Mechanical Systems and Signal Processing 209 (2024) 111125

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

journal homepage: www.elsevier.com/locate/ymssp

Multi-objective dynamic coordinated Adaptive Cruise Control for

intelligent electric vehicle with sensors fusion
Dongmei Wu a,b,c , Baobao Qiao a,b,c , Changqing Du a,b,c ,∗, Yifan Zhu a,b,c , Fuwu Yan a,b,c ,
Changsheng Liu d , Yang Li e , Jun Li f
a
Hubei Key Laboratory of Advanced Technology for Automotive Components, Wuhan University of Technology, Wuhan 430070, China
b
Foshan Xianhu Laboratory of the Advanced Energy Science and Technology Guangdong Laboratory, Foshan 528200, China
c
Hubei Research Center for New Energy & Intelligent Connected Vehicle, Wuhan University of Technology, Wuhan 430070, China
d College of Computer Science and Technology, Zhejiang University, Hangzhou 310027, China
e Technical Center of Dongfeng Commercial Vehicle, Wuhan 430056, China
f Tsinghua Intelligent Vehicle Design and Safety Research Institute, Tsinghua University, Beijing 100084, China

ARTICLE INFO ABSTRACT

Communicated by J.E. Mottershead Adaptive Cruise Control (ACC) is a multi-objective control problem, where the balance of the
ride comfort and safety performance of vehicles is a challenge. This paper proposes a multi-
Keywords:
objective dynamic coordinated ACC based on variable weight model predictive control (MPC).
Adaptive cruise control
Variable weight
The weight coefficients are regulated according to the inter distance and relative speed between
Sensors fusion the front vehicle and ego vehicle through a fuzzy controller. To improve the object detection
Model predictive control accuracy, the decision-level sensors fusion scheme based on millimeter wave radar and camera
Electric vehicle is designed. The upper and lower layers of the hierarchical ACC controller are implemented.
The proposed ACC control system is tested on a smart electric vehicle. The dynamic coordinated
ACC can increase the tracking accuracy of the safe inter distance to enhance the vehicle safety
in scenarios where there is possible risk of collision, such as when the inter distance is small
and the relative speed is negative. In some safe conditions, the dynamic coordinated ACC can
reduce the acceleration and jerk to improve the ride comfort. The performance of driving safety
and ride comfort are dynamically coordinated and adapted to the changing driving conditions
through the implementation of the dynamic ACC. The ACC has the ability to adapt to these
changes and respond accordingly.

1. Introduction

The Adaptive Cruise Control(ACC) system automatically adjusts the vehicle longitudinal speed based on the behavior of other
vehicles [1]. It relies on sensors such as radar and camera to detect the speed and distance of the object vehicle. As an important
Advanced Driving Assistance System (ADAS), ACC has been widely used in production vehicles [2]. Currently, ACC is designed to
meet various performance criteria, including tracking safety, ride comfort, and energy efficiency [3–5]. However, achieving dynamic
coordination between these different control objectives is a critical challenge due to the complexity and variability of the traffic
environment.

∗ Corresponding author at: Hubei Key Laboratory of Advanced Technology for Automotive Components, Wuhan University of Technology, Wuhan 430070,
China.
E-mail addresses: [email protected] (D. Wu), [email protected] (B. Qiao), [email protected] (C. Du), [email protected] (Y. Zhu),
[email protected] (F. Yan), [email protected] (C. Liu), [email protected] (Y. Li), [email protected] (J. Li).

https://doi.org/10.1016/j.ymssp.2024.111125
Received 27 June 2023; Received in revised form 5 December 2023; Accepted 7 January 2024
Available online 18 January 2024
0888-3270/© 2024 Elsevier Ltd. All rights reserved.
D. Wu et al. Mechanical Systems and Signal Processing 209 (2024) 111125

Nomenclature

𝑑, 𝑑𝑠𝑎𝑓 𝑒 Actual and safe inter space

𝑣ℎ , 𝑣𝑓 Speeds of host vehicle and front
𝜏1 , 𝜏1 Coefficients of the time headway vehicle
𝑎ℎ , 𝑎𝑓 Accelerations of host vehicle and
𝑙 Minimum safe inter space front vehicle
𝜂 Efficiency of the driving system
𝑇𝑒 , 𝑇𝑑𝑒𝑠 Actual and desired output torques of electric motor
𝑖0 Transmission ratio of the main reducer
𝑟 Wheel radius
𝑚 Vehicle mass
𝑓 Coefficient of rolling resistance
𝜃 Road slope angle
𝐶𝐷 Coefficient of air resistance
𝐴 Cross sectional area of vehicle
𝜌𝐴 Air density
𝛿 Rotational mass conversion factor
𝑎𝑑𝑒𝑠 Desired acceleration
𝐾𝑏 Ratio of braking force to braking pressure
𝑃 , 𝑃𝑑𝑒𝑠 Actual and desired braking pressure
𝛥𝑑 Inter distance error
𝛥𝑣 Relative speed of host vehicle and front vehicle
𝐾𝐿 , 𝑇𝐿 Gain and time constant of the inertial system
𝑢 Control input
𝑇𝑠 Discrete time step
𝐾 Feedback coefficient matrix of LQR controller
𝜇𝑥 , 𝜇𝑦 Confidence levels in longitudinal and lateral detection
𝐺0 Threshold of Mahalanobis distance
𝑇𝐸 Detection period of the ESR
𝑁𝐶 , 𝑁𝑃 Prediction and control horizons
𝑢∗ Optimal control input
𝑤 System disturbance
𝛥𝑢 Control increment
𝑄𝐿𝑄𝑅 Weight matrix of state variables for LQR controller
𝑄𝑑,𝐿 , 𝑄𝑣,𝐿 , 𝑄𝑎,𝐿 Weight factors of state variables for LQR controller
𝑅𝐿 Weight factor of control input for LQR controller
𝑄̃ Weight matrices of state variables in prediction horizon for MPC controller
𝑅̃ Weight matrices of control increment in the control horizon for MPC controller
𝑄𝑑 , 𝑄𝑣 , 𝑄𝑎 Weight factors of state variables for MPC controller
𝑅𝑗𝑒𝑟𝑘 Weight factor of control increment for MPC controller
𝛥𝑑min , 𝛥𝑑max Limits of the inter distance error
𝛥𝑣min , 𝛥𝑣max Limits of the relative speed
𝑣min , 𝑣max Speed limits
𝑎min , 𝑎max Acceleration limits
𝛥𝑢min , 𝛥𝑢max Limits of the control increment
𝜀1 , 𝜀2 , 𝜀3 , 𝜀4 , 𝜀5 Relaxation factors
𝑟𝛥𝑑 , 𝑟𝛥𝑣 , 𝑟𝛥𝑎 , 𝑟𝑎 , 𝑟𝑣 Relaxation coefficients of the lower limits for hard constraints
min min min min min
𝑟max , 𝑟𝛥𝑣
𝛥𝑑 𝛥𝑎 𝑎 𝑣
max , 𝑟max , 𝑟max , 𝑟max Relaxation coefficients of the upper limits for hard constraints
𝛾1 , 𝛾2 , 𝛾3 , 𝛾4 , 𝛾5 Weight factors of relaxation factors
𝑤𝑑 , 𝑤𝑣 , 𝑤𝑎 Outputs of fuzzy controller
𝜃𝑑 , 𝜃𝑣 , 𝜃𝑎 Correction coefficients of the weight factors

In previous research, the control objective is mainly regulated by switching to the different control modes when the driving
condition changes. Typically, there are several control modes, such as constant speed cruise mode, following mode, avoid-collision
mode, etc. For instance, ACC control the vehicle for improving the ride comfort and energy efficient in the constant cruise mode.

2
D. Wu et al. Mechanical Systems and Signal Processing 209 (2024) 111125

𝑎𝑟 Acceleration threshold
𝐷 Mahalanobis distance between ESR and camera targets
𝑥𝐸 , 𝑥𝐶 , 𝑥𝐹 Horizontal distances between sensor targets and vehicle
𝑦𝐸 , 𝑦𝐶 , 𝑦𝐹 Lateral distances between sensor targets and vehicle

In following mode, ACC prioritizes driving safety. In Ref. [6], the multi-objective optimization problem is divided into two modes
according to the Time To Collision (TTC): performance-oriented mode and safety-oriented mode. However, even within a particular
control mode, specific driving conditions like inter-distance and relative speed might not be constant. Consequently, current
coordination methods are unable to respond rapidly to these changing conditions. Therefore, a multi-objective dynamic coordinated
ACC is required to adjust control objectives more promptly.
The hierarchical controller typically used in ACC consists of a lower level controller and an upper level controller [7]. The upper
level controller is responsible for vehicle motion control, determining the desired acceleration based on environmental information.
The lower level controller controls the driving and braking actuators to track the desired acceleration. External sensors such as lasers,
cameras, or radars provide surrounding information to the environment perception system. For the decision making in the upper level
controller, several control theories have been adopted such as Proportion Integral Differential (PID), Sliding Mode Control (SMC)
and Model Predictive Control (MPC). Recently, the machine learning algorithms has also been widely researched for ACC [8,9].
In particular, MPC is a suitable method for handling multiple objectives and constraints and has been gradually applied to ACC in
recent years [10,11].
Millimeter wave radar are commonly used for target detection in current ACC system [12]. Radar can provide accurate radial
distance and velocity of objects, as well as detect obstacles [13]. However, relying solely on radar detection may result in missed
targets and low accuracy of tangential distance and velocity measurements. In Ref. [14], the relevant distance has been resiliently
and accurately estimated for the first time to address the cyber attack posing on automated vehicles to fuse data from multi-connected
automated vehicles. Other alternative is to fuse sensor data such as camera and radar on the ego vehicle. The fusion of millimeter
wave radar and camera are researched recently, which can be used for other ADAS functions. The camera has obvious advantages
in object recognition, and can accurately detect the lateral movement of the object [15], but it is affected by lighting, rain, and
fog. Combining the information from cameras and radar can provide more accurate information about the environment [16,17].
According to the level of fusion, the methods of multi-sensor fusion can be divided into three types: pixel-level fusion, feature-level
fusion, and decision-level fusion [18].pixel-level fusion fuses the raw data from radar and vision to obtain new, more informative raw
data [19]. Many feature-level fusion techniques define a region of interest (ROI) from radar measurements and validate the detection
results in camera image data using various image processing techniques [20]. The decision-level fusion uses target information
independently by radar and vision sensors. For low-speed enclosed experimental environments, decision-level fusion offers lower
cost and higher efficiency. Therefore, this paper utilizes decision-level fusion to detect the target information for ACC.
ACC can be formulated as a multi-objective optimization problem based on MPC, where various criteria can be integrated. In
Ref. [21], the cost function is formulated to minimize the speed tracking error, the acceleration increment and the acceleration. In
Ref. [22], the optimization objective includes tracking capability, high fuel economy, and good driver desired response. Three items
of cost function are combined using a linear weighting method. Ref. [23] formulates a cost function and constraint conditions to
achieve coordination among fuel economy, tracking capability, and ride comfort. It also incorporates sensitivity functions for relative
velocity and distance error into the cost function to describe expected control gain of the driver over tracking errors. For the above
research, the weighting matrices of each cost items are not presented in detail or set as constant for the specific control system.
Weight factors should be regulated dynamically in a multi-objective optimization problem to coordinate performance requirements.
Ref. [24] optimizes the weights of the MPC using personal driving data, solved by the Particle Swarm Optimization algorithm.
However, this data-driven optimization method is unsuitable for real vehicles. In Ref. [25], the tracking and economic weights
are switched among only four fixed combinations based on driving conditions and vehicle speed. However, this approach lacks
comprehensive consideration of variable weights and may not be suitable for all traffic scenarios.
According to the above review, it is critical for ACC to dynamically coordinate various performance requirements to adapt to a
complex traffic environment. The current control methods based on control mode switching or step weights need improvement in
coordinating multi-objective more efficiently. In this paper, a multi-objective dynamic coordinated ACC is proposed using a variable
weight model predictive control. The main contributions of this paper are as follows:

