METHOD FOR PREDICTING CAR-FOLLOWING BEHAVIOR UNDER APOLLO PLATFORM

METHODE POUR PREDIRE UN COMPORTEMENT DE SUIVI DE VEHICULE DANS LA PLATEFORME APOLLO

English Abstract

Disclosed is a method for predicting a car-following behavior under the Apollo platform; during the construction process disclosed herein, the dynamic information and the static information during vehicle driving are obtained through understanding of the scene; the desired distance and reaction time are obtained by capturing the driver's behavior features, and the defuzzification process of the following-car model is improved using the heuristic search algorithm, and after computing based on the fuzzy inference model, the safety- and- comfort-based optimal solution of the following-car acceleration range is obtained. Meanwhile, the model is tested and validated using the Apollo simulation platform to ensure accuracy and utility of the model. The method as disclosed solves the issues that the conventional schemes utilizing fuzzy theories and artificial neural networks only focus on the velocities and accelerations of the leading car and the following car, as well as the interval therebetween, without consideration of driving environments.

French Abstract

Un procédé pour prédire un comportement de suivi de véhicule dans la plateforme Apollo est décrit. Au cours du procédé de construction décrit dans la présente, les informations dynamiques et les informations fixes lors de la conduite dun véhicule sont obtenues grâce à la compréhension de la scène. La distance souhaitée et le temps de réaction sont obtenus en capturant les particularités du comportement du conducteur, et le procédé de clarification du modèle de suivi de véhicule est amélioré à laide de lalgorithme de recherche heuristique. De plus, après le calcul en fonction du modèle dinférence obscure, la solution optimale de la fourchette daccélération de suivi de véhicule basée sur la sécurité et le confort est obtenue. Entre-temps, le modèle est mis à lessai et validé à laide de la plateforme de simulation Apollo afin dassurer lexactitude et lutilité de ce premier. Le procédé, tel quil est décrit, ne résout que les questions des schémas conventionnels utilisant des théories obscures et des réseaux neuronaux artificiels, les vélocités et accélérations de lautomobile davant et de lautomobile qui suit cette dernière, ainsi que lintervalle entre ces dernières, sans se préoccuper des environnements de conduite.

Claims

I/We Claim:
1. A method for controlling a car-following behavior under the Apollo
platform,
specifically comprising:
Step 1: differentiating scene information in an autonomous driving process of
a vehicle
into static information and dynamic information, and importing the static
information and the
dynamic information into Dreamview of the Apollo platform to construct a road
scene;
Step 2: capturing the following-car driver's behavior features in the car-
following state,
computing a desired distance through a dynamics equation based on the driving
data of the
following-car driver, and fitting out a reaction time distribution function of
the driver under the
influence of velocity difference and relative distance using a polynomial
regression method;
Step 3: first, performing fuzzification processing to the captured behavior
feature data of
the following-car driver using an improved fuzzy inference vehicle model;
second, selecting a
membership function based on analysis of the following-car driver behavior
features, and
formulating a fuzzy rule library; third, performing fuzzy inference using the
Mamdani model;
finally, improving defuzzification using heuristic learning to enhance
solution efficiency;
Step 4: predicting the following-car acceleration a using the improved vehicle
inference
model, i.e., the predicted acceleration value; introducing the computed
predicted acceleration
value a and the real acceleration a into the desired safe distance equation,
the ratio a between
the desired distance D' of the predicted acceleration value a and the desired
distance D of the
real acceleration a, and substituting a as the parameter factor into the
desired safe distance
equation to control feedback adjustment; and
changing a velocity of a first vehicle to change a following distance between
the first
vehicle and a second vehicle based on the safe distance equation, after step
4.
2. The method for controlling a car-following behavior under the Apollo
platform
according to claim 1, wherein Step 1 specifically comprises:
obtaining three-dimensional information and motion information of a traffic
scene,
wherein the three-dimensional information of the traffic scene refers to
static information in the
corresponding scene information and the motion information of the traffic
scene refers to
dynamic information in the scene information; preliminarily constructing a
topological structure
19

of the scene, wherein the topological information of the scene includes
information such as the
number of surrounding vehicles, the lanes occupied by surrounding vehicles,
and the distance
from road edge; inputting such information into Dreamview via a corresponding
interface of
Apollo; configuring paths to specific modules based on the table of Module
Output Interface
Standards provided by the simulated environment, and performing, by respective
modules in the
standard, environment construction with reference to the traffic flow and the
simulated
environment resulting from understanding of the scene.
3. The method for controlling a car-following behavior under the Apollo
platform
according to claim 1, wherein Step 2 specifically comprises:
Step 2.1 computing the desired distance:
let the maximum threshold spacing for the following-car driver to receive the
stimulus of
the leading car be Hmax , and the desired following spacing of the following
car within the
<IMG>
spacing ; the desired spacing should guarantee that when the leading
car abruptly
stops with the maximum deceleration, the following-car driver's post-reaction
braking can safely
avoid collision; the condition for preventing collision is:
<IMG>
where(nIG)denotes safe spacing, <IMG> is the velocity of the following car at
moment t,
r denotes the driver's reaction time, L denotes the vehicle length, k denotes
the allowed buffer
spacing between the head of the following car and the tail of the leading car
(i.e., followed car)
after stop, k is a constant, a. and a.+1 are maximum decelerations of the
leading and following
cars in the car-following behavior;
IMG:.
it is seen from the equation above that the safe spacing i5 dynamically
correlated
with the velocities of the leading and following cars; the driver desired
spacing is:
<IMG>

