See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/3902454

Data fusion of four ABS sensors and GPS for an enhanced localization of car-like
vehicles

Conference Paper  in  Proceedings - IEEE International Conference on Robotics and Automation · February 2001
DOI: 10.1109/ROBOT.2001.932839 · Source: IEEE Xplore

CITATIONS READS
91 276

4 authors, including:

Philippe Bonnifait Paul Crubille

Université de Technologie de Compiègne Université de Technologie de Compiègne
131 PUBLICATIONS   1,828 CITATIONS    11 PUBLICATIONS   208 CITATIONS   

SEE PROFILE SEE PROFILE

Dominique Meizel
University of Limoges
103 PUBLICATIONS   978 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Thesis/HEUDIASYC View project

All content following this page was uploaded by Philippe Bonnifait on 16 May 2014.

The user has requested enhancement of the downloaded file.

 Data Fusion of Four ABS Sensors and GPS for an Enhanced
Localization of Car-like Vehicles
Philippe Bonnifait, Pascal Bouron, Paul Crubillé, Dominique Meizel

Heudiasyc UMR CNRS 6599

Université de Technologie de Compiègne. France.
[email protected]

Abstract with an encoder measuring the driving wheel angle. The

A localization system using GPS, ABS sensors and a fusion of all these measurements uses an Extended
driving wheel encoder is described and tested through Kalman Filter. A special attention is dedicated to the
real experiments. A new odometric technique using the sampling process that is customized to the GPS. This is
four ABS sensors is presented. Due to the redundancy of done by taking both the displacement and the time as
the measurements, the precision is better than the one of progression variables. Experiments show the efficiency of
differential odometry using the rear wheels only. The this technique in comparison with differential odometry
sampling is performed when necessary and when a GPS using only the sensors of the rear wheels. The complete
measurement is performed. This implies a noticeable localizer in then presented in section 3. Experiments
reduction of the GPS latency, simplifying thus the data- performed with a laboratory car quantify the precision
fusion process and improving the quality of its results. obtained when the GPS is available and when satellite
masks occur.
1 INTRODUCTION
Real-time continuous localization is essential for many 2 ODOMETRY WITH ABS SENSORS
applications concerning outdoors vehicles [6-1-9]. It The purpose of odometry is to build an incremental
consists of estimating the vehicle location in a global model of the motion using measurements of the
digital map. GPS is a very attractive solution because it is elementary wheel rotations. In this section, we first
affordable and convenient. Moreover, since May 2000, introduce the generic integration process whose inputs are
the SA degradation has been turned off, and the natural theoretical quantities (not directly measured). In a second
precision (a few meters) is sufficient for usual navigation time, these inputs are estimated using the measurements
tasks. Nevertheless, GPS still suffers from satellite masks of all the encoders. This is one major contribution of the
occurring in urban environments, under bridges, tunnels paper, the second one being the sampling strategy.
or in forests. GPS appears then as an intermittent 2.1 Integration process
positioning system that demands the help of a dead- Consider a car-like vehicle (figure 2). The mobile frame
reckoning system. Commercially available solutions use M is chosen with its origin M attached to the center of
an odometer, a gyro and sometimes a magnetic compass the rear axle. The x-axis is aligned with the longitudinal
[1] to maintain an estimation of the position during axis of the car. At time tk, the vehicle position is
satellites masks. On the other hand, braking in modern represented by the (xk,yk) Cartesian coordinates of M in a
cars is assisted with ABS systems that utilize angular
world frame W. The heading angle is denoted θk.
encoders attached to the wheels. In this case, the sensors
Let Mk and Mk+1 be two successive positions. Supposing
basically measure the wheel speeds. We propose
the road is perfectly planar and horizontal, as the motion
hereafter to use them to estimate traveled distances.
is locally circular, we have (figure 1):
This paper presents a new odometric technique that can
advantageously replace the commercially available dead- ∆ = ρ.ω (1)
reckoning techniques that use supplementary sensors like where ∆ is the length of the circular arc followed by M, ω
gyros [2]. The models presented in the sequel are real the elementary rotation of the mobile frame, and ρ the
odometric models and not discretized kinematics models radius of curvature and I the instantaneous center of
like in [7]. Assumptions are made on the elementary rotation.
motions and geometric relationships are expressed to Supposing the car is moving forward, the variation on the
provide relations between the rotations of the wheels and position is expressed as:
the displacements. ∆x = Μ k Μ k +1 . cos(θ k + ω/2)
The paper is organized as follows: section 2 describes an 
odometric technique using four ABS sensors together ∆y = Μ k Μ k +1 . cos(θ k + ω/2)
From basic Euclidean geometry, we know that Observation equation
{∆ ≈ |MkMk+1|} up to the second order. The integration The driving wheel encoder (which measures ψ) provides
process is then: a relation between ∆ and ω. Figure 2 gives:
 x k +1 = xk + ∆. cos(θ k + ω/2) L
 tan(ψ) = (3)
= y k + ∆. sin(θ k + ω/2) ρ
 y k +1 (2)
θ = θk +ω
Rewriting equations (1) and (3):
 k +1
ω
tan(ψ) = L. (4)
y0
yk
∆
instantaneous xk+1 Let consider the situation described in figure (3). Since
center of rotation the rear wheels do not turn, we directly have:
∆RL = ω.(ρ-e) = ∆-e.ω
I
ω (5)
∆RR = ω.(ρ+e) = ∆+e.ω (6)
ω/2
xk
Fk+1
ρ
ω
M k+1

