances which
should be e↵ectively compensated. Various research have been done to estimated the vehi-
cle parameters and the environmental disturbances including the vehicle mass, the tire-road
friction, and the road slope. Some research have been performed to estimate the vehicle mass
and the road slope via the Extended Kalman filtering [27], the Extended Kalman filtering
with Model Predictive Controller [59], and the Recursive Least Square method [57]. In the
previous research, the vehicle mass and the road slope are simultaneously estimated. It is
assumed, however, that the external disturbances such as the aerodynamic resistance and
the rolling resistance are known. Also, the torque generated by the engine or the brake is
assumed to be known or able to be estimated from a model. In this thesis it is assumed
that the Multi-Sensing-Hub (MSH) [52] is used to directly measure the tire force. With
the force measurement, a recursive least square method with a vector-type forgetting factor
with a constant trace is used to estimate not only the vehicle mass and the road slope but
also the aerodynamic coefficient. The estimated vehicle mass is utilized to calculate the
desired torque from the desired acceleration. The gravity force due to the road slope and
the aerodynamics resistance are compensated via a model based disturbance compensator.
Assuming perfect disturbance compensation and lower level controllers, the longitudinal
vehicle dynamics can be modeled as a double integrator whose input is the acceleration.
The desired acceleration is determined by an upper level feedback controller. To design
a feedback controller, stability is one of the most important thing to consider. In case
of ACC, there is a chance that the vehicles equipped with the same ACC system make a
platoon and string stability should also be considered. To prevent the string instability, a
constant time-headway method is used to calculate the desired relative distance to the lead
vehicle [42]. PID (Proportional-Integral-Di↵erential) controller is one of the most popular of
feedback controllers and various methods have been studied to tune the feedback gains [50,
17, 16]. In this chapter, the Linear Quadratic (LQ) optimal control theory is used to find
an optimal state feedback control gains when the host vehicle follows a lead vehicle with a
constant time-headway. The state feedback controller is equivalent to a PD (Proportional-
Di↵erential) controller for second order systems and the PID gains can be found by the
Linear Quadratic Controller with Integrator(LQI).
This chapter is organized as follows. A vehicle model including the longitudinal vehicle
body dynamics, an engine model, a tire model, and a brake model is introduced in Section
2.2. In Section 2.3, a recursive least square method is introduced to estimate the vehicle
parameters and external disturbances. A model based disturbance compensator is used to
compensate the external disturbances. In Section 2.4, an LQ optimal controller is studied
to find feedback control gains. Section 2.5 summarizes the chapter.
CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 9

!
Figure 2.1: Free body diagram of the longitudinal vehicle model

2.2 Vehicle Model

2.2.1 Longitudinal Vehicle Body Dynamics
This section presents a longitudinal dynamics model of the vehicle body. Referring to Figure
2.1, the governing equation of the longitudinal vehicle model is given by:

mẍ = Fxf + Fxr Faero Rxf Rxr mg sin ✓ (2.1)

where m is the mass of the vehicle, x is the longitudinal position of the vehicle, Fxf and Fxr
are the tire forces at front and rear tires, Faero is the aerodynamic resistance, Rxf and Rxr
are the rolling resistances at front and rear tires, and ✓ is the slope angle of the road.

Tire Force
Tire forces, Fxf and Fxr , are controlled by the engine and the brake torque, however are
generated via road-tire interaction. For this reason, the tire model is extremely important
in vehicle control. There are many tire models including many empirical models [39]. While
they have been successfully integrated with vehicle dynamics in many applications, they
are static models and are not suited for transient phenomena in particular at low speeds.
The Relaxation Length Tire (RLT) model was introduced to address this problem [29]. The
Dynamic Deflection Tire (DDT) model has been developed by S.L. Koo, et al for better
prediction and description of the tire characteristics, especially at low speeds [22]. In case of
Stop-and-Go, the longitudinal oscillation during deceleration and stopping with the brake,
which is related to the comfort for passengers, should be considered. Figure 2.2 which is
adopted from [22] shows that the longitudinal oscillation can be predicted by using the DDT
model.
CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 10

!
Figure 2.2: Stopping a vehicle with a constant brake torque: Experimental results vs simu-
lation results based on vehicle model with DDT tire model (a) longitudinal acceleration; (b)
pitch rate [22]

The front and rear tire forces can be derived from the DDT model. The governing
equation is given by:
Flong = Dlong ˙ x + Clong x (2.2)
where Dlong and Clong are the damping and the spring constant of the tire, respectively. x
represents the longitudinal tire deflection, which is defined by:

x = xcp xw (2.3)
where xcp is the longitudinal displacement of the contact patch and xw is the longitudinal
displacement of the wheel. xw is assumed to be the same with the displacement of the
vehicle.
The kinematic equation of xcp can be expressed as:
ẋcp = R0 (1 + )! (2.4)
where R0 is the e↵ective radius of the tire and ! is the tire angular velocity. The longitudinal
slip is defined by:
x
= (2.5)
long
where long is the tire relaxation length.

Torque-Tire Force Relation

The governing equation for the torque and tire force is given by:
Iw !˙ = ⌧ R0 Flong (2.6)
CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 11

where Iw is the inertia of the wheel.

Aerodynamic Resistance
The aerodynamic resistance on a vehicle can be represented by:
1
Faero = ⇢Cd Af |vx + vwind |(vx + vwind ) (2.7)
2
where ⇢ is the mass density of air, Cd is the aerodynamic draw coefficient, Af is the front
projected area of the vehicle, vx is the longitudinal velocity of the vehicle, and vwind is the
wind velocity.

Rolling Resistance
The rolling resistance of each tire is proportional to the normal force of the each tire: i.e.

Rx = f F z (2.8)

where f is the rolling resistance coefficient, and Fz is the normal force. The normal forces
for front and rear tire are respectively:
Faero haero mẍh mgh sin ✓ + mglr cos ✓
Fzf = (2.9)
lf + lr
Faero haero + mẍh + mgh sin ✓ + mglf cos ✓
Fzr = (2.10)
lf + lr

where haero is the height where the aerodynamic resistance is applied, h is the height of the
center of gravity of the vehicle, and lf and lr are the lengths from the center of gravity of
the vehicle to the front or the rear tire axle, respectively.

2.2.2 Powertrain and Engine Model

Previously, many engine models have been developed and studied. In view that we are
primarily interested in the motion of the vehicle body, however, too detailed engine model
is not required. In this thesis, a discrete event model [30, 31] is utilized in simulations.

Engine Model Structure

Figure 2.3 shows the interaction inside the powertrain model. The input to the entire
powertrain model is the desired engine torque, and the output is the torque delivered to
wheels. The first block of the powertrain is the engine torque controller. In this thesis, a
simple model based controller which calculates the appropriate throttle angle for the desired
engine torque based on the current engine state is used. The throttle angle and angular
 d#desire# 4#
PID# τ desire
d#rel# Mode#
Decision# v#desire#
PID#

R0ω
CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 12
Block#Diagram4#Detail#Engine#
τ engine τt τ wheel
ThroSle#
τ desire Engine# angle# Torque#
Torque# Engine# Transmission#
Converter#
Controller#
ωe ωt ω wheel

Figure 2.3: Powertrain model block diagram

MECHANICAL#SYSTEMS#CONTROL#LABORATORY,#UC#Berkeley#

Gear Ratio Remark

1st 2.8:1
2nd 1.5:1
Final gear with a ratio of 4:1 is additionally attached after each gear.
3rd 1:1
4th 0.7:1
Table 2.1: Typical gear ratios of an automatic transmission

velocity of the engine determine the output torque of the engine block [30, 31]. Then, the
engine torque goes through the torque converter and transmission to the wheels.
The torque generation process of a combustion engine consists of four di↵erent strokes:
intake, compression, combustion, and exhaust strokes. The dynamics of each stroke are
di↵erent from the others. Each stroke is changed to the next stroke after the engine rotates
180 degrees. Since, the engine rotation speed is not constant, the interval between the strokes
is not fixed and the dynamics of the engine is hard to describe in a time domain. So the
discrete event model describes the operation of the engine as events occur.

Transmission
Since the torque production of a vehicle engine is limited in magnitude, the engine rotates
very fast and the transmission gears reduces the speed while amplifying the torque. The
transmission is regarded as a gear box which has 3 to 6 di↵erent gears inside it. At steady
states, the transmission works like a gear with a fixed ratio. During transients, according to
the engine speed and the throttle angle, gear is shifted to maintain the engine RPM in an
adequate range. Typical gear ratios of an automatic transmission are given in Table 2.1.
At steady states, the torque transmitted to the wheel is:
1
⌧wheel = ⌧t (2.11)
R
where R is the gear ratio including the final gear. i.e. at 1st gear, R = 2.8 ⇥ 4
The speed of transmission is:
1
!t = !wheel (2.12)
R
CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 13

Figure 2.4: Braking System [13]

As shown in (2.11) and (2.12), at a lower gear, engine RPM is higher than a higher gear.
However, the transmitted torque at the wheel may be larger at a lower gear.
The dynamics during a gear change is studied in [9], which is complex. An alternative
way to express the gear change dynamics [42], which is simpler but adequate for the present
study is:
1
T ⌧˙wheel + ⌧wheel = ⌧t (2.13)
R
1
T !˙ t + !t = !wheel (2.14)
R
where T is a time constant. The initial value of ⌧wheel is zero when the gear change is
initiated. For !t , the initial value is R1old !wheel where Rold is the previous gear ratio.

2.2.3 Brake Model

Unlike the conventional cruise control system, the ACC system with Stop-and-Go should
utilize the brake to make the host vehicle decelerate. To design and evaluate the brake
control system, a brake model which can better represent the characteristics of the actual
brake system is required for simulations. In this thesis, the brake model introduced in [13] is
used. However, unlike [13], it will be assumed that there is no additional pressure cylinder
in the brake system and the brake torque is determined only by the brake pedal force.
CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 14

0
desired
actual

-50
torque[N m]

-100

-150
0 5 10 15 20 25 30 35 40 45 50
time[sec]

Figure 2.5: Simulation results with fuzzy logic: brake torque

Brake System
Figure 2.4 which is adopted from [13] depicts the structure of an automotive brake system.
When the driver wants to decelerate, the driver presses the brake pedal. The force from
the driver is amplified by the vacuum booster. The amplified force increases the pressure
of the master cylinder which is connected to the brakes of four wheels. Between the master
cylinder and the wheel brakes there are Anti-lock Brake System / Traction Control System
actuators. In this thesis, the piston masses are ignored and single state hydraulic system
model is used.

Brake Controller
In [13], an additional hydraulic actuator was introduced and a sliding controller was designed
to control the pressure of the brake pad cylinder and the brake torque. In this thesis, it is
assumed that no additional actuator is available. Instead, a Disturbance Observer (DOB)
based feedback controller is utilized. Also, to prevent the e↵ect of the vacuum booster
hysteresis, a fuzzy switching logic is used. The lower level controller shows a reasonably
good performance as shown in Figure 2.5. The brake model is implemented in MATLAB
with the parameters adopted from [13].
The simulation results show that the performance of the DOB is acceptable. Also it is
not necessary to modify the brake hydraulic system. However, an additional actuator in the
brake pedal is required. One main drawback of the DOB is that, due to the nonlinearity of
CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 15

the brake system, the Q-filter should be designed very conservatively and the performance
is worsened with more conservative choices of the cuto↵ frequency. The Q-filter should be
tuned later with experimental data. Also, the critical force to compensate the spring forces
inside the vacuum booster is necessary for the fuzzy switching logic and should be also tuned
by experiments.

2.3 Model Based Disturbance Compensator

Among the environmental disturbances, the aerodynamic resistance is dominant at high
speeds, while the gravity force due to the road slope can be the primary disturbance at low
speeds. The disturbances are hardly measurable and is related to the vehicle parameters
such as the vehicle mass. Thus, we have conducted a simulation study to estimate the
vehicle mass, the aero dynamic coefficient and the road slope by a least square method.
The estimated values are used in a model based disturbance compensator to compensate
the external disturbances. Simulation studies are performed to show the performance of the
disturbance compensator.

2.3.1 Recursive Least Square Method

Model Description
The longitudinal vehicle dynamics of a front wheel driven vehicle may be expressed as:

mẍ = Ftire mg sin ✓ Caero |ẋ|ẋ (2.15)

where m is the mass of the vehicle, Ftire is the force from the tire to the vehicle body
including the rolling resistance, g is the gravity acceleration, ✓ is the road slope, Caero is
the aerodynamic coefficient, and x is the longitudinal position of the vehicle. Among the
variables, Ftire can be measured by a Multi Sensing Hub (MSH) [52], ẋ can be obtained from
the rear (undriven) wheel angular velocity which can be measured by an encoder, and ẍ can
be measured by an accelerometer or estimated from ẋ. With these measurements and/or
estimates, we would like to estimate or identify parameters M , Caero , and ✓.
(2.15) is a model suited for parameter identification, in particular when Ftire can be
directly measured by MSH. If it is not possible to use the MSH and measure the tire force
directly, we will need to measure or know the engine torque and the brake torque as well as
other quantities such as the vehicle speed.

Recursive Least Square Problem

In the least square method, the system equation is represented as:

y(k) = a1 x1 (k) + a2 x2 (k) + a3 x3 (k) = ✓T '(k) (2.16)

CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 16

⇥ ⇤T ⇥ ⇤T
where ✓ = a1 a2 a3 , '(k) = x1 (k) x2 (k) x3 (k) , and k denotes time. From
(2.15), the variables are defined as:

y(k) = a(k) (2.17)

h iT
1 R02
✓ = m
C
m aero
sin ✓ (2.18)
⇥ ⇤
'(k) = Ftire (k) ! 2 (k) g (2.19)

where R0 is the e↵ective radius of the tire and ! is the angular velocity of the tire. At time
ˆ
k, we would like to find the best estimate of ✓, ✓(k), that minimizes the cost function
k h
X i2
J= y(i) ✓ˆT (k)'(k 1) (2.20)
i=1

The necessary condition for minimum is:

k h
X i
@J
= 2 y(i) ✓ˆT (k)'(i 1) '(i 1) = 0 (2.21)
@ ✓ˆ i=1

ˆ
Solving this equation for ✓(k), we obtain
k
X
ˆ = F (k)
✓(k) '(i 1)y(i) (2.22)
i=1

where, F (k) is the gain matrix, given by

" k
# 1
X
T
F (k) = '(i 1)' (i 1) (2.23)
i=1

(2.22) is the batch formula for the least square estimate. For real time applications, the
recursive formula given in the following form is better suited.
ˆ + 1) = ✓(k)
✓(k ˆ + [correction term] (2.24)

To obtain the recursive least square formula, we first note

k+1
X
1
F (k + 1) = '(i 1)'T (i 1) = F 1
(k) + '(k)'T (k) (2.25)
i=1
CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 17

(2.25) is in a recursive form. We then note

k+1
X
ˆ + 1) = F (k + 1)
✓(k '(i 1)y(i) (2.26)
i=1
" k #
X
= F (k + 1) '(i 1)y(i) + '(k)y(k + 1) (2.27)
i=1
ˆ + '(k)y(k + 1)]
1
= F (k + 1)[F (k)✓(k) (2.28)
ˆ + '(k)y(k + 1)]
= F (k + 1)[ F 1 (k + 1) '(k)'T (k) ✓(k) (2.29)
(2.30)

Thus we obtain
ˆ + 1) = ✓(k)
✓(k ˆ + F (k + 1)'(k)[y(k + 1) ✓ˆT (k)'(k)] (2.31)

By applying matrix inversion lemma to Eq. (2.25),

F (k)'(k)'T (k)F (k)

F (k + 1) = F (k) (2.32)
1 + 'T (k)F (k)'(k)

(2.31) and (2.32) give the recursive least square formula. With the least square method, we
can still get the same result. However, as the number of data is increased, the size of the
data matrix is getting larger and the calculation time will be significantly increased. In that
case, the real time application is almost impossible. With the recursive least square method
[24], the very new data are used to update the estimates and the calculation time remains
the same for each iteration.
For time varying parameters, we can introduce a forgetting factor, , in the cost function
to reduce the e↵ect of older data on the current estimate of the parameter vector:
k
X h i2
J= k i
y(i) ✓ˆT (k)'(k 1) (2.33)
i=1

where 0 <  1. In this case, the adaptation gain is updated by:


1 F (k)'(k)'T (k)F (k)
F (k + 1) = F (k) (2.34)
+ 'T (k)F (k)'(k)

Constant Trace Method

Recursive least square method with a forgetting factor is e↵ective when the dynamic param-
eters are not fixed but time varying, e.g. the road slope. However, with the forgetting factor
noises can be dominant and the parameter estimates may drift in case of signals with small
CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 18

amplitude. To overcome this problem, a constant trace method is introduced. Without a

forgetting factor, the adaptation gain matrix is updated by:
1 1
F (k + 1) = F (k) + '(k)'T (k) (2.35)

Its trace is given by:

1 1
tr[F (k + 1)] = tr[F (k)] + |'(k)|2 (2.36)
With a non-zero '(k), the trace of F 1 (k) always increases and consequently the trace of
F (k) decreases. Since the update law of the estimated parameters is
ˆ + 1) = ✓(k)
✓(k ˆ + F (k + 1)'(k)[y(k + 1) ✓ˆT (k)'(k)] (2.37)
ˆ is updated lesser in magnitude with a smaller F (k). This is the reason why the forgetting
✓(k)
factor is introduced. With a forgetting factor, the gain matrix is updated by:

1 F (k)'(k)'T (k)F (k)
F (k + 1) = F (k) (2.38)
+ 'T (k)F (k)'(k)
Its trace is given by:

1 F (k)'(k)'T (k)F (k)

tr[F (k + 1)] = tr[F (k)] 2 + 'T (k)F (k)'(k)
(2.39)

The forgetting factor prevents the gain matrix from decreasing to zero. The gain matrix,
however, may blow up if '(k) is not persistently exciting. A constant trace method maintains
the trace of the gain matrix at a constant to prevent it from either blowing up or decreasing
to zero. The update law is given by:

1 F (k)'(k)'T (k)F (k)
F (k + 1) = F (k) T
(2.40)
1 (k) 1 (k)/ 2 (k) + ' (k)F (k)'(k)

where 1 (k) is updated by:

1 F (k)'(k)'T (k)F (k)

1 (k) = tr[F (k) ] (2.41)
tr[F (0)] ↵ + 'T (k)F (k)'(k)

where ↵ = 1 / 2 is a constant value. (2.40) and (2.41) guarantee that the trace of the gain
matrix is maintained constant and does not blow up nor decrease to zero.

Vector-type Forgetting Factor

With the constant trace method, we can get a more stable estimation result. However, if we
apply a same forgetting factor for all parameters, there is an undesirable excitation in the
estimates of constant parameters like the vehicle mass when only the road slope is changing.
If the estimator has a prior information that the vehicle mass and the aerodynamic coefficient
CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 19

are time invariant, this problem may be prevented. A vector-type forgetting factor can make
this happen [65].
The vector-type forgetting factor method applies di↵erent forgetting factors for di↵erent
parameters. For frequently varying parameters, smaller forgetting factors will be applied to
update the parameters more sensitively. In the vector-type forgetting factor method, the
gain matrix is updated by:

F (k)'(k)'T (k)F (k)
F (k + 1) = ⇤(k) F (k) ⇤(k) (2.42)
1 + 'T (k)F (k)'(k)

where the forgetting factor matrix ⇤(k) is given by:

2 1 3
(k)
0 0
6 10 1
0 7
⇤(k) = 4 2 (k)
5 (2.43)
1
0 0 3 (k)

Each forgetting factor i (k)’s are updated by:

s
1
i (k) = gi "o (k+1)2
(2.44)
1+ 1+'T (k)F (k)'(k)

where "o (k) is an a-priori estimation error, and gi ’s are constants for each parameters. Larger
gi ’s are used for frequently varying parameters to make the forgetting factor smaller.
However, the vector-type forgetting factor method is too sensitive to the values of gi ’s.
Also, the estimation errors are too large if a sufficient amount of excitation is not given. To
resolve these problems, the constant trace method can be introduced again. While using
a vector-type forgetting factor, the trace of the gain matrix is maintained constant. With
this constant trace method with a vector-type forgetting factor, the estimator works better.
Figures 2.6 and 2.7 show the simulation results with di↵erent values of gi ’s, and the results
are almost the same. Also in Figures 2.8 and 2.9, it is seen that even without enough
excitations, the estimator works well.