the larger one of the two distances as derived using the max function is the
current driver
desired distance;
Step 2.2: computing the reaction time:
first, computing the reaction time r based on the time sequence data of
variations of the
leading car acceleration and the following car acceleration; owing to
different reaction time for
different individuals, the corresponding reaction time may be inferred;
then, each reaction time corresponds to a set of data (velocity difference,
relative distance
(Av, Ax)), within the same reaction time, the leading car and the following
car are paired; this
velocity differences in the set of data include: velocity change and relative
distance within the
reaction time of each vehicle, wherein the velocity change refers to the
velocity difference of
each vehicle, while the relative distance is obtained through a relative
distance equation based on
the velocity difference and the acceleration difference;
finally, fitting out, by polynomial regression, the driver's reaction time
distribution
function under the influence of velocity difference and relative distance,
wherein the reaction
time distribution function has different function expressions for individual
following-car drivers,
e.g., exponential function, Sigmoid function.
4. The method for controlling a car-following behavior under the Apollo
platform
according to claim 3, wherein Step 3 specifically comprises:
Step 3.1 defining input parameters and output parameters:
let the computation equation of the velocity difference be:
<IMG>
the offset difference refers to the difference between the inter-vehicle
distance <IMG>
moment t and the following-car desired spacing cam), the equation being as
follows:
<IMG>
with the velocity difference and the offset difference as the input parameters
and the
angular velocity of the following car as the output parameter, the velocity
difference and offset
difference of the following car in the data set are computed based on the road
traffic driving data
21

set, and the velocity difference, offset difference and the range of
acceleration are found by
statistics;
Step 3.2 Fuzzification
with the velocity difference and the offset difference as the input parameters
in the fuzzy
inference system and the acceleration of the following car as the output
parameter, each input
parameter and each output parameter have 7 levels, which are represented as
N3, N2, N1, ZE,
P1, P2, and P3, respectively; for the velocity difference in the input
parameter, the level P3
represents that the value of the velocity difference is positive and largest;
levels P2 and P1
represent that the value of the velocity difference is positive but gradually
smaller; level ZE
represents that the value of velocity difference is 0, while levels N1, N2,
and N3 represent that
the values of velocity difference are negative and gradually smaller; for the
same reason, the
seven levels of another input parameter offset value and output parameter
acceleration are
identical to the above.
Step 3.3 Selecting a Membership Function:
let X *be an accurate value and A* represent a converted fuzzy set; then the
trigonometric membership function is:
<IMG>
where, it is seen from the trigonometric membership function distribution
diagram that,
O; when <IMG> , the trigonometric membership function fuzzy set becomes a
fuzzy
single value; the larger 0-, the less the influence of variation of <IMG> ,
i.e. when (7 is
large enough, the method offers a strong enough anti-interference capability;
Step 3.4 establishing a fuzzy rule library:
the fuzzy relationship is <IMG>
wherein: <IMG>
22

<IMG> <IMG>
in the equation, lenotes a dimensional vector formed by the matrix
i.e., if A and h, then e'; therefore, when inputting A, h, then
<IMG>
the conversion relationships of 7 levels of the velocity differences, offset
differences, and
accelerations of the two vehicles in car-following state are expressed with a
conditional
statement, thereby establishing 49 fuzzy rules; when the velocity differences
and offset
differences of the vehicles are all at level N3, it indicates that the
following car is in a very safe
driving state, the velocity of the leading car is greater than the velocity of
the following car, and
the actual inter-vehicle spacing is greater than the desired inter-vehicle
spacing; in order to
maintain a car-following state with respect to the leading car, the following
car driver needs to
accelerate to reduce the distance from the leading car so as to achieve a
desired distance;
therefore, with constant reduction of the offset difference, the driver's
acceleration decreases;
Step 3.5: Fuzzy inference:
with the Mamdani model as the fuzzy inference model, by resolving the smaller
one of
Cartisan products of fuzzy sets A and B in the Mamdani model, it is derived
that:
PR, (X, .3)) = PA ()OA /-1B (Y)
Step 3.6: Defuzzification:
Step 3.6.1: obtaining the initial solution anowof the acceleration of the
following car by
the center-of-gravity method, i.e., the following car acceleration a; then,
letting the taboo table H
be empty, i.e., H = ;
Step 3.6.2: if a termination condition is satisfied, jumping to step 3.6.4;
otherwise,
selecting a candidate set Can ¨N(an w)satisfying the taboo requirement from
the neighboring
domain N(II, anmv) of the initial solution an', and then jumping to step
3.6.3;
wherein the taboo requirement is: neighboring domain dal', satisfying
<IMG>
23

the termination condition is: when the local optimal solutions resulting from
twice
iterations do not change any more, or the difference between the evaluation
functions of the
twice optimal solutions is not large, stopping iteration;
Step 3.6.3: selecting a solution anext with the best evaluation value from the
candidate set,
updating the taboo table <IMG> ), and setting it as the currently
optimal solution
<IMG> ; then shifting to step 2;
Step 3.6.4: output the computation result anOW, and stopping searching.
5. The method for controlling a car-following behavior under the Apollo
platform
according to claim 4, wherein the fuzzy rule in Step 3.4 is constructed as
follows:
the value ranges of the velocity difference Av., offset difference 811, and
acceleration cl.+1
are divided evenly into 7 levels: P3, P2, Pl, ZE, N1, N2, N3, wherein P3
denotes the positive
maximum value; for the velocity difference Av., when the velocity difference
Av. is P3, it
indicates that the velocity difference between the leading car and the
following car is very large;
P2 and P3 denote that the values of the velocity difference Av., the offset
difference 811, and the
acceleration a.+1 decrease gradually; ZE denotes 0, while N1, N2, and N3
indicate that the
velocity difference Av., the offset difference 811, and the acceleration a.+1
are negative values
that decrease gradually; therefore, the fuzzy rule with the velocity
difference Av. and the offset
difference 811 between the two vehicles is established as such:
let the set <IMG>
given that Av. and 8. each have 7 states, 49 acceleration states are obtained.
6. The method for controlling a car-following behavior under the Apollo
platform
according to claim 4, wherein in Step 4, the feedback adjustment is controlled
based on a
feedback adjustment equation below:
<IMG>
24