∆y H θk ∆F

Mk
∆ virtual front wheel
∆x Fk
O x0
D
Figure 1: elementary displacement between 2 samples Mk+1
D
The question is now to estimate the ∆ and ω.
2.2 Estimating ∆ and ω ω
∆ RR
Consider the four wheels vehicle sketched on figure 2. ∆
ψ
x
∆RL
ψL ψ ψR ω
ψ
I
virtual front wheel Mk
ρ
D Figure 3: elementary displacement of a car in a turning
DL manoeuvre
Commonly, in the technique called differential odometry
DR L
[9], ∆ and ω are computed just using the measurements
(∆RR, ∆Rl):
ψ ∆RR+∆RL ∆RR-∆RL
M ∆= ω=
I 2 2.e
The measurements of the front wheels are more difficult
ρ e to use because their orientation with respect to the mobile
Figure 2: geometry of car in a turning manoeuvre frame is not constant. In order to simplify, let consider
Let us denote e the half-track, L the wheel-base and ψ the the virtual front wheel. The length of the circular arc it
steering angle of the virtual front wheel of the bicycle follows between two samples is (Fig 3):
model. This angle is directly proportional to the angle of ∆F = D.ω
the steering wheel. Now,
In order to develop an odometric model, we assume that, L
between two samples, the wheels do not slip and that the D=
sin(ψ)
distances e and L are known and constant: this is a
By using (3), we have:
simplified view of the phenomenon occurring at the
tire/ground contact zone. ρ
D=
∆ and ω can be estimated using the steering angle ψ and cos(ψ)
the elementary distances traveled by each wheel. In the By multiplying each side by ω:
sequel, ∆RL and ∆RR denote the distances traveled between ∆F .cos(ψ) = ∆ (7)
two samples by the rear wheels and ∆FL and ∆FR those The same procedure can be applied to each front wheel; it
traveled by the front wheels. yields (ψL and ψR are shown on figure 2):
 L L Sk+1/k = Sk/k + Qγ
tan(ψL) = and tan(ψR) = . (8)
ρ-e ρ+e
Estimation stages
ρ can be eliminated in (8) by using relation (3). One gets:
1) As the first observation equation is not defined when ∆
ψL = atan
tan(ψ).L 
and ψR = atan
tan(ψ).L  equal zero (see eq. 13), we only use it when the speed of
(9)
L-e.tan(ψ) L+e.tan(ψ) the car is high enough.
Finally, the adaptations of equation (7) for each wheel Moreover, this equation being non-linear, we compute the
are: Jacobian matrix with respect to the state:
∆FL.cos(ψL) = ∆-e.ω δz1
C1 =   =  2
(10) -L.ω L 
∆FR.cos(ψR) = ∆+e.ω (11)  δζ   ∆ ∆ 
Equations (4, 5, 6, 10, 11) define a redundant and non- The updated state and variance matrix are then given by:
linear system that links the unknown quantities (∆, ω) to Kk= Sk+1/k.C1t.(C1.Sk+1/k.C1t+R2)-1
the measured variables (∆FL, ∆FR, ∆RL, ∆RR, ψ):
ζk+1 = ζk + Kk.(z2 – C1.ζk)
 ω
 tan (ψ ) = L. Sk+1/k+1 = (I22 - Kk.C1).Sk+1/k
∆
 ∆RL = ∆ − e.ω
 (12) 2) The four other observations provide estimations as
 ∆RR = ∆ + e.ω
 follows:
 ∆FL .cos(ψ L ) = ∆ − e.ω for i=2 to 5,
∆ .cos(ψ ) = ∆ + e.ω
 FR R Ci = [ 1 +/-e ]
Defining an equivalent measurement vector (13), Kk= Sk+1/k+1.Cit.(Ci.Sk+1/k.Cit+Ri)-1
equation (12) takes the usual form of observation
ζk+1 = ζk+1 + Kk.(zi – Ci.ζk+1)
equation (14).
Sk+1/k+1 = (I22 - Kk.Ci).Sk+1/k+1
z = [tan(ψ), ∆RL, ∆RR, ∆FL.cos(ψL), ∆FR.cos(ψR)]T (13) end for.
z = h(ζ) (14) Finally, each estimation of (∆, ω) is used in the model (2)
to provide an odometric location (starting from a given
State estimation of (∆∆, ω) position and heading). This technique is called odometric
Because of presence of unavoidable slippage and EKF in the sequel.
modeling errors, we propose to use all the measurements
2.3 Sampling the odometric EKF
to estimate the state vector ζ=[∆, ω]Τ. The redundancy
Commonly, odometric models are sampled with respect
should reduce the effects of the unknown disturbances.
to time. One should notice that they can be sampled with
Since the observation equation (14) is not linear, we
respect to the traveled distance. As a matter of fact, the
propose to apply an Extended Kalman Filter (EKF),
sampling rate must be such that the elementary motion is
interpretable here as a Weighted Least Squares estimator.
circular. Actually, as long as the steering angle is
This technique presents also the advantage to estimate the
constant, the motion is circular. Therefore, the model
covariance of the estimation error, which will be useful
should be sampled when a significant variation of the
for further estimating the localization [x, y, θ]Τ by EKF.
steering angle is detected during a displacement.
Since the models we use are odometric ones and that no
This sampling strategy presents also the interest to be
acceleration nor forces are supposed to be known, the
well adapted to the physical nature of the process: if the
variation of the state vector is modeled by a Wiener
car is motionless, the model is not sampled and the
process whose input γk is a zero-mean white noise with a estimation error logically does not increase.
Qγ covariance matrix: On the other hand, the position (x, y) in the model (2) is
ζk+1 = ζk + γk not observable [3] from the encoder measurements (13).
Since the equivalent measurements that are the One can use a GPS receiver to correct the unavoidable
components of z (eq. 13) are obtained from five different drift of the odometric EKF. This is nowadays an
sensors, we can assume that the covariance matrix R of z affordable sensor, but its latency can induce unacceptable
is diagonal. The EKF estimation can therefore be split errors when the speed is high. The GPS latency is the
into five independent computations, each one involving a delay between the moment all the visible satellites
scalar division in place of inverting a 5x5 matrix. measurements are performed and the moment the position
Denote S the variance of the estimation error. The EKF is output and received by the onboard computer. The
works in six steps: transmission delay is unavoidable but the latency can be
reduced to its minimum by using the analogous signal
Prediction stage "1PPS" available on the Motorola VP ONCORE used in
our experiments. This signal rises each time the satellites with differential odometry using rear wheels only (dotted
are sampled. Thus, we propose to trig the odometric EKF lines) and with the true trajectory (thin line).
using the "1PPS" signal.
point of departure
In summary, the sampling is performed as following:
- each time a GPS measurement is performed,
lap 1 lap 2
- each time the car has traveled one meter and the
(m ) (m )
steering angle has changed more than 0.5 degree.
The two sampling processes are asynchronous. The -1 5 0 0 -1 5 0 0