2.3.2 Disturbance Compensator

With the estimated values, it is possible to estimate the disturbances based on the previously
introduced model used in the parameter estimation. The three disturbance models are given
by:
1
Faero = ⇢Cd Af |vx + vwind |(vx + vwind ) (2.45)
2
Fgrav = mg sin ✓ (2.46)
Froll = f Fz = f mg (2.47)
CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 20

1500

Mest [kg]
1400

1300
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
k
1
Caero est

-1
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
k
5
thetaest [degree]

-5
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
k

Figure 2.6: Simulation result with g1 = g2 = g3 = 1

1500
Mest [kg]

1400

1300
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
k
1
Caero est

-1
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
k
5
thetaest [degree]

-5
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
k

Figure 2.7: Simulation result with g1 = g2 = 0.1, g3 = 1

CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 21

3000

Mest [kg] 2000

1000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
k
1
Caero est

0.5

0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
k
5
thetaest [degree]

-5

-10
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
k

Figure 2.8: Simulation result where there is not enough excitations: with a vector-type
forgetting factor without the constant trace method

3000
Mest [kg]

2000

1000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
k
1
Caero est

0.5

0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
k
10
thetaest [degree]

-5
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
k

Figure 2.9: Simulation result where there is not enough excitations: with a vector-type
forgetting factor and a constant trace method
CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 22

20

velocity[m/s]
lead
10
with DC
w/o DC
0
0 10 20 30 40 50 60 70
time[sec]
road slope[deg]

0
0 10 20 30 40 50 60 70
time[sec]
0.2 with DC
w/o DC
error[m]

−0.2
0 10 20 30 40 50 60 70
time[sec]

Figure 2.10: Simulation results with and without the disturbance compensator

where Faero is the aerodynamic resistance, Fgrav is the gravitational resistance, and Froll is the
rolling resistance. 12 ⇢Cd Af is the aerodynamic coefficient, m is the vehicle mass, and ✓ is the
road slope which are estimated by the recursive least square method. vx is the longitudinal
velocity of the vehicle and vwind is the longitudinal wind velocity which is unknown and is
ignored. f is the rolling resistance coefficient and assumed to be known.
The output from the disturbance compensator is the desired additional acceleration and
is given by:
1
adc = (F̂aero + F̂grav + F̂roll ) (2.48)
m̂
This additional acceleration command to compensate for disturbances is added to the feed-
back command:
adesired = af eedback + adc (2.49)

2.3.3 Simulation Results

Simulations are performed to examine the e↵ects of the disturbance compensator. Figure
2.10 shows the simulation results with and without the disturbance compensator. As shown
in the figures, with the disturbance compensator the tracking error is reduced especially
when the road slope is changing.
CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 23

2.4 Feedback Controller for Vehicle Following

2.4.1 Desired Distance
The most significant di↵erence between a normal cruise control system and an Advanced
Cruise Control (ACC) system is that the ACC system must have a function to maintain
the desired distance from the leading vehicle, while the sole function of the normal cruise
control system is to maintain the speed of the vehicle at a desired value. Defining and
calculating the desired distance which may depend on the state of the vehicle itself as well
as its environment, is one of the crucial tasks in developing the ACC system.
When designing the feedback controller, not only the stability of the single system but
also the string stability of the vehicle platoon should be considered. String stability should
be considered for a platoon of vehicles equipped with identical ACC systems. If the feedback
controller is not properly designed, when the very first vehicle of the platoon accelerates or
decelerates, the distance errors increases along the vehicle stream. To prevent it, the string
stability should be guaranteed.
There are many ways to determine the desired distance. The simplest choice to implement
for the desired distance is a constant distance. This method is very simple but has many
problems. Intuitively, it is hard to determine a proper value of the constant distance. If
the distance is set small, safety becomes an issue at high speeds. On the other hand, a too
large distance is inefficient at low speeds. In city driving, large distances will waste very
large spaces between cars. Furthermore, the constant distance method can cause a string
instability issue [42] or vehicle to vehicle communication is necessary to guarantee the string
stability [47].
Instead of the constant distance, the constant time-headway policy is widely used to
calculate the desired distance. Time-headway is the time to take for the host vehicle to
collide with the lead vehicle when the lead vehicle is suddenly stopped and the host vehicle
maintains its speed. If the desired distance is calculated such that the time-headway remains
constant, the string stability is guaranteed [42]. In that case, the desired distance is given
as:
ddesired = thw vhost (2.50)
where thw is the time-headway and vhost is the velocity of the host vehicle.
However, (2.50) is not valid for complete stops since the desired distance is zero when the
host vehicle is completely stopped. Instead, a safety distance which should be maintained
between two vehicles when both the following and preceding vehicles are stopped can be
introduced[18]. In that case, the desired distance is given as:

ddesired = thw vhost + dof f set (2.51)

where dof f set is the safety distance. Both the time-headway and the safety distance for a
complete stop may vary according to the road condition and the preference of the driver.
CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 24

2.4.2 Tuning of Feedback Gains

In this section, a Linear Quadratic (LQ) Optimal Control scheme will be studied as a way
to find the optimal feedback gains.

Problem Description
In this section, a single host vehicle following a lead vehicle is considered. Each vehicle
will be modeled as a double integrator whose input is the acceleration. The desired relative
distance is given by (2.51). However, the safety distance will be ignored to make the problem
linear. The safety distance can be added later after the state feedback gains are calculated.
For the LQ problem, the output of the system is selected as the error. The state space
equations of the system are:

Ẋ(t) = AX(t) + BU (t) (2.52)

Y (t) = CX(t) (2.53)
⇥ ⇤T ⇥ ⇤T
X(t) = xl (t) x(t) vl (t) v(t) , U (t) = al (t) a(t) (2.54)
2 3 2 3
0 1 1 0 0 ⇥ ⇤
A = 4 0 0 0 5, B = 4 1 0 5, C = 1 0 thw (2.55)
0 0 0 0 1
where x, v, and a are the position, the velocity, and the acceleration of the host vehicle,
respectively and xl , vl , and al are the position, the velocity, and the acceleration of the lead
vehicle, respectively.

Linear Quadratic Optimal Control

If the plant is time invariant, (A, B) is controllable, and (A, C) is observable, the solution of
the stationary LQ problem exists for the performance index
Z
1 1 T
J= X (t)QX(t) + U T (t)RU (t)dt (2.56)
2 t0

where Q = C T C.
However with the C given in (2.55), (A, C) is not observable. To make it observable with
a minimum modification of the performance index, a new C is given as:

1 0 thw
C= (2.57)
0 ✏o 0
The number ✏o is a very small number that can be further reduced depending on the available
computational resolution.
The input penalty matrix R should be positive definite. Also, it is assumed that we cannot
control the lead vehicle acceleration al . However if we ignore the lead vehicle acceleration, the
CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 25

solution of the stationary LQ problem is not guaranteed since in that case (A, B) becomes
not controllable or stabilizable. Instead of ignoring the lead vehicle acceleration, we can
penalize it a lot so that the magnitude of it is limited to be very small (the actual lead
vehicle acceleration will depend on the driver and this can be considered as a disturbance).
To do that, R can be selected as:

1/✏c 0
R= (2.58)
0 1

where is a design parameter to tune. The number ✏c is a very small number that can be
further reduced depending on the available computational resolution. The state feedback
control law is given as:
U (t) = R 1 B T P+ X(t) = KX(t) (2.59)
where P+ is the positive definite solution of the algebraic Riccati equation

AT P + P A P BR 1 B T P + C T C = 0 (2.60)

With thw = 2, = 1, and ✏o = ✏c = 1e 6, the state feedback gain is calculated to be:


1.0102e 6 2.5061e 6 4.4949e 6
K= (2.61)
1.0000 0.4495 2.4495

As desired, the gains for the lead vehicle acceleration are very small. The feedback control
law can be written as:

a(t) = (thw v(t) (xl (t) x(t))) 0.4495( (vl (t) v(t))) (2.62)

Assuming v(t) is constant, the control law is:

a(t) = err(t) 0.4495err(t)

˙ (2.63)

where err(t) = thw v(t) (xl (t) x(t)). As shown in (2.63), the state feedback law is the
same as the PD controller.

Linear Quadratic Optimal Control with Integrator

To add an integral controller to the LQ controller, a Linear Quadratic Optimal Control with
Integrator (LQI) is introduced. The plant is given by (2.52) ⇠ (2.55) where the matrix C is
modified as (2.57). The cost function of the LQI problem is:
Z 1
J= [Y (t) r]T Qy [Y (t) r] + W T (t)RW (t)dt (2.64)
t0

where Qy and R are symmetric and positive definite weight matrices, r is a constant set
point vector, and W is the derivative of U .
CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 26

The new state vector is defined as X̃(t) = [ E T (t) Ẋ T (t) ]T where E(t) = Y (t) r.
Then the state equation is:
 
d 0 C 0
X̃ = X̃(t) + W (t) = ÃX̃(t) + B̃W (t) (2.65)
dt 0 A B

The cost function is:

Z 1 
T Qy 0
J= X̃ (t) X̃(t) + W T (t)RW (t)dt (2.66)
t0 0 0

1 0
and the problem reduces to a standard LQ problem. With Qy = , (Ã, B̃) is con-
0 ✏o
1
trollable and (Ã, Qy1 [ I2⇥2 02⇥3 ]) is observable and the Riccati equation has a stationary
positive definite solution. 
1 0 1e6 0
With Qy = and R = , the PID control law is:
0 1e 6 0 1
Z t
a(t) = 0.9804err(t) 0.4806err(t)
˙ err(⌧ )d⌧ (2.67)
t0

2.4.3 Simulation Results

Figure 2.11 shows the simulation result comparing the PD (LQ) 2.63 and the PID (LQI)
2.67 controllers. Both controllers show stable performances while PID shows better tracking
performance when the lead vehicle accelerates or decelerates. Under our assumption, the
lead vehicle acceleration is considered as a disturbance. Since the Integral controller has the
disturbance rejection property, it shows better performance especially when the lead vehicle
acceleration is nonzero.
Figure 2.12 shows the simulation result of a vehicle platoon. In the platoon, the first
lead vehicle is assumed to be driven by a human driver. The following vehicles are equipped
with the identical ACC systems. For simplicity, the engine and the brake are assumed to
be first order and no tire dynamics are included. As shown in the figure, the designed PID
controller is string stable and the peak value of the error decreases along the vehicle stream.

2.5 Summary
In this chapter, a longitudinal vehicle model was introduced. External disturbances such as
the aerodynamic drag, the rolling resistance, and the gravity force due to the road slope were
considered. To better predict the behavior of the vehicle while preventing too complicated
structure, the discrete event engine model [30, 31], the single state hydraulic brake model
[13], and the dynamic deflection tire model [22] were included in the vehicle model.
CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 27

16
lead
14 PID
velocity[m/s] PD
12

10

8
0 10 20 30 40 50 60 70
time[sec]

1
PID
0.5 PD
error[m]

−0.5

−1
0 10 20 30 40 50 60 70
time[sec]

Figure 2.11: Simulation results with the PID and PD controllers

12

10
velocity[m/s]

8 lead
#1
6
#2
#3
4
0 20 40 60 80 #4 100
time[sec] #5
−3
x 10
5
error[m]

−5
0 20 40 60 80 100
time[sec]

Figure 2.12: Simulation results of a vehicle platoon with the PID controller
CHAPTER 2. VEHICLE MODEL AND BASELINE CONTROLLER 28

A recursive least square method was used to estimate the vehicle mass, the road slope,
and the aerodynamic resistance. A vector-type forgetting factor was utilized to estimate the
parameter where some of the parameters were changing frequently while others were changing
very slowly. The constant trace method was also used to achieve more stable performance
with the forgetting factor especially when there was not enough excitation. A model based
disturbance compensator was designed to compensate the external disturbances.
Linear Quadratic (LQ) optimal control was studied to find the optimal feedback gains
for a single vehicle following case. The vehicle was modeled as a double integrator and the
state space realization matrices were slightly modified to make the system controllable and
observable. The designed LQ controller was nothing but a PD controller. Linear Quadratic
control with Integrator (LQI) was introduced to add the integral action. The LQI controller
showed better tracking performance than the LQ controller when the lead vehicle accelerated
or decelerated. The constant time-headway policy was used to calculate the desired distance
and the simulation result of a vehicle platoon showed string stable performances.
 29

Chapter 3

Virtual Lead Vehicle Scheme

3.1 Introduction
The ACC system should be able to perform two di↵erent tasks according to the surrounding
situation. If there is a lead vehicle detected, the ACC system should follow the lead vehicle
maintaining an appropriate relative distance. When no lead vehicles are detected in front of
the host vehicle, the ACC system should maintain the desired velocity like the conventional
cruise control system. It is important that not only both modes should operate properly
and stably, but also the transient performance should be smooth and safe when the mode is
switched from one to the other.
A mode switching scheme is one method which deals with the presence of a lead vehicle.
The control algorithm switches between the speed control and the distance control modes
based on the relative distance to the lead vehicle [63, 58]. The speed control algorithm is used
when no lead vehicle is detected or the lead vehicle is farther than a specified relative distance.
When the lead vehicle is close, the distance control algorithm is used. One downside of this
method is that two di↵erent feedback loops are required: velocity and distance control loops.
If conventional PID feedback controllers are used, integrator windup must also be properly
taken into consideration.
Another drawback of the mode switching scheme is that the transient maneuver is not
considered. When the speed control algorithm is activated, if a new lead vehicle cuts in
from a side lane or a slower lead vehicle approaches from the front, the mode is abruptly
changed to the distance control which can cause unpleasant motion of the host vehicle. A
model predictive control method was studied to control the transitional maneuver of the
ACC system [4]. In [4], the optimal motion of the host vehicle was calculated by solving an
optimization problem over a time horizon and the initial optimal control input was utilized as
the control command. The model predictive controller was appropriate only for the distance
control mode or the transition from the speed control to the distance control mode. Still, a
separate speed controller was required to control the host vehicle during the speed control
mode. Also, the model predictive control requires significant computational e↵orts and a
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 30

fast processor is necessary for real time application.

A virtual lead vehicle scheme [20] is a method which can replace the complicated switching
mechanism. When no lead vehicle is detected, the controller generates a virtual vehicle whose
speed is the same with the desired speed. When there is a lead vehicle, the speed and the
position of the virtual lead vehicle are set to be those of the actual vehicle. The host vehicle
always follows the virtual lead vehicle with an appropriate distance control algorithm (i.e.
a PID controller) and no switching scheme between the speed and the distance control is
necessary.
Additional advantage of the virtual lead vehicle scheme is that the motion of the host
vehicle can be smoothly controlled when a new lead vehicle cuts in or the current lead vehicle
cuts out. If the host vehicle follows the actual lead vehicle, the cutting in/out of the lead
vehicle is a step change of the error signal. However, if the host vehicle follows the virtual
lead vehicle and the speed and the position of the virtual lead vehicle merge smoothly from
those of the old lead vehicle to those of the new lead vehicle, the host vehicle also moves
smoothly.
Since the virtual vehicle has no internal dynamics or delay, the position and the velocity
of it can be set without any physical limitations. However, a sudden change of the position
or velocity of the virtual lead vehicle can cause an undesirable motion of the host vehicle
which can make the passengers uncomfortable. To prevent this, the virtual lead vehicle is
modeled as a double integrator whose input is the acceleration and outputs are the velocity
and the position. A smooth switching can be achieved by controlling the virtual lead vehicle
with limited acceleration and jerk.
This chapter is organized as follows. In Section 3.2, the modes of the ACC system and the
concept of the virtual lead vehicle scheme are introduced. In Section 3.3, the linear quadratic
controller is designed to control the virtual lead vehicle. In Section 3.4, variable weights are
utilized for the linear quadratic controller to better control the virtual lead vehicle during
transient motion. The simulation results are shown to support the suggested ideas. The
stability of the suggested controller is analyzed in Section 3.5. In Section 3.6, a Disturbance
Observer is utilized to estimate and compensate the acceleration of the lead vehicle. Section
3.7 summarizes the chapter.

3.2 Concept of the Virtual Lead Vehicle Scheme

The ACC system should be able to perform two modes: speed control and distance control.
The modes depend on the desired and actual distance between the lead and host vehicles.
One way to implement these two modes is by designing two di↵erent controllers and switching
between them. However this can cause various problems. Another way is the virtual lead
vehicle scheme which makes the switching between the speed control algorithm and the
distance control algorithm unnecessary [20]. With the virtual lead vehicle scheme, the control
system structure is simplified.
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 31

! Figure 3.1: Mode switching logic

3.2.1 Mode Decision of the ACC system

As discussed in previous chapters, when the host vehicle follows a lead vehicle the desired
distance is given by:
ddesired = thw vhost + dof f set (3.1)
where thw is the time headway, vhost is the velocity of the host vehicle, and dof f set is the
safety distance to be maintained in case of a complete stop. The time headway is the time
to take for the host vehicle to collide with the lead vehicle when the lead vehicle is suddenly
stopped and the host vehicle maintains its speed. The desired time headway is decided by
the driver when the ACC system is activated. Using (3.1), a simple scheme to determine the
proper mode is given by: ⇢
DC drel  ddesired
mode = (3.2)
SC drel > ddesired
where DC is the distance control, SC is the speed control, and drel is the relative distance
to the lead vehicle. Using this scheme, however, may cause chattering in the algorithm if the
distance between the lead and host vehicle is too close to the desired distance. To prevent
chattering, marginal distances are introduced as:
⇢
SC ! DC drel dm1  ddesired
mode = (3.3)
DC ! SC drel + dm2 > ddesired

where dm1 and dm2 are marginal distances. This introduces hysteresis in the mode switching
as shown in Figure 3.1. If there is no lead vehicle, drel is infinite and the speed control mode
is utilized.
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 32

3.2.2 Mode Switching Scheme

Figure 3.2 shows the block diagram of the mode switching scheme. The input signal to the
distance control algorithm is the distance error which is given by:
⇢
drel ddesired mode = DC
e1 = (3.4)
0 mode = SC

The input signal to the speed control algorithm is the speed error which is given by:
⇢
vdesired vhost mode = SC
e2 = (3.5)
0 mode = DC

where vdesired is the desired speed set by the driver. There is an additional switch to cancel
the output of the inactivated mode:
⇢
a1 mode = DC
ad = (3.6)
a2 mode = SC

If PID controllers are used for both algorithms, the integral controller of the inactivated
mode should be initialized when the mode is switched.

3.2.3 Virtual Lead Vehicle Scheme

Figure 3.3 shows the conceptual sketch of the virtual lead vehicle scheme. In the virtual
lead vehicle scheme, the host vehicle follows the virtual lead vehicle with a distance control
algorithm in the absence of the actual lead vehicle. When a lead vehicle is detected inside
the desired relative distance and the speed of the lead vehicle is slower than the desired
speed, the position and the velocity of the virtual lead vehicle is set to be those of the actual
lead vehicle. When no lead vehicle is detected or if the speed of the lead vehicle is too fast,
the virtual lead vehicle scheme prevents switching to the speed control algorithm by limiting
the velocity of the virtual lead vehicle to the desired speed. In that case, the position of the
virtual lead vehicle is updated by integrating the velocity.
As a result, the speed control algorithm is not utilized with the virtual lead vehicle scheme
and the overall controller structure has only one loop which is the distance control algorithm
as described in Figure 3.4.

3.3 Control of the Virtual Lead Vehicle in Transient

In this section, a double integrator model will be suggested as the model of the virtual
lead vehicle and a linear quadratic controller will be designed to control the speed and the
position of the virtual lead vehicle in transient motions. The control scheme deciding the
speed and position of the virtual lead vehicle during the speed control and the distance
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 33

!
Figure 3.2: Block diagram of the mode switching scheme

!
Figure 3.3: Concept of the virtual lead vehicle scheme
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 34

!
Figure 3.4: Block diagram of the virtual lead vehicle scheme

control algorithms are described in Section 3.2.3. However, when the mode is changed from
one to the other due to cutting in or out of the lead vehicle, the speed and the position of the
virtual lead vehicle should be smoothly change from those of the old lead vehicle to those of
the new lead vehicle. A double integrator model and a linear quadratic control scheme are
strong candidates to achieve this desired behavior.

3.3.1 Double Integrator Model for the Virtual Lead Vehicle

When the current lead vehicle cuts out from the current lane or a new lead vehicle cuts in
from a side lane, there are step changes in the position and the speed of the lead vehicle.
If the host vehicle directly follows the actual lead vehicle, undesirable transient oscillation
or overshoots can be generated. With the virtual lead vehicle scheme, however, the virtual
vehicle‘s position and speed are controlled to merge smoothly with those of the new lead
vehicle. Hence, the host vehicle will also move smoothly.
In this thesis, the goal is to design a controller which makes the virtual lead vehicle
moves smoothly from an initial state to the desired state. The position and the speed of
the virtual lead vehicle can be set arbitrarily by the controller. However, the host vehicle
should react slowly and smoothly given any initial conditions to avoid producing unnecessary
uncomfortable motions. The virtual lead vehicle is considered as a double integrator whose
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 35

input is the acceleration. The continuous time model of the virtual lead vehicle is:
    
ẋvl (t) 0 1 xvl (t) 0 xvl (t)
= + avl (t) = A + Bavl (t) (3.7)
v̇vl (t) 0 0 vvl (t) 1 vvl (t)
where xvl , vvl , and avl are the position, speed, and the acceleration of the virtual lead vehicle,
respectively.
The task of the controller is to make the virtual vehicle follow the actual lead vehicle
smoothly. In this thesis, to follow means that both the position and the velocity of the
virtual vehicle converge to those of the actual vehicle. The errors are defined as:
 
ex (t) xvl (t) xal (t)
= (3.8)
ev (t) vvl (t) val (t)
where xal and val are the position and the speed of the actual lead vehicle, respectively. The
error dynamics is:  
ėx (t) ex (t)
=A + B(avl (t) aal (t)) (3.9)
ėv (t) ev (t)
where aal is the acceleration of the actual lead vehicle. It is assumed that the relative distance
and velocity of the actual lead vehicle are measurable. Since the position and velocity of the
virtual lead vehicle is either known and fixed, or dependent on the actual lead vehicle, the
relative distance and the relative velocity between the virtual and the actual lead vehicle can
be computed. However, the behavior of the lead vehicle is unknown and hence the actual
lead vehicle acceleration aal is considered as a disturbance.