(1) A multi-objective dynamic coordination idea that utilizes a fuzzy controller is proposed to regulate the weight factors of MPC
according to the inter distance and relative speed between the front vehicle and ego vehicle. This approach can overcome
the non-real-time and sudden changes in the process of multi-objective coordination of existing methods.
(2) To improve the object detection accuracy, the decision-level sensors fusion scheme based on millimeter wave radar and
camera is developed. Compared with the current radar-based ACC system, this solution is conducive to improving control
performance and integration with other ADAS functions.
(3) At present, the MPC method is rarely applied in the actual ACC system because of the high calculation load. In this paper,
the proposed MPC-based ACC system is tested on a real vehicle, which promotes the application and development of MPC in
real vehicles.

3
D. Wu et al. Mechanical Systems and Signal Processing 209 (2024) 111125

Fig. 1. Hierarchical control framework of ACC.

The remaining parts of the paper are organized as follows: Section 2 introduces the dynamic coordinated ACC framework and
modeling, while Section 3 presents decision-level sensors fusion of camera and radar. In Section 4, the variable weight MPC method
is adopted to realize the ACC upper level control, followed by simulation and analysis of the ACC model using MATLAB/Simulink
in Section 5. Finally, Section 6 conducts tests on the vehicle to verify the proposed algorithm.

2. Control framework and modeling of ACC

2.1. Control framework of ACC based on MPC

Currently, the hierarchical control framework stands as the mainstream architecture of an ACC system, consisting of upper and
lower controllers as illustrated in Fig. 1. The upper controller determines the expected longitudinal acceleration based on the motion
information of the host vehicle and the target vehicle obtained from the perception layer, including speed, acceleration, vehicle
spacing, and relative speed. Firstly, two control modes are distinguished based on the mode switching strategy: the cruise control
(CC) mode and the follow cruise mode (FC). When the sensor detects the absence of a target vehicle ahead, the vehicle switches
into CC mode. Conversely, when the sensor detects the presence of a target vehicle ahead, the vehicle switches to FC mode. The
expected inter space is determined according to the safe space model. Then, the MPC controller is formulated and solved to obtain
the desired acceleration. The weight factors of the MPC is regulated based on the fuzzy algorithm to meet the multiple objectives.
The lower controller receives the desired acceleration signal and control the driving system or braking system to realize the expected
acceleration. The switching logic of the driving and braking is initially utilized to select the actuators: driving motor or brake. Then,
the desired acceleration is transformed into the control input of the actuators based on the inverse longitudinal dynamics model: the
expected throttle opening or expected braking pressure. Finally the longitudinal control of the vehicle is completed and the vehicle
state registers as feedback to the control system. The sensing layer is responsible for collecting the signals of wheel speed, vehicle
speed, acceleration, relative speed, relative distance, etc. It serves as the data input of the decision-making layer. In this paper, the
information about the target vehicle is measured by the sensor fusion of a camera and millimeter-wave radar. The states of the host
vehicle are measured by the inertial measurement unit (IMU).

2.2. Safe vehicle spacing model

The safe spacing between vehicles is a crucial input for the upper decision-making algorithm of the ACC system. The widely-used
variable safe spacing model determines the safe inter distance based on the speed of the host vehicle and constant time headway,
which can be expressed through the following equation:

𝑑𝑠𝑎𝑓 𝑒 = 𝜏 ⋅ 𝑣ℎ + 𝑙 (1)

where 𝑑𝑠𝑎𝑓 𝑒 is the safe inter space, 𝜏 is the time headway (THW), 𝑙 is minimum safe vehicle spacing, 𝑣ℎ is the speed of host vehicle.
Considering the impact of relative speed on the THW, this paper improved the above variable spacing model as follows:

𝑑𝑠𝑎𝑓 𝑒 = 𝜏1 ⋅ 𝑣ℎ + 𝜏2 ⋅ (𝑣ℎ − 𝑣𝑓 ) + 𝑙 (2)

where 𝜏1 and 𝜏2 are the positive constants which can be well calibrated by the work in Ref. [26], which presents the comprehensive
data processing pipeline for traffic leveraging the advanced cooperative perception technique for the first time. 𝑣𝑓 is the speed of
front vehicle. In this paper, 𝜏1 = 𝜏2 = 1.

4
D. Wu et al. Mechanical Systems and Signal Processing 209 (2024) 111125

Fig. 2. Fusion architecture diagram.

2.3. Vehicle dynamics model

The dynamic equation of the electric vehicle researched in this paper, under driving conditions, can be expressed as Eq. (3).
𝜂𝑇𝑒 𝑖0 1
= 𝑚𝑔𝑓 cos 𝜃 + 𝐶𝐷 𝐴𝜌𝐴 𝑣2 + 𝑚𝑔 sin 𝜃 + 𝛿𝑚𝑎 (3)
𝑟 2
where 𝜂 is the efficiency of the driving system, 𝑇𝑒 is the output torque of electric motor, 𝑖0 is the transmission ratio of the main
reducer, 𝑟 is the wheel radius, 𝑚 is the vehicle mass, 𝑓 is the rolling resistance coefficient, 𝐶𝐷 is the air resistance coefficient, 𝐴 is
the cross sectional area of vehicle, 𝜌𝐴 is the air density. 𝜃 is road slope angle. 𝛿 is rotational mass conversion factor. As there is no
flywheel in in-wheel motor electric vehicle, 𝛿 is set to 1 in this paper.
Under braking conditions, the dynamic equation of the vehicle is:
1
𝛿𝑚𝑎 = −(𝐾𝑏 𝑃 + 𝑚𝑔𝑓 cos 𝜃 + 𝐶 𝐴𝜌 𝑣2 + 𝑚𝑔 sin 𝜃) (4)
2 𝐷 𝐴
where 𝐾𝑏 is the linear coefficient, 𝑃 is braking pressure of the master cylinder.

3. Multi-sensors fusion for object detection

3.1. Fusion algorithm framework

Perception system comprises one millimeter-wave radar and one camera in this paper. The millimeter-wave radar is Electronic
Scanning Radar (ESR) produced by Delphi. The camera is the ME560 produced by Mobileye. Compared to the ME560 camera,
ESR can detect more targets while ME560 can only detect pedestrians and the car’s rear. Regarding accuracy of detection, ME560
provides a more reliable lateral distance and target type, while ESR excels in detecting longitudinal distance and target speed.
After receiving the radar and camera sequences, the data fusion node synchronizes them and fuses their advantages to improve the
stability and reliability of the target detection system. The sensors fusion architecture is shown in Fig. 2.
The signals transmitted by the millimeter-wave radar and camera to the industrial computer need to be unpacked and filtered.
Valid target information will be sent to the ROS network. Millimeter-wave radars and cameras also require the vehicle’s speed and
yaw rate signals, which are supplied by the IMU and transformed to the ROS network. In this paper, the Socket CAN interface to
receive and send CAN signals are written in C++, CAN-Pack/Unpack module is built by Simulink, which will be converted to C++
code. The above code will be compiled together with the data processing algorithm code to generate ROS nodes.

3.2. Data preprocessing for radar and camera

Before performing sensor fusion, preprocessing is necessary for both the millimeter-wave radar and the camera signals. For the
ESR signal, filtering is applied to remove empty target, targets outside the specified range, and invalid targets. When the ESR fails to
detect a target, it generates an empty target with a velocity of 81.91 m/s. In this case, the TRACK_STATUS is set to ‘‘No Target’’ to
filter out empty target. Setting appropriate detection ranges for the longitudinal distance and lateral distance reduces the amount of
preprocessing information. In this paper, the detection ranges for longitudinal distance and lateral distance are defined as (0, 30) m
and (−3, 3) m, respectively. Invalid targets include targets that cannot be continuously detected in the radar data and targets with
significant state changes. This paper selects longitudinal distance, lateral distance, and longitudinal velocity as state parameters. A
target can be considered valid if the ESR detects the same ID in 5 consecutive cycles and the differences in longitudinal distance,
lateral distance, and longitudinal velocity between adjacent cycles are less than 0.7 m, 0.7 m, and 0.3 m/s, respectively. Otherwise,
it will be considered invalid.
After the filtering process, the five nearest data points are selected for each valid target, and the mean filter algorithm is applied
for the smoothing process, resulting in the initial sequence from the ESR. Due to uncertainty factors, there may be frame loss and

5
D. Wu et al. Mechanical Systems and Signal Processing 209 (2024) 111125

changes in target ID during the transmission of ESR data, leading to target omissions. Therefore, after obtaining the initial sequence
from the ESR, the state of the missed targets is estimated based on Eq. (5).
⎡𝑠𝑛+1|𝑛 ⎤ ⎡1 𝑇𝐸 𝑇𝐸 2 ∕2⎤ ⎡ 𝑠𝑛 ⎤
⎢𝑣 ⎥ = ⎢0 1 𝑇𝐸 ⎥ ⎢𝑣𝑛 ⎥ (5)
⎢ 𝑛+1∣𝑛 ⎥ ⎢ ⎥⎢ ⎥
⎣𝑎𝑛+1∣𝑛 ⎦ ⎣0 0 1 ⎦ ⎣ 𝑎𝑛 ⎦
where 𝑇𝐸 represents the detection period of the ESR. 𝑠𝑛 , 𝑣𝑛 , and 𝑎𝑛 are distance, velocity and acceleration of the current cycle,
respectively. 𝑠𝑛+1∣𝑛 , 𝑣𝑛+1∣𝑛 , and 𝑎𝑛+1∣𝑛 represent the estimated values based on the current cycle. The estimated state values are
compared with the state values of other targets in the sequence, utilizing the target state parameters of distance, velocity, and
azimuth. If the differences between the estimated values and state values of targets in the sequence are less than 1.3 m, 0.25 m/s,
and 12◦ , respectively, it is considered that the target and the missed target are the same. For the camera, if no target is detected,
no signal is sent. Therefore, the preprocessing for the camera does not include filtering out empty targets, while the remaining
processing steps and methods are similar to those of ESR.