sampling frequency is not constant and higher than 1Hz.

It can reach several Hz while entering in a roundabout for
example.
-2 0 0 0 -2 0 0 0
-1 5 0 0 -1 0 0 0 -1 5 0 0 -1 0 0 0
(m ) (m )
2.4 Tuning of the filter
For the car used in the experiments (a Citroën Xantia
Break see photo on figure 4), the sources of errors that (m ) lap 3 lap 4
affect the observation (14) are numerous. The first ones (m )
result from the resolutions of the ABS encoders (1.93 cm -1 5 0 0 -1 5 0 0
for ∆RL, ∆RR, ∆FL, ∆FR) and of the driving wheel encoder
(2.π/4096 rad for ψ). The other ones are due to the
approximate nature of the model (12) (for instance,
distances e and L are not exactly constant). -2 0 0 0 -2 0 0 0
-1 5 0 0 (m ) -1 0 0 0 -1 5 0 0 (m ) -1 0 0 0
Since there are several uncorrelated causes of error, we
can suppose that inaccuracy can be globally represented Figure 5: tops views of the experiments
by an additive zero-mean temporally uncorrelated noise. Figures (6) and (7) show longitudinal and lateral errors
(with respect to the DGPS) as functions of the linear
abscissa, making the benefit of the odometric EKF
clearer.
It can be seen from these plots that the use of all the
sensors increase significantly the precision. This can be
explained by the fact that the redundancy in average
allows a better estimation. For example, slips which
disturb the front wheels may not affect the rear ones.
Moreover, other experiments, not reported here, have
Figure 4: the car used in the experiments shown that the drift of the estimation of the heading of the
In order to tune the EKF, we have computed position odometric EKF is in the same order as the one of a dead-
errors in comparison with a post-processed DGPS. The reckoning technique using an odometer and a vibrating
tuning of the filter is assumed to be correct when the gyro (a British Aerospace VSG2000).
errors are consistent with the 3 σ bounds (99% of
400 200
confidence bounds). lap 1 lap 2