3.3.2 Linear Quadratic Control

The goal is to make the states in (3.9) converge zero. One promising way to control this
system is doing state feedback using the steady state Linear Quadratic (LQ) optimal con-
troller. The cost function is a quadratic sum of the position error, the velocity error, and
the acceleration given as:
Z
1 1 2 2 2
J= x ex (t) + v ev (t) + a avl (t) dt (3.10)
2 t0
Without the constraints, the problem can be formulated as a stationary LQ problem.
Z
1 1
J= X(t)T QX(t) + avl (t)T Ravl (t)dt (3.11)
2 t0

⇥ ⇤T x 0
where X(t) = ex (t) ev (t) , Q = , and R = a . Since the system is time
0 v
invariant, controllable, and observable, the optimal solution of the stationary LQ problem is
given as:
avl (t) = R 1 B T P+ X(t) (3.12)
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 36

where P+ is the positive (semi-)definite solution of the algebraic Riccati equation

AT P + P A P BR 1 B T P + Q = 0 (3.13)

In this case, there are only two states and the Riccati equation can be algebraically solved.
The feedback control law is given as:
 q r q
aLQ (t) = x
2 x
+ v X(t) (3.14)
a a a

The solution given in (3.14) is for the case without constraints. For actual implementation,
however, acceleration and jerk have to be constrained to limit passenger discomfort. These
constraints are given by:
amin  avl (t)  amax (3.15)
jmin  jvl (t)  jmax (3.16)
The acceleration of the virtual lead vehicle is given as:

avl (t) = max{min(aLQ , amax , a(t t) + jmax t), amin , a(t t) + jmin t} (3.17)

where aLQ is the solution of the unconstrained problem given in (3.14), and t is the con-
troller sampling time.

3.3.3 Comparison with the Model Predictive Control

In (3.15) ⇠ (3.17), the acceleration and the jerk constraints are applied to only the current
time and not the whole time horizon. Ideally, the constraints should be applied to the whole
time horizon. To do so the Model Predictive Control (MPC) method can be applied. The
cost function for MPC is the same as (3.10) and the inequality constraints described in (3.15)
⇠ (3.16) are applied to the whole time horizon. Figure 3.5 compares the simulation results
with MPC and LQ control. As shown in the figure, the two results are very similar to each
other. However, the MPC method requires a lot more calculation e↵orts than the LQ. As a
result, using the MPC method will require a faster processor for real time application. Since
the results are almost identical, the LQ method can efficiently replace the MPC scheme and
will be used in this thesis.

3.4 Linear Quadratic Control With Variable Weights

In the cost function of the LQ problem given in (3.10), the weights x , v , and a are
parameters to tune by the controller designer. The state feedback gains are defined as
functions of these weights. To control the motion of the virtual lead vehicle smoothly, state
dependent weights are suggested and simulation result shows better performance in the
viewpoint of passenger comfort and fuel efficiency [21].
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 37

12
lead
10 MPC

velocity [m/s]
LQ
8

4
0 5 10 15 20 25
time [sec]

300
position [m]

200

100

0
0 5 10 15 20 25
time [sec]

!
Figure 3.5: Simulation results with MPC and LQ control: Position and velocity of the virtual
lead vehicle

3.4.1 Tuning of the LQ Weights

The weights of the cost functions can be tuned using the return di↵erence equation. The
closed loop poles of the system are the left half plane solution of the return di↵erence equation
[23] :
v 2 x
s4 s + =0 (3.18)
a a
The symmetric root locus is shown in Figure 3.6 varying a from 0 to 1. Considering the
passenger’s comfort and the fuel efficiency, the damping should be large to prevent a large
overshoot. Also it is desirable to have the real part of the poles to be as small as possible
for faster responses. The point marked in Figure 3.6 is such a point and the weights are:

x = 1, v = 10, a = 25 (3.19)

3.4.2 Simulation Results with Constant Weights

Figure 3.7 shows the simulation results with the constant weights given in (3.19). The
simulation describes the case when the current lead vehicle cuts out and a new lead vehicle
is detected. As shown in the figure, the virtual lead vehicle initially accelerates to chase
the new lead vehicle and then decelerates to the speed of the new lead vehicle. The initial
acceleration is constrained by the acceleration and the jerk constraints. However, the large
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 38

Root Locus
0.2

0.15

System: sys
0.1
Gain: 0.04
Pole: -0.447 + 7.52e-009i
Damping:
0.05 1
Overshoot (%): 0
Imaginary Axis

Frequency (rad/sec): 0.447

0

-0.05

-0.1

-0.15

-0.2
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
Real Axis

! 3.6: The symmetric root locus - Arrows indicate decreasing

Figure a

velocity overshoot is undesirable. To prevent the overshoot, the acceleration is penalized

more and the simulation results show that the velocity overshoot is reduced as a is increased.
However, if a is increased, the response of the closed loop system gets slower. This can
be also found in Figure 3.6 where two poles are moving close to the origin if a is increased
more than the critical value. Slow response can make the driver uncomfortable especially
when a new lead vehicle cuts in and the relative distance to the new lead vehicle is closer
than the desired distance. Another solution is to increase v . In that case, when the velocity
error is positive, the feedback term of the velocity error is negative and the acceleration can
be reduced. However, when a new lead vehicle cuts in and the position error is positive, the
large v can cause undesirable system behavior.

3.4.3 LQ with Variable Weights

To prevent the large velocity overshoot while maintaining a fast response in the cutting
in situation, variable weights are introduced. The desired performance of the virtual lead
vehicle is described in Table 3.1. When the position error is negative and the velocity error
is positive, it is desirable to maintain the speed to prevent a large velocity overshoot. To do
that the position error should be less penalized while the velocity error and the acceleration
should be penalized more. When the position error is positive and the velocity error is
negative, which describes the case when a lead vehicle faster than the virtual lead vehicle
cuts in, the virtual lead vehicle should smoothly decelerate and merge to the new lead vehicle.
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 39

15
actual lead
lambdaa=10

velocity [m/s]
lambdaa=25
10
lambdaa=50

5
0 5 10 15 20 25
time [sec]

300
position [m]

200

100

0
0 5 10 15 20 25
time [sec]

!
Figure 3.7: Simulation results with constant weights position and velocity of the virtual
lead vehicle

The position error should be more penalized than the speed error while the acceleration is
penalized more to make the motion smooth. When both the position and the velocity errors
are positive, it is a dangerous situation and it is desirable to decelerate the host vehicle
quickly. In that case, both errors should be penalized while the acceleration should be less
penalized to ensure a fast response.
Since the desired values of the weights are changing according to the sign of the errors,
the variable weights are suggested as:
2
x = x0 (1 + tan 1 ( x ex (t))) (3.20)
⇡
2
v = v0 (1 + tan 1 ( v ev (t))) (3.21)
⇡
4
a = a0 (1 2
tan 1 ( x ex (t))tan 1 ( v ev (t))) (3.22)
⇡
The arctangent function is introduced instead of the sign function to make it continuous and
’s are constants describing the slope of the arctangent function near the origin.

3.4.4 Simulation Results with Variable Weights

Figures 3.8 ⇠ 3.10 show the simulation results with the variable weights given in (3.20) ⇠
(3.22). As shown in the figure, the velocity peak is reduced with the variable weights. Figure
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 40

Sign of errors Desired performance Desired x Desired v Desired a

ex < 0, ev > 0 Do not accelerate small large large
ex > 0, ev < 0 Decelerate smoothly large large
ex > 0, ev > 0 Decelerate fast large large small
Table 3.1: Desired performance and weights

3.9 shows the weights. As shown in Figure 3.9, when the velocity error becomes positive, the
weights on the velocity and the acceleration are increased to prevent the velocity overshoot.
Figure 3.10 shows the velocity and the fuel consumption of the host vehicle. The engine
and the brake dynamics are assumed to be first order with time constants 0.5 and 0.04
respectively. The fuel consumption is calculated with a VT model [2]. As shown in Figure
3.10, with the variable weights, the large velocity overshoot is avoided and fuel efficiency is
improved. The square sums of the acceleration and the jerk, which can represent a measure
of passenger comfort, are 0.2014[m2 /sec4 ] and 0.0814[m2 /sec6 ] respectively with variable
weights, and 0.4003[m2 /sec4 ] and 0.7638m[2 /sec6 ] respectively with the constant weights.
Figures 3.11 ⇠ 3.13 show the simulation results with di↵erent situation where a slower
lead vehicle cuts in. As shown in the figures, with variable weights, the motions of the virtual
lead vehicle and the host vehicle are smoother. Also the fuel efficiency is better than the
constant weights. The square sums of the acceleration and the jerk are 0.1050[m2 /sec4 ] and
0.0586[m2 /sec6 ] respectively with the variable weights, and 0.1527[m2 /sec4 ] and 0.1865[m2 /sec6 ]
respectively with the constant weights.

3.5 Stability Analysis of the Linear Quadratic

Controller with Variable Weights
Since the weights in the cost function of the linear quadratic controller are state depen-
dent, the stability and the robust features of the general linear quadratic controller are not
guaranteed anymore. Also the stability of the suggest scheme should be proved.
The closed loop system dynamics is given by:

ėx (t) = ev (t) (3.23)

s v s
u
u
x (ex ) t2 x (ex ) v (ev )
ėv (t) = ex (t) + ev (t) (3.24)
a (ex , ev ) a (ex , ev ) a (ex , ev )

q r q
Since x
a
and 2 x
a
+ v
a
are always positive, the equilibrium point of the system is only
the origin and the origin is a stable equilibrium point. However this does not guarantee that
the solution converges to zero with an arbitrary initial condition. To show this, a phase
plane analysis is performed.
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 41

12
actual
10 Variable weights
velocity [m/s] Constant weights
8

4
0 5 10 15 20 25 30 35 40 45 50
time [sec]

400

300
position [m]

200

100

0
0 5 10 15 20 25 30 35 40 45 50
time [sec]

!
Figure 3.8: Simulation results when a lead vehicle cuts out position and velocity of the
virtual lead vehicle

20
lambdax

10

0
0 5 10 15 20 25 30 35 40 45 50
time [sec]
20
lambdav

10

0
0 5 10 15 20 25 30 35 40 45 50
time [sec]
1000
lambdaa

500

0
0 5 10 15 20 25 30 35 40 45 50
time [sec]

Figure !3.9: Simulation results when a lead vehicle cuts out variable weights
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 42

12
Variable weights
10 Constant weights
velocity [m/s]
8

4
0 5 10 15 20 25 30 35 40 45 50
time [sec]

0.015
fuel consumption [Gal]

0.01

0.005

0
0 5 10 15 20 25 30 35 40 45 50
time [sec]

!
Figure 3.10: Simulation results when a lead vehicle cuts out position and velocity of the
host vehicle

12
actual
Variable weights
velocity [m/s]

10
Constant weights

6
0 5 10 15 20 25
time [sec]

300
position [m]

200

100

0
0 5 10 15 20 25
time [sec]

!
Figure 3.11: Simulation results when a lead vehicle cuts in position and velocity of the
virtual lead vehicle
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 43

20

lambdax
10

0
0 5 10 15 20 25
time [sec]
20
lambdav

10

0
0 5 10 15 20 25
time [sec]
1000
lambdaa

500

0
0 5 10 15 20 25
time [sec]

Figure !3.12: Simulation results when a lead vehicle cuts in variable weights

12
Variable weights
velocity [m/s]

Constant weights
10

6
0 5 10 15 20 25
time [sec]
-3
x 10
8
fuel consumption [Gal]

0
0 5 10 15 20 25
time [sec]

!
Figure 3.13: Simulation results when a lead vehicle cuts out position and velocity of the
host vehicle
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 44

10

2
v[m/s]

-2

-4

-6

-8

-10
-10 -8 -6 -4 -2 0 2 4 6 8 10
x[m]

! Figure 3.14: Phase plane plot of the closed loop system

Figure 3.14 shows the phase plane plot of the closed loop system. As expected from
(3.23) the solution trajectory moves to the positive x direction in the first and the second
quadrants, and to the negative x direction in the third and the forth quadrants. Also as
expected from (3.24) the solution trajectory goes to the negative v direction in the first
quadrant and to the positive v direction in the third quadrant. As a result the solution will
meet the x axis at some point. The solution from a point on the positive x axis will meet
the positive x axis again. If the second point is closer to the origin than the first point, it
can be concluded that all initial points converge to the origin and the system is stable.
Since the system is nonlinear, it is hard to find the solution explicitly. However it is
possible to find a linear switching system which always behaves worse than the original
system. From (3.23) and (3.24), the slope on the phase plane is:
s v s
u
dev (e ) e (t) u
x x x t2 x (ex ) v (ev )
= + (3.25)
dex a (ex , ev ) ev (t) a (ex , ev ) a (ex , ev )

The convergence is slower if the slope is large (steep when positive and flat when negative).
With the weights given in (3.20) ⇠ (3.22), the behavior of the system is always better than
the worst case system given by:

ėx (t) = ev (t) (3.26)

ėv (t) = a1 ex (t) a2 ev (t) (3.27)
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 45

10 10

v[m/s]

v[m/s]
5 5

0 0
-10 -5 0 0 5 10
x[m] x[m]

0 0
v[m/s]

v[m/s]
-5 -5

worst
original
-10 -10
-10 -5 0 0 5 10
x[m] x[m]

!
Figure 3.15: Phase plane plot of the closed loop system at each quadrant comparing the
worst case gains and the original gains

8 q
>
> x0
x > 0, v > 0
>
> q a0
>
>
< 2 x0
x > 0, v < 0
a1 = q a0
(3.28)
>
> x (xmin )
x < 0, v > 0
>
> q a0
>
>
: x0
x < 0, v > 0
a0
8 r q
>
>
>
> 2 x0
+ a0 v0
x > 0, v > 0
>
> r
a0
>
> q
>
>
< 2 2 x0a0 + v2(vmin
a0
)
x > 0, v < 0
a2 = r q (3.29)
>
> x (xmin )
>
>
> 2 + v (va0 min )
x < 0, v > 0
>
> r q a0
>
>
>
: 2 x (xmin )
+ 2 v0a0 x < 0, v > 0
2 a0

where xmin and vmin are the minimum position error and the minimum velocity error, re-
spectively. In our case vmin is 15[m/sec] and xmin is 32[m].
Figure 3.15 shows the simulation results comparing the worst case gains and the original
gains. As shown in the figures, with the worst case gains the performance is worse at all four
quadrants in the sense of converging to the origin.
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 46

10

2
v[m/s]

-2

-4

-6

-8

-10
-10 -8 -6 -4 -2 0 2 4 6 8 10
x[m]

! Phase plane plot of the closed loop system with the worst case gains
Figure 3.16:

Figure 3.16 shows the simulation result with the worst case system described in (3.26) ⇠
(3.29). The initial point is (10, 0) and the simulation results shows that the next intersection
with the positive x axis is closer to the origin. Since the system is linear at each quadrant,
the sequence of intersections with the positive x axis converges to the origin and any initial
states will converge to the origin. The original system behaves better than the worst case
system at each quadrant, i.e. the solution of the original system is confined by the solution
of the worst case system, and the states will converge to the origin faster than the worst case
system.

3.6 Disturbance Observer for the Virtual Lead

Vehicle
In previous sections, the virtual lead vehicle scheme was introduced and studied as a switch-
ing scheme between the distance control and the velocity control modes. Previously, the
acceleration of the actual lead vehicle was considered as a disturbance and assumed to be
zero which is generally not true. The feedback controller eventually compensated the e↵ect
of the nonzero lead vehicle acceleration, however with a time delay. In this section a distur-
bance observer (DOB) is introduced which can efficiently estimate and compensate for the
actual lead vehicle acceleration.
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 47

Figure 3.17: Block diagram of the speed and the position errors between the virtual and the
actual lead vehicles

3.6.1 Model of the Virtual Lead Vehicle

Figure 3.17 describes the block diagram of the speed and the position errors between the vir-
tual and the actual lead vehicles when an actual lead vehicle exists in a detectable range. The
subscript v and l stand for the virtual lead vehicle and the actual lead vehicle, respectively.
x, v, and a mean the position, the velocity, and the acceleration, respectively.
In the previous section, the linear quadratic controller calculated the acceleration of
the virtual lead vehicle av based on the velocity error and the position error. The lead
vehicle acceleration al was considered as a disturbance and assumed to be zero. Without
a disturbance compensation method, nonzero lead vehicle accelerations will cause velocity
and position errors, which are eventually compensated by the feedback controller. However,
there is a time delay with the compensation. This can be very dangerous particularly if
the lead vehicle decelerates very fast. If the lead vehicle and the host vehicle are equipped
with communication devices, the control commands from the driver or the cruise control
system (if equipped and activated) of the lead vehicle can be transmitted and utilized by
the controller of the host vehicle. In general driving situations, it is hard to assume that
the lead vehicle is equipped with a communication device compatible with that of the host
vehicle. In that case, a DOB constructed for the error model in Figure 3.17 can be utilized
to estimate and compensate for the lead vehicle acceleration.

3.6.2 Disturbance Observer

Figure 3.18 describes the block diagram of a typical DOB [37, 55]. G(z) is the transfer
function of the actual plant, Gn (z) is the nominal plant transfer function, and r is the relative
degree of the plant. Q filter is usually a low pass filter which handles the measurement noise
and the model discrepancy. In general, the actual plant dynamics is not precisely known,
and the inverse of the nominal plant model is used in DOB.
The actual dynamics of the plant in the present problem is that of a double integrator.
Thus, in the present problem, there is no model uncertainty i.e. G(z) = Gn (z). We have
two outputs, the velocity error and the position error. When we utilize the position error as
the output, noting that the input is the acceleration error, the discrete time plant transfer
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 48

Figure 3.18: Block diagram of the DOB

function is:
t2 (z + 1)
Gn (z) = (3.30)
2(z 1)2
where t is the sampling time. The relative degree r is 1. Furthermore, d and u⇤ in Figure
3.18 are respectively al and af eedback , where af eedback is the desired acceleration of the virtual
lead vehicle calculated by the feedback controller. With the estimated disturbance, the
acceleration of the virtual lead vehicle is:

av = af eedback + âl (3.31)

where âl is the estimated lead vehicle acceleration by DOB. If the measurement noise is
neglected, the Q-filter can be set to 1 since there is no model uncertainty.
Figures 3.19 ⇠ 3.22 compare the simulation results with and without the DOB when there
are initial velocity and position errors and the actual lead vehicle accelerates or decelerates.
As shown in the figures, the DOB helps the host vehicle react faster to the lead vehicle
acceleration and reduces or practically eliminates the overshoot or the undershoot. Since
there is no model uncertainty, the closed loop of the DOB is stable with Q(z) = 1.
However in real circumstances, the measurement signals are contaminated by noise. With
no model uncertainty and Q(z) = 1, the transfer function of the closed loop system with the
DOB is:

Gaf eedback !e (z) = G(z) (3.32)

Gal !e (z) = (1 z 1 )G(z) (3.33)
Gn!e (z) = z 1 (3.34)

where e is the position error. Figure 3.23 shows the block diagram of the closed loop system.
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 49

14

13

12
velocity [m/s]

11

10

9 lead
w/ DOB
w/o DOB
8

7
0 5 10 15 20 25 30 35 40 45 50
time [sec]

Figure 3.19: Simulation results with and without the DOB when the lead vehicle accelerates:
velocity

20
w/ DOB
w/o DOB

15
position error[m]

10

-5
0 5 10 15 20 25 30 35 40 45 50
time [sec]

Figure 3.20: Simulation results with and without the DOB when the lead vehicle accelerates:
position error
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 50

10
lead
9 w/ DOB
w/o DOB
8

7
velocity [m/s]

1
0 5 10 15 20 25 30 35 40 45 50
time [sec]

Figure 3.21: Simulation results with and without the DOB when the lead vehicle decelerates:
velocity

20
w/ DOB
w/o DOB

15
position error[m]

10

-5
0 5 10 15 20 25 30 35 40 45 50
time [sec]

Figure 3.22: Simulation results with and without the DOB when the lead vehicle decelerates:
position error
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 51

Figure 3.23: Block diagram of the DOB

250
lead
200 virtual

150

100

50
velocity[m/s]

-50

-100

-150

-200

-250
0 5 10 15 20 25 30 35 40 45 50
time[sec]

Figure 3.24: Simulation with the measurement noise

(3.32) and (3.33) mean that the transfer function from the control input to the output
is the same with the original open loop transfer function and the disturbance is efficiently
rejected. These are the typical benefits of the DOB. However (3.34) means that the noise
is not rejected at all. As a result, the position of the virtual lead vehicle is contaminated
by the measurement noise and the velocity, which is the time derivative of the position, and
becomes very noisy.
A simulation study is performed to check the e↵ect of the noise. In the simulation the
velocity and the position of the lead vehicle are contaminated by zero-mean white Gaussian
noises. It is assumed that the noises of the position and the velocity measurements are
independent. As shown in Figure 3.24, the velocity of the virtual lead vehicle is very noisy
as expected. This also causes a problem for the feedback controller of the host vehicle. If
the feedback controller has a derivative controller, the control signal would be very noisy.
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 52

11
lead
10 virtual

7
velocity [m/s]

1
0 5 10 15 20 25 30 35 40 45 50
time [sec]

Figure 3.25: Simulation results with measurement noise using the velocity measurement:
velocity