3.3. Coordinate transformation and time synchronization

Due to the different installation positions of the radar and the camera, the spatial deviation of the measurement values of the
two sensors will be generated. In terms of coordinate transformation, this paper mainly considers the differences between the two
sensors in the longitudinal distance and the lateral distance.
Given that the CAN signal period differs between the ESR and the camera, it is necessary to ensure that the time interval between
the transmission of the two sensor sequences falls within an acceptable range. This paper uses the message synchronization function
supplied by ROS to synchronize the two target sequences. Once the camera and radar data processing ROS nodes receive the final
target sequence, the current time information is appended to the time stamp within the Header subclass of this class. Then the
message synchronization function will handle two sequences with the closest time interval.

3.4. Target matching

In decision-level fusion, the detection of targets by radar and camera is independent, so targets in two sets of sequences need to
be matched. This paper adopts location information as the matching parameter and utilizes Mahalanobis Distance (MD) to quantify
the degree of match between the two targets, considering varying reliabilities for different parameters. The calculation formula for
MD is defined as follows:
{ [ ]𝑇
𝑉 = 𝑥𝐸 − 𝑥𝐶 , 𝑦𝐸 − 𝑦𝐶
(6)
𝐷2 = 𝑉 𝑇 𝑆 −1 𝑉 ≤ 𝐺0
where 𝑥𝐸 , 𝑦𝐸 , 𝑥𝐶 , 𝑦𝐶 are the horizontal and lateral distances of the radar and camera targets, respectively. 𝐷 is the MD of two targets,
𝑆 is the covariance matrix corresponding to longitudinal distance and lateral distance, 𝐺0 is the threshold of MD. When the MD is
greater than the threshold, it is considered that the two targets are not generated by the same source.
In this paper, Global Nearest Neighbor (GNN) algorithm is used for target matching. After the two sequences are obtained, the
MD of all cases where the targets of the two sequences match each other will be calculated. The matching scheme with the smallest
sum of MD will be selected, which means the targets matching is completed.

3.5. Data fusion

After target matching, two scenarios arise based on the size of the MD in the paired data: MD exceeding the threshold and MD
falling below the threshold. When the MD is below the threshold, a weighted evaluation fusion algorithm is employed for data
fusion, as depicted in the following equation:
{
𝑥𝐹 = 𝜇𝑥 ⋅ 𝑥𝐸 + (1 − 𝜇𝑥 ) ⋅ 𝑥𝐶
(7)
𝑦𝐹 = 𝜇𝑦 ⋅ 𝑦𝐸 + (1 − 𝜇𝑦 ) ⋅ 𝑦𝐶
where 𝑥𝐹 , 𝑦𝐹 are the horizontal and lateral distances of the fusion target, respectively. 𝜇𝑥 and 𝜇𝑦 represent the confidence levels of
millimeter wave radar in longitudinal and lateral detection, with values of 0.8 and 0.1 used in this paper. This choice is based on
the fact that ESR provides higher accuracy in measuring longitudinal distances but performs poorly in detecting lateral motion of
objects.
For paired data with MD exceeding the threshold, the target state detected by ESR is initially determined according to the
following equation:
{
0 𝑚 < 𝑥𝐸 < 20 𝑚
(8)
− 0.8 𝑚 < 𝑦𝐸 < 0.8 𝑚
When the detected ESR target state does not satisfy the equation above, it is classified as debris located on either side of the
vehicle, such as roadside trees. In such cases, only the target data from the camera will be outputted. Conversely, when the ESR
target state satisfies the equation above, it is recognized as an object capable of influencing the vehicle’s movement. To ensure
safety, the ESR target data will be outputted.

6
D. Wu et al. Mechanical Systems and Signal Processing 209 (2024) 111125

4. Multi-objective dynamic coordinated ACC based on variable weight MPC

4.1. Predictive model

The model of ACC can be described as follows:

⎧𝛥𝑑̇ = 𝛥𝑣 − (𝜏 + 𝜏 )𝑎 + 𝜏 𝑎
⎪ 1 2 ℎ 2 𝑓
⎪𝛥𝑣̇ = 𝑎𝑓 − 𝑎ℎ
⎨ (9)
⎪ 𝐾𝐿 1
⎪𝑎̇ ℎ = 𝑇 𝑎𝑑𝑒𝑠 − 𝑇 𝑎ℎ
⎩ 𝐿 𝐿

where 𝛥𝑑 = 𝑑 − 𝑑𝑠𝑎𝑓 𝑒 is the distance error, 𝑑 is the inter space, 𝛥𝑣 = 𝑣𝑓 − 𝑣ℎ is the relative speed, 𝑎ℎ is the real acceleration of the
host vehicle, 𝑎𝑓 is the real acceleration of the front vehicle, 𝑎𝑑𝑒𝑠 is the desired acceleration. 𝐾𝐿 = 0.98 is the gain of the inertial
system, 𝑇𝐿 = 0.53 is the time constant of the inertial system. As this paper mainly focuses on the research of ACC control algorithms,
the selection of 𝐾𝐿 and 𝑇𝐿 is based on the empirical values of relevant vehicle models [6].
The state space model of ACC model is established as follows:
{
𝑋̇ = 𝐴𝑎 𝑋 + 𝐵𝑎 𝑢 + 𝐺𝑎 𝜔
(10)
𝑌 = 𝐶𝑎 𝑋
[ ]𝑇
where 𝑋 = 𝛥𝑑, 𝛥𝑣, 𝑎ℎ , 𝑌 = 𝑋, 𝑢 = 𝑎𝑑𝑒𝑠 , 𝑤 is the system disturbance, 𝑤 = 𝑎𝑓 ,

⎡0 1 −(𝜏1 + 𝜏2 )⎤ ⎡ 0 ⎤ ⎡1 0 0⎤ ⎡𝜏2 ⎤
𝐴𝑎 = ⎢0 0 −1 ⎥ , 𝐵𝑎 = ⎢ 0 ⎥ , 𝐶𝑎 = ⎢0 1 0⎥ , 𝐺𝑎 = ⎢ 1 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣0 0 −1∕𝑇𝐿 ⎦ ⎣𝐾𝐿 ∕𝑇𝐿 ⎦ ⎣0 0 1⎦ ⎣0⎦
The controllability discriminant matrix of the state space equation is [𝐵, 𝐴𝐵, 𝐴2 𝐵]. Given that the rank of [𝐵, 𝐴𝐵, 𝐴2 𝐵] is 3, it
can be concluded that the system exhibits complete controllability.

4.2. Design of LQR controller

The following mode of ACC system is a typical linear quadratic optimal control problem. To facilitate comparison with the
dynamic ACC strategy, this paper introduces the design of Linear Quadratic Regulator (LQR) controller. The goal of following mode
is to reduce the distance error (𝛥𝑑), the relative speed (𝛥𝑣) and acceleration (𝑎ℎ ), therefore the cost function of LQR is designed as
follows:
∞
1
𝐽𝐿 = (𝑋 𝑇 𝑄𝐿 𝑋 + 𝑢𝑇 𝑅𝐿 𝑢) (11)
2 ∫0
where 𝑄𝐿 = 𝑑𝑖𝑎𝑔[𝑄𝑑,𝐿 , 𝑄𝑣,𝐿 , 𝑄𝑎,𝐿 ] is the weight coefficient matrix of the state variable, and 𝑄𝑑,𝐿 , 𝑄𝑣,𝐿 , 𝑄𝑎,𝐿 are the weight factors
of 𝛥𝑑, 𝛥𝑣 and 𝑎ℎ , respectively. 𝑅𝐿 is the weight factor of the control input. This paper takes into account the safety, following, and
ride comfort of the ACC system. After conducting multiple tests, the values for 𝑄𝑑,𝐿 , 𝑄𝑣,𝐿 , 𝑄𝑎,𝐿 and 𝑅𝐿 are set to 1, 1, 1 and 30,
respectively, in both simulation and experiment. The optimal control input is obtained through calculation:

𝑢∗ = −𝐾𝑋 (12)

where 𝐾 = [𝑘1 , 𝑘2 , 𝑘3 ] is optimal control feedback coefficient matrix, and 𝑘1 = −0.1826, 𝑘2 = −0.4586, 𝑘3 = 0.3816 are solved by
Matlab.

4.3. Formulation of MPC problem

In this paper, Forward Euler Method is used for discretization. Specifically, the method takes the following form:

𝑋(𝑘 + 1) = (𝐼 + 𝐴𝑎 𝑇𝑠 )𝑋(𝑘) + 𝑇𝑠 𝐵𝑎 𝑢(𝑘) + 𝑇𝑠 𝐶𝑎 𝑤(𝑘) (13)

where 𝑇𝑠 is the discrete time step, in this paper, 𝑇𝑠 = 0.02 s.