200
2.5 Results of the odometric EKF (m )
0
(m )
The experimental results presented in this section were 0

obtained using stored data of the sensors. The sampling -2 0 0

-2 0 0
of the odometric EKF has been done has presented in
section 2.2.3. The rough measurements of the GPS -4 0 0 -4 0 0
0 1000 2000 3000 0 1000 2000 3000
receiver VP ONCORE were stored and then processed in s (m ) s (m )
association with a fixed receiver in order to produce the 200 200
lap 3 lap 4
true positions of the car (Kinematic Differential GPS on
L1 with the software Jupiter). The precision is estimated 0 0
(m ) (m )
to be better than two meters while the car is moving.
The initial heading has been computed with an error less -2 0 0 -2 0 0

than 0.5 degrees.

The odometric EKF has been tested on four laps of -4 0 0 -4 0 0
0 1000 2000 3000 0 1000 2000 3000
2.4 km long, at the maximum allowed speed of 50 km/h. s (m ) s (m )
Results are shown on figure (5) (thick line) and compared Figure 6: longitudinal errors as functions of the distance
 400 400 does not change. This is interesting when compared with
lap 1 lap 2

200 200
a dead reckoning system using a gyro that will suffer
(m ) (m ) from drifting in the same condition. When the vehicle
0 0
will move again, the localizer will start with a non
-2 0 0 -2 0 0 detoriorated estimation of the heading.
-4 0 0 -4 0 0 3.3 Tuning of the filter
0 1000 2000 3000 0 1000 2000 3000
s (m ) s (m )
The errors on position measured by GPS (with the SA
400 400 degradation off) are temporally correlated. We have
lap 3 lap 4
estimated a first order Auto-Regressive model [5] of the
200 200
noise. The time-constant of this AR model is
(m ) (m )
0 0 approximately 2 hours [4]. This value exceeds the time-
constant desired for the filter from several degrees of
-2 0 0 -2 0 0
magnitude. It is thus not good to define an augmented
-4 0 0 -4 0 0 state that incorporate this AR model.
0 1000 2000 3000 0 1000 2000 3000
s (m ) s (m ) One practical solution to take into account the correlation
Figure 7: lateral errors as functions of the distance of the noise consists to increase the variances of the
observations. If not, the filter would yield erroneous too
3 SENSOR FUSION OF GPS AND ODOMETRY confident estimations [4]. In the experiments presented in
the sequel, we have supposed that the variances Rx and
When a GPS position is available, a correction of the Ry of the observations were constant and, experimentally,
odometric estimation is performed using an Extended we have used a 5 ratio for the augmentation
Kalman Filter formalism with noisy input. (Rx=5*1.222 m and Ry=5*2.752 m). The first rough GPS
3.1 Architecture measurement is used to initialize the position.
The odometric EKF computes an estimation of ζ=(∆, ω) 3.4 Results
and of the covariance matrix S. This data is used as the The global localizer has been tested on the four laps of
input of a second EKF estimating (x,y,θ). The the experiment of section 2.5 (9.6 km long run). Errors
measurements (x,y) of the GPS (in a local ground frame) are computed thanks to the post-processed DGPS.
are used as the observations. It is obvious that the GPS Three characteristics are tested in this experiment:
data change with the number of visible satellites. One 1) The normal operation with all measurements
usual way to take into account this non stationarity available,
consists to inflate the state covariance matrix each time 2) seven small GPS masks of ten seconds (for
the constellation changes. 70<t<210 s), similar to the ones observed in town
1 PPS GPS circuits,
(VP ONCORE)
3) a large one of 5 minutes (for 900<t<1200 s) to test
(x,y) the robustness of the error computation.
ABS
ζ (x,y,θ) Figures 9, 10, 11 and 12 present the results (the dotted
computer odometric S Localiser P lines are the 3 σ estimated bounds).
+ EKF EKF
steering
angle 300
V