where, a denotes the feedback adjustment equation, based on which the feedback
adjustment may be controlled.
7. A method for controlling a car-following behavior under the Apollo platform
application executed on a computer system, specifically comprising:
Step 1: differentiating scene information in an autonomous driving process of
a vehicle
into static information and dynamic information, and importing the static
information and the
dynamic information into Dreamview of the Apollo platform to construct a road
scene;
Step 2: capturing the following-car driver's behavior features in the car-
following state,
Step 3: executing computer executable instructions stored in a storage of the
computer
system, the computer executable instructions, when executed on a processor of
the computer
system causing the computer system to:
compute a desired distance through a dynamics equation based on the driving
data of the
following-car driver, and fit out a reaction time distribution function of the
driver under the
influence of velocity difference and relative distance using a polynomial
regression method;
perform fuzzification processing to the captured behavior feature data of the
following-
car driver using an improved fuzzy inference vehicle model;, select a
membership function based
on analysis of the following-car driver behavior features, and formulating a
fuzzy rule library;
perform fuzzy inference using the Mamdani model; and, perform defuzzification
using heuristic
learning to enhance solution efficiency; and
predict the following-car acceleration a using the improved vehicle inference
model, i.e.,
the predicted acceleration value; introduce the computed predicted
acceleration value a..i and the
real acceleration a into the desired safe distance equation, the ratio a
between the desired
distance D' of the predicted acceleration value ...!i and the desired distance
D of the real
acceleration a, and substitute a as the parameter factor into the desired
safe distance equation
to control feedback adjustment; and
changing a velocity of a first vehicle to change a following distance between
the first
vehicle and a second vehicle based on the safe distance equation, after step
3.

Description

Method for Predicting Car-following behavior Under Apollo Platform
FIELD
[0001] The present disclosure relates to car-following behaviors, and more
particularly
relates to a method for predicting a car-following behavior under the Apollo
platform.
BACKGROUND
[0002] Traffic refers to the conveyances and environment for travel. With
improvement of
economic level, development of scientific and technical level, and
acceleration of
urbanization, the people's living standards are also improved, which imposes
higher demands
on travel provision and quality. However, growth and development of first and
second-tier
cities boost constant increase of urban population density; more and more
urban citizens buy
automobiles, causing urban traffic loads heavier year on year. Due to limits
of city size and
road network capacity, urban traffic becomes heavy and jammed, which seriously
affects
travel and commuting of urban residents. To handle the ever-serious traffic
jam, the
government has launched a series of policies including: vigorously developing
public
transportations, restricting use of private vehicles based on license plate
numbers, practicing
plate-number lottery or auction, charging street-parking, etc., but none of
such moves
succeeds in deterring car-parc surge. Car following is a common phenomenon in
road traffic,
especially in traffic jam which makes it impossible to change a lane or
overtake. Therefore,
study of car-following behaviors helps understand traffic-flow
characteristics.
[0003] The theory of following-car model emerged in 1950s. According to the
then model,
measured vehicle data information was fitted to obtain a mathematical
equation. However, a
following-car model obtained based on that method has certain limitations. For
example,
when the data change, the model would become unsuitable any more, which does
not
facilitate promotion and extension of the model. Therefore, in recent years, a
plurality of
models have been proposed focusing on internal causes of car-following
behaviors, which
significantly enriches studies on the traffic-flow theory.
[0004] However, influenced by multiple sources of information, a driver's
decision-making
and judgment process exhibits a complex non-linear modality during driving,
and the driver's
psychological decision cannot be described with a simple mathematical
expression. Fuzzy
theories and artificial neural networks show certain operational advantages in
handling
complex non-linear issues and also exhibit a good learning capacity under big
data samples.
CA 3065617 2019-12-19

Therefore, the fuzzy theory and artificial neural network are often used for
simulating driving
behaviors under different environments. However, the current schemes utilizing
fuzzy
theories and artificial neural networks only focus on the velocities and
accelerations of the
leading car and the following car, as well as the spacing therebetween,
without considering
driving environments.
[0005] In April 2017, Baidu released its open platform Apollo for autonomous
driving; after
iterations of multiple versions, the platform has been enabled for
localization, sensing,
decision, and simulation. Apollo may help its partners in the automotive and
autonomous
driving industries to quickly develop a set of their own autonomous driving
systems in
consideration of vehicles and hardware systems. In the Apollo simulation
environment,
environment information including traffic signs, index lines, and the
relationships with
surrounding vehicles may be inputted into Dreamview via corresponding
interfaces to thereby
construct a driving environment. Besides, the Apollo platform is further
enabled for validating
the following-car model and optimizing the relevant algorithm through a 3D
visual interface.
SUMMARY
[0006] An object of the present disclosure is to provide a method for
predicting a
car-following behavior under the Apollo platform, which solves the issues that
the
conventional schemes utilizing fuzzy theories and artificial neural networks
only focus on the
velocities and accelerations of the leading car and the following car, as well
as the spacing
therebetween, but fail to consider driving environments.
[0007] The present disclosure relates to a method for predicting a car-
following behavior
under the Apollo platform, specifically comprising:
[0008] Step 1: differentiating scene information in an autonomous driving
process of a
vehicle into static information and dynamic information, and importing the
static information
and the dynamic information into Dreamview of the Apollo platform to construct
a road
scene;
[0009] Step 2: capturing the following-car driver's behavior features in the
car-following
state, computing a desired distance through a dynamics equation based on the
driving data of
the following-car driver, and fitting out a reaction time distribution
function of the driver
under the influence of velocity difference and relative distance using a
polynomial regression
method;
2
CA 3065617 2019-12-19