To prevent this, the velocity error may be utilized as the output of the system instead of
the position error. The transfer function from the acceleration to the velocity is:
t
Gn (z) = (3.35)
z 1
Figures 3.25 ⇠ 3.27 show the simulation results with the plant given by (3.35). As shown
in the figures, the position signal of the virtual lead vehicle is very smooth and the velocity
of the virtual lead vehicle is a little noisy. Since the feedback controller of the host vehicle
utilizes the relative position error, the noise of the velocity signal does not a↵ect much.
However, the acceleration of the virtual lead vehicle, which can potentially be utilized as a
feedforward control signal, is very noisy as shown in Figure 3.27. To prevent this, the Q-filter
is designed to be a low pass filter.
Figures 3.28 and 3.29 show the simulation results with a Q-filter, given by:
✓ ◆3
0.01
Q(z) = (3.36)
1 0.99z 1
The velocity of the virtual lead vehicle is smoother. The acceleration is a little noisy,
but compared to the case without Q-filter (Figure 3.27) the noise level has significantly been
reduced. For smoother acceleration, the cuto↵ frequency of the Q filter can be lowered.
Doing so, however, slows the response and consequently reduces the performance of the
DOB. The Q filter should be tuned based on the properties of the actual measurement noise.
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 53

20

15
position error[m]

10

-5
0 5 10 15 20 25 30 35 40 45 50
time [sec]

Figure 3.26: Simulation results with measurement noise using the velocity measurement:
position error

25

20

15

10
acceleration [m/s 2]

-5

-10

-15

-20

-25
0 5 10 15 20 25 30 35 40 45 50
time [sec]

Figure 3.27: Simulation results with measurement noise using the velocity measurement:
acceleration of the virtual lead vehicle
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 54

10
lead
9 virtual

7
velocity[m/s]

1
0 5 10 15 20 25 30 35 40 45 50
time[sec]

Figure 3.28: Simulation results with a Q-filter: velocity

0.1

-0.1
acceleration [m/s 2]

-0.2

-0.3

-0.4

-0.5

-0.6
0 5 10 15 20 25 30 35 40 45 50
time [sec]

Figure 3.29: Simulation results with a Q-filter: acceleration of the virtual lead vehicle
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 55

10
lead
9 virtual
w/ avl
8 w/o avl

7
velocity [m/s]

1
0 5 10 15 20 25 30 35 40 45 50
time [sec]

Figure 3.30: Simulation results showing the vehicle behavior with and without the feedfor-
ward control when the lead vehicle decelerates: the lead and virtual vehicle‘s velocity

3.6.3 Utilizing the Lead Vehicle Acceleration

The acceleration of the virtual lead vehicle can be utilized as a feedforward control signal for
the host vehicle. Previously, the desired acceleration of the host vehicle was calculated by
a feedback controller. The acceleration of the lead vehicle is added to the feedback control
signal as:
adesired = af eedback + avl (3.37)
where adesired , af eedback , and avl are the desired acceleration, the feedback control signal, and
the acceleration of the virtual lead vehicle, respectively.
Figures 3.30 ⇠ 3.33 show the simulation results of the virtual and the host vehicle with
and without using the virtual lead vehicle acceleration as a feedforward control signal. As
shown in the figures, when the virtual lead vehicle acceleration is utilized as a feedforward
signal, the response of the host vehicle is faster and the error is smaller. Another advantage
of utilizing the virtual lead vehicle acceleration is that when the lead vehicle decelerates, the
position error of the host vehicle remains positive during most of the time. Positive position
error means that the relative distance between the host and the lead vehicles is larger than
desired and is hence more preferable than the negative errors. In the simulations, it is
assumed that the lower level controllers, i.e. the brake controller and the engine controller,
make the host vehicle accelerate or decelerate exactly with the desired acceleration.
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 56

0.4
w/ avl
0.3 w/o avl

0.2

0.1
position error[m]

-0.1

-0.2

-0.3

-0.4

-0.5
0 5 10 15 20 25 30 35 40 45 50
time [sec]

Figure 3.31: Simulation results showing the vehicle behavior with and without the feedfor-
ward control when the lead vehicle decelerates: the host vehicle‘s position

14

13

12
velocity [m/s]

11

10
lead
virtual
9 w/ avl
w/o avl

7
0 5 10 15 20 25 30 35 40 45 50
time [sec]

Figure 3.32: Simulation results showing the vehicle behavior with and without the feedfor-
ward control when the lead vehicle accelerates: the lead and virtual vehicle‘s velocity
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 57

1
w/ avl

0.8 w/o avl

0.6
position error[m]

0.4

0.2

-0.2

-0.4
0 5 10 15 20 25 30 35 40 45 50
time [sec]

Figure 3.33: Simulation results showing the vehicle behavior with and without the feedfor-
ward control when the lead vehicle accelerates: the host vehicle‘s position

3.6.4 Extended State DOB

In the previous section, only the velocity measurement is utilized to estimate the lead ve-
hicle acceleration. If the position measurement is equally good or better than the velocity
measurement in the sense of noise contamination, an estimation method better than DOB
may be devised. For this purpose, a Kalman filter with augmented states can be designed.
The lead vehicle acceleration is assumed to be continuous and be generated by a stochastic
process as:
al (k + 1) = al (k) + w(k) t (3.38)
The augmented system model is:
2 3 2 1
32 3 2 1 2 3 2 3
ex (k + 1) 1 t 2
t2 ex (k) 2
t 0
4 ev (k + 1) 5 = 4 0 1 t 5 4 ev (k) 5 + 4 t 5 av (k) + 4 0 5 w(k) (3.39)
al (k + 1) 0 0 1 al (k) 0 t
2 3
 ex (k) 
1 0 0 4 nx (k)
Y (k) = ev (k) 5 + (3.40)
0 1 0 nv (k)
al (k)
where ex and ev are the position error and the velocity error, and al and av are the ac-
celeration of the lead vehicle and the virtual lead vehicle, respectively. w, nx , and nv are
the zero mean process noise, the position measurement noise, and the velocity measurement
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 58

noise, respectively. They are uncorrelated and their covariances are ⇤ww , ⇤nx nx , and ⇤nv nv ,
respectively.
Assuming that the process noise and the measurement noises are stationary, a steady
state Kalman filter can be designed to estimate the augmented states. In DOB, the output
of the Q-filter is the estimated disturbance which in this case is the lead vehicle accelera-
tion. In the Extended State DOB(ESDOB) the lead vehicle acceleration is included in the
augmented states and estimated by the Kalman filter. Like the DOB, the estimated lead
vehicle acceleration is added to the feedback acceleration.
The DOB introduced in the previous sections utilizes only the velocity measurement.
The Kalman filter designed above utilizes both the velocity and the distance measurements.
It can be expected that the performance of the Kalman filter would be better than the
DOB since it uses more information. However such an expectation holds only if the position
measurement is as good as or better than the velocity measurement in terms of measurement
noise. This does not mean that the covariance of nx is smaller than that of nv . Since the
velocity is related to the position by a di↵erentiator given by:

x(k + 1) x(k)
v(k + 1) = (3.41)
t
The covariance of the position measurement noise is translated to the velocity measurement
noise by:
2
⇤⇤nv nv = ⇤n n (3.42)
t2 x x
If the covariance of the position measurement is fairly smaller than the velocity measure-
ment noise covariance, i.e. ⇤⇤nv nv < ⇤nv nv , the augmented Kalman filter should show better
performance than DOB.
Table 3.2 compares simulation results of di↵erent DOBs with same noise data where
the position measurement is better than the velocity measurement in the sense of noise
covariances. In the simulations, initial velocity and position errors are zero and the lead
vehicle decelerates. The root mean square (RMS) errors of the position and the velocity are
used to compare the performance and the RMS jerk of the estimated lead vehicle acceleration
is used to measure how much the control signals are noisy. Two di↵erent Q-filters are used
to compare both the performance and the noise cancellation.

✓ ◆2
0.01
Q1 (z) = 1
(3.43)
1 0.99z
✓ ◆3
0.01
Q2 (z) = 1
(3.44)
1 0.99z

The augmented Kalman filter is given by:

Ỹ o (k) = Y (k) C X̂ o (k) (3.45)

CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 59

RMS Position Error [m] RMS Velocity Error [m/s] RMS Jerk [m/s3 ]
DOB with Q1 0.0910 0.0252 2.4855
DOB with Q2 0.1316 0.0345 0.9976
ESDOB 0.0760 0.0203 0.9911
Table 3.2: Simulation results comparing the DOB and the Extended State DOB(ESDOB)

X̂(k) = X̂ o (k) + M C T [CM C T + V ] 1 Ỹ o (k) (3.46)

X̂ o (k + 1) = AX̂(k) + Bav (k) (3.47)
⇥ ⇤T
where X is the augmented state, i.e. X(k) = ex (k) ev (k) al (k) and M is the station-
ary a-prioi state estimation covariance which is the solution of:

AM AT AM C T [CM C T + V ] 1 CM AT M + Bw W BwT = 0 (3.48)

A, B, C, and Bw are from the system model:

2 3 2 1 2 3 2 3
1 t 1
t2 t  0
2 2 1 0 0
4
A= 0 1 t 5 ,B = 4 t 5 ,C = , Bw = 4 0 5 (3.49)
0 1 0
0 0 1 0 t

The input noise covariance and the measurement noise covariance are:

⇤ nx nx 0
W = ⇤ww , V = (3.50)
0 ⇤n v n v

This Kalman filter is called the extended state DOB in this thesis.
As shown in the simulation results, the performance of extended state DOB is better
than the typical DOB for both the disturbance compensation and the noise cancellation.
However, the extended state DOB is useful only when the position measurement is minimally
contaminated by noise. Note that with a short sampling time, ⇤nx nx must be very small
for ⇤⇤nv nv < ⇤nv nv to be hold. For further progress and implementation, more information
about the measurement sensor noise is required to check the e↵ectiveness of the extended
state DOB.

3.7 Summary
In this chapter, the virtual lead vehicle scheme was introduced. Instead of following the
actual lead vehicle, the host vehicle followed the virtual lead vehicle generated by the ACC
system. When there was a lead vehicle, the speed and the position of the virtual lead vehicle
were set to be the same with the speed and the position of the actual lead vehicle. When
no lead vehicle was detected, the speed of the virtual lead vehicle was set to be the desired
CHAPTER 3. VIRTUAL LEAD VEHICLE SCHEME 60

speed and the switching between the distance control and the velocity control modes was
prevented.
Also the virtual lead vehicle scheme was utilized for smooth transient during cutting
in/out of a lead vehicle. When the lead vehicle cut out or a new lead vehicle cut in from
a side lane, the virtual lead vehicle merged smoothly to the new lead vehicle and the host
vehicle followed the virtual lead vehicle and moved smoothly. A linear quadratic controller
with variable weights was designed to control the virtual vehicle in transient. With the
variable weights, the virtual vehicle was controlled to move smoothly when it was a safe
situation and react faster when it was a dangerous situation. The stability of the linear
quadratic control with variable weights was proved by a phase portrait method.
In the transient motion, the virtual lead vehicle was modeled as a double integrator and
the lead vehicle acceleration was assumed to be zero which is not always true. A disturbance
observer (DOB) was utilized to estimate and compensate the actual lead vehicle acceleration.
When the lead vehicle acceleration was estimated and compensated by adding a feedforward
control input to the virtual lead vehicle, the virtual lead vehicle showed faster and safer
performance. Extended state disturbance observer was studied for the case when both the
position and the velocity measurements were utilized in the DOB. Extended state DOB
showed better performance than the original DOB when the position measurement noise
covariance was smaller than the equivalent velocity measurement noise covariance.
 61

Chapter 4

Optimal Profile Generation for

Stop-and-Go

4.1 Introduction
While the conventional ACC is intended to assist the driver at relatively higher speeds, the
ACC system with Stop-and-Go is intended to be used at low speeds and should have a
capability to stop the vehicle completely if there is a stopped lead vehicle [38]. Since it is
expected to be utilizable at low speed in urban driving situations, the ACC system with
Stop-and-Go should be able to properly react to frequent cutting in and out of lead vehicles
and this feature is studied with the virtual lead vehicle scheme in 3. Another necessary
feature is the complete stop.
Various research have been done about the ACC with Stop-and-Go. Fuzzy control algo-
rithms were used in [11] and [32]. Linear quadratic optimal controller with mode switching
between the speed control algorithm and the distance control algorithm was used in [63,
64]. Sliding controller was utilized in [1] and [6]. Various feedback control laws based on
the velocities of the lead and the host vehicles and the relative distance were also used [28,
41]. Explicit Model Predictive Control was also applied to ACC with Stop-and-Go in [34].
In most of the previous research, the ACC with Stop-and-Go systems were designed and
verified that they were utilizable in certain speed range including low speeds. However, the
passenger comfort was not considered during a complete stop.
As studied in Section 2.4, the desired relative distance to the lead vehicle is calculated
from the time-headway and the safety distance. i.e. ddesired = thw vhost + dof f set With the
constant time-headway, the string stability is guaranteed. However, it can cause a problem
when the lead vehicle is stopped completely. When the lead vehicle is stopped completely,
the host vehicle should be stopped dof f set behind the lead vehicle. However at the moment
when the lead vehicle is stopped, the desired distance, which is proportional to the velocity
of the host vehicle, is larger than dof f set and the controller tries to maintain larger distance
to the lead vehicle than it actually should. Figure 4.1 shows the simulation result for a
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 62

12
lead
host
10

8
velocity[m/s]

−2
0 10 20 30 40 50
time[sec]

Figure 4.1: Velocity of the lead and the host vehicle in case of a complete stop

complete stop when a PID controller is used and the desired relative distance is determined
by the constant time-headway. As shown in the figure, when the lead vehicle is stopped,
initially the host vehicle still decelerates very fast and maintains very slow velocity until it
stops completely. It takes more than 10 seconds for the host vehicle to completely stop after
the lead vehicle has stopped. This motion can make the driver feel uncomfortable and can
cause more wear on brakes.
To prevent this unnecessary motion when the lead vehicle completely stops, an optimal
acceleration profile should be calculated and the PID feedback controller should be replaced
by an alternative controller which has a better tracking capability. Sliding controller is
a possible candidate. However, the optimization problem should be solved in real time to
calculate the optimal profile for the sliding controller. Fortunately, the optimization problem
formulated in this chapter is a quadratic programming problem and can be converted to a
linear complementarity problem. The linear complementarity problem can be solved fast
with the Lemke algorithm if the dimension of the problem is small. In this thesis, a multi-
resolution formulation is suggested to reduce the dimension of the problem without changing
the fundamental characteristic of the problem and to solve it very fast.
This chapter is organized as follows. Optimization problems are formulated to find the
optimal relative distance profile during a complete stop and the optimal velocity profile
during a starting motion in Section 4.2. In Section 4.3, a sliding mode controller is designed
for a complete stop and the virtual lead vehicle scheme is also utilized for the starting motion.
In Section 4.4, algorithms solving the optimization problem are studied and multi-resolution
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 63

Figure 4.2: Sketch of the complete stop

formulation is suggested to solve the problem in real time. Section 4.5 summarizes the
chapter.

4.2 Optimal Relative Distance Profile during a

Complete Stop
4.2.1 Optimization Problem Formulation
To find the optimal trajectory, an optimization problem is formulated and solved. In this
section, a linear model is used to formulate the problem into a Quadratic Programming (QP)
optimization problem.
The complete stop scenario is described in Figure 4.2. The lead vehicle is completely
stopped in front of the host vehicle at position xlead . dof f set is the desired relative distance
to the lead vehicle when the host vehicle is completely stopped behind the lead vehicle. The
relation between the position, the velocity, and the acceleration of the host vehicle is a double
integrator given by:
Z t
v(t) = v(0) + a(⌧ )d⌧ (4.1)
0
Z t
x(t) = x(0) + v(⌧ )d⌧ (4.2)
0

where x(t), v(t) and a(t) are the position, the velocity, and the acceleration at time t, respec-
tively. Using (4.1) and (4.2), however, makes the problem infinite dimensional and generally
hard to solve. To get the discrete time formulation, it is assumed that the acceleration is
constant for each time step t. Then the double integrator can be expressed as:
    1 2
xk+1 1 t xk t
= + 2 ak (4.3)
vk+1 0 1 vk t
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 64

where xk and vk are the position and the velocity at each time, i.e. xk = x(k t), and the
ak is the constant acceleration during the time step, i.e. a(t) = ak for (k 1) t  t < k t.
The time horizon should be fixed as a constant Tf < 1 to make the problem finite
dimensional and to find the solution such that the host vehicle stops in finite time after a
reasonable time duration. The initial conditions x(0) and v(0) are assumed to be known,
and the final conditions are x(Tf ) = xlead dof f set , v(Tf ) = 0, and a(Tf ) = 0. The cost
which is minimized in the optimization problem should penalize not only the stopping time
but also the acceleration and jerk to improve the passenger comfort. There should also be
maximum and minimum limits for the acceleration and the jerk. One simple cost function
choice to make the optimization problem easily solvable is to use a quadratic cost function
that penalizes distance error, acceleration, and jerk given by:
J(x, a, j) = xl k22 + a kak22 + j kjk22
x kx (4.4)
⇥ ⇤T ⇥ ⇤T
where x xl = x1 xl x2 xl · · · xN xl , xl = xlead dof f set , a = a1 a2 · · · aN ,
⇥ ⇤T
and j = j1 j2 jN 1 . jk is the jerk which is given as jk = (ak+1 ak )/ t and
N = Tf / t. Since the jerk can be represented as a linear function of the acceleration, the
cost function can be represented only as a function of the acceleration by:
J(x, a) = x kx xl k22 + aT a IN ⇥N + jJ
T
J a (4.5)
where 2 3
1 1 ··· 0
0
. . . .. 7
1 6
6 0 1 1 . 7
J= 6 .. .. .. .. 7 (4.6)
t4 . . . . 0 5
0 ··· 0 1 1
Since, the cost function in (4.5) is quadratic in terms of the acceleration and is strictly
convex, it can be solved in finite time[7]. The algorithm will be reviewed in Section 4.4. To
reduce the time to completely stop, x xl should converges to zero quickly. To enforce it,
l0 norm can be introduced. The l0 norm of a vector is defined as the number of nonzero
elements of it. If the l0 norm of a vector is penalized in the optimization problem, the
solution becomes sparse i.e. many elements are zero. However, the l0 norm of a vector is
not convex and it is not guaranteed that the optimization problem is solvable. Instead the
l1 norm which is a convex relaxation of the l0 norm can be used. The l1 norm of a vector is
defined as:
XN
ksk1 = |si | (4.7)
i=1
1
By penalizing the l norm of the relative distance in the cost function, the relative distance
of the solution becomes sparse. Since the relative distance should always be larger than or
equal to zero, penalizing the l1 norm of the relative distance works as a time penalization.
The modified cost function is:
J(x, a) = aT a IN ⇥N + jJ
T
J a+ x kx xl k1 (4.8)
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 65

Considering the passenger comfort, the inequality constraints for the acceleration and
the jerk are introduced. Also, to prevent unreasonable motions, i.e. moving backward, the
inequality constraints for the velocity are introduced. Also, the linear model equation and
the initial conditions are added as equality constraints.

xi  xl (4.9)

vi 0 (4.10)
amin  ai  amax (4.11)
jmin  ji  jmax (4.12)
    1
xk+1 1 t xk t2
= + 2 ak (4.13)
vk+1 0 1 vk t
    1
x1 1 t x0 t2
= + 2 a0 (4.14)
v1 0 1 v0 t
where x0 , v0 , and a0 are the initial position, velocity, and acceleration of the host vehicle.
The optimization problem is given with (4.8) and (4.9) ⇠ (4.14). However it can be
simplified to be solved efficiently. The problem has 3N variables, i.e. x 2 RN , v 2 RN ,
and a 2 RN . However (4.8) is only a function of the acceleration while the position and the
velocity are related through (4.13) and (4.14). The position vector and the velocity vector
can be expressed as an affine function of the acceleration vector as (4.15) and (4.16) to reduce
the problem dimension.
2 3 2 1 t2 0 0 · · · 0
3
x0 + v0 t 2
6 x0 + 2v0 t 7 6 3 1 .. .. 7
6 7 6 t2 t2 0 . . 7
6 x0 + 3v0 t 7 6 5 2 6 2 2
. 7
x=6 7+6 2 t 3
t 2 1
t 2 . . 0 7 a , X0 + Xa (4.15)
6 .. 7 6 2 2 7
4 . 5 4 .. ... .. .. 7
. . . 0 5
x0 + N v 0 t 2N 1
t2 2N 3 t2 · · · 3 t2 1 t2
2 2 2 2
2 3 2 3
v0 t 0 ··· 0
6 7 6 . .. 7
6 v0 7 6 t t .. . 7
v=6 ..7+6 . .. . . 7 a , V0 + V a (4.16)
4 .5 4 .. . . 0 5
v0 t t ··· t
From (4.9), kx xl k1 in the cost function can be represented as 1TN (x xl ) where
⇥ ⇤T
1N = 1 · · · 1 2 RN ⇥1 . Also from (4.10), the velocity is always greater than or equal
to zero and hence, (4.9) can be reduced to a single constraint xN xl  0. Finally, the
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 66

x[m]
2

0
0 1 2 3 4 5
time[sec]
1
l−1
v[m/s]

l−2
0.5

0
0 1 2 3 4 5
time[sec]
0.5
a[m/s ]
2

−0.5
0 1 2 3 4 5
time[sec]

Figure 4.3: Solution of the QP comparing l 1 and l 2 norms, x = 1, a = 1, j = 0.5

problem can be expressed as:

minimize aT ( a IN ⇥N + jJ
T
J)a T
x 1N Xa
a2RN
subject to (X0 + Xa)N  xl
(V 0 + V a)i 0 (4.17)
amin  ai  amax
jmin  (Ja)i  jmax

The final problem is in a form of a linear constrained Quadratic Program (QP).