The discretization form of ACC model is as follows:
{
𝑋(𝑘 + 1) = 𝐴𝑏 𝑋(𝑘) + 𝐵𝑏 𝑢(𝑘) + 𝐺𝑏 𝑤(𝑘)
(14)
𝑌 (𝑘) = 𝐶𝑏 𝑋(𝑘)
Among them:

⎡1 𝑇𝑠 −(𝜏1 + 𝜏2 ) ⋅ 𝑇𝑠 ⎤ ⎡ 0 ⎤ ⎡𝜏2 ⋅ 𝑇𝑠 ⎤
⎢ ⎥ ⎢ ⎥
𝐴𝑏 = ⎢0 1 −𝑇𝑠 ⎥ , 𝐵𝑏 = ⎢ 0 ⎥ , 𝐺𝑏 = ⎢⎢ 𝑇𝑠 ⎥⎥ , 𝐶𝑏 = 𝐶𝑎
⎢ 𝑇 ⎥ ⎢ 𝐾𝐿 𝑇 𝑠 ⎥ ⎣ 0 ⎦
⎣0 0 1 − 𝑇𝑠 ⎦ ⎣ 𝑇𝐿 ⎦
𝐿

7
D. Wu et al. Mechanical Systems and Signal Processing 209 (2024) 111125

Considering the existence of jerk which is an important indicator for evaluating comfort in constraints, the control input in
Eq. (14) is rewritten into a control increment which can be used to evaluate the value of jerk. The state equation of control increment
form is as follows:
{
𝜉(𝑘 + 1|𝑘) = 𝐴𝑐 𝜉(𝑘) + 𝐵𝑐 𝛥𝑢(𝑘) + 𝐺𝑐 𝑤(𝑘)
(15)
𝜂(𝑘) = 𝐶𝑐 𝜉(𝑘)
Among them:

𝛥𝑢(𝑘) = 𝑢(𝑘) − 𝑢(𝑘 − 1),

⎡𝛥𝑑(𝑘 + 1)⎤ ⎡1 0 0 0⎤
⎢ ⎥ [ ] [ ] [ ] ⎢ ⎥
𝛥𝑣(𝑘 + 1) ⎥ 𝐴𝑏 𝐵𝑏 𝐵𝑏 𝐺𝑏 0 1 0 0⎥
𝜉(𝑘 + 1|𝑘) = ⎢ , 𝐴𝑐 = , 𝐵𝑐 = , 𝐺𝑐 = , 𝐶𝑐 = ⎢
⎢ 𝑎0 (𝑘 + 1) ⎥ 0𝑚×𝑛 𝐼𝑚 𝐼𝑚 0𝑚 ⎢0 0 1 0⎥
⎢ 𝑢(𝑘) ⎥ ⎢0 0 0 1⎥⎦
⎣ ⎦ ⎣
In this paper, the prediction horizon 𝑁𝑃 = 20 and the control horizon 𝑁𝐶 = 3 are selected. In order to derive the prediction
equation, the control input beyond the control horizon is set to 0, i.e. 𝛥𝑢(𝑘 + 𝑖) = 0, 𝑖 = 𝑁𝑐 , 𝑁𝑐 + 1, … , 𝑁𝑝 − 1. It is assumed that the
disturbance remains in the prediction horizon, i.e. 𝑤(𝑘) = 𝑤(𝑘 + 𝑖), 𝑖 = 1, 2, … , 𝑁𝑝 . When the state variables and control increment
at time step 𝑘 are known, the system output within the prediction horizon can be estimated according to Eq. (15):

𝑌 (𝑘) = 𝐴𝑑 𝜉(𝑘) + 𝐵𝑑 𝛥𝑈 (𝑘) + 𝐺𝑑 𝑊 (𝑘) (16)

Among them:
⎡ 𝜂(𝑘 + 1|𝑘) ⎤ ⎡ 𝐴𝑐 ⎤ ⎡ 𝐶𝑐 𝐵𝑐 0 ⋯ 0 ⎤
⎢ ⎥ ⎢ 𝐴2 ⎥ ⎢ 𝐶 𝐴 𝐵 𝐶𝑐 𝐵𝑐 ⋯ 0 ⎥
𝜂(𝑘 + 2|𝑘)
𝑌 (𝑘) = ⎢ ⎥ , 𝐴𝑑 = ⎢ 𝑐 ⎥ , 𝐵𝑑 = ⎢ 𝑐 𝑐 𝑐 ⎥
⎢ ⋮ ⎥ ⎢ ⋮ ⎥ ⎢ ⋮ ⋮ ⋱ ⋮ ⎥
⎢𝜂(𝑘 + 𝑁 |𝑘)⎥ ⎢ 𝑁𝑝 ⎥ ⎢ 𝑁𝑝 −1 𝑁𝑝 −2 𝑁𝑝 −𝑁𝑐 ⎥
⎣ 𝑝 ⎦ ⎣𝐴𝑐 ⎦ ⎣𝐶𝑐 𝐴𝑐 𝐵𝑐 𝐶𝑐 𝐴𝑐 𝐵𝑐 ⋯ 𝐶𝑐 𝐴𝑐 𝐵𝑐 ⎦
⎡ 𝛥𝑢(𝑘) ⎤ ⎡ 𝐺𝑐 0 ⋯ 0⎤ ⎡ 𝑤(𝑘) ⎤
⎢ ⎥ ⎢ 𝐴 𝐺 𝐺𝑐 ⋯ 0⎥ ⎢ ⎥
𝛥𝑢(𝑘 + 1) ⎥ 𝑤(𝑘 + 1)
𝛥𝑈 (𝑘) = ⎢ ,𝐺 = ⎢ ⎥ , 𝑊 (𝑘) = ⎢ ⎥
𝑐 𝑐
⎢ ⋮ ⎥ 𝑑 ⎢ ⋮ ⋮ ⋱ ⋮⎥ ⎢ ⋮ ⎥
⎢𝛥𝑢(𝑘 + 𝑁 − 1)⎥ ⎢ 𝑁𝑝 −1 𝑁𝑝 −2 ⎥ ⎢𝑤(𝑘 + 𝑁 − 1)⎥
⎣ 𝑐 ⎦ ⎣ 𝑐
𝐴 𝐺𝑐 𝐴 𝑐 𝐺𝑐 ⋯ 𝐺𝑐 ⎦ ⎣ 𝑝 ⎦
In order to achieve multi-objective optimization of safety, following, the ride comfort and obtain the optimal control input, the cost
function is constructed according to the control objectives as follows:

𝐽 = 𝑌 (𝑘)𝑇 𝑄𝑌
̃ (𝑘) + 𝛥𝑈 (𝑘)𝑇 𝑅𝛥𝑈
̃ (𝑘) (17)

where:
⎧𝑄̃ = 𝑑𝑖𝑎𝑔(𝑄 𝑄 ⋯ 𝑄)𝑁𝑝
⎪
⎪𝑅̃ = 𝑑𝑖𝑎𝑔(𝑅 𝑅 ⋯ 𝑅)𝑁𝑐
⎨
⎪𝑄 = 𝑑𝑖𝑎𝑔(𝑄𝑑 𝑄𝑣 𝑄𝑎 𝑄𝑎 )
⎪
⎩𝑅 = 𝑅𝑗𝑒𝑟𝑘
𝑄̃ and 𝑅̃ are the weight matrices of the state variables in the prediction horizon and the control increment in the control horizon,
respectively. 𝑄𝑑 , 𝑄𝑣 , 𝑄𝑎 and 𝑅𝑗𝑒𝑟𝑘 are the weight factors corresponding to distance error, relative speed, acceleration and jerk of the
host vehicle, respectively.
To improve the safety and following of ACC system, constraints have been incorporated in the following equation to avoid
excessive distance error and relative speed:
{
𝛥𝑑min ≤ 𝛥𝑑(𝑘) ≤ 𝛥𝑑max
(18)
𝛥𝑣min ≤ 𝛥𝑣(𝑘) ≤ 𝛥𝑣max
where 𝛥𝑑min = 0 m, 𝛥𝑑max = 5 m, 𝛥𝑣min = −3 m/s, and 𝛥𝑣max = 3 m/s are the lower and upper limits of distance error and relative
speed, respectively. Notably, the values for all parameters in MPC constraints are initially derived based on practical experience and
subsequently optimized through multiple simulation tests to obtain the final values.
Considering traffic regulations, the speed of the host vehicle is limited:

𝑣min ≤ 𝑣(𝑘) ≤ 𝑣max (19)

where 𝑣min = 0 m/s and 𝑣max = 40 m/s are the lower and upper limits of speed, respectively.
Longitudinal acceleration and jerk are closely related to the ride comfort in ACC. To improve the ride comfort, we restrict the
acceleration and jerk as follows:
{
𝑎min ≤ 𝑎ℎ (𝑘) ≤ 𝑎max
(20)
𝑎̇ min ≤ 𝑎̇ ℎ (𝑘) ≤ 𝑎̇ max

8
D. Wu et al. Mechanical Systems and Signal Processing 209 (2024) 111125

where 𝑎min = −5 m/s2 and 𝑎max = 5 m/s2 are the lower and upper limits of acceleration, respectively. 𝑎̇ min , 𝑎̇ max are the lower and
upper limits of jerk, respectively. In this paper, the control increment described in Eq. (15) can be used to evaluate jerk, thereby
𝛥𝑢min = −0.06 m/s2 , 𝛥𝑢max = 0.06 m/s2 could serve as substitutes for 𝑎̇ min , 𝑎̇ max .
Considering the above optimization objectives, the constraints of the MPC problem can be defined as follows:
⎧𝛥𝑑min ≤ 𝛥𝑑(𝑘) ≤ 𝛥𝑑max
⎪
⎪𝛥𝑣min ≤ 𝛥𝑣(𝑘) ≤ 𝛥𝑣max
⎪
⎨𝛥𝑢min ≤ 𝛥𝑢(𝑘) ≤ 𝛥𝑢max (21)
⎪
⎪𝑎min ≤ 𝑎ℎ (𝑘) ≤ 𝑎max
⎪𝑣
⎩ min ≤ 𝑣(𝑘) ≤ 𝑣max
Driving conditions can be complex, and sometimes it may be infeasible for a controller to satisfy all the prescribed constraints.
Consequently, finding an optimal or even a feasible solution for the objective function becomes a challenge. To address this issue,
the relaxation factors can be introduced to soften the constraints. The resulting constraint conditions are as follows:
⎧𝛥𝑑min + 𝜀1 𝑟𝛥𝑑 ≤ 𝛥𝑑(𝑘 + 𝑖|𝑘) ≤ 𝛥𝑑max + 𝜀1 𝑟𝛥𝑑 max
min
⎪
⎪𝛥𝑣min + 𝜀2 𝑟𝛥𝑣
min
≤ 𝛥𝑣(𝑘 + 𝑖|𝑘) ≤ 𝛥𝑣max + 𝜀 2 𝑟 𝛥𝑣
max
⎪ 𝛥𝑢 𝛥𝑢
⎨𝛥𝑢min + 𝜀3 𝑟min ≤ 𝛥𝑢(𝑘 + 𝑖|𝑘) ≤ 𝛥𝑢max + 𝜀3 𝑟max (22)
⎪ 𝑎 𝑎
⎪𝑎min + 𝜀4 𝑟min ≤ 𝑎ℎ (𝑘 + 𝑖|𝑘) ≤ 𝑎max + 𝜀4 𝑟max
⎪𝑣 + 𝜀 𝑟𝑣 ≤ 𝑣(𝑘 + 𝑖|𝑘) ≤ 𝑣 𝑣
⎩ min 5 min max + 𝜀5 𝑟max

where 𝜀1 , 𝜀2 , 𝜀3 , 𝜀4 , 𝜀5 are relaxation factors with values greater than 0, 𝑟𝛥𝑑 , 𝑟𝛥𝑣 , 𝑟𝛥𝑎 , 𝑟𝑎 , 𝑟𝑣 are relaxation coefficients of the lower
min min min min min
𝛥𝑑 𝛥𝑣 𝛥𝑎 𝑎 𝑣
limits for the hard constraints, 𝑟max , 𝑟max , 𝑟max , 𝑟max , 𝑟max are relaxation coefficients of the upper limits for the hard constraints.
In order to adjust the ‘‘soft and hard’’ degree of the constraints, different weights of relaxation factors can be added to the cost
function as follows:

𝐽 (𝑌 (𝑘), 𝛥𝑈 (𝑘), 𝜀) = 𝐽 (𝑌 (𝑘), 𝛥𝑈 (𝑘)) + 𝜀𝑇 𝛾𝜀 (23)

where 𝜀 = 𝑑𝑖𝑎𝑔[𝜀1 , 𝜀2 , 𝜀3 , 𝜀4 , 𝜀5 ], 𝛾 = 𝑑𝑖𝑎𝑔[𝛾1 , 𝛾2 , 𝛾3 , 𝛾4 , 𝛾5 ] is weight matrix of relaxation factors. All the values of 𝛾1 , 𝛾2 , 𝛾3 , 𝛾4 and 𝛾5
are set to 20.
Eq. (23) can be converted to the quadratic programming form:
⎧ 1 𝑇 𝑇
⎪min( 𝑥 𝐻𝑥 + 𝐹 𝑥)
⎨ 2 (24)
⎪𝐴𝜙 𝑥 ≤ 𝐵𝜙
⎩
Among them:
⎡𝑀 𝑈 ⎤
𝛯max ⎡ 𝑎max − 𝑎(𝑘) ⎤
[ ] [ ] ⎢ ⎥ ⎢ ⎥
̃ 𝑑 + 𝑅)
2(𝐵𝑑𝑇 𝑄𝐵 ̃ 0𝑁𝑐 ×5 2𝐵𝑑𝑇 𝑄̃ 𝑇 𝐸 ⎢ −𝑀
𝑈
𝛯min ⎥ ⎢ 𝑎min + 𝑎(𝑘) ⎥
, 𝐴𝜙 = ⎢
𝑌 ⎥ ⎢ ⎥
𝐻= ,𝐹 = , 𝐵𝜙 =
05×𝑁𝑐 𝛾 05×1 ⎢ 𝐵𝑑 𝛯max ⎥ ⎢ 𝑌max − 𝐴𝑑 𝜉(𝑘) − 𝐺𝑑 𝑊 (𝑘) ⎥
⎢ 𝑌 ⎥ ⎢ ⎥
⎣−𝐵𝑑 𝛯min ⎦ ⎣−𝑌min + 𝐴𝑑 𝜉(𝑘) + 𝐺𝑑 𝑊 (𝑘)⎦
𝑈
⎧𝛯 = 𝑟𝑎max 𝛶𝑁𝑐 ×1 × [0, 0, 0, 1, 0]
⎡1 0 ⋯ 0⎤ ⎪ max
[ ] ⎢ ⎥ 𝑈 𝑎
𝛥𝑈 1 1 ⋯ 0⎥ ⎪𝛯min = 𝑟min 𝛶𝑁𝑐 ×1 × [0, 0, 0, 1, 0]
𝑥= , 𝐸 = 𝐴𝑑 𝜉(𝑘) + 𝐺𝑑 𝑊 (𝑘), 𝑀 = ⎢ ,⎨ 𝑌
𝜀 ⎢⋮ ⋮ ⋱ 0⎥ ⎪𝛯max = [𝑟𝛥𝑑 𝛥𝑣 𝛥𝑢 𝑎
max , 𝑟max , 𝑟max , 𝑟max ] ⊗ (𝛶𝑁𝑝 ×1 × [1, 1, 1, 1, 0])
⎢1 1 ⋯ 1⎥⎦ ⎪ 𝑌
⎣ 𝛥𝑑 𝛥𝑣 𝛥𝑢 𝑎
⎩𝛯min = [𝑟min , 𝑟min , 𝑟min , 𝑟min ] ⊗ (𝛶𝑁𝑝 ×1 × [1, 1, 1, 1, 0])
𝛶 is a vector whose elements are all 1, ⊗ is the Kronecker product of the two.
Within each sampling period, the control increment sequence and relaxation factor sequence that meet the constraint conditions
can be obtained by solving. Then, the first element 𝛥𝑢(𝑘) of the control increment sequence is added to the control input 𝑢(𝑘 − 1) at
the previous time to obtain the optimal control input 𝑢∗ (𝑘) at the current time.

4.4. Variable weight MPC based on fuzzy control

The weight factors in MPC indirectly reflect the expectations of decision planning algorithm for the performance of vehicle.
Constant weight factors fail to adapt to complex and variable traffic scenarios. To address this issue, a real-time weight adjustment
strategy is proposed by adjusting the weights based on the distance error and speed error. There is no direct relationship between the
state variables and their corresponding weights, while fuzzy control excels in handling fuzzy and uncertain information. This paper
presents a real-time weight adjustment strategy based on fuzzy control inspired by the intuitive and effective fuzzy logic research
in Ref. [27] for the key state (sideslip angle, velocity and attitude) estimation. The inputs of the controller are the distance error
𝛥𝑑 and the speed error 𝛥𝑣, and the outputs are 𝑤𝑑 , 𝑤𝑣 and 𝑤𝑎 , which correspond to the weight factors 𝑄𝑑 , 𝑄𝑣 and 𝑄𝑎 in the MPC

9
D. Wu et al. Mechanical Systems and Signal Processing 209 (2024) 111125

Fig. 3. The membership functions of inputs and outputs.

controller, as shown in Eq. (24). The basic domains for the inputs and outputs, which refers to the ranges of the inputs and outputs,
are specified as follows: 𝛥𝑑 ∈ [−30 m, 30 m], 𝛥𝑣 ∈ [−20 m∕s, 20 m∕s], 𝑤𝑑 ∈ [0, 1.5], 𝑤𝑣 , 𝑤𝑎 ∈ [0, 1]. For the objective functions of
MPC, the relative relationship between the weight factors is more important than the specific value. As a result, one of the four
weight factors 𝑅𝑗𝑒𝑟𝑘 = 2 is fixed and the other three weight factors are regulated.

⎧𝑄𝑑 = 𝜃𝑑 ⋅ 𝑤𝑑
⎪
⎨𝑄𝑣 = 𝜃𝑣 ⋅ 𝑤𝑣 (25)
⎪
⎩𝑄𝑎 = 𝜃𝑎 ⋅ 𝑤𝑎
where 𝜃𝑑 , 𝜃𝑣 , 𝜃𝑎 are the correction coefficients of the three weight factors 𝑄𝑑 , 𝑄𝑣 and 𝑄𝑎 , respectively. The introduction of 𝜃𝑑 , 𝜃𝑣
and 𝜃𝑎 enables the adjustment of the weight factor ranges for dynamic ACC, providing a choice to adjust the relative importance of
various performance aspects throughout the entire driving process. In the simulation section of this paper, 𝜃𝑑 , 𝜃𝑣 , 𝜃𝑎 are set to 5,5,5,
respectively, resulting in weight factor ranges of 𝑄𝑑 ∈ [0, 7.5] and 𝑄𝑣 , 𝑄𝑎 ∈ [0, 5]. Due to the deviation between the vehicle model
and the actual vehicle, the weight factors used for simulation and that for experiment are not exactly the same. 𝜃𝑑 , 𝜃𝑣 , 𝜃𝑎 are set to
10, 10, 20 in the experimental section, resulting in weight factor ranges of 𝑄𝑑 ∈ [0, 15], 𝑄𝑣 ∈ [0, 10] and 𝑄𝑎 ∈ [0, 20].
First, the controller fuzzifies the inputs to transform them into fuzzy linguistic variables. The fuzzy linguistic variables for the
inputs and outputs are defined as follows:
{
𝛥𝑑, 𝛥𝑣 ∶ {𝑆𝑁, 𝑀𝑁, 𝑍𝑂, 𝑀𝑃 , 𝑆𝑃 }
(26)
𝑤𝑑 , 𝑤𝑣 , 𝑤𝑎 ∶ {𝑇 , 𝑆, 𝑀, 𝐿}
where 𝑆𝑁, 𝑀𝑁, 𝑍𝑂, 𝑀𝑃 and 𝑆𝑃 represent significantly negative, modestly negative, zero, significantly positive and modestly
positive, respectively. 𝑇 , 𝑆, 𝑀, 𝐿 represent tiny, small, moderate and large, respectively.
The membership function is employed to quantify the degree of proximity between a variable and its corresponding fuzzy
linguistic variable. A higher value signifies a greater membership to the associated fuzzy linguistic variable. Four widely utilized
types of membership functions are prevalent in practical engineering: triangular, Gaussian, trapezoidal, and sigmoid. When the state
variable deviates significantly from the equilibrium point of system, it is advisable to select a smooth and highly stable membership
function. Conversely, when the state variable approaches the equilibrium point, a steep and highly sensitive membership function
is preferred. Consequently, a smooth sigmoid membership function is chosen for fuzzy linguistic variables 𝑆𝑁, 𝑆𝑃 , 𝑇 and 𝐿. And a
steeper triangular membership function is employed for fuzzy linguistic variables 𝑀𝑁, 𝑍𝑂, 𝑀𝑃 , 𝑆 and 𝑀 as depicted in Fig. 3.
The fuzzy rules establish a connection between the fuzzified inputs and outputs. Their formulation is primarily based on
engineering experience. In this paper, the fuzzy rules are formulated based on the control logic of the ACC system as follows:

(1) When the inter distance is small and the speed of preceding vehicle is smaller than that of the host vehicle (i.e. both 𝛥𝑑 and
𝛥𝑣 are significantly negative), the driving situation becomes dangerous. In such scenarios, the ACC system must promptly

10
D. Wu et al. Mechanical Systems and Signal Processing 209 (2024) 111125

Table 1
The fuzzy rule table.
𝛥𝑑∖𝛥𝑣 𝑆𝑁 𝑀𝑁 𝑍𝑂 𝑀𝑃 𝑆𝑃
𝑆𝑁 {𝐿, 𝐿, 𝑇 } {𝐿, 𝐿, 𝑇 } {𝐿, 𝑀, 𝑆} {𝐿, 𝑆, 𝑀} {𝐿, 𝑆, 𝐿}
𝑀𝑁 {𝑀, 𝐿, 𝑇 } {𝑀, 𝑀, 𝑆} {𝑀, 𝑀, 𝑀} {𝑀, 𝑆, 𝑀} {𝑀, 𝑆, 𝐿}
𝑍𝑂 {𝑀, 𝑀, 𝑆} {𝑆, 𝑀, 𝑆} {𝑆, 𝑀, 𝑀} {𝑆, 𝑆, 𝑀} {𝑆, 𝑇 , 𝐿}
𝑀𝑃 {𝑆, 𝑀, 𝑆} {𝑇 , 𝑀, 𝑆} {𝑇 , 𝑆, 𝑀} {𝑇 , 𝑆, 𝑀} {𝑇 , 𝑆, 𝐿}
𝑆𝑃 {𝑇 , 𝑀, 𝑀} {𝑇 , 𝑀, 𝑀} {𝑇 , 𝑆, 𝐿} {𝑇 , 𝑇 , 𝑀} {𝑇 , 𝑇 , 𝐿}

respond by adjusting the speed of the host vehicle, increasing the distance from the front vehicle, and synchronizing its speed
with the front vehicle. Therefore, 𝑤𝑑 and 𝑤𝑣 should be large, while 𝑤𝑎 should be tiny.
(2) When the inter distance is small and the speed of the front vehicle is larger than that of the host vehicle (i.e. 𝛥𝑑 is significantly
negative and 𝛥𝑣 is significantly positive). The host vehicle experiences deceleration due to the inter distance error, while the
relative speed results in acceleration. Given the close distance and uncontrolled motion of the front vehicle, this paper aims
to enhance safety by increasing 𝑤𝑑 and 𝑤𝑎 . This approach ensures a gradual deceleration of the host vehicle to guarantee
safety.
(3) When the inter distance is large and the speed of the front vehicle is larger than the that of the host vehicle (i.e. both 𝛥𝑑 and
𝛥𝑣 are significantly positive), the ACC system can lead to quick acceleration of the host vehicle and result in a reduction in
ride comfort as it attempts to maintain pace with the front vehicle. Thus, it is necessary to decrease 𝑤𝑑 and relative velocity
error 𝑤𝑣 , while increasing 𝑤𝑎 under such circumstances.
(4) When the inter distance is large and the speed of the front vehicle is smaller than that of the host vehicle (i.e. 𝛥𝑑 is significantly
positive and 𝛥𝑣 is significantly negative), the host vehicle experiences acceleration due to the inter distance error, while
simultaneously experiencing deceleration since the preceding vehicle is moving slower. This study aims to enhance ride
comfort by implementing a gradual deceleration strategy for the host vehicle. Therefore, it is necessary to increase 𝑤𝑣 and
𝑤𝑎 .

Based on the control logic above, the corresponding fuzzy rule table is obtained and presented in Table 1. According to the fuzzy
rules and the membership degrees of the inputs, the fuzzy inference process is performed to obtain the fuzzy output. Subsequently,
the output is defuzzified to obtain a precise value. In this paper, the centroid method is employed for defuzzification. Finally, the
relationship of outputs and inputs can be shown in Fig. 4.
In summary, the weight factors of MPC are regulated by the fuzzy control algorithm. As the driving condition changes, the input
𝛥𝑑 and 𝛥𝑣 are changed. Then the weight factors are adjusted by the fuzzy controller and updated in the MPC problem. After solved
by the interior-point algorithm, the desired acceleration of the host vehicle is output by the motion control layer. The algorithm
framework of variable weight MPC is shown as follows:

Algorithm 1: variable weight MPC

Inputs
The inter distance 𝑑, speeds of preceding vehicle 𝑣𝑓 and host vehicle 𝑣ℎ , longitudinal accelerations of host
vehicle 𝑎ℎ and preceding vehicle 𝑎𝑓 at the current time T.
Outputs
Optimal control input 𝑢∗ at T.
While t = T do
(1) Calculate the safe inter distance 𝑑𝑠𝑎𝑓 𝑒 , 𝛥𝑑 and 𝛥𝑣 based on states at the current time;
(2) Obtain weight factors 𝑄𝑑 , 𝑄𝑣 , 𝑄𝑎 , 𝑅𝑗 based on fuzzy controller;
(3) Establish the system matrices as Eq. (16);
(4) Solve the quadratic programming problem defined in Eq. (24) by the interior-point method to
calculate control increment sequence 𝛥𝑈 and relaxation factor sequence 𝜀;
(5) Add the first element of the control increment sequence to the control input at the previous time
to calculate the optimal control input 𝑢∗ at the current time;
return 𝑢∗

4.5. Vehicle control in lower controller

The combination of the feedforward control and feedback control is applied for the vehicle control in the lower level of ACC. The
control scheme is shown in Fig. 5. To achieve the desired acceleration set by the upper controller, the lower controller of the ACC
system computes the appropriate throttle or braking pressure for the vehicle. Establishing a longitudinal inverse dynamics model
for the vehicle is crucial in this process.
The inverse vehicle model can be created from the vehicle model in Section 2.3 to establish a feedforward control. The inverse
vehicle model is presented in Eqs. (27), (28) for driving and braking conditions, respectively.
𝑚𝑎𝑑𝑒𝑠 + 𝑚𝑔𝑓 cos 𝜃 + 12 𝐶𝐷 𝐴𝜌𝐴 𝑣2 + 𝑚𝑔 sin 𝜃
𝑇𝑑𝑒𝑠 = 𝑟 (27)
𝑖0 𝜂

11
D. Wu et al. Mechanical Systems and Signal Processing 209 (2024) 111125

Fig. 4. The relationship of inputs and outputs.

Fig. 5. Lower controller diagram of ACC.

| 1 |
|−𝑚𝑎𝑑𝑒𝑠 − 2 𝐶𝐷 𝐴𝜌𝐴 𝑣2 − 𝑚𝑔 sin 𝜃 − 𝑚𝑔𝑓 cos 𝜃 |
𝑃𝑑𝑒𝑠 = | | (28)
𝐾𝑏
To minimize the discrepancy between the inverse vehicle dynamics model and the actual vehicle model and enhance the
accuracy and robustness of the system, this paper incorporates a PID feedback controller to regulate the deviation between the
desired acceleration and the actual acceleration. Following numerous repeated experiments, the proportional coefficient 𝐾𝑝 , integral
coefficient 𝐾𝑖 , and differential coefficient 𝐾𝑑 of the PID controller are set to 5, 15 and 0, respectively.
When the lower controller receives the expected acceleration, it makes a decision on switching between driving and braking.
It is essential to determine the acceleration threshold, which represents the acceleration that results from slope resistance, rolling
resistance, and air resistance during driving:
1
𝑎𝑟 = −𝑓 cos 𝜃 − 𝑔 sin 𝜃 − 𝐶 𝐴𝜌 𝑣2 (29)
2𝑚 𝐷 𝐴
where 𝑎𝑟 is the acceleration threshold.
However, when the expected acceleration fluctuates near the acceleration threshold, it may cause frequent switching, resulting in
a negative impact on ride comfort and fuel economy. This paper sets the acceleration boundary value at 𝑎𝑟 + 𝛥𝑎 and the deceleration
boundary value at 𝑎𝑟 − 𝛥𝑎, where 𝛥𝑎 = 0.1 m/s2 . Thus a transition region is formed as shown in Fig. 6.
Then the driving and braking switching logic is:

(1) When 𝑎𝑑𝑒𝑠 > 𝑎𝑟 + 𝛥𝑎, the vehicle is in drive mode;

(2) When 𝑎𝑑𝑒𝑠 < 𝑎𝑟 − 𝛥𝑎, the vehicle is in braking mode;
(3) When 𝑎𝑟 − 𝛥𝑎 < 𝑎𝑑𝑒𝑠 < 𝑎𝑟 + 𝛥𝑎, the vehicle mode will not be changed.

5. Simulation and results

In order to verify the proposed control method, simulation is conducted first before experiment on real vehicle. The combination
simulation platform is constructed based on Carsim and Matlab/Simulink. In Carsim, the host vehicle model is established and
various driving conditions are configured. The ACC control algorithm is modeled in Matlab/Simulink. The main parameters of the
host vehicle model are presented in Table 2. To compare with the constant weight MPC method and variable weight MPC method,
the LQR algorithm is also established to control the inter distance and relative speed for ACC.

12
D. Wu et al. Mechanical Systems and Signal Processing 209 (2024) 111125

Fig. 6. Switching boundary of driving and braking.