[0010] Step 3: first, performing fuzzification processing to the captured
behavior feature
data of the following-car driver using an improved fuzzy inference vehicle
model; second,
selecting a membership function based on analysis of the following-car driver
behavior
features, and formulating a fuzzy rule library; third, performing fuzzy
inference using the
Mamdani model; finally, improving defuzzification using heuristic learning to
enhance
solution efficiency; and
[0011] Step 4: predicting the following-car acceleration a using the improved
vehicle
inference model, i.e., the predicted acceleration value; introducing the
computed predicted
acceleration value a and the real acceleration a into the desired safe
distance equation, the
ratio a between the desired distance D' of the predicted acceleration value a
and the
desired distance D of the real acceleration a, and substituting a as the
parameter factor
into the desired safe distance equation to control feedback adjustment.
[0012] Preferably, wherein Step 1 specifically comprises: obtaining three-
dimensional
information and motion information of a traffic scene, wherein the three-
dimensional
.. information of the traffic scene refers to static information in the
corresponding scene
information ,and the motion information of the traffic scene refers to dynamic
information in
the scene information; preliminarily constructing a topological structure of
the scene, wherein
the topological information of the scene includes information such as the
number of
surrounding vehicles, the lanes occupied by surrounding vehicles, and the
distance from road
edge; inputting such information into Dreamview via a corresponding interface
of Apollo;
configuring paths to specific modules based on the table of Module Output
Interface
Standards provided by the simulated environment, and performing, by respective
modules in
the standard, environment construction with reference to the traffic flow and
the simulated
environment resulting from understanding of the scene.
.. [0013] Preferably, wherein Step 2 specifically comprises:
[0014] Step 2.1 computing the desired distance: let the maximum threshold
spacing for the
following-car driver to receive the stimulus of the leading car be H., and the
desired
following spacing of the following car within the spacing Hmax be h,(t) ; the
desired
spacing should guarantee that when the leading car abruptly stops with the
maximum
deceleration, the following-car driver's post-reaction braking can safely
avoid collision; the
condition for preventing collision is:
3
CA 3065617 2019-12-19

he(t) +1(t) =r+L+k+x n+12 (t) x2 n (t)
2an+1 2an+1
where he (1) denotes safe spacing,x+1(t) is the velocity of the following car
at moment t ,
denotes the driver' s reaction time, L denotes the vehicle length, k denotes
the allowed
buffer spacing between the head of the following car and the tail of the
leading car (i.e.,
followed car) after stop, k is a constant, an and an+1 are maximum
decelerations of the
leading and following cars in the car-following behavior; it is seen from the
equation above
,
that the safe spacing h(1) is dynamically correlated with the velocities of
the leading and
following cars; the driver desired spacing is:
/r n2 (t) ,L + k)
D(t)= max(x1(t) = x12 (t) x
+ L + k + ________________________ n+
2an+1 2an+1
the larger one of the two distances as derived using the max function is the
current driver
desired distance;
100151 Step
2.2: computing the reaction time: first, computing the reaction time r based
on the time sequence data of variations of the leading car acceleration and
the following car
acceleration; owing to different reaction time for different individuals, the
corresponding
reaction time may be inferred; then, each reaction time corresponds to a set
of data (velocity
difference, relative distance (Ay' Ax)); within the same reaction time, the
leading car and the
following car are paired; this velocity differences in the set of data
include: velocity change
and relative distance within the reaction time of each vehicle, wherein the
velocity change
refers to the velocity difference of each vehicle, while the relative distance
is obtained through
a relative distance equation based on the velocity difference and the
acceleration difference;
finally, fitting out, by polynomial regression, the driver's reaction time
distribution function
under the influence of velocity difference and relative distance, wherein the
reaction time
distribution function has different function expressions for individual
following-car drivers,
e.g., exponential function, Sigmoid function.
100161 Preferably, wherein Step 3 specifically comprises:
[0017] Step 3.1 defining input parameters and output parameters: let the
computation
4
CA 3065617 2019-12-19