Figure 4.3 shows the solutions of the QP with l 1 and l 2 norms of the x xl . As
shown in the figure, with l 1 norm, the velocity converges to zero faster. This comes from
the sparsity of the x kx xl k1 . Figure 4.4 shows the solutions with di↵erent x ’s. With
very larger x , the solution is very similar to the bang-bang control which only utilizes the
maximum or the minimum values of the acceleration and the jerk. With larger x or smaller
a and j , the trajectory is similar to aggressive driving while with smaller x or larger
a and j , it is more like conservative driving. This weights can be tuned based on the
preference of the driver.
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 67

x[m]
2

0
0 1 2 3 4 5
time[sec]
1
=1
x
v[m/s]

0.5 = 100
x

0
0 1 2 3 4 5
time[sec]
1
a[m/s2]

−1
0 1 2 3 4 5
time[sec]

Figure 4.4: Solution of the QP comparing x = 1 and x = 100, a = 1, j = 0.5

4.2.2 Velocity Profile of a Starting Motion

Once the vehicle completely stops, when the lead vehicle resumes its motion the host vehicle
should also resume its motion and follow the lead vehicle. However, due to safety and
comfort issue, it is not preferable to directly follow the lead vehicle. Instead, an optimal
velocity profile can be designed and the host vehicle or the virtual lead vehicle can start
their motion with the optimal profile. In this case the acceleration, the jerk, and the velocity
error should be considered in the cost function:
2 2
minimize a kak2 + j kjk2 + v kV a v f k1
a2RN
subject to (V a)i  vf (4.18)
amin  ai  amax
jmin  (Ja)i  jmax

where vf is the desired final velocity and V is the matrix defined in (4.16).
Figure 4.5 shows the simulation result with various j ’s. The initial velocity, the initial
acceleration, and the desired final velocity are 0[m/sec], 0[m/sec2 ], and 5[m/sec], respec-
tively. The maximum acceleration is given by 1[m/sec2 ], and the maximum magnitude of
the jerk is 2[m/sec3 ]. The final time is 10 sec. As shown in Figure 4.5 with larger j , the
jerk is penalized more and as a result the acceleration profile becomes smoother. However,
it will make the velocity converge to its final value a little slower.
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 68

40

x[m]
20

0
0 2 4 6 8 10
time[sec]
10
v[m/s]

0
0 2 4 6 8 10
time[sec] = 0.5
j
2 = 0.005
j
a[m/s2]

−2
0 2 4 6 8 10
time[sec]

Figure 4.5: Optimal profile generation result for starting motion

4.3 Sliding Control and Virtual Lead Vehicle Scheme

In Section 4.2, the optimal relative distance profile during a complete stop and the optimal
velocity profile during a starting motion were designed. The design problem to follow is track-
ing the desired optimal profiles, for which sliding mode control is promising. Sliding mode
control is a tracking control method which guarantees a convergence to the desired states in
a finite time [48]. In this section, a sliding mode controller for velocity tracking is designed
and verified by simulations. To remove a high-frequency oscillation due to the discontinuity
of sliding mode controller, a smooth sliding mode controller is designed. Simulation results
show that the smooth sliding mode controller shows a good tracking performance. When
the vehicle starts moving after a complete stop, the optimal velocity profile can be utilized
with the sliding controller. However, in many cases it is expected that there is an actual
lead vehicle moving slower than the desired speed. To react this lead vehicle, instead of the
sliding controller, a virtual lead vehicle utilizes the optimal velocity and acceleration.

4.3.1 Sliding Controller Design for a Complete Stop

To design a sliding mode controller, the dynamics of the system should be examined. For
simplicity, it is assumed that only the brake is utilized during the stopping motion, and the
dynamics of the brake is first order:
1
Gbrake (s) = (4.19)
⌧b s + 1
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 69

where ⌧b is the time constant of the brake. The state space equations of the system are given
by:
x1 = drelative = d0 x (4.20)
ẋ1 = x2 = ẋ = v (4.21)
ẋ2 = x3 = ẍ = a (4.22)
... 1 1
ẋ3 = x4 = x = u+ a (4.23)
⌧b ⌧b
where d0 is the relative distance to the stopped lead vehicle, x is the position of the host
vehicle from the initial position, v is the velocity of the host vehicle, a is the acceleration of
the host vehicle, and u is the control input. The objective of the sliding control is to make
the x1 track the previously designed desired trajectory. The relative degree of the system is
three, and we can define a sliding surface as:
✓ ◆2
d
S= + (ẋ1 d˙d ) (4.24)
dt
where dd is the desired relative distance trajectory. The sliding condition is represented by:
S Ṡ  ⌘|S| (4.25)
To satisfy this condition, the time derivative of the sliding surface can be chosen as:
Ṡ = ⌘sgn(S) (4.26)
Since, Ṡ is a function of the control input u, the control input is determined from (4.26) as:
✓ ◆
... 1
d d + a + 2 ( d¨d a) + ( d˙d v) + ⌘sgn(S)
2
u = ⌧b (4.27)
⌧b
(4.27) gives the sliding control law.

4.3.2 Simulation Results for a Complete stop

Figures 4.6 and 4.7 show the simulation result of the sliding mode controller with = 0.1
and ⌘ = 1. As shown in Figure 4.6, the sliding mode controller let the relative distance
follow the desired trajectory very accurately. However, there is an undesirable oscillation
in the acceleration and the control input as shown in Figure 4.7. Also the relative distance
error does not converge exactly to zero but keeps oscillating near zero and that makes the
control signal nonzero as shown in Figure 4.7. This is due to the sign function in (4.27)
which introduces discontinuity in the control input.
To prevent the undesired oscillation, a smooth sliding controller can be introduced [48].
Instead of the sign function, a saturation function is used:
✓ ✓ ◆◆
... 1 ¨ 2 ˙ S
u = ⌧b d d + a + 2 ( dd a) + ( dd v) + ⌘sat (4.28)
⌧b
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 70

2
relative distance
1.8 desired relative distance

1.6

1.4

1.2
d[m]

0.8

0.6

0.4

0.2

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
time [sec]

Figure 4.6: Simulation results with the sliding mode controller: relative distance

0.2
acceleration[m/s 2]

-0.2

-0.4

-0.6
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
time [sec]

0.2
control signal[m/s 2]

-0.2

-0.4

-0.6
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
time [sec]

Figure 4.7: Simulation results with the sliding mode controller: acceleration and control
command
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 71

2
relative distance
1.8 desired relative distance

1.6

1.4

1.2
d[m]

0.8

0.6

0.4

0.2

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
time [sec]

Figure 4.8: Simulation results with the smooth sliding mode controller: relative distance

where the saturation function is:

8
< 1 x 1
sat(x) = x 1<x<1 (4.29)
:
1 x 1

in (4.28) defines the boundary layer thickness. Large reduces the oscillation but
worsens the guaranteed precision [48]. In the simulation = 0.1 is used.
Figures 4.8 and 4.9 show the result with the smooth sliding control law. The tracking
performance is almost the same as the sliding mode controller with the sign function, while
there is no undesirable oscillation.

4.3.3 Virtual Lead Vehicle for a Starting Motion

Sliding Mode Controller for Starting Motion
The sliding mode control scheme can be applied for a starting motion. The velocity profile
is designed by solving an optimization problem introduced in Section 4.2.2. For the starting
motion, it is assumed that only the engine is utilized and the dynamics of the engine is first
order:
1
Gengine (s) = (4.30)
⌧e s + 1
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 72

0.2

acceleration[m/s 2]
0

-0.2

-0.4

-0.6
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
time [sec]

0.2
control signal[m/s 2]

-0.2

-0.4

-0.6
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
time [sec]

Figure 4.9: Simulation results with the smooth sliding mode controller: acceleration and
control command

where ⌧e is the time constant of the engine. The state space equations of the system are
given by:

x1 = v (4.31)
ẋ1 = x2 = a (4.32)
1 1
ẋ2 = x3 = ȧ = u a (4.33)
⌧e ⌧e
The relative degree of the system is two, and we can define a sliding surface as:
✓ ◆
d
S= + (ẋ1 v̇d ) (4.34)
dt

where vd is the desired velocity. The sliding surface can be chosen as:

Ṡ = ⌘sgn(S) (4.35)

To prevent the undesired oscillation, the smooth sliding control input is:
✓ ✓ ◆◆
1 S
u = ⌧e v̈d + a + (v̇d a) ⌘sat (4.36)
⌧e
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 73

16

14
actual
desired
12

10

0
0 5 10 15 20 25 30 35 40 45 50

Figure 4.10: Simulation result for a starting motion with a sliding mode controller

Simulation Results and Virtual Lead Vehicle Implementation for a Starting

Motion
Figure 4.10 shows the simulation result with the sliding mode controller. As shown in the
figure, the host vehicle follows the desired velocity profile very accurately. In this simulation,
the desired final velocity is 15[m/sec]. However, in the actual driving situation, the velocity
of the lead vehicle can be slower than the desired velocity and in that case if the lead vehicle
is close to the host vehicle, the host vehicle should follow the lead vehicle rather than tracking
the desired velocity profile. One possible solution is introducing a switch which utilizes the
sliding controller when the lead vehicle is far ahead and uses the PID controller when the
lead vehicle is close. However, this method requires an additional switch and makes the
structure of the system more complicated. Instead of using a switch, the virtual lead vehicle
scheme unnecessitates such a switch.
Figure 4.11 shows the simulation result with a virtual lead vehicle driven at the optimal
velocity profile. As shown in the figure, the host vehicle follows the virtual lead vehicle
and the velocity, the acceleration, and the jerk profiles of the host vehicle are similar to the
optimal profiles. The virtual lead vehicle scheme has its benefit especially when there is an
actual lead vehicle driven slowly.
Figures 4.12 and 4.13 show the simulation result of the virtual lead vehicle scheme when
there is a slower lead vehicle. As shown in the figure, initially, the virtual lead vehicle
accelerates according to the optimal velocity profile. However, when the actual lead vehicle
is detected the virtual lead vehicle decelerates and follows the actual lead vehicle. The
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 74

20

v[m/s]
10 virtual lead
host
0
0 5 10 15 20 25 30 35 40 45 50
time[sec]
2
virtual lead
a[m/s 2]

1 host

0
0 5 10 15 20 25 30 35 40 45 50
time[sec]
0.5 virtual lead
jerk[m/s 3]

host
0

-0.5
0 5 10 15 20 25 30 35 40 45 50
time[sec]

Figure 4.11: Simulation result with a virtual lead vehicle utilizing the optimal velocity profile

host vehicle follows the virtual lead vehicle and eventually follows the actual lead vehicle
smoothly.

4.4 Real Time Calculation of Quadratic

Programming Problem
Since the optimization problems introduced in the previous sections are all convex and linear
or quadratic programming problems, they are solvable using general convex optimization
tools like CVX. However, CVX is a tool designed for general convex optimization problems
and is too large and slow to embed in a controller. Quadratic programming problems with
linear equality and linear inequality constraints can be transformed to linear complementary
problems if the matrix representing the quadratic term in the cost function is positive definite.
The linear complementary problem can be solved by Lemke algorithm which shows faster
computation speed if the variable is not too large [12].
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 75

4
v[m/s]

actual lead
1
virtual lead
host

0
0 5 10 15 20 25 30 35 40 45 50
time [sec]

Figure 4.12: Simulation result with a virtual lead vehicle and a slow lead vehicle: velocity

14

12

10
distance error [m]

0
0 5 10 15 20 25 30 35 40 45 50
time [sec]

Figure 4.13: Simulation result with a virtual lead vehicle and a slow lead vehicle: distance
error
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 76

4.4.1 Algorithms to Solve the Constrained QP

The optimization problem (4.17) can be expressed as a standard quadratic programming
problem with linear constraints as:
1 T
minimize y Qy + f T y
y 2 (4.37)
subject to Aineq y bineq  0
where Q is a positive definite symmetric matrix.
Since (4.37) is (strictly) convex, it can be solved e↵ectively. One approach is the interior
point algorithm which is generally used to solve convex optimization problems, and another
is the Lemke algorithm which can solve the corresponding linear complementarity problem
of (4.17). In this section, both algorithms will be introduced and compared.

Interior Point Algorithm (Barrier Method) [7]

The interior point algorithm (barrier method) is one of the most popular algorithms solving
convex optimization problems. The interior point algorithm converts the problem with
inequality constraints to a unconstrained problem by introducing the log barrier function
instead of the inequality constraints and starting with a strictly feasible initial point. It is a
reliable method and has some better features than the Lemke algorithm especially for real
time application.
Generally, a constrained optimization problem is more difficult to solve than a prob-
lem without constraints. Some problems like Linear Programming (LP) problems can be
solved by the simplex method without modifying the constraints. However it is hard to find
the algorithm which can generally solve optimization problems with inequality constraints.
Instead, the problems are converted to an unconstrained problem which can be solved by
many di↵erent algorithms such as the steepest descent method or the Newton method. The
interior point method introduces the log barrier function to the cost instead of the inequality
constraints. In the quadratic programming problem described in the previous section, the
linear inequality constraints are:
Aineq y bineq  0 (4.38)
⇥ ⇤T
where Aineq = aineq,1 · · · aineq,m . The log barrier function for these inequality con-
straints is: m
X
(y) = log( aTineq,j y + bineq,j ) (4.39)
j=1

Due to the property of the log function, if the inequality constraints are about to be violated,
the log barrier function becomes arbitrarily large. The log barrier function is added to the
cost function to make the problem with only equality constraints as:
1
minimize t( y T Qy + f T y) + (y) (4.40)
y 2
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 77

t in the cost function is the weight between the original cost and the log barrier function.
Initially, t is small and is increased as the iteration goes on. With a very large t, the solution
is almost the same with the original problem.
The unconstrained optimization problem (4.40) can be solved by the Newton method.
In the Newton method, the gradient and the Hessian are used to calculate the searching
direction. From (4.39) and (4.40), the gradient and the Hessian are:
✓ ◆
1 T
r t( y Qy + f y) + (y) = t(Qy + f ) + ATineq d
T
(4.41)
2
✓ ◆
1 T
r t( y Qy + f y) + (y) = tQ + ATineq diag(d)2 Aineq
2 T
(4.42)
2
where di = 1/(bineq,i aTineq,i y). For each iteration, the gradient and the Hessian are calculated
and the descent direction is given as:

vk = (tQ + ATineq diag(d)2 Aineq ) 1

t(Qy + f ) + ATineq d (4.43)

The solution point is updated by:

y k+1 = y k + ↵k v k (4.44)

where ↵k is the step size at k-th iteration. There are many di↵erent ways to decide the step
size such as the constants step size, the line search, and Armijo rule. In this section, Armijo
rule [5] is used. After each iteration, the t is increased and the iteration stops when the
di↵erence between the new solution and the previous solution is smaller than the tolerance
or the iteration number reaches its maximum value. This procedure is called Phase II.
Since the log function is not defined for the values smaller than or equal to zero, the
Phase II should start with a strictly feasible initial point. To find the strictly feasible point,
another optimization problem is solved, which is called Phase I. The problem is:

minimize s
y,s
(4.45)
subject to Aineq y bineq  s1m
⇥ ⇤T
where 1m = 1 · · · 1 2 Rm⇥1 .
The Phase I problem can also be solved by the same scheme with Phase II where the
initial point can be y = 0, s = max( bi ). It is not necessary to solve the Phase I completely
and the iteration can be stopped when the value of s becomes negative. If the optimal s
is nonnegative, the problem is infeasible. In that case, the constraints needs to either be
relaxed, i.e. bineq,new = µbineq , µ > 1, or reformulated.
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 78

Lemke Algorithm
A QP problem with linear constraints can be converted to a linear complementarity problem
which can be efficiently solved by the Lemke algorithm. The Lemke algorithm utilizes
tableaux and pivoting as a linear programming solver and is efficient especially when the
problem size is small.
Since the problem is convex with only linear constraints, strong duality holds and the
Karush-Kuhn-Tucker (KKT) conditions are sufficient for the primal and dual optimal point.
The KKT conditions for the problem (4.17) are:

Aineq y ⇤ bineq  0 (4.46)

⇤
0 (4.47)
⇤ T ⇤
i aineq,i y bineq,i = 0 f or i = 1, 2, ..., m (4.48)
Qy ⇤ + f + ATineq ⇤
=0 (4.49)
where aTineq,i s are rows of Aineq . Since Q is strictly positive definite, x⇤ can be represented as
affine functions of ⇤ and the KKT conditions (4.46) ⇠ (4.49) are reduced to:
⇤
0 (4.50)
⇤
M +q 0 (4.51)
⇤ ⇤
(M + q)T =0 (4.52)
where
M = Aineq Q 1 ATineq (4.53)
q = bineq + Aineq Q 1 f (4.54)
The problem described by (4.50) ⇠ (4.52) is in the form of a standard linear complemen-
tarity problem.
Lemke algorithm is a method solving the linear complementarity problem by utilizing the
tableaux and the pivoting scheme. The complexity of the Lemke algorithm is about twice
of Gaussian elimination [12] which is efficient when the size of the problem is small. In this
research, the lemke algorithm program introduced in [12] is modified and used.

Comparing Interior Point Algorithm and Lemke Algorithm

Table 4.1 compares the calculation time of the Lemke algorithm and the interior point algo-
rithm. As shown in the table, the calculation time of the interior point algorithm increases
almost linearly with the problem size while that of the Lemke algorithm increases expo-
nentially. So, the interior point method has its advantage in the calculation time when the
size of the variable is large. However if the dimension of the problem is small, the Lemke
algorithm can solve the problem faster than the interior point method.
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 79

XXX
XXX t/Tf
XXX 0.05 / 10 0.05 / 5 0.1 / 10 0.1 / 5
I.C. XXX
1 / 0.5 / -0.1 8.745 / 2.176 0.4363 / 0.7434 1.008 / 0.8485 0.1756 / 0.9189
2 / 1 / -0.2 7.818 / 1.941 0.2906 / 0.7234 0.8616 / 0.6474 0.0830 / 0.9787
2/1/0 8.313 / 1.804 0.3528 / 0.6694 0.8750 / 0.6917 0.1048 / 0.9707
5 / 3 / -0.7 25.86 / 1.819 0.4796 / 0.5480 1.406 / 0.7114 0.1589 / 0.9041
Table 4.1: Calculation time [sec] of the Lemke algorithm / Interior point algorithm (Initial
Condition : d0 [m],v0 [m/s],a0 [m/s2 ])

Observing the result in the last column, it can be observed that the calculation time
of the interior point algorithm is longer than the second last column whose dimension is
larger. If checking the calculation time of the Phase I and Phase II of the interior point
method separately, it is found that the majority of the time is consumed in Phase I to find
the feasible initial point. Since the interior point method should start with a strictly feasible
initial point, the shorter time horizon and the longer time step make it hard to find a strictly
feasible point due to the acceleration and jerk constraints. Hence, it takes longer for the
interior point method to solve the problem.
However, it does not necessarily mean that the Lemke algorithm is better than the
interior point method when the initial point is close to the infeasible region. One drawback
of the Lemke algorithm is that the performance of the Lemke algorithm is poor if the initial
condition is poorly conditioned. If the initial distance is very short and the initial velocity is
fast, the host vehicle cannot stop exactly at the desired point with the limited acceleration
and jerk. In that case, the initial condition is infeasible and no solution exists for the
optimization problem. However, the Lemke algorithm cannot distinguish the infeasible initial
condition until it solves the problem and finds out that the solution is infeasible. Also, for
some initial conditions which are feasible but close to being infeasible, the solution of the
Lemke algorithm does not converge. Figures 4.14 and 4.15 show the simulation results with
the initial conditions (2[m],1[m/sec],0.05[m/sec2 ] ) and t = 0.25[sec] which is very close
to an infeasible problem. Although the interior point took longer to solve the problem, the
Lemke algorithm failed to solve the problem.
Also, due to the log barrier, the intermediate solutions of the interior point algorithm are
always feasible. Figures 4.14 and 4.15 show that the intermediate solutions of the interior
point methods are always feasible while those of the Lemke algorithm are infeasible. Since
the intermediate solutions of the interior point algorithm are always feasible, the interior
point algorithm has an advantage for real time applications. Also, Figures 4.14 and 4.15
show that the the interior point algorithm converges exponentially to the final solution while
the Lemke algorithm converges linearly. In the real time application, even if the calculation
is not finished on time, the intermediate solution at that time can be utilized since it is still
feasible and close to the optimal solution. For real time applications, the accuracy of the
solution can be sacrificed to reduce the calculation time. To do that, the iteration number
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 80

0
relative distance [m]

-1

-2

-3
Final Solution

-4

-5

-6

-7
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
time [sec]

Figure 4.14: Intermediate and final solutions of the Lemke algorithm

1.8

1.6

1.4

1.2
relative distance [m]

0.8

0.6

0.4

0.2

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
time [sec]

Figure 4.15: Intermediate and final solutions of the interior point algorithm
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 81

or the calculation time can be limited or the algorithm can be terminated once the tolerance
has been met. However those schemes are not applicable with the Lemke algorithm since
the intermediate solutions are infeasible.
One of the biggest advantages of the Lemke algorithm is that it is very fast when the
size of the problem is small. In this thesis, instead of using the intermediate solution, the
size of the problem will be reduced and the Lemke algorithm will be utilized. The scheme
to reduce the problem size will be introduced in the next section.