Table 2
Parameters of the host vehicle model.
Parameters Description Value
𝑚 Vehicle mass (kg) 250
𝜂 Efficiency of the driving 0.92
system
𝑖0 Transmission ratio of main 4.1
reducer
𝑟 Wheel radius (mm) 225
𝐴 Cross sectional area of vehicle 1.2
(m2 )
𝜌𝐴 Air density (kg/m3 ) 1.206
𝐾𝑏 Ratio of braking force to 1.635
braking pressure
𝑓 Rolling resistance coefficient 0.018
𝐶𝐷 Air resistance coefficient 0.30

5.1. Case 1: front car cut-in

For simulation, a scenario is selected where the front vehicle cuts in. Initially, when there is no vehicle on the road ahead, the
host vehicle accelerates from 15 m/s to 25 m/s, which is the target speed for cruise control mode. Then a vehicle cut in at 7 s and
keeps driving at a speed of 15 m/s. For comparison, the LQR-based ACC and the constant weight ACC and the dynamic ACC are
applied. For the constant weight ACC, the weight factors are set as 𝑄𝑑 = 5, 𝑄𝑣 = 5, 𝑄𝑎 = 5. The results are shown in Fig. 7.
Fig. 7(a) illustrates that the host vehicle maintains a constant acceleration until 7 s when no target vehicle is detected in front,
and the ACC system is in cruise control mode. When the front vehicle is detected at 7 s, the ACC system changes to following
mode. Three strategies can decrease the speed to match the speed of the front car at 15 m/s. However, the maximum deceleration
controlled by dynamic ACC is −5.5 m/s2 , which is higher compared to −4 m/s2 for the constant weight ACC and −4.4 m/s2 for
the LQR-based ACC as depicted in Fig. 7(b). Fig. 7(c) indicates that the minimum distance controlled by dynamic ACC is 10.4 m,
while those controlled by LQR-based ACC and constant weight ACC are 8.2 m and 8.4 m, respectively. Additionally, during periods
when the actual distance is less than the safe distance, the distance error of dynamic ACC is also smaller than in the other two
strategies according to Fig. 7(d). LQR and constant weight ACC demonstrate similar levels of ride comfort and safety performance.
Notably, in this situation, safety is more important than ride comfort. In terms of safety, dynamic ACC outperforms the other two
methods. Fig. 7(e) gives the weight factors of the dynamic ACC. 𝑄𝑑 increases sharply to guarantee the following safety and 𝑄𝑣
increases slightly to improve speed tracking accuracy, 𝑄𝑎 decreases to enable large changes in acceleration. Consequently, from 7
s–9 s, ride comfort may be sacrificed for enhancing safety. After 10 s, 𝑄𝑑 and 𝑄𝑣 decrease due to a reduction of tracking errors in
inter distance and speed, 𝑄𝑎 increases to enhance ride comfort.
In summary, when driving conditions shift and safety decreases (e.g. when a cut-in occurs), dynamic ACC can increase the weight
𝑄𝑑 for safer inter-distance tracking. Compared to the other two strategies, dynamic ACC regulates vehicle acceleration and speed
more efficiently. As a result, the host vehicle can be controlled to optimize driving safety with larger acceleration and inter distance.

5.2. Case 2: approaching the front car

The second simulation scenario involves the host vehicle approaching the front vehicle. The initial speed of the host vehicle is set
at 30 m/s, with the preceding vehicle maintaining a constant speed of 25 m/s ahead, and the distance between them is 100 m. Since
the experiment in this paper is conducted on a low-speed platform, the vehicle speeds in this simulation scenario are deliberately
set at a significantly higher level to validate the performance of the proposed dynamic ACC on highways. When the ACC system
detects the front vehicle, it switches to following mode. The constant weights for this scenario are 𝑄𝑑 = 5, 𝑄𝑣 = 5, and 𝑄𝑎 = 5,
which is the same as Section 5.1. The results of this simulation case are shown in Fig. 8.

13
D. Wu et al. Mechanical Systems and Signal Processing 209 (2024) 111125

Fig. 7. Simulation results for the case of front car cut-in.

Initially, the host vehicle accelerates to reach the safe distance. As the distance decreases and the speed increases, host vehicle will
decelerate. As shown in Fig. 8(a), in terms of speed tracking performance, the dynamic ACC with a maximum speed of 30.9 m/s
significantly outperforms the other two strategies with maximum speeds of 35.1 m/s and 38.0 m/s, respectively. According to
Fig. 8(b), the maximum acceleration and deceleration values for dynamic ACC are 0.54 m/s2 and −0.79 m/s2 , respectively. In
contrast, the values for LQR-based ACC are 4.85 m/s2 and −1.39 m/s2 , respectively. Lastly, the constant weight ACC exhibits
acceleration and deceleration values of 4.46 m/s2 and −2.56 m/s2 , respectively. From Fig. 8(c) and Fig. 8(d), it can be seen that all
three strategies can ensure good safety. Notably, the ride comfort is important as the driving condition is rather safe. Employing the
constant weight ACC and LQR-based ACC leads to significant fluctuation in acceleration, which negatively impacts ride comfort.
However, dynamic ACC can achieve smoother control of acceleration by adjusting weight factors. As a result, the dynamic ACC
enhances ride comfort while ensuring safety. The weight factors of the dynamic ACC are introduced in Fig. 8(e). The weight factor
𝑄𝑑 stays at 0.22 before 7.5 s due to the large inter-distance. 𝑄𝑎 and 𝑄𝑣 are greater than 𝑄𝑑 to improve ride comfort. As the vehicle
gradually approaches a safe distance, 𝑄𝑑 increases to prioritize driving safety. Hence, the dynamic ACC can balance following safety
and ride comfort adaptively based on the driving condition.
In comparison to LQR-based ACC and constant weight ACC, dynamic ACC can more effectively and smoothly regulate vehicle
acceleration and speed. This allows the host vehicle to be controlled with smaller acceleration, leading to improved ride comfort
without compromising driving safety.

6. Experiments and results

In this paper, the proposed ACC algorithm is arranged on a low-speed experimental platform, as shown in Fig. 9. The vehicle for
experiment is an electric vehicle, which can be controlled by-wire in driving, braking and steering. The perception system consists

14
D. Wu et al. Mechanical Systems and Signal Processing 209 (2024) 111125

Fig. 8. Simulation results for the case of approaching the front car.

of NAV992 GNSS/INS integrated navigation system and a millimeter-wave radar ESR produced by Delphi and a camera ME560
produced by Mobileye. An industry computer (Model: Nuvo-5095GC (GTX1050, 512 g SSD)) is utilized as the control unit. The ACC
control algorithm is run on the ROS platform, which is integrated with MATLAB/Simulink. The sensors fusion algorithm is built by
C++ in ROS. The upper level and lower level control strategies of ACC are established in MATLAB/Simulink. The Simulink model
is combined with the ROS algorithm for calculation.

6.1. Sensors fusion validation

Validation experiment for sensors fusion is about a person crossing the road. Fig. 10 presents the lateral distance and longitudinal
distance of the object detected by the millimeter-wave radar ESR without preprocessing. It can be seen that the ESR performs
poorly at detecting the lateral movement of objects. Due to the lack of filtering and estimation processing, the lateral distance and
longitudinal distance of the ESRTarget1 are unstable, resulting in multiple instances of missed target detection.
Fig. 11 shows the results of the preprocessed ESR target, camera target, and fusion target. After preprocessing, the stability of
ESRTarget1 is improved and the jump of the distance signal can be eliminated by estimation. FusionTarget1 obtained by ESRTarget1
and CameraTarget1 has no frame loss and mutation in the longitudinal and lateral distance.

6.2. Test for different weight factors

The first case of the ACC experiment involves following the front vehicle. The objective is to validate the impact of weight
factors on control results, for which two sets of constant weight factors are compared: [𝑄𝑑 , 𝑄𝑣 , 𝑄𝑎 ] = [10, 10, 10] and [𝑄𝑑 , 𝑄𝑣 , 𝑄𝑎 ] =

15
D. Wu et al. Mechanical Systems and Signal Processing 209 (2024) 111125

Fig. 9. The experimental platform.

Fig. 10. ESR targets without preprocessing in the first experiment that a person crossing the road.

Fig. 11. Targets preprocessed in the first experiment that a person crossing the road..

[10, 10, 25]. For this experiment, a vehicle runs ahead as the proceeding vehicle. To provide consistent working conditions when
testing different control systems, this paper utilize the perceptual data from the first experiment (i.e., the speed and displacement
of the leading vehicle, and the initial inter distance) in subsequent experiments to emulate identical driving conditions. The initial
inter distance is 4.3 m. The experiment results are shown in Fig. 12.
Figs. 12(a)–12(c) show the speed, the inter distance and the distance error controlled by ACC with different weight factors,
respectively. It is observed that the speed and distance tracking errors of ACC with 𝑄𝑎 set to 25 is larger. This implies that increasing
the weight factor for acceleration 𝑄𝑎 lowers the following performance. Conversely, it can be stated that increasing 𝑄𝑑 and 𝑄𝑣 is
beneficial to improve the following performance.

16
D. Wu et al. Mechanical Systems and Signal Processing 209 (2024) 111125

Fig. 12. Comparative experimental results of ACC with different weight factors.

Fig. 12(d) gives the acceleration controlled by ACC with different weight factors. It can be seen that the vehicle longitudinal
acceleration on real road is rather sensitive with significant jitter. This is affected by the vehicle motion attitude, road unevenness
and the limitations of sensor measure accuracy. However, it is important to note that despite this jitter, the overall change trend
and equivalent results can represent the results of different control methods. It is observed that ACC with a higher value of 𝑄𝑎 leads
to reduced acceleration. This finding suggests that increasing 𝑄𝑎 effectively decreases acceleration, thereby enhancing ride comfort.
Based on the results above, it can be concluded that tracking speed and safe distance is more effective when 𝑄𝑑 and 𝑄𝑣 are
relatively large. However, speed and distance fluctuate greatly with high acceleration, which results in poor ride comfort. On the
other hand, increasing 𝑄𝑎 improves ride comfort with less acceleration, but reduces following accuracy. These results are consistent
with the control principle discussed in Section 4. Because it is challenging to balance different performance criteria using a constant
weight ACC, adaptive regulation of variable weights is necessary.

6.3. Test for dynamic ACC

The second case of ACC experiment is also the following of front car. In this case, the dynamic ACC is compared with the
LQR-based ACC and the constant weight ACC with 𝑄𝑑 , 𝑄𝑣 , 𝑄𝑎 set to 10, 10, 10, respectively. Similarly, To ensure consistent driving
conditions, the perception data from the first set of experiments is used in the subsequent experiments. Initially, the target stops
14 m ahead of the host vehicle and the host vehicle accelerates from 0 m/s. The experiment results are shown in Fig. 13. This paper
utilizes MATLAB to calculate essential data for each strategy during its driving period, as presented in Table 3. Notably, the jerk
indices are obtained by calculating the maximum and average absolute values of jerk and the maximum distance error is calculated
during the time interval in which the distance error is less than 0.
From Fig. 13(a), the driving time of the LQR-based ACC is approximately 10 s, while that of other strategies is around 8 s.
Additionally, the LQR-based ACC demonstrates lower average speed and deceleration compared to other strategies. These findings
indicate that the LQR-based ACC focuses on longer driving time to achieve reduced deceleration and enhanced ride comfort during
deceleration. However, it also reveals a slower response of the LQR-based ACC and a longer time for the host vehicle to match the
speed of the front vehicle. From 5 to 8 s, the inter distance becomes smaller than the safe distance. To ensure safety, the dynamic ACC
increases the weight 𝑄𝑑 and decreases 𝑄𝑎 , which leads to slightly compromised ride comfort in comparison to the LQR-based ACC.
During the acceleration period, it is evident from Table 3 and Fig. 13(b) that dynamic ACC outperforms the other two strategies in
terms of multiple ride comfort indices, including average acceleration, maximum jerk and average jerk. These findings demonstrate
that dynamic ACC can effectively prioritize ride comfort in driving scenarios where safety is ensured and ride comfort is crucial.
According to Table 3 and Fig. 13(c)-(d), the dynamic ACC strategy outperforms the other two strategies in terms of safety indices,
including minimum inter distance and maximum distance error. While the constant ACC performs the worst. When the inter distance
becomes small and distance error is negative, the implementation of dynamic ACC can effectively enhance safety.