equation of the velocity difference be:
Avn = vn ¨ vn+,
n
the offset difference refers to the difference between the inter-vehicle
distance AL(t)at
moment t and the following-car desired spacing D(t) , the equation being as
follows:
6 = ALn (i) ¨ D n (t)
with the velocity difference and the offset difference as the input parameters
and the angular
velocity of the following car as the output parameter, the velocity difference
and offset
difference of the following car in the data set are computed based on the road
traffic driving
data set, and the velocity difference, offset difference and the range of
acceleration are found
by statistics;
[0018] Step 3.2 Fuzzification: with the velocity difference and the offset
difference as the
input parameters in the fuzzy inference system and the acceleration of the
following car as the
output parameter, each input parameter and each output parameter have 7
levels, which are
represented as N3, N2, Ni, ZE, PI, P2, and P3, respectively; for the velocity
difference in the
input parameter, the level P3 represents that the value of the velocity
difference is positive and
largest; levels P2 and PI represent that the value of the velocity difference
is positive but
gradually smaller; level ZE represents that the value of velocity difference
is 0, while levels
Ni, N2, and N3 represent that the values of velocity difference are negative
and gradually
smaller; for the same reason, the seven levels of another input parameter
offset value and
output parameter acceleration are identical to the above.
[0019] Step 3.3 Selecting a Membership Function: let x*be an accurate value
and A*
represent a converted fuzzy set; then the trigonometric membership function
is:
Ix_
RA ) {(1 X* 1)
0a - x* s
x - x >
where, it is seen from the trigonometric membership function distribution
diagram that,
a > ; when Ix 1> , the
trigonometric membership function fuzzy set becomes a fuzzy
single value; the larger a, the less the influence of variation of x* to
PA*(x), i.e. when a
5
CA 3065617 2019-12-19

is large enough, the method offers a strong enough anti-interference
capability;
[0020] Step 3.4 establishing a fuzzy rule library:
the fuzzy relationship is
¨ ¨
wherein: A x: = 74(x) A lAY),in the equation, [A x B]T denotes a dimensional
vector formed
AxB ¨
by the matrix , i.e., if A b
and ' , then therefore, when inputting h
', then
r -
C =[A X B.}
the conversion relationships of 7 levels of the velocity differences, offset
differences, and
accelerations of the two vehicles in car-following state are expressed with a
conditional
statement, thereby establishing 49 fuzzy rules; when the velocity differences
and offset
differences of the vehicles are all at level N3, it indicates that the
following car is in a very
safe driving state, the velocity of the leading car is greater than the
velocity of the following
car, and the actual inter-vehicle spacing is greater than the desired inter-
vehicle spacing; in
order to maintain a car-following state with respect to the leading car, the
following car driver
needs to accelerate to reduce the distance from the leading car so as to
achieve a desired
distance; therefore, with constant reduction of the offset difference, the
driver's acceleration
decreases;
[0021] Step 3.5: Fuzzy inference: with the Mamdani model as the fuzzy
inference model,
by resolving the smaller one of Cartisan products of fuzzy sets A and B in the
Mamdani
model, it is derived that:
(xs Y) = fiA (x) A PB (Y)
[0022] Step 3.6: Defuzzification.
[0023] Step 3.6.1: obtaining the initial solution a of the acceleration of the
following car
by the center-of-gravity method, i.e., the following car acceleration a; then,
letting the taboo
table H be empty, i.e., H 0 .
.. [0024] Step 3.6.2: if a termination condition is satisfied, jumping to step
3.6.4; otherwise,
n
selecting a candidate set Can _ N(a")satisfying the taboo requirement from the
neighboring domain N(H , an") of the initial solution an" , and then jumping
to step 3.6.3;
6
CA 3065617 2019-12-19

wherein the taboo requirement is: neighboring domain a"n , satisfying
a"" e N(H,a"")and a"" H ;the termination condition is: when the local optimal
solutions
resulting from twice iterations do not change any more, or the difference
between the
evaluation functions of the twice optimal solutions is not large, stopping
iteration.
[0025] Step 3.6.3: selecting a solution anew with the best evaluation value
from the
now
candidate set, updating the taboo table H = H L..)Can _N(a) , and setting it
as the currently
ext
optimal solution anom ¨ an ; then shifting to step 2.
[0026] Step 3.6.4: output the computation result anou , and stopping
searching.
[0027] Preferably, wherein the fuzzy rule in Step 3.4 is constructed as
follows: the value
ranges of the velocity difference Avn , offset difference s", and acceleration
a,,1 are
divided evenly into 7 levels: P3, P2, Pl, ZE, N1, N2, N3, wherein P3 denotes
the positive
maximum value; for the velocity difference Avn , when the velocity difference
Avn is P3, it
indicates that the velocity difference between the leading car and the
following car is very
large; P2 and P3 denote that the values of the velocity difference Avn , the
offset difference
En, and the acceleration an+1 decrease gradually; ZE denotes 0, while Ni, N2,
and N3
indicate that the velocity difference Avn , the offset difference En, and the
acceleration a0+1
are negative values that decrease gradually; therefore, the fuzzy rule with
the velocity
difference Avn and the offset difference En between the two vehicles is
established as such:
let the set Q=11\13, N2, Ni, ZE, PI, P2, P31, Avnnsr, > a+I Avn e Q EnEQ a,, e
Q
n
3 9
given that Avn and n each have 7 states, 49 acceleration states are obtained.
[0028] Preferably, wherein in Step 4, the feedback adjustment is controlled
based on a
feedback adjustment equation below:
xn'+,2(t) 2(t),L + k)
k(o=a=max(xn+,=r +L+k+
2aõ, 2an+,
where, a denotes the feedback adjustment equation, based on which the feedback
adjustment
may be controlled.
[0029] The method for predicting a car-following behavior under the Apollo
platform
according to present disclosure offers the following beneficial effects: (1)
by testing and
7
CA 3065617 2019-12-19