4.4.2 Multi-resolution Optimization Problem Formulation

In the previous section, the interior point method and the Lemke algorithm are suggested as
candidates for the solving the quadratic programming problem (4.37). As shown in 4.1, the
dimension of the problem has a significant e↵ect on the calculation time. In this section, a
multi-resolution problem formulation will be introduced as a method to find a suboptimal
solution in real time. The multi-resolution scheme has been used to reduce the number of
meshes for computational fluid dynamics [53] and to compress the size of the image in image
processing [44]. However, it has not been utilized in the area of real time optimization or
real time trajectory generation. The suggested multi-resolution scheme is also utilizable for
Model Predictive Control problems.

Applications of the Optimal Solution

In previous sections, sliding control and virtual lead vehicle scheme were suggested as candi-
date control schemes which can control the vehicle to follow the optimal velocity or position
profile. Both control schemes utilize the optimal profile as a trajectory that the output of
the system should follow. The optimal profile and its time derivatives are used in the sliding
controller. When sliding controller is used, the optimization problem is solved only once and
the solution is saved and utilized for the following time steps. One thing to notice is that in
both controller, only the first part of the optimization problem is utilized at the time when
the optimization problem is solved.

Multi-resolution Problem Formulation

As shown in the previous section, the calculation time significantly depends on the dimension
of the optimization problem. If the controller time step t is smaller, the dimension of the
optimization problem becomes larger and it takes longer to solve the problem. However,
since both controller utilize the first part of the optimal solution, it is possible to decrease
the dimension of the optimization problem by using larger time step for later time horizon.
    1 2
xk+1 1 t xk t
= + 2 ak (4.55)
vk+1 0 1 vk t
(4.55) was utilized in previous sections and it was assumed that the accelerations are constant
during each time step t. In that case the time step was a fixed constant and the dimension
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 82

Figure 4.16: Sketch of the multi-resolution problem concept

of the problem was Tf / t. However if a smaller time step is used for the initial time horizon
and a longer time step for later time, the problem size can be decreased. Figure 4.16 shows
the concept of the multi-resolution problem. If the time horizons is 5 second and a constant
time step of 0.2 second is used, the size of the problem is 25. However, if the time step is 0.2
second for the first 1 second and 1 for the later 4 seconds, the size of the problem becomes
9.
The optimization problem is in a same form with (4.37), however the matrices in the
weighting function and the constraints are modified according to the multi-resolution time
step.

Applications and Simulation Results

Figure 4.17 compares the solution with a uniform time step and a multi-resolution formu-
lation. The uniform time step has a sampling time of 10[msec]. For the multi-resolution
formulation, 10[msec] is used for the first 100[msec], 100[msec] is used from 0.1[sec] to 1[sec],
and 1[sec] is used from 2 to 5[sec]. The calculation time were 3.9729[sec] for the uniform
time step and 0.0045[sec] for the multi-resolution formulation. As shown in the figure, the
optimal solutions are di↵erent in the later time steps, but are very similar in the first 0.1[sec]
where the time steps are the same.
When the sliding controller is used, the desired trajectory should be given. If the desired
trajectory is calculated in real time, the optimization problem should be solved once at the
very first time step to calculate the desired trajectory. However at the very first time step,
only the values of the very first part of the trajectory are utilized and the overall solution of
the optimization problem is saved as a desired trajectory for later usage. Since the solution
of the later time horizon does not need to be calculated very accurately at that very first time
step, the multi-resolution formulation is also applicable to calculate the desired trajectory
for the sliding controller.
Figure 4.18 describes the concept of using the multi-resolution problem formulation to
solve for the optimal trajectory. As shown in the figure, a shorter time step (same with the
control time step) is used for the first part of the time horizon while a longer time step is
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 83

2
uniform

x[m]
1 multi

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
time[sec]
1
uniform
v[m/s]

0.5 multi

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
time[sec]
0.5 uniform
multi
a[m/s 2]

-0.5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
time[sec]

Figure 4.17: Solutions of the optimization problems with uniform time step and multi-
resolution time step

used for the later part. At the next time step, the time horizon with the shorter time step
is shifted but is overlapped with the previous time horizon with the shorter time step. The
solution of the overlapped time from the previous problem is applied as equality constraints
for the next problem. Note that at this moment, the controller still utilizes the trajectory
from the first problem solved in the previous time step. However the controller solves the
optimization problem for later time. The overlap is to prevent the discontinuity of the final
solution.
Figure 4.19 compares the solutions of the uniform time step and the multi-resolution time
step. 10[msec] sampling time is used for the uniform time step. For the multi-resolution
formulation, 10[msec] is used for the first 0.2[sec], 0.1[sec] is used from 0.2[sec] to 1[sec], and
1[sec] is used for the later time horizon. Only the solutions of the first 0.2[sec] time horizon
are saved and used. The calculation time of the uniform time horizon problem was 3.9842[sec]
and the average calculation time of the multi-resolution problems was 0.005039[sec]. As
shown in the figure, two solutions are similar.
Figures 4.20 and 4.21 show the simulation result with the sliding controller whose desired
trajectory is calculated in real time. As shown in the figures, with the sliding controller the
vehicle stops completely faster than with the PID controller.
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 84

Figure 4.18: Concept of Multi-resolution formulation to calculate the optimal trajectory for
the sliding control

2
uniform
x[m]

1 multi

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
time[sec]
1
uniform
v[m/s]

0.5 multi

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
time[sec]
0.5 uniform
multi
a[m/s 2]

-0.5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
time[sec]

Figure 4.19: Solutions of the optimization problems with uniform time step and multi-
resolution time step
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 85

8
lead
sliding
7 PID

4
velocity[m/s]

−1

−2
0 5 10 15 20 25 30 35 40 45 50
time[sec]

Figure 4.20: Simulation result with a sliding controller in case of a complete stop: velocity
[m/sec]

380
lead
sliding
PID
375

370

365
position[m]

360

355

350

345

340
0 5 10 15 20 25 30 35 40 45 50
time[sec]

Figure 4.21: Simulation result with a sliding controller in case of a complete stop: position
[m]
CHAPTER 4. OPTIMAL PROFILE GENERATION FOR STOP-AND-GO 86

4.5 Summary
In this chapter, optimal trajectories for a complete stop and a starting motion were calculated
from a constrained Quadratic Programming (QP) problem. The constant time-headway
policy caused the too slow motion when the vehicle stopped completely. A QP problem was
formulated to find the optimal velocity profile for the complete stop. l1 norm was introduced
as a convex relaxation of the l0 norm. Thanks to the sparse property, the l1 norm made the
velocity converge to zero faster.
Sliding controller was designed to control the host vehicle to follow the optimal trajectory
during a complete stop. Smooth response was achieved by introducing saturation function
instead of the sign function. For the starting motion, the optimal trajectory was utilized by
a virtual lead vehicle to better react to a slow lead vehicle.
Interior point method and Lemke algorithm were studied as the methods to solve the
optimization problem. The interior point method showed faster performance than the Lemke
algorithm when the size of the problem was large. Also, the interim solutions of the interior
point method were feasible. However, the Lemke algorithm was faster than the interior point
method when the size of the problem was small and the calculation time was almost linear
to the size of the problem.
To solve the optimization problem in real time, a multi-resolution formulation was sug-
gested to reduce the problem size and shorten the calculation time. The multi-resolution
formulation was utilized to calculate the optimal trajectory in real time for sliding con-
troller. With the sliding controller, the host vehicle stopped completely faster than with a
PID controller whose desired relative distance was calculated by the constant time-headway
policy.
 87

Chapter 5

Cooperative Adaptive Cruise Control

5.1 Introduction
Cooperative Adaptive Cruise Control(CACC) is an enhanced version of the ACC. The main
di↵erence of the CACC from the ACC is that CACC utilizes communication between the
vehicles and/or between the vehicle and the road structure. This communication allows the
control system on any single vehicle to get information about other vehicles in the platoon
like the acceleration, the velocity, the throttle control command, and the brake control
command. Recently, various research have been done about the CACC in the area of traffic
control and communication. Many previous research show that the CACC has the potential
to improve both the traffic flow [3] and the string stability [33].
Many existing research on the CACC system follow a hierarchical controller structure
where the system consists of an upper lever controller and a lower level controller [8]. The
upper level controller determines the desired acceleration based on the information acquired
from the sensors and through inter vehicle communications. Based on the desired accelera-
tion, the lower level controller determines the throttle and/or brake control command. Most
of the research about the CACC have been focused on designing the upper level control.
One approach for the longitudinal control of the CACC is utilizing the front vehicle
acceleration in the feedforward controller [3, 33, 43]. Another approach is building a world
model based on the information from multiple vehicles and find an appropriate acceleration
command based on the world model [8, 25, 40]. In existing studies, the accelerations were
determined either based on the states of the lead vehicle which causes the largest potential
danger [8], or from a learning algorithm [25]. Another longitudinal control approach was to
find the desired acceleration based on the velocity and the distance errors from some vehicles
in the upstream of the vehicle string [45]. The CACC is also used for the lateral control and
the case when the vehicles merge in intersections [40, 61, 19].
In this chapter, only the longitudinal control problem will be considered. Instead of
utilizing only the front vehicle acceleration or determining the desired acceleration based on
the largest potential danger, the vehicle platoon itself is considered as a single system whose
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 88

inputs are the accelerations of all the vehicles in the platoon. In Section 2.4, the Linear
Quadratic optimal control problem was solved to find the feedback gains. In this chapter, a
vehicle platoon is modeled as a set of double integrator and the LQ problem is solved. As a
result, the optimal feedback control law is a state feedback control law where the states are
the relative positions and velocities of all vehicles in the platoon.
When the vehicle platoon is considered as a single system, cutting in and out of vehicles
can be an issue. Virtual lead vehicle scheme was studied in Section 3 in case of a single host
vehicle. The virtual lead vehicle scheme is also utilizable in case of a vehicle platoon. Instead
of utilizing the actual relative distance, the relative distance to the virtual lead vehicle is
utilized when a vehicle cuts in or out to prevent a sudden acceleration and deceleration.
Each vehicle in a platoon can measure the relative distance to the front vehicle and its own
velocity. However, the other relative distances and velocities should be transmitted through
wireless communication networks. In this chapter, only vehicle-to-vehicle communication will
be considered and it will be assumed that there is no communication between the vehicles
and the road structure. The information shared by each vehicle will be the relative distance
to the lead vehicle, the velocity of the lead vehicle, and the velocity and the acceleration of
itself. The wireless network, however, is usually lossy and it is not guaranteed that all the
data are communicated for every single control time step. In this chapter, various control
schemes when the data are lost are compared by simulations.
This chapter is organized as follows. In Section 5.2, an LQ optimal control problem is
solved for a vehicle platoon to find feedback control gains. In Section 5.3, the virtual lead
vehicle scheme is utilized to control the platoon react smoothly to cutting in or out vehicles.
In Section 5.4, packet loss during wireless communications is considered and various control
scheme are compared by simulations. Section 5.6 summarizes the chapter.

5.2 Linear Quadratic Optimal Controller for a

Vehicle Platoon
In Section 2.4, the LQ optimal control problem was formulated and solved for one host
vehicle and one lead vehicle. In this section, the problem will be solved for a platoon of
vehicles equipped with identical CACC systems.

5.2.1 Problem Description

In this section, a vehicle platoon with length N is considered. The first vehicle is a lead
vehicle driven by a human driver and N 1 vehicles equipped with identical CACC systems
are following the lead vehicle. The linear model can be formulated as:

Ẋ(t) = AX(t) + BU (t) (5.1)

Y (t) = CX(t) (5.2)
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 89

2 3
xl (t) x1 (t)
6 x1 (t) x2 (t) 7
6 .. 7
6 7
6 . 7
6 7
6 x 2 (t) xN 1 (t) 7
X(t) = 6 N 7 (5.3)
6 vl (t) 7
6 7
6 v1 (t) 7
6 .. 7
4 . 5
vN 1 (t)
⇥ ⇤T
U (t) = al (t) a1 (t) · · · aN 1 (t) (5.4)
2 3
thw v1 (t) (xl (t) x1 (t))
6 .. 7
Y (t) = 4 . 5 (5.5)
thw vN 1 (t) (xN 2 (t) xN 1 (t))

2 3
1 1 0 ··· 0
6 . .. 7
6 0(N 1)⇥(N 1)
0 1 1 .. . 7
6 .. ... ... ... 7
A = 6 . 0 7 (5.6)
6 7
4 0 ··· 0 1 1 5
0N ⇥(N 1) 0N ⇥N

0(N 1)⇥N
B = (5.7)
IN ⇥N
⇥ ⇤
C = I(N 1)⇥(N 1) 0(N 1)⇥1 thw I(N 1)⇥(N 1) (5.8)
where xi , vi , and ai are the position, the velocity, and the acceleration of the ith vehicle.

5.2.2 Linear Quadratic Optimal Control

If the plant is time invariant, (A, B) is controllable, and (A, C) is observable, the solution of
the stationary LQ problem exists for the performance index
Z
1 1 T
J= X (t)QX(t) + U T (t)RU (t)dt (5.9)
2 t0
where Q = C T C.
However with the C given in (5.8), (A, C) is not observable. To make it observable, while
not modifying the performance index, a new C is given as:

I(N 1)⇥(N 1) 0(N 1)⇥1 thw I(N 1)⇥(N 1)
C= (5.10)
01⇥(N 1) ✏o 01⇥(N 1)
As mentioned in Section 2.4, the number ✏o is a very small number that can be further
reduced depending on the available computational resolution.
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 90

Similar to Section 2, the input penalty matrix R is:


1/✏c 01⇥(N 1)
R= (5.11)
0(N 1)⇥1 I(N 1)⇥(N 1)

where is a design parameter to tune. The number ✏c is a very small number that can be
further increased depending on the available computational resolution. The state feedback
control law is given as:
U (t) = R 1 B T P+ X(t) = KX(t) (5.12)
where P+ is the positive definite solution of the algebraic Riccati equation

AT P + P A P BR 1 B T P + C T C = 0 (5.13)

For example, with N = 5, thw = 2, = 1, and ✏o = ✏c = 1e 5 the state feedback gain

is calculated to be:
2 1.1e 6 4.6e 7 4.3e 7 1.8e 7 5.4e 6 4.7e 6 7.7e 7 8.5e 8 4.2e 8 3
0.9952 0.0974 0.0098 1.2e 4 0.4726 2.4788 0.1996 0.0256 0.0048
K=4 0.0960 0.9906 0.0967 0.0074 0.0765 0.1996 2.4685 0.2019 0.0252 5
0.0190 0.0942 0.9912 0.0914 0.0085 0.0256 0.2019 2.4662 0.2043
0.0023 0.0160 0.0903 0.9958 0.0042 0.0048 0.0252 0.2043 2.4391
(5.14)
As described in (5.14), for each vehicle in the platoon, the largest gains are applied
for its own velocity and the relative distance to its lead vehicle. Relatively higher gain is
also applied for the velocity of the lead vehicle. With these three gains, the control law is
similar to a PD controller. Additionally, the velocity of the vehicle right behind, the relative
distances of the lead vehicle to its lead vehicle, and the relative distance to the vehicle behind
are also multiplied to relatively larger gains and considered in the feedback control law. The
first row which will be related to the lead vehicle acceleration will be rounded o↵ to zero.

5.2.3 Simulation Result

Figures 5.1 ⇠ 5.3 show the simulation results comparing the CACC system designed in the
previous section and the ACC system designed in Section 2.4. In the simulation, each vehicle
is modelled as a double integrator whose input is the acceleration. Aerodynamic drag and
rolling resistance are included in the model and the brake/engine is modeled as a first order
system. Disturbance observer is utilized for each vehicle to compensate the disturbances
and the unmodeled actuator dynamics. In the simulation, a vehicle platoon of 6 vehicles is
considered. The lead vehicle is assumed to be driven by a human driver and the following
vehicles are assumed to be controlled by a CACC system. Solid lines are from the simulation
result when the CACC controller with the feedback gain given in (5.14) is used and dashed
lines are from the simulation result when each host vehicle utilizes the ACC controller whose
gain was from the LQ controller designed in Section 2.4.
Figure 5.1 shows the velocities. As shown in the figure, the lead vehicle decelerates
from 10[m/sec] to 5[m/sec], maintains the speed, and accelerates back to 10[m/sec]. The
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 91

11

10

9
velocity[m/s]

7
lead
#1
6 #2
#3
#4
5
#5

4
0 10 20 30 40 50
time[sec]

Figure 5.1: Simulation results with LQ controllers: velocity (solid lines: with CACC, dashed
lines: with ACC)

velocities with the ACC and CACC are very similar, however the errors are di↵erent as shown
in Figures 5.2. As shown in Figure 5.2, the peak error of the vehicle #1 is reduced with the
CACC compared to the result with the ACC. As described before, the CACC additionally
considered the velocities and the relative distances of other vehicles in the platoon. That
makes the vehicle react faster and the errors of the remaining vehicles are more negative
when the lead vehicle decelerates and more positive when the lead vehicle accelerates while
the errors with the ACC system is the opposite. Negative error means that the relative
distance to the front vehicle is longer than the desired distance and hence is more desirable
and safer than a positive error especially when the lead vehicle is decelerating. When the
lead vehicle is accelerating, safety is not a big issue and positive error is acceptable and also
desirable since it means faster reaction. In other words, it is better for the sign of the error
to be the same with the sign of the lead vehicle acceleration. Figure 5.3 shows the plot of
the product of the error and the lead vehicle acceleration. As shown in the figure, with the
CACC the values are positive in most of time while the values are mostly negative with the
ACC.

5.3 Cut In/Out of a Vehicle in a Platoon

As studied in the previous section, the feedback gain of a CACC system depends on the
number of vehicles in the platoon. When a vehicle cuts into the platoon, or cuts out from
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 92

0.015
#1
#2
#3
0.01
#4
#5

0.005
error[m]

−0.005

−0.01

−0.015
0 10 20 30 40 50
time[sec]

Figure 5.2: Simulation results with LQ controllers: error (solid lines: with CACC, dashed
lines: with ACC)

0.015
#1
#2
#3
0.01
#4
#5
error * acclead [m2/sec2]

0.005

−0.005

−0.01

−0.015
0 10 20 30 40 50
time[sec]

Figure 5.3: Simulation results with LQ controllers: product of the error and the lead vehicle
acceleration (solid lines: with CACC, dashed lines: with ACC)
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 93

15
lead
1
2
14
3
4
5
13
velocity[m/s]

12

11

10

9
0 10 20 30 40 50
time[sec]

Figure 5.4: Simulation results when a vehicle cuts out from the platoon

the platoon, there is a sudden change in the states of the platoon and that can cause some
undesirable acceleration or deceleration if there is no proper transient control scheme. As
shown in Section 3, the virtual lead vehicle scheme is useful for a single vehicle ACC system
when the lead vehicle cuts out or another vehicle cuts in between the host and the lead
vehicles. In this section, the virtual lead vehicle scheme is utilized for smooth transient
motion control when a vehicle cuts in to or out from the platoon.

5.3.1 Problem Description

Figure 5.4 shows the simulation result for the case where a vehicle cuts out from the platoon.
As described in Figure 5.5, a vehicle platoon of 6 vehicles is considered in the simulation
where the lead vehicle is moving with a constant speed. At 20 second, the vehicle #3 cuts
out from the platoon. Without any transient controller, there is a sudden large step change
in the error for the vehicle #4. That makes the vehicle #4 accelerate suddenly and also the
vehicle #5 since it follows the vehicle #4. Also, when the vehicle #3 cuts out the relative
distance between the vehicles #2 and #4 becomes very larger than it should be. That makes
the vehicle #2 decelerate to compensate the error. The deceleration of the vehicle #2 is very
undesirable and unnecessary since it does not improve safety while worsens the fuel efficiency.
Also it can make the passengers feel very uncomfortable since the passengers may not know
the cutting out of the vehicle behind.
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 94

Lead%

#1%

#2%

#3%

#4%

#5%

Figure 5.5: Sketch of the cutting out scenario

5.3.2 Virtual Lead Vehicle Scheme

For the case described in Figure 5.5, the state of the LQ formulation before the cutting out
is given as: 2 3
xl (t) x1 (t)
6 x1 (t) x2 (t) 7
6 7
6 x2 (t) x3 (t) 7
6 7
6 x3 (t) x4 (t) 7
6 7
6 x4 (t) x5 (t) 7
6 7
X(t) = 6 6 v l (t) 7
7 (5.15)
6 v1 (t) 7
6 7
6 v2 (t) 7
6 7
6 v (t) 7
6 3 7
4 v4 (t) 5
v5 (t)
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 95

When the vehicle #3 cuts out, the new state is:

2 3
xl (t) x1 (t)
6 x1 (t) x2 (t) 7
6 7
6 x2 (t) x4 (t) 7
6 7
6 x4 (t) x5 (t) 7
6 7
X(t) = 6
6 vl (t) 7
7 (5.16)
6 v 1 (t)
7
6 7
6 v2 (t) 7
6 7
4 v4 (t) 5
v5 (t)

Since the relative distance x2 (t) x4 (t) is very large when the cutting out happens, the
vehicle #4 accelerates and #2 decelerates to reduce the relative distance. To prevent this, the
virtual lead vehicle scheme is introduced. Let xv,i (t) the position of the virtual lead vehicle
followed by the vehicle #i. Assuming that the initial positions and the initial velocities of
the virtual lead vehicles are the same with those of the actual lead vehicles, (5.15) can be
represented as: 2 3 2 3
xl (t) x1 (t) xv,1 (t) x1 (t)
6 x1 (t) x2 (t) 7 6 xv,2 (t) x2 (t) 7
6 7 6 7
6 x2 (t) x3 (t) 7 6 xv,3 (t) x3 (t) 7
6 7 6 7
6 x3 (t) x4 (t) 7 6 xv,4 (t) x4 (t) 7
6 7 6 7
6 x4 (t) x5 (t) 7 6 xv,5 (t) x5 (t) 7
6 7 6 7
X(t) = 66 v l (t) 7=6
7 6 v l (t) 7
7 (5.17)
6 v (t) 7 6 v (t) 7
6 1 7 6 1 7
6 v (t) 7 6 v (t) 7
6 2 7 6 2 7
6 v (t) 7 6 v (t) 7
6 3 7 6 3 7
4 v4 (t) 5 4 v4 (t) 5
v5 (t) v5 (t)
Similarly, the new state (5.16) is also represented as:
2 3
xv,1 (t) x1 (t)
6 xv,2 (t) x2 (t) 7
6 7
6 xv,4 (t) x4 (t) 7
6 7
6 xv,5 (t) x5 (t) 7
6 7
X(t) = 6 6 vl (t) 7
7 (5.18)
6 v1 (t) 7
6 7
6 v2 (t) 7
6 7
4 v4 (t) 5
v5 (t)

Comparing (5.16) and (5.18), with the virtual lead vehicle scheme only the relative dis-
tances between the virtual lead vehicles and the host vehicles are utilized. That prevents
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 96

the sudden change of the relative distance is the state and makes the state to be continuous
if the virtual lead vehicle is properly controlled to merge to the actual lead vehicle smoothly
when a vehicle cuts in or out.