17
D. Wu et al. Mechanical Systems and Signal Processing 209 (2024) 111125

Table 3
Comparison of essential data for three strategies.
LQR-based ACC Constant weight ACC Dynamic ACC
Driving time (s) 9.62 7.64 7.92
Average acceleration (m/s2 ) 0.591 0.793 0.312
Average deceleration (m/s2 ) −0.304 −0.553 −0.334
Maximum jerk (m/s3 ) 19.74 10.12 6.78
Average jerk (m/s3 ) 0.96 1.10 0.62
Average speed (m/s) 0.976 1.635 1.073
Minimum inter distance (m) 2.880 1.963 3.229
Maximum distance error (m) −1.186 −2.213 −0.890

Fig. 13. Comparative experimental results of ACC with different strategies.

Fig. 13(e) gives the weight factors of dynamic ACC. It is evident that the variable weighting strategy provides a large 𝑄𝑎 to
improve driving comfort when the inter distance is large. As the inter distance decreases, 𝑄𝑑 increases and 𝑄𝑎 decreases to improve
the following performance.
Based on the detected target information, the proposed dynamic ACC strategy can adjust the weight factors according to the
tracking error of the relative distance and speed. In scenarios where safety is ensured and ride comfort is crucial, dynamic ACC
increases 𝑄𝑎 and decreases 𝑄𝑑 to reduce the acceleration and jerk, resulting in better ride comfort. When the inter distance becomes
small and distance error is negative, dynamic ACC decreases 𝑄𝑎 and increases 𝑄𝑑 to ensure safety. Hence, the dynamic ACC provides
a performance balance between driving safety and ride comfort. Additionally, the dynamic ACC has the capability to automatically
adapt to changes in driving conditions, thus improving the apprehensive performance of vehicle under variable driving conditions.

18
D. Wu et al. Mechanical Systems and Signal Processing 209 (2024) 111125

7. Conclusions

This paper researches the multi-objective dynamic coordination and sensors fusion of ACC. The weight coefficients of MPC are
regulated according to the inter distance and relative speed between the front vehicle and ego vehicle through a fuzzy controller.
The decision-level sensors fusion scheme based on millimeter wave radar and camera is developed. The hierarchical ACC controller
is designed and tested both by simulation and road experiment. The results indicate that the proposed ACC system coordinates the
driving safety and ride comfort according to the driving conditions timely. Besides, the accuracy of object detection is improved by
multi-sensors fusion.
This paper provides a practical solution for ACC improvement to adapt to the complex dynamic traffic environment. However,
only two main control objectives, driving safety and ride comfort, are considered in this paper. In future work, the energy efficient
and driving time can be further considered in the multi-objective dynamic coordination. Furthermore, more tests on the real road
will be applied to verify the performance of the proposed algorithm.

CRediT authorship contribution statement

Dongmei Wu: Conceptualization, Methodology, Project administration, Supervision, Writing – review & editing. Baobao Qiao:
Formal analysis, Investigation, Software, Writing – original draft. Changqing Du: Conceptualization, Investigation, Methodology.
Yifan Zhu: Conceptualization, Investigation, Methodology. Fuwu Yan: Conceptualization, Investigation, Methodology. Chang-
sheng Liu: Conceptualization, Investigation, Methodology. Yang Li: Conceptualization, Investigation. Jun Li: Conceptualization,
Investigation.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared
to influence the work reported in this paper.

Data availability

Data will be made available on request.

Acknowledgments

This work was financially supported by the National Natural Science Foundation of China (51975434, 51705383).

References

[1] M. Althoff, S. Maierhofer, C. Pek, Provably-correct and comfortable adaptive cruise control, IEEE Trans. Intell. Veh. 6 (1) (2021) 159–174.
[2] B. Gao, K. Cai, T. Qu, Y. Hu, H. Chen, Personalized adaptive cruise control based on online driving style recognition technology and model predictive
control, IEEE Trans. Veh. Technol. 69 (11) (2020) 12482–12496.
[3] F. Lei, Y. Bai, W. Zhu, J. Liu, A novel approach for electric powertrain optimization considering vehicle power performance, energy consumption and ride
comfort, Energy 167 (2019) 1040–1050.
[4] X. Lin, D. Görges, Robust model predictive control of linear systems with predictable disturbance with application to multiobjective adaptive cruise control,
IEEE Trans. Control Syst. Technol. 28 (4) (2020) 1460–1475.
[5] J. Chen, Y. Ye, Q. Wu, R. Langari, C. Tang, Low-cost and high-performance adaptive cruise control based on inertial-triggered mechanism and multi-objective
optimization, IEEE Trans. Veh. Technol. 72 (6) (2023) 7279–7289.
[6] Y. Zhang, M. Xu, Y. Qin, M. Dong, L. Gao, E. Hashemi, MILE: Multi-objective integrated model predictive adaptive cruise control for intelligent vehicle,
IEEE Trans. Ind. Inform. (2022) 1–9.
[7] W. Cao, S. Liu, J. Li, Z. Zhang, H. He, Analysis and design of adaptive cruise control for smart electric vehicle with domain-based poly-service loop delay,
IEEE Trans. Ind. Electron. 70 (1) (2023) 866–877.
[8] S. Boddupalli, A.S. Rao, S. Ray, Resilient cooperative adaptive cruise control for autonomous vehicles using machine learning, IEEE Trans. Intell. Transp.
Syst. 23 (9) (2022) 15655–15672.
[9] H. Hailemichael, B. Ayalew, L. Kerbel, A. Ivanco, K. Loiselle, Safety filtering for reinforcement learning-based adaptive cruise control, IFAC-PapersOnLine
55 (24) (2022) 149–154.
[10] Y. Jia, R. Jibrin, D. Görges, Energy-optimal adaptive cruise control for electric vehicles based on linear and nonlinear model predictive control, IEEE
Trans. Veh. Technol. 69 (12) (2020) 14173–14187.
[11] S. Anand, S. Ohol, Modelling and simulation of adaptive cruise control and overtake assist system, Mater. Today: Proc. 72 (2023) 1353–1360.
[12] B. Zhu, Y. Sun, J. Zhao, S. Zhang, P. Zhang, D. Song, Millimeter-wave radar in-the-loop testing for intelligent vehicles, IEEE Trans. Intell. Transp. Syst.
23 (8) (2021) 11126–11136.
[13] Z. Meng, X. Xia, R. Xu, W. Liu, J. Ma, HYDRO-3D: Hybrid object detection and tracking for cooperative perception using 3D lidar, IEEE Trans. Intell.
Veh. 8 (8) (2023) 4069–4080.
[14] X. Xia, R. Xu, J. Ma, Secure cooperative localization for connected automated vehicles based on consensus, IEEE Sens. J. 23 (20) (2023) 25061–25074.
[15] R.K. Satzoda, S. Lee, F. Lu, M.M. Trivedi, Vision-based front and rear surround understanding using embedded processors, IEEE Trans. Intell. Veh. 1 (4)
(2016) 335–345.
[16] A. Rangesh, M.M. Trivedi, No blind spots: Full-surround multi-object tracking for autonomous vehicles using cameras and lidars, IEEE Trans. Intell. Veh.
4 (4) (2019) 588–599.
[17] X.W.L. Chehri, Multi-sensor fusion technology for 3D object detection in autonomous driving: A review, IEEE Trans. Intell. Transp. Syst. (2023) 1–18.

19
D. Wu et al. Mechanical Systems and Signal Processing 209 (2024) 111125

[18] R. Gravina, P. Alinia, H. Ghasemzadeh, G. Fortino, Multi-sensor fusion in body sensor networks: State-of-the-art and research challenges, Inf. Fusion 35
(2017) 68–80.
[19] S. Li, X. Kang, L. Fang, J. Hu, H. Yin, Pixel-level image fusion: A survey of the state of the art, Inf. Fusion 33 (2017) 100–112.
[20] J. Jo, S.J. Lee, K.R. Park, I.-J. Kim, J. Kim, Detecting driver drowsiness using feature-level fusion and user-specific classification, Expert Syst. Appl. 41 (4)
(2014) 1139–1152.
[21] M. Zhu, H. Chen, G. Xiong, A model predictive speed tracking control approach for autonomous ground vehicles, Mech. Syst. Signal Process. 87 (2017)
138–152.
[22] S. Li, K. Li, R. Rajamani, J. Wang, Model predictive multi-objective vehicular adaptive cruise control, IEEE Trans. Control Syst. Technol. 19 (3) (2011)
556–566.
[23] S.E. Li, Q. Guo, S. Xu, J. Duan, S. Li, C. Li, K. Su, Performance enhanced predictive control for adaptive cruise control system considering road elevation
information, IEEE Trans. Intell. Veh. 2 (3) (2017) 150–160.
[24] W. Lim, S. Lee, J. Yang, M. Sunwoo, Y. Na, K. Jo, Automatic weight determination in model predictive control for personalized car-following control, in:
IEEE Access, IEEE Access 10 (2022) 19812–19824.
[25] C. Pan, A. Huang, J. Wang, L. Chen, J. Liang, W. Zhou, L. Wang, J. Yang, Energy-optimal adaptive cruise control strategy for electric vehicles based on
model predictive control, Energy 241 (2022) 122793.
[26] X. Xia, Z. Meng, X. Han, H. Li, T. Tsukiji, R. Xu, Z. Zheng, J. Ma, An automated driving systems data acquisition and analytics platform, Transp. Res. C
151 (2023) 104120.
[27] X. Xia, P. Hang, N. Xu, Y. Huang, L. Xiong, Z. Yu, Advancing estimation accuracy of sideslip angle by fusing vehicle kinematics and dynamics information
with fuzzy logic, IEEE Trans. Veh. Technol. 70 (7) (2021) 6577–6590.

20