validating algorithm feasibility with the Apollo simulation platform and
applying a learning
mechanism, model accuracy is improved; (2) by adding reaction time varied
dependent on the
driver's different stages into the desired distance equation and introducing a
parameter factor
and tuning the driver's desired distance, the simulation result of the model
is more
approximate to the car-following behavior of a real driver, empowering the
model to
adaptively adjust the feedback control; (3) the defuzzification process of the
car-following
behavior model is improved using a heuristic search algorithm, thereby finding
the optimal
solution for safety- and-comfort-based acceleration.
BRIEF DESCRIPTION OF THE DRAWINGS
[0030] In order to make more likely to be clearly understood that the content
of the present
invention, the following specific embodiment according to the present
invention with the
accompanying drawings which, for the present invention as described in further
detail,
wherein :
[0031] Fig. 1 shows the method for predicting a car-following behavior under
the Apollo
platform flow diagram;
[0032] Fig. 2 shows a schematic diagram of a safe spacing for emergent
braking;
[0033] Fig. 3 shows a flow diagram of a fuzzy inference system;
[0034] Fig. 4 shows a trigonometric membership function diagram;
[0035] Fig. 5 shows a fussy inference diagram of the Mamdani model; and
[0036] Fig. 6 shows a Dreamview interface for environment simulation.
DETAILED DESCRIPTION OF EMBODIMENTS
[0037] Hereinafter, the present disclosure will be described in detail through
preferred
embodiments with reference to the accompanying drawings.
[0038] The present disclosure provides a method for predicting a car-following
behavior
under the Apollo platform, wherein a structured description of a scene is
first formed based on
understanding of the scene to give restrictions in motion geometry and
physical dimensions of
the vehicle; by inputting environment information such as traffic signs, index
lines, and the
relationships with surrounding vehicles into the Dreamview of the Apollo
platform, a real
road scene is constructed; then, the driving behaviors of the following-car
driver are divided
8
CA 3065617 2019-12-19

into 4 stages: perception stage, inference stage, decision stage, and
execution stage. During
the perception stage, the obtained ambient information is processed to compute
the desired
spacing of the following-car driver and the velocity difference at the current
moment; during
the inference stage, the following-car driver's action plan is inferred based
on a fuzzy
interference rule; during the decision stage, the leading-car driver's safety-
and-comfort-based
optimal acceleration is obtained by defuzzifying the following-car driver's
action plan; and
during the executing stage, the following-car driver implements vehicle
velocity change from
in-brain decision to hand manipulation. Finally, the predicting method is
tested and validated
using the Apollo simulation platform to ensure accuracy and utility of the
predicting method.
[0039] The present disclose provides a method for predicting a car-following
behavior under
the Apollo platform, a flow diagram of which is shown in Fig. 1, specifically
comprising:
[0040] Step 1: differentiating scene information in an autonomous driving
process of a
vehicle into static information and dynamic information, and importing the
static information
and the dynamic information into Dreamview of the Apollo platform to construct
a road scene,
specifically including: obtaining three-dimensional information of a traffic
scene and motion
information, wherein the three-dimensional information of the traffic scene
refers to static
information in the corresponding scene information and the motion information
of the traffic
scene refers to dynamic information in the scene information; preliminarily
constructing a
topological structure of the scene, wherein the topological information of the
scene includes
information such as the number of surrounding vehicles, the lanes occupied by
surrounding
vehicles, and the distance from a road edge; inputting such information into
Dreamview via a
corresponding interface of Apollo; configuring paths to specific modules based
on the Table
of Module Output Interface Standards ( Table 3)provided by the simulation
environment, and
performing, by respective modules in the Standard, environment construction
with reference
to the traffic flow and simulated environment information resulting from
understanding of the
scene, as shown in Table 1 below:
[0041] Table 1 Table of Module Output Interface Standards Provided by
Simulation
Output data
the simulation will provide:
such as the
position, orientation, linear¨ velocity
/apollo/localizati position and
Localization linearacceleration, angular
velocity in
on/pose orientation of _ _
the following pose
car
9
CA 3065617 2019-12-19

Output data
such as
the simulation will provide:
positions,
id,position, theta,velocity, length,width,
/apollo/percepti orientations,
on/obstacles velocities, height,type, polygon_point in
Perception shapes, and etc. PerceptionObstacle
of respective
obstacles
/apollo/percepti Output traffic the simulation will provide:
on/traffic_light light signals color,id and tracking_time in
TrafficLight,
Output data
such as the
/apollo/canbus/c velocity and the simulation will provide:
CAN bus
hassis drive mode of speed_mps
the following
car
the simulation will provide:
Output the
/apollo/routing_ Routing Response, including the
planned
Router navigation
response result navigation route from the current
position
to the destination
100421 Table 3 Table of Decision Planning Module Interface Standards
Module Topic Description Fields
Output the
The developer must provide:
planned
(1) timestamp_sec in Header
/apollo/planni trajectory of the
Planning (2) v,a, relative_time in
ng following car in
TrajectoryPoint
a future period
(3) x,y,z,theta,kappa in PathPoint
of time
Output The developer optionally outputs:
/apollo/predict respective trajectories in
PredictionObstacle,
Prediction obstacles and which may be used for
displaying
ion
their predicted predicted trajectories of
trajectories respective obstacles
Output
The developer optionally outputs:
decisions with
MainDecision and
respect to
/apollo/decisio ObjectDecisions, which may be
Decision various
used for displaying the main
obstacles and
decisions and the decisions with
the main
respect to respective obstacles
decisions
100431 Step 2: capturing the following-car driver's behavior features in car-
following state,
computing a desired distance of the driver through a dynamics equation based
on the driving
data of the following-car driver, and fitting out a reaction time distribution
function of the
CA 3065617 2019-12-19