Figure 8: architecture of the localization system 200

This architecture can be seen as a "loosely coupled fusion

system" [8]. 100
(m )
3.2 Heading observability
0
The heading θ is not directly measured since the GPS is
used as a position sensor. Nevertheless, by studying the
-1 0 0
state observability of the non linear system [3], it is easy
to verify that the observability condition is verified when
-2 0 0
the speed of the car is non-zero.
In order to assess what happens in this case, let consider a
-3 0 0
car when it stops: the measurement ω becomes zero 0 200 400 600 800 1000 1200 1400
t (s)
without any noise (no signal come from the ABS
encoders) and the odometric prediction of the heading Figure 9: x error and 3 σ estimated bounds
 80 Figure 12 indicates that the tuning of the filter is well
60
done since the estimated 3 standard deviation is
consistent with the error. Naturally, the position error is
40 mainly the one of the GPS and therefore is not white.
20
(m ) 4 CONCLUSION
0
This work has presented the development of a
-2 0 localization system and its experimentation with real data.
It is a very cheap localizer, since it uses natural GPS
-4 0
receiver and ABS sensors available on most modern cars.
-6 0 Thanks to an adequate sampling, the proposed odometric
method is computed when necessary and in a way well
-8 0
0 200 400 600 800 1000 1200 1400 adapted to a GPS correction. The experiments have
t (s)
proved that the use of all the ABS sensors increases the
Figure 10: y error and 3 σ estimated bounds precision of the positioning system in comparison with
(m )
the use of only the two ones associated with the rear
20 wheels. This dead-reckoning technique is sufficient to
maintain a precise estimation of the position during small
10 GPS masks. During large masks, we think it will produce
an efficient estimation of the local trajectory which
0
should be useful for the map-matching process with a
precise digital map. The future work concerns the study
of an automatic calibration of the wheels circumference.
-1 0
As a matter of fact, even it not very sensible to the
variation of tires pressure, the circumference changes
-2 0 with the wear of the tires.

80 100 120 140 160 180 200

References
t (s) [1] E. Abbott and D. Powell. "Land-Vehicle Navigation
Figure 11: x error during the small masks using GPS". Proc. of the IEEE, Vol. 87, NO. 1, Jan. 99.
6 [2] B. Barshan and H.F. Durrant-Whyte. "Inertial
navigation systems for mobile robots". IEEE Trans. on
4 Rob. and Auto. Vol. 11, n°3, June 95. pp. 328-342.
[3] Ph. Bonnifait, G. Garcia. "Design and Experimental
2 Validation of an Odometric and Goniometric Locali-
zation System for Outdoor Robot Vehicles". IEEE Trans.
(m )
0 on Rob. and Auto. Vol. 14, n° 4, 98, pp. 541 - 548.
[4] Ph. Bonnifait, P. Bouron. "Statistical Modeling of the
-2 GPS and DGPS signals". Technical report
UTC/Heudiasyc, n° 2296. May 2000.
-4 [5] S. Cooper and H. Durrant-Whyte. "A Kalman Filter
Model for GPS Navigation of Land Vehicles". IROS.
-6 Sept. 94. Munich. pp. 157-163.
300 400 500 600 700 800
t (s)
[6] J. Guivant, E. Nebot, S. Baiker. "High accuracy
Navigation Using Laser Range Sensors in outdoors
Figure 12: x error when the GPS is ok. Applications". ICRA2000. San Francisco. pp. 3817-3822.
Even if they are computed using a linearized model, the [7] S. Julier, H. Durrant-Whyte. "Process Models For
large mask shows that the errors are well estimated since The High-Speed Navigation of Road Vehicles". ICRA.
the confidence bounds are consistent with the errors 1995. pp. 101-105.
(figures 9 and 10). [8] D. Mc Neil. "Multi-rate Sensor Fusion for GPS
Figures 11 and 12 focus on interesting parts of figure 9. Navigation Using Kalman Filtering". M. Sc. Report.
On figure 11, it should be noticed that, each time a mask Virginia Polytechnic Institute. May, 99.
occurs, the error is consistent with the estimated [9] Y. Zhao. "Vehicle Location and Navigation Systems".
confidence. This realistic situation gives an idea of the Artech House. 1997.
actual precision (a few meters) of the localization.

View publication stats