5.3.3 Simulation Results

Simulations are performed to show the performance of the virtual lead vehicle scheme. In
the simulations, each vehicle has its own virtual lead vehicle. At steady state, the position
and the velocity of the virtual lead vehicle are the same with the position and the velocity of
the actual lead vehicle. When a vehicle cuts in or cuts out, the virtual lead vehicle merges
to the new lead vehicle by the controller designed in Chapter 3. The state feedback control
law is utilized with the control gains found in Section 5.2.
Figure 5.6 shows the simulation result with the virtual lead vehicle scheme. As shown
in the figure, the vehicle #4 does not accelerates much when the cutting out happens.
Instead, it follows the virtual lead vehicle which smoothly merges into the vehicle #2. The
peak velocity and the convergence time depends on the virtual lead vehicle whose controller
gains can be adjusted based on the preference of the driver. Also, the vehicle #2 does not
decelerate unnecessarily. Figures 5.7 and 5.8 show the simulation result when two vehicles
cut out from the platoon at the same time. As shown in the figures, with the virtual lead
vehicle scheme there is no excessive acceleration and the velocities and the positions of the
vehicles converge smoothly to their desired values.
Figures 5.10 ⇠ 5.12 show the simulation results when a vehicle cuts into the platoon. As
described in Figure 5.9, initially a vehicle platoon of 5 vehicles is considered where the lead
vehicle is moving with a constant speed. At 20 second, vehicle #3 cuts into the platoon from
the side lane. Figure 5.10 shows the result without any transient controller. As shown in the
figure, without any transient control scheme, vehicles #3 ⇠ #5 decelerates rapidly. Also,
the vehicle #2 accelerates to increase the relative distance between the vehicles #2 and #3.
It is very undesirable since it can make the passengers feel uncomfortable since they may
not know that the vehicle cut in.
As shown in Figures 5.11 and 5.12, the vehicles #3 ⇠ #5 does not excessively decelerate
with the virtual lead vehicle scheme. Instead they decelerate smoothly to increase the relative
distances. Once the relative distances are increased enough, the vehicle accelerates and
maintains the speed same with the lead vehicle. Also, it can be observed that the vehicle
#2 does not unnecessarily accelerate. In this simulation, it is assumed that the speed of the
vehicle #3 is the same with other vehicles in the platoon. Also the CACC system of the
vehicle #3 is activated right after the cutting in. If the speed of the vehicle that cuts in is
slower than the speed of the platoon, it can be a dangerous situation and a separate safety
feature such as the anti collision system should be utilized.
Comparing (5.17) and (5.18), the length of the state vector is changed after the cutting
out. It is the same for the cutting in case. This can make the feedback control command be
discontinuous when the cutting in or out occurs. To prevent the discontinuities, some tran-
sient control scheme can be introduced. For example, a fuzzy controller can be used whose
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 97

12
lead
1
2
11.5
3
4
5
11
velocity[m/s]

10.5

10

9.5

9
0 10 20 30 40 50
time[sec]

Figure 5.6: Simulation results when a vehicle cuts out from the platoon (Solid line: Actual
vehicle, Dashed line: Virtual lead vehicle

12
lead
1
2
3
11.5 4
5

11
velocity[m/s]

10.5

10

9.5

9
0 10 20 30 40 50 60 70 80 90 100
time[sec]

Figure 5.7: Simulation results when two vehicles cut out from the platoon (Solid line: Actual
vehicle, Dashed line: Virtual lead vehicle
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 98

1200
lead
1
2
1000
3
4
5

800
position[m]

600

400

200

0
0 10 20 30 40 50 60 70 80 90 100
time[sec]

Figure 5.8: Simulation results when two vehicles cut out from the platoon (Solid line: Actual
vehicle, Dashed line: Virtual lead vehicle

Lead%

#1%

#2%

#3%

#4%

#5%

Figure 5.9: Sketch of the cutting in scenario

CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 99

11

10

9
velocity[m/s]

lead
1
8
2
3
4
5
7

5
0 5 10 15 20 25 30 35 40 45 50
time[sec]

Figure 5.10: Simulation results without the virtual lead vehicle when a vehicle cuts into the
platoon

11

10.5

10

9.5

9
velocity[m/s]

8.5

8 lead
1
2
7.5
3
4
7 5

6.5

6
0 5 10 15 20 25 30 35 40 45 50
time[sec]

Figure 5.11: Simulation results when a vehicle cuts into the platoon (Solid line: Actual
vehicle, Dashed line: Virtual lead vehicle
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 100

600

500

400
position[m]

300

lead
200 1
2
3
4
100
5

0
0 5 10 15 20 25 30 35 40 45 50
time[sec]

Figure 5.12: Simulation results when a vehicle cuts into the platoon (Solid line: Actual
vehicle, Dashed line: Virtual lead vehicle

membership functions depend on the error of the virtual lead vehicles. However, this scheme
makes the control structure complicated while showing little performance improvement. As
shown in the previous simulation results, the discontinuities are hardly shown when cutting
in or out occurs and will be ignored in this thesis.

5.4 CACC in Lossy Network Condition

Since the CACC system utilizes wireless communication, there is a possibility that some data
are lost during the transmission through the wireless network. Recently, with the increase
of wireless applications, various research have been done in the area of system control over
lossy networks. Some of the research are related to designing state estimators with the data
packet loss [46, 14, 62, 49] or the time delay [36, 35] in wireless networks. Also, feedback
controllers were designed considering the packet loss between the controller and the actuator
[46, 35, 15, 26, 66].
In case of the CACC, the data are transmitted between vehicles in wireless networks.
However, the ACC controller in each vehicle is connected to the actuator (or lower level
controllers) in a wired manner. So it is reasonable to assume that there is no dropped control
command from the controller to the actuator. Also, sensor measurements are transferred to
the controller in wired manner inside each vehicle. So some of the data are transferred to
the controller without any data loss. In most cases, those data are related to the states of
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 101

the lead and the host vehicles and are usually important. Since the data which can be lost
in the wireless network are far less important, it can be ignored or be given a minimum level
of attention to compensate the data loss. In this section, various control and estimation
schemes are suggested and compared by simulations.

5.4.1 Problem Description

In this section, it is assumed that sensors in each vehicle measure the relative distance to
the lead vehicle, the velocity of the lead vehicle, and the velocity of the host vehicle itself.
The measurements are transmitted to other vehicles in the platoon via wireless network. It
is assumed that the data can be lost during the transmission over the wireless network with
a fixed probability and it can be represented as:
⇢ s
r Xi (k) if i,j = 1
Xij (k) = (5.19)
n/a if i,j = 0
where Xijr (k) is the data from the ith vehicle received by the j th vehicle, Xis (k) is the data
sent from the ith vehicle, and i,j is a Bernoulli random variable with P ( i,j = 1) = ¯ . It
is also assumed that i,j ’s are independent for all i’s and j’s. Note that the lost data (n/a)
can be distinguished from zero values. For simplicity, a discrete time system is considered
and it is assumed that there is no time delay through the communication.
In the simulations, no measurement and process noises are considered.

5.4.2 Control Schemes for Dropped Data

For ith vehicle, the feedback control law is given by:

ades,i (k) = Ki X(k) (5.20)

where Ki is the ith row of the feedback gain matrix from the LQ optimal control problem,
and X(k) is the state vector given in (5.3). In X(k), xi 1 (k) xi (k), vi 1 (k), and vi (k) are
measured by sensors and other data are transmitted by wireless networks and can be lost.
If some of the data are lost, there are various ways to respond.

Case 1 Set the lost dataQ to be zeros.

Case 2 ades,i (k) = 0 if j i,j = 0.
⇢ Q
Ki,CACC X(k) if i,j =1
Case 3 ades,i (k) = Qj .
Ki,ACC X(k) if j i,j =0
Case 4 Estimate the lost data.

For Case 1, vj (k) can be transmitted from the j th and j + 1th vehicles to the ith vehicle.
vj (k) is set to be zero only if both i,j and i,j+1 are zeros. For Case 2, the desired acceleration
is set to be zero when any of the data are lost. For Case 3, if any of the data are lost during
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 102

0.1
Case 2
Case 3
0.05 Case 4
no loss
ACC
0

−0.05
error[m]

−0.1

−0.15

−0.2

−0.25
0 10 20 30 40 50
time[sec]

Figure 5.13: Simulation results over lossy network: error for vehicle #1

the transmission, the ACC control law is utilized instead of the CACC. It is possible since
all the data used in the ACC control law are measurable with the sensors in the ith vehicle.
For Case 4, the state is estimated by X̂ o (k + 1) = (A BK)X̂(k) and X̂(k + 1) is updated
by replacing the elements of a-priori estimate X̂ o (k + 1) with the data measured by sensors
or transmitted successfully through the network.

5.4.3 Simulation Results

The simulation set-up is the same with the simulation in Section 5.2. The data loss rate is
assumed to be 5%, i.e. ¯ = 0.95. 4 cases suggested in the previous section are simulated and
compared with the CACC without any data loss and with the case where all the vehicles
utilizes identical ACC controllers. Case 1 shows the worst performance which is a lot worse
than the other cases and the result with Case 1 will not be shown in the figures for better
comparison.
Figures 5.13 ⇠ 5.17 show the simulation results. As shown in the result, Case 2 shows
the worst performance. It shows the largest error and also the benefit of the CACC, reacting
faster and making the sign of the error to be the opposite of the lead vehicle acceleration,
is lost. Case 4, in this case, shows almost the same performance with the CACC system
without data loss.
As shown in the figures, Case 3 shows the best performance. The error with Case 3 is
even smaller than the case without data loss and also has opposite sign with the lead vehicle
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 103

0.06
Case 2
Case 3
Case 4
0.04
no loss
ACC

0.02
error[m]

−0.02

−0.04

−0.06
0 10 20 30 40 50
time[sec]

Figure 5.14: Simulation results over lossy network: error for vehicle #2

0.06
Case 2
Case 3
Case 4
0.04
no loss
ACC

0.02
error[m]

−0.02

−0.04

−0.06
0 10 20 30 40 50
time[sec]

Figure 5.15: Simulation results over lossy network: error for vehicle #3
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 104

0.05
Case 2
0.04 Case 3
Case 4
no loss
0.03 ACC

0.02

0.01
error[m]

−0.01

−0.02

−0.03

−0.04
0 10 20 30 40 50
time[sec]

Figure 5.16: Simulation results over lossy network: error for vehicle #4

0.05
Case 2
0.04 Case 3
Case 4
0.03 no loss
ACC
0.02

0.01
error[m]

−0.01

−0.02

−0.03

−0.04

−0.05
0 10 20 30 40 50
time[sec]

Figure 5.17: Simulation results over lossy network: error for vehicle #5
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 105

0.01
Case 3
0.008 Case 4
no loss
0.006 ACC

0.004

0.002
error[m]

−0.002

−0.004

−0.006

−0.008

−0.01
0 10 20 30 40 50
time[sec]

Figure 5.18: Simulation results over lossy network with ¯ = 0.50: error for vehicle #3

acceleration. As described before, the error with the opposite sign with the lead vehicle
acceleration is the benefit of the CACC which considers the errors of the front vehicles in
the platoon. In Case 3, the vehicle only utilizes the measurement and does not consider
the error in front when the data are lost. This reduces the e↵ect of CACC and fortunately
reduces the magnitude of the error.
However it is not guaranteed that the performance of Case 3 is better than CACC without
any data loss when the data loss rate is larger. Figure 5.18 shows the simulation result with
a worse loss rate. The loss rate is 50% in the simulation. When the loss rate is higher, Case
3 loses the benefit of the CACC system and the error has the same sign with the lead vehicle
acceleration while Case 4 still shows similar performance with the CACC without data loss.

5.4.4 Estimation with Kalman Filter

In the previous section, the estimation scheme showed the best performance. However, no
measurement noises were considered during the estimation process. Actually, the measured
signals are contaminated by noises. Also, the covariance of each noise are di↵erent. For
example, the velocity of the j th vehicle was assumed to be measured by the j th vehicle with
encoders and by the j + 1th vehicle with the radar. However, actually the radar measures the
relative velocity between the j th and j + 1th vehicles and the velocity of the j + 1th vehicle
should be added to it. By doing that, the velocity measurement from the j + 1th vehicle has
larger measurement noise than the measurement from the j th vehicle. Kalman filter is one
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 106

of the most popular estimation scheme when the data are contaminated by noise. It can
consider the di↵erent noise covariances and find the optimal state estimation.

Problem Description
The system model is basically the same with the model used in Section 5.2. However, process
noises and measurement noises are introduced. Also, as described before, it is assumed that
the velocity of the j th vehicle is measured by the j th vehicle with encoders and by the j + 1th
vehicle with radar. Considering the data loss, the discrete time state space equations are
considered as:

X(k + 1) = AX(k) + BU (k) + Bw W (k) (5.21)

Y (k) = CX(k) + V (k) (5.22)

2 3
xl (k) x1 (k)
6 x1 (k) x2 (k) 7
6 .. 7
6 7
6 . 7
6 7
6 x 2 (k) xN 1 (k) 7
X(k) = 6 N 7 (5.23)
6 vl (k) 7
6 7
6 v1 (k) 7
6 .. 7
4 . 5
vN 1 (k)
⇥ ⇤T
U (k) = al (k) a1 (k) · · · aN 1 (k) (5.24)
2 3T
xl (k) x1 (k)
6 .. 7
6 . 7
6 7
6 xN 2 (k) xN 1 (k) 7
6 7
6 v1 (k) 7
6 . 7
Y (k) = 6
6 .. 7
7 (5.25)
6 v (k) 7
6 N 1 7
6 vl (k) 7
6 7
6 . 7
4 .. 5
vN 2 (k)
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 107

⇥ ⇤T
W (k) = wl (k) w1 (k) · · · wN 1 (k) (5.26)
2 3
vr,d,1 (k)
6 .. 7
6 . 7
6 7
6 vr,d,N 1 (k) 7
6 7
6 ve,1 (k) 7
6 .. 7
V (k) = 6
6 . 7
7 (5.27)
6 v 7
6 e,N 1 (k) 7
6 v (k) 7
6 r,v,1 7
6 .. 7
4 . 5
vr,v,N 1 (k)
(5.28)

where wi is the process noises for the ith vehicle. vr,d,i , vr,v,i , and ve,i are the measurement
noises of the distance measurement by the radar, the front vehicle velocity measurement by
the radar, and the host vehicle velocity measurement by the encoder in ith vehicle, respec-
tively.
The state space matrices are given as:
2 3
t t 0 ··· 0
6 . .. 7
6 I(N 1)⇥(N 1) 0 t t .. . 7
6 . . . . 7
A = 6 .. .. .. .. 0 7 (5.29)
6 7
4 0 ··· 0 t t 5
0N ⇥(N 1) IN ⇥N
2 1 2 1 3
2
t 2
t2 0 ··· 0
6 1 1 ... .. 7
6 0 2
t2 2
t2 . 7
6 .. .. .. .. 7
B = Bw = 6 . . . . 0 7 (5.30)
6 7
4 0 ··· 0 1
t2 1
t2 5
2 2
tIN ⇥N
2 3
I(N 1)⇥(N 1) 0(N 1)⇥N
C = 4 0(N 1)⇥1 I(N 1)⇥(N 1)
5 (5.31)
02(N 1)⇥(N 1)
I(N 1)⇥(N 1) 0(N 1)⇥1

In this thesis, the noises are assumed to be stationary zero mean white and independent
to each other. Also, the covariances of radar measurement noises are assumed to be larger
than the encoder measurement noises.

Kalman Filter
The standard a-posteriori Kalman Filter equations are:
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 108

X̂(k) = X̂ o (k) + F (k)Ỹ o (k) (5.32)

X̂ o (k + 1) = AX̂(k) + BU (k) (5.33)
Ỹ o (k) = Y (k) C X̂ o (k) (5.34)

F (k) = M (k)C T (CM (k)C T + Svv (k)) 1 (5.35)

M (k + 1) = AM (K)AT + Bw Sww (k)BwT
AM (k)C T (CM (k)C T + Svv (k)) 1 CM (k)AT (5.36)
where Sww (k) and Svv (k) are the noise covariance matrices for the process and the measure-
ment noises. Since the lead vehicle acceleration is assumed to be zero, the actual lead vehicle
acceleration is considered as a process noise. So, Sww (k) is set as a diagonal matrix. The
diagonal elements are all the same except the (1, 1) element which is larger than others.
The measurement noise covariance matrix Svv (k) is stationary. However, when some of
the data are lost through the wireless communication, the lost data are set to be zeros and
the error covariances are set to be very large numbers. For example, Yj (k), YN 1+j (k), and
Y2(N 1)+j (k) are from the j th vehicle. If the data are lost, the data are set to be zeros and
(j, j), (N 1 + j, N 1 + j), and (2(N 1) + j, 2(N 1) + j) elements of Svv (k) are set to
be very large so that the lost data are not considered in the Kalman Filter.

5.4.5 Simulation Result

Simulations are performed to verify the performance of the Kalman Filter. Simulation results
with 4 di↵erent methods are compared: (1) KF w/o loss Kalman Filter designed in the
previous section is utilized while no data are lost. (2) ACC ACC system is utilized for each
single vehicle. (3) CACC w/KF Kalman Filter designed in the previous section is utilized
with 5% data loss rate. (4) Case 3 Kalman Filter is utilized when no data are lost. However,
ACC system is utilized when data are lost. The loss rate is 5%.
For (2) and (4), when the ACC system is utilized, the relative distance to the lead vehicle
and the velocities of the lead and the host vehicles are measured by each vehicle. However,
a Kalman Filter with a single lead and a single host vehicle model is utilized to compensate
the noises.
Figures 5.19 ⇠ 5.23 show the simulation results with 5% loss rate. As shown in the result,
with the Kalman Filter the performance is very similar to the case when no data are lost.
The overall results are similar to the case without noises and for some cases Case 3 shows
better performance than the CACC with Kalman Filter. However, similar to the result
without noises, it is not guaranteed that the performance of Case 3 is better than CACC
with Kalman Filter when the data loss rate is larger. Figure 5.24 shows the simulation result
with a worse loss rate. The loss rate is 50% in the simulation. When the loss rate is higher,
Case 3 loses the benefit of the CACC system and the performance is even worse than the
ACC system.
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 109

0.04
KF w/o loss
0.03 ACC
CACC w/KF
Case 3
0.02

0.01

0
error[m]

−0.01

−0.02

−0.03

−0.04

−0.05
0 10 20 30 40 50
time[sec]

Figure 5.19: Simulation results with Kalman Filter over lossy network: error for vehicle #1

0.025
KF w/o loss
0.02 ACC
CACC w/KF
0.015 Case 3

0.01

0.005
error[m]

−0.005

−0.01

−0.015

−0.02

−0.025
0 10 20 30 40 50
time[sec]

Figure 5.20: Simulation results with Kalman Filter over lossy network: error for vehicle #2
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 110

0.015

0.01

0.005

0
error[m]

−0.005

−0.01
KF w/o loss
ACC
−0.015
CACC w/KF
Case 3
−0.02
0 10 20 30 40 50
time[sec]

Figure 5.21: Simulation results with Kalman Filter over lossy network: error for vehicle #3

0.02
KF w/o loss
ACC
0.015 CACC w/KF
Case 3
0.01

0.005
error[m]

−0.005

−0.01

−0.015

−0.02
0 10 20 30 40 50
time[sec]

Figure 5.22: Simulation results with Kalman Filter over lossy network: error for vehicle #4
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 111

0.025
KF w/o loss
0.02 ACC
CACC w/KF
Case 3
0.015

0.01

0.005
error[m]

−0.005

−0.01

−0.015

−0.02
0 10 20 30 40 50
time[sec]

Figure 5.23: Simulation results with Kalman Filter over lossy network: error for vehicle #5

0.025
KF w/o loss
0.02 ACC
CACC w/KF
0.015 Case 3

0.01

0.005
error[m]

−0.005

−0.01

−0.015

−0.02

−0.025
0 10 20 30 40 50
time[sec]

Figure 5.24: Simulation results with Kalman Filter over lossy network with ¯ = 0.50: error
for vehicle #3
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 112

5.5 Limitation of the Centralized Controller

The LQ controller designed in the previous sections is a centralized controller. The perfor-
mance of the LQ controller was good, but the controller may not be practical for implemen-
tations. In the LQ controller, it is assumed that the controller have the information about all
the vehicles in the platoon. However, the CACC control law is embedded in each vehicle and
there is no central control system which can control all the vehicles. Thus, synchronization
of the controllers distributed among the vehicles in a platoon is an issue. To utilized the LQ
control law, the controllers in each vehicle should have the information such as the length
of the platoon and the location of itself in the platoon. Failure of an embedded controller
in one vehicle may cause an unacceptable situation. If the length of the platoon is too long,
the wireless communication signal may not reach the entire platoon.
For implementations, a decentralized controller is more practical than the centralized
control design. The decentralized controller is embedded in each vehicle and may not have all
the information of the entire vehicle platoon. However it can utilize the sensor measurements
and receive some information from the vehicles inside its communication range. Unlike the
centralized LQ controller, the control law of the decentralized design should be independent
of the location in the platoon, and it may depend on the parameters set by the driver.
When the decentralized controller is designed, the errors of other vehicles in the platoon
are not considered anymore and the overall performance of the vehicle platoon may be wors-
ened with the decentralized controller. String stability must be maintained in decentralized
designs. The performance of the centralized control design presented in this section can be
utilized as one of the performance standards for the design of a CACC system.