driver under the influence of velocity difference and relative distance using
a polynomial
regression method, specifically including steps of:
[0044] Step 2.1 computing the desired distance:
[0045] Let the maximum threshold spacing for the following-car driver to
receive the
stimulus from the leading car be Hmax , the desired following spacing of the
following car
within the spacing H max be kW. The desired spacing should guarantee that when
the
leading car abruptly stops with the maximum deceleration, the following-car
driver's
post-reaction braking can safely avoid collision; the safe spacing under the
brake condition is
schematically shown in Fig.2; the condition for collision avoidance is:
XI 2(t) ,/ 2
he(t) xn' +1(t) = r + L + k + n+1 A'n
2an+1 2an+1
where he (t) denotes safe spacing, n+1 (t) is the velocity of the following
car at moment t,
r denotes the driver's reaction time, L denotes the vehicle length, k denotes
the allowed
buffer spacing between the head of the following car and the tail of the
leading car after stop,
k is a constant, au and an+1 are maximum decelerations of the leading and
following cars
in the car-following behavior;
[0046] It is seen from the equation above that the safe spacing he(1) is
dynamically
correlated with the velocities of the leading and following cars; the driver
desired spacing is:
i n2(t) , L +k)
D(t)= max(x X+12 (t) X
,c+1(t)=r+L+k+ n
2aõ 1 2an+1
The larger one of the two desired spacings in the equation as derived using
the max function
is the current driver desired distance;
[0047] Step 2.2: computing the reaction time(fitting out a reaction time
distribution function
of the driver under the influence of velocity difference and relative distance
using a
polynomial regression method):
[0048] first, computing the reaction time r based on the time sequence data of
acceleration
variations of the leading car and the following car; owing to different
reaction time for
different individuals, the corresponding reaction time may be inferred;
11
CA 3065617 2019-12-19

[0049] then, each reaction time corresponds to a set of data (velocity
difference, relative
distance (Ay' Ax)); within the same reaction time, the leading car and the
following car are
paired; the velocity difference in the data set include: velocity change and
relative distance
within the reaction time of each vehicle, wherein the velocity change refers
to the velocity
difference of each vehicle, while the relative distance is obtained through a
relative distance
equation based on the velocity difference and the acceleration difference;
[0050] Finally, fitting out, by polynomial regression, the driver's reaction
time distribution
function under the influence of velocity difference and relative distance,
wherein the reaction
time distribution function has different function expressions for individual
following-car
drivers, e.g., exponential function, Sigmoid function.
[0051] As shown in Figs. 3, 4, and 5, Step 3 as following: first, performing
fuzzification
processing to the captured behavior feature data of the following-car
driver,i.e. accurate data,
using an improved fuzzy inference vehicle model; second, selecting a
membership
function ,i.e.data source,based on analysis of the following-car driver
behavior features, and
formulating a fuzzy rule library; next, performing fuzzy inference using the
Mamdani
model,i.e.Fuzzy reasoning Computing Center; finally, improving defuzzification
using
heuristic learning to enhance solution efficiency, specifically including
steps of:
[0052] Step 3.1 defining input parameters and output parameters:
[0053] Let the computation equation of the velocity difference be:
Avõ = võ ¨v,1
the offset difference refers to the difference between the inter-vehicle
distance AL n(t) at
moment t and the following-car desired spacing D(t) , the equation being:
6 n(t) = ALn(t) D n(t)
with the velocity difference and the offset difference as input parameters and
the angular
velocity of the following car as the output parameter, the velocity difference
and offset
difference of the following car in the data set is computed based on the road
traffic driving
data set, and the velocity difference, the offset difference and the range of
acceleration are
found by statistics;
[0054] Step 3.2 Fuzzification
12
CA 3065617 2019-12-19

[0055] With the velocity difference and the offset difference as the input
parameters in the
fuzzy inference system and the acceleration of the following car as the output
parameter, each
input parameter and each output parameter have 7 levels, which are represented
as N3, N2,
Ni, ZE, PI, P2, and P3, respectively, as shown in Table2. For the velocity
difference in the
input parameters, the level P3 represents that the value of the velocity
difference is positive
and largest; levels P2 and PI represent that the value of the velocity
difference is positive but
gradually smaller; level ZE represents that the value of velocity difference
is 0, while levels
N1, N2, and N3 represent that the values of velocity difference are negative
and gradually
smaller; for the same reason, the seven levels of another input parameter
offset value and
output parameter acceleration are identical to the above.
[0056] Table 2 Fuzzy Rule Correspondence Table
Avn
en
N3 N2 Ni ZE PI P2 P3
N3 N3 N3 N2 N2 Ni ZE ZE
N2 N3 N2 N2 Ni ZE ZE P1
Ni N2 N2 Ni Ni ZE P1 P1
ZE N2 N1 NI ZE P1 P1 P2
PI N2 Ni ZE ZE PI P2 P2
P2 Ni ZE ZE P1 P2 P2 P3
P3 P1 P1 P1 P1 P2 P3 P3
[0057] Step 3.3 Selecting a Membership Function:
[0058] Let x*be an accurate value and 4* represent a converted fuzzy set; then
the
trigonometric membership function is:
= {(1 ¨ IX - X* ) x - xl s a
cj
A
O
where, as shown in Fig. 4, it is seen from the trigonometric membership
function distribution
diagram that, a > 0; when Ix ¨ x I > o-, the trigonometric membership function
fuzzy set
becomes a fuzzy single value; the larger a, the less the influence of
variation of x* on
i.e. When a is large enough, the method offers a strong enough anti-
interference
13
CA 3065617 2019-12-19