5.6 Summary
In this chapter, the Cooperative Adaptive Cruise Control(CACC) system which utilizes the
vehicle-to-vehicle wireless communication was studied. A Linear Quadratic (LQ) optimal
control problem was formulated for a platoon of vehicles and solved to find the optimal
state feedback control law. The performance of the CACC system were compared to the
ACC system via simulations. The CACC system additionally utilized the error of the front
vehicles in the platoon and reacted faster to the lead vehicle acceleration. As a result, the
error with the CACC system had the same sign with the lead vehicle acceleration which
means safer performance when the lead vehicle decelerates and faster when the lead vehicle
accelerates.
The virtual lead vehicle scheme was also utilized in case of the CACC system. Without
the virtual lead vehicle scheme, the following vehicles accelerated suddenly and also the ve-
hicle in front decelerated to compensate the sudden large step change of the error when a
vehicle cuts out from the platoon. Similar result was shown when a vehicle cuts into the pla-
toon. With the virtual lead vehicle scheme, the acceleration or deceleration of the following
vehicles were decreased and the front vehicle did not decelerate or accelerate unnecessarily.
CHAPTER 5. COOPERATIVE ADAPTIVE CRUISE CONTROL 113

Since the wireless network, which is lossy, is utilized in the CACC, control schemes when
the data are lost during the transmission were studied. 4 di↵erent control schemes were
compared via simulations. The scheme in which the ACC control law was utilized when
the data were lost and the CACC law was utilized when no data were lost showed the best
performance with a low data loss rate. However the performance worsened as the loss rate
was increased. When the lost data were estimated from the linear closed loop model, the
performance was similar to the case when no data were lost. Also the estimation scheme
showed more robust performance when the loss rate was increased. A Kalman Filter was
introduced for the case when the measurement data were contaminated by noise and showed
similar performance with the case when no data were lost.
 114

Chapter 6

Concluding Remarks and Open Issues

6.1 Concluding Remarks

This thesis suggested and studied several control schemes for the Adaptive Cruise Control
systems, including (1) smooth reaction of the host vehicle to the cutting in and out of
lead vehicles, (2) real-time optimal profile generation for stop-and-go motions, (3) optimal
feedback controller design, and (4) Cooperative Adaptive Cruise Control (CACC) system.
Among the suggested schemes, original contributions of this thesis are (1) virtual lead
vehicle scheme, (2) LQ optimal controller design with variable weights, (3) multi-resolution
formulation of a Quadratic Programs for real-time applications, and (4) design of a CACC
controller considering upstream vehicles. Details are discussed separately below.

6.1.1 Smooth Reaction to Cutting In and Out

The virtual lead vehicle scheme was studied in Chapter 3. It was originally introduced as a
method to prevent mode switching between the distance control and the speed control. With
a smooth motion control of the virtual vehicle, the scheme could provide smooth transient
motion control of the host vehicle when a lead vehicle cuts out or a new lead vehicle cuts
in. It was also utilized in Chapter 5 for the CACC system. With the virtual lead vehicle
scheme, some undesirable motions were prevented when a vehicle cuts into or out from the
platoon of vehicles equipped with identical CACC systems.

6.1.2 Real-time Optimal Profile Generation for Stop-and-Go

Optimal profiles for stopping and starting motions were studied in Chapter 4. The optimiza-
tion problems were formulated into constrained Quadratic Programs. Interior Point method
and Lemke algorithm were utilized to solve the optimization problems. To solve the problems
and find the solutions in real time, a multi-resolution formulation was suggested. With the
multi-resolution formulation, the size of the problem was decreased and the problem could
be solved very fast with the Lemke algorithm. The solutions of the optimization problems
CHAPTER 6. CONCLUDING REMARKS AND OPEN ISSUES 115

were utilized in the sliding control for the stopping motion and in the virtual lead vehicle
scheme for the starting motion.

6.1.3 Optimal Feedback Controller Design

The optimal feedback controllers were designed and used as feedback controllers for both
the host and the virtual vehicles. In Chapter 2, a Linear Quadratic (LQ) optimal controller
was designed for a single host and a single lead vehicle. In Chapter 3, variable weights were
utilized for the LQ optimal controller for the virtual lead vehicle. With the variable weights,
the motion of the virtual lead vehicle was controlled to be smooth when there were no safety
threats while the motion was still responsive and fast when a dangerous situation occurs.
The LQ optimal controller was also applied for a platoon of vehicles equipped with identical
CACC systems in Chapter 5.

6.1.4 Cooperative Adaptive Cruise Control

The Cooperative Adaptive Cruise Control system was studied in Chapter 5. An LQ optimal
controller was designed. Compared to the general ACC system, the CACC system showed
faster and safer performance. The virtual lead vehicle scheme was applied for the case when
vehicles cut into or out from the platoon. Since the CACC system utilized wireless commu-
nications, lossy network conditions were considered. In lossy network, the estimation scheme
showed good and also robust performance compared to the cases where zero output or ACC
control law were utilized when the data were lost through the wireless communication. A
Kalman Filter was introduced for the case when the measurement signals were contaminated
by noise. The noise covariance was set to be very large when the data were lost. The designed
Kalman Filter showed similar performance with the case when no data were lost.

6.2 Open Issues

In this thesis, several control schemes have been discussed regarding ACC systems. However,
there are some issues that remain and is an area to explore for future work.

6.2.1 Lower Level Control and Brake/Engine Switching Logic

This thesis focused on designing and analyzing the upper level controller. With better lower
level controllers, the acceleration of the vehicle body will better follow the desired acceleration
from the upper level controller and the performance of the overall system can be improved.
Also, if the engine/brake dynamics are considered in the design of the upper level controller,
the desired acceleration signal can be determined by the upper level controller so that it can
be better followed by the actual engine and brake.
CHAPTER 6. CONCLUDING REMARKS AND OPEN ISSUES 116

When considering the engine/brake dynamics, an appropriate scheme for switching be-
tween the engine and the brake should be studied. Since the ACC system is highly related
to the comfort of the passengers, frequent switchings between the engine and the brake are
undesirable. A fuzzy controller which can imitate the driver’s habit can be a good candidate
for the switching logic.

6.2.2 Mixed Traffic CACC

In Chapter 5, the CACC system was studied. However it was assumed that all the vehicles
are equipped with identical CACC systems. If the platoon is mixed with vehicle equipped
with the CACC system and vehicles driven manually, the performance of the CACC system
may decrease or CACC may be not appropriate. However, if the number of manually driven
vehicles are relatively smaller compared to the number of vehicles equipped with the CACC
systems, there may be enough data to estimate the behaviors of the manually driven vehicles.
Then the manually driven vehicles can be modeled and included in the CACC system to
enhance the overall performance.

6.2.3 Implementation and Verification

In this thesis, all the performance verifications and evaluations have been done via simula-
tions. Even though the simulator includes realistic models of the vehicle body, the engine,
the brake, and the environments, it cannot perfectly emulate the behavior of actual vehicles.
To better evaluate the suggested control schemes and to find potential problems that were
not considered in this thesis, implementation of the control algorithms and experiments with
actual vehicle systems are necessary.
 117

Bibliography

[1] T. Acarman, Y. Liu, and U. Ozguner. “Intelligent cruise control stop and go with and
without communication”. In: Proceedings of the 2006 American Control Conference
(2006).
[2] K. Ahn et al. “Microscopic Fuel Consumption and Emission Models”. In: Proceedings
of the 78th Annual Meeting of the Transportation Research Board (1999).
[3] B. Arem, C. Driel, and R. Visser. “The impact of cooperative adaptive cruise control
on traffic-flow characteristics”. In: IEEE Transactions on Intelligent Transportation
Systems 7.4 (2006).
[4] V. Bageshwar, W. Garrard, and R. Rajamani. “Model Predictive Control of Transi-
tional Maneuvers for Adaptive Cruise Control Vehicles”. In: IEEE Transactions on
Vehicle Technology 53.5 (2004), pp. 1573–1585.
[5] D. Bertsekas. Nonlinear Programming. Athena Scientific, 1999.
[6] Y. Bin et al. “Longitudinal acceleration tracking control of vehicular stop-and-go cruise
control system”. In: Proceedings of the 2004 IEEE Internation Conference on Network-
ing, Sensing & Control (2004).
[7] S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge, 2004.
[8] D. Bruin et al. “Design and test of a cooperative adaptive cruise control system”. In:
IEEE Intelligent Vehicles Symposium (2004).
[9] D. Cho and J.K. Hedrick. “Automotive powertrain modeling for conrol”. In: ASME
Journal of Dynamic Systems, Measurement and Control 111 (1989), pp. 568–576.
[10] S.C. Davis, S.W. Diegel, and R.G. Boundy. Transportation Energy Data Book: Edition
30. Oak Ridge National Laboratory, 2011.
[11] Z. Eizad and L. Vlacic. “A control algorithm and vehicle model for stop & go cruise
control”. In: IEEE Intelligent Vehicles Symposium (2004).
[12] R. Fletcher. Practical Methods of Optimization. Wiley, 1987.
[13] J. Gerdes. “Decoupled design of robust controllers for nonlinear systems: as motivated
by and applied to coordinated throttle and brake control for automated highways”. In:
PhD Dissertation, University of California, Berkeley (1996).
BIBLIOGRAPHY 118

[14] V. Gupta et al. “Optimal LQG control across a packet-dropping link”. In: Systems
Control Lett. (2005).
[15] C.N. Hadjicostis and R. Touri. “Feedback contol utilizing packet dropping network
links”. In: Proc. IEEE Conf. Decision and Control 2 (2002), pp. 1205–1210.
[16] S. Ibaraki and M. Tomizuka. “H optimization of fixed structure controller”. In: Proc.
Of the 2000 International Mechanical Engineering Congress and Exposition (2001).
[17] S. Ibaraki and M. Tomizuka. “Tuning of a hard disk drive servo controller using fixed-
structure H controller optimization”. In: Journal of Dynamic Systems, Measurement,
and Control 123 (2001), pp. 31–55.
[18] P. Ioannou and C.C. Chien. “Intelligent Cruise Control”. In: IEEE Trans. on Vehicular
Technology 42.4 (1993), pp. 657–672.
[19] S. Kata et al. “Vehicle control algorithms for cooperative driving with automated ve-
hicles and intervehicle communications”. In: IEEE Transactions on Intelligent Trans-
portation Systems 3.3 (2002).
[20] S.G. Kim, M. Tomizuka, and K.H. Cheng. “Mode Switching and Smooth Motion Gen-
eration of the Adaptive Cruise Control System by a Virtual Lead Vehicle”. In: IFAC
Symposium on Control in Transportation Systems (2009).
[21] S.G. Kim, M. Tomizuka, and K.H. Cheng. “Smooth Motion Control of the Adaptive
Cruise Control System with Linear Quadratic Control with Variable Weights”. In:
ASME Dynamic Systems and Control Conference (DSCC) (2010).
[22] S.L. Koo, H.S. Tan, and M. Tomizuka. “Impact of Tire Compliance Behavior to Vehicle
Longitudinal Dynamics and Control”. In: American Control Conference (2007).
[23] H. Kwakernaak and R. Sivan. Linear Optimal Control Systems. New York: Wiley-
Interscience, 1972.
[24] YD. Landau. System identification and control design. Englewood Cli↵s, NJ: Prentice
Hall, 1990.
[25] J. Laumonier, C. Desjardins, and B. Chaib-draa. “Cooperative adaptive cruise control:
a reinforcement learning approach.” In: in Proceedings of 4th Workshop on Agents in
Traffic And Transportation, AAMAS’06 (2006).
[26] Q. Ling and M. Lemmon. “Optimal dropout compensation in networked control sys-
tems”. In: IEEE Conf. Decision and Control (2003).
[27] P. Lingman and B. Schmidtbauer. “Road slope and vehicle mass estimation using
Kalman filtering”. In: Vehicle System Dynamics 37 supp. (2003), pp. 12–23.
[28] J.J. Martinez and C. Canudas-de Wit. “A safe longitudinal control for adaptive cruise
control and stop-and-go scenarios”. In: IEEE Transactions on Control Systems Tech-
nology 15.2 (2007), pp. 246–258.
BIBLIOGRAPHY 119

[29] J.P. Maurice and H.B. Pacejka. “Relaxation length behavior of tyres”. In: Vehicle
system dynamics 27 supplement (1997).
[30] T. Nagata and M. Tomizuka. “Engine torque control based on discrete event model
and disturbance observer”. In: Proceedings of the ASME Int. Mech, Eng, Congr. Expo.
(2007).
[31] T. Nagata and M. Tomizuka. “Robust engine torque control by discrete event distur-
bance observer”. In: Proceedings of the 17th IFAC World Congress (2008).
[32] J.E. Naranjo et al. “ACC+Stop&Go maneuvers with throttle and brake fuzzy control”.
In: IEEE Transactions on Intelligent Transportation Systems 7.2 (2006).
[33] G. Naus et al. “Cooperative adaptive cruise control, design and experiments”. In:
American Control Conference (2010).
[34] G. Naus et al. “Explicit MPC design and performance evaluation of an ACC Stop-&-
Go”. In: Proceedings of the 2008 American Control Conference (2008).
[35] J. Nilsson. “Real-time control systems with delays”. In: Ph.D. dissertation, Dept. Au-
tomatic Control, Lund Inst. Technol. (1998).
[36] J. Nilsson, B. Bernhardsson, and B. Wittenmark. “Stochastic analysis and control of
real-time systems with random time delays”. In: Automatica 34.1 (1998), pp. 57–64.
[37] K. Ohnishi. “A new servo method in mechatronics”. In: Transaction of Japanese So-
ciety of Electrical Engineering 107-D (1987), pp. 83–86.
[38] H. Ohtsuka and L. Vlacic. “Stop & Go vehicle longitudinal model”. In: IEEE 5th
International Conference on Intelligent Transportation Systems (2002).
[39] H.B. Pacejka. Tire and vehicle dynamics. Oxford: Society of Automotive Engineers /
Butterworth-Hienemann, 2002.
[40] Z. Papp, C. Brown, and C. Bartels. “World modeling for cooperative intelligent vehi-
cles”. In: IEEE Intelligent Vehicles Symposium (2008).
[41] M. Persson et al. “Stop & Go controller for adaptive cruise control”. In: Proceedings
of the 2004 IEEE Internation Conference on Control Applications (1999).
[42] R. Rajamani. Vehicle Dynamics and Control. Springer, 2006.
[43] R. Rajamani and S.E. Shladover. “An experimental comparative study of autonomous
and co-operative vehicle-follower control systems”. In: Journal of Transportation Re-
search, Part C Emerging Technologies 9.1 (2001), pp. 15–31.
[44] A. Said and W.A. Pearlman. “An image multiresolution representation for lossless and
lossy compression”. In: IEEE Transaction on Image Processing 5.9 (1996).
[45] J. Schakel, B. Arem, and B. Netten. “E↵ects of cooperative adaptive cruise control on
traffic flow stability”. In: 13th International IEEE Annual Conference on Intelligent
Transportation Systems (2010).
BIBLIOGRAPHY 120

[46] L. Schenato et al. “Foundations of Control and Estimation Over Lossy Networks”. In:
Proceedings of IEEE 95.1 (2007), pp. 163–187.
[47] S. Sheikholeslam and C.A. Desoer. “Longitudinal control of a platoon of vehicles”. In:
Proc. of 1990 American Control Conference (1990), pp. 291–296.
[48] J.-J. E. Slotine and W. Li. Applied Nonlinear Control. Prentice Hall, 1991.
[49] S. Smith and P. Seiler. “Estimation with lossy measurements: Jump estimators for
jump systems”. In: IEEE Trans. Autom. Control 48.12 (2003), pp. 1453–1464.
[50] H. Stearns, S. Mishra, and M. Tomizuka. “Iterative tuning of feedforward controller
with force ripple compensation for wafer stage”. In: Advanced Motion Control, 2008.
10th IEEE International Workshop (2008).
[51] Y. Takae et al. “A study of drivers’ trust in a low-speed following system”. In: SAE
Technical Paper doi:10.4271/2005-01-0430 (2005).
[52] T. Takizawa et al. “Load measuring device for rolling bearing unit and load measuring
rolling bearing unit”. In: U.S. Patent 7 320 257 (2008).
[53] C.K.W. Tam and K.A. Kurbatskii. “Multi-size-meshmulti-time-step dispersion-relation-
preserving scheme for multiple-scales aeroacoustics problems”. In: International Jour-
nal of Computational Fluid Dynamics 17 (2003), pp. 119–132.
[54] H.S. Tan, R. Rajamani, and W.B. Zhang. “Demonstration of an automated highway
platoon system”. In: Proc. of 1998 American Control Conference 3 (1998), pp. 1823–
1827.
[55] T. Umeno and Y. Hori. “Robust speed control of DC servomotors using modern two
degrees of freedom controller design”. In: IEEE Transactions on Industrial electronics
38.5 (1991), pp. 363–368.
[56] A. Vahidi and A. Eskandarianm. “Research Advances in Intelligent Collision Avoidance
and Adaptive Cruise Control”. In: IEEE Transactions on Intelligent Transportation
Systems 4.3 (2003), pp. 143–153.
[57] A. Vahidi, A. Stefanopoulou, and H. Peng. “Recursive least squares with forgetting for
online estimation of vehicle mass and road grade: theory and experiments”. In: Vehicle
System Dynamics 43.1 (2005), pp. 31–55.
[58] P. Venhovens, K. Naab, and B. Adiprasito. “Stop and Go Cruise Control”. In: Inter-
nation Journal of Automotive Technology 1.2 (2000), pp. 61–69.
[59] V. Winstead and V. Kolmanovsky. “Estimation of road grade and vehicle mass via
model predictive control”. In: Proceedings of the 2005 IEEE Conference on Control
Applications (2005).
[60] J.Y. Wong. Theory of ground vehicles. 2nd. John Wiley and Sons, Inc., 1993.
[61] Q. Xu et al. “E↵ects of vehicle-vehicle/ roadside-vehicle communication on adpative
cruise controlled highway systems”. In: IEEE 56th In Vehicular Technology Conference,
2002. Proceedings. VTC 2002-Fall 2 (2002), pp. 1249–1253.
BIBLIOGRAPHY 121

[62] Y. Xu and J. Hespanha. “Estimation under controlled and uncontroller communica-

tions in networked control systems”. In: Proceedings of IEEE Conference Decision and
Control (2005), pp. 842–847.
[63] K. Yi, J. Hong, and Y.D. Kwon. “A Vehicle Control Algorithm for Stop-and-Go Cruise
Control”. In: Journal of Automobile Engineering, Proceedings of the Institution of Me-
chanical Engineers 1.215 (2001), pp. 1099–1115.
[64] K. Yi et al. “Implementation and vehicle tests of a vehicle stop-and-go cruise control
system”. In: Proceedings of the Institution of Mechanical Engineers, Part D: Journal
of Automobile Engineering 216.7 (2002), pp. 537–544.
[65] N. Yoshitani and A. Hasegawa. “Model-based control of strip temperature for the
heating furnace in continuous annealing”. In: IEEE Transactions on Control Systems
Technology 6 (1998), pp. 146–156.
[66] N. Yu et al. “Stabilization of networked control system approach”. In: IEEE Int. Symp.
Computer Aided Control Systems Design (2004), pp. 362–367.