capability;
[0059] Step 3.4 establishing a fuzzy rule library:
[0060] The fuzzy relationship is
wherein: x = A(x) A B(y) ,
.
in the equation, [A x BIT denotes a dimensional vector formed by the matrix
"m t , .e.,
if A and B, then e'; therefore, when inputting A, :6, then
e = [A x B
]r
the conversion relationships of the 7 levels of velocity difference, offset
difference, and
acceleration of each of the two vehicles in the car-following state are
expressed with a
.. conditional statement, thereby establishing 49 fuzzy rules, as shown in
Table 2; when the
velocity differences and offset differences of the two vehicles are all at
level N3, it indicates
that the following car is in a very safe driving state, the velocity of the
leading car is greater
than the velocity of the following car, and the actual inter-vehicle spacing
is greater than the
desired inter-vehicle spacing; in order to maintain a state of following the
leading car, the
following car driver needs to accelerate to reduce the distance from the
leading car so as to
achieve a desired distance; therefore, with constant reduction of the offset
difference, the
driver's acceleration decreases;
[0061] Step 3.5: Fuzzy Inference:
[0062] With the Mamdani model as the fuzzy inference model, the inputs and
outputs of the
Mamdani model are all fuzzy amounts; de-fuzzification is needed after the
fuzzy inference;
when the Max-Min operator is adopted, the fuzzy inference is shown in Fig. 5,
wherein by
resolving the smaller one of Cartisan products of fuzzy sets A and B in the
Mamdani model, it
is derived that:
I'RC'Y) = (x) A pB(y)
[0063] Step 3.6: Defuzzification:
[0064] Step 3.6.1: obtaining the initial solution an" of the acceleration of
the following car
14
CA 3065617 2019-12-19

by the center-of-gravity method, i.e., the following car acceleration a; then,
letting the taboo
table H be empty, i.e., H = 0 ;
[0065] The center-of-gravity method is a method of taking, as the clear value,
the element
corresponding to the center-of-gravity of the area enclosed by the fuzzy set
membership
function curve and the basic variable axis, which is a commonly used
defuzzification method.
[0066] The basic idea of the taboo search algorithm is to use a taboo table to
record trapping
into local optimal solutions through searching historical records, such that
in the next search,
repeated selection of searching local extremum points is forbidden using the
information in
the taboo table, so as to jump out of the local optimal points, which
facilitates obtaining a
global optimal solution.
[0067] Step 3.6.2: if a termination condition is satisfied, jumping to step
3.6.4; otherwise,
now
Can _N(a )
selecting a candidate set
satisfying the taboo requirement from the
N( H "}") now
neighboring domain ,a of
the initial solution a , and then jumping to step 3.6.3;
wherein the taboo requirement is that the neighboring domain acan should
satisfy
a' e N(H, )and a' 0 H .
, the termination condition is: when the local optimal
solutions resulting from twice iterations do not change any more, or the
difference between
the evaluation functions of the two optimal solutions is not large, stopping
iteration;
[0068] Step 3.6.3: selecting a solution amy' with the best evaluation value
from the
candidate set, updating the taboo table H = H uCan _N(an'') , and setting it
as the currently
optimal solution anon ¨ anext ; then shifting to step 2;
[0069] Step 3.6.4: outputting the computation result a , and stopping
searching.
[0070] The evaluation value is obtained by dividing the time headway (TH) by
the predicted
acceleration. In equivalent driving conditions, the larger the time headway,
the smaller the
possibility of collision with the leading car. Further, in consideration of
comfort, the absolute
value of the acceleration generally shall not be too large; a too large
acceleration would cause
discomfort of passengers in the following car. Therefore, the acceleration is
in reverse
proportion to the time headway.
[0071] In step 3.4, the fuzzy rule is established in the following manner:
CA 3065617 2019-12-19

[0072] The value ranges of the velocity difference Avn , offset difference en
, and
acceleration an+, are divided evenly into 7 levels: P3, P2, Pl, ZE, Ni, N2,
N3, wherein P3
denotes the positive maximum value; for the velocity difference Ayn , when the
velocity
difference Avn is P3, it indicates that the velocity difference between the
leading car and the
following car is very large; P2 and P3 denote that the values of the velocity
difference Avn,
the offset difference , and
the acceleration an+, decrease gradually; ZE denotes 0, while
Ni, N2, and N3 indicate that the velocity difference Ayn, the offset
difference , and the
acceleration an+, are negative values that decrease gradually; therefore, the
fuzzy rule for the
velocity difference Ayn and the offset difference en between the two vehicles
is established
as such:
Av n an+, n
[0073] Let the set Q={N3, N2, NI, ZE, PI, P2, P3}, n AvEQ sEQ
aõ,, E Q; given that Avn and en each have 7 states, 49 acceleration states are
obtained.
[0074] Step 4: predicting the following-car acceleration a using the improved
vehicle
inference model, i.e., the predicted acceleration value; substituting the
computed predicted
acceleration value a and the real acceleration a into the desired safe
distance equation,
computing the ratio a between the desired distance D' of the predicted
acceleration value
U and the desired distance D of the real acceleration a, and substituting a as
the
parameter factor into the desired safe distance equation to control feedback
adjustment,
wherein the feedback adjustment equation is specified below:
xn'+12(t) 2(t), L + k)
he(t) = a=max(xn.,=r + L + k +
2a1 2an+1
where, a denotes the feedback adjustment equation, through which the feedback
adjustment
may be controlled.
[0075] During the construction process as disclosed herein, the dynamic
information and the
static information during vehicle driving are obtained through understanding
of the scene; the
desired distance and reaction time are obtained by capturing the driver's
behavior features,
and the defuzzification process of the following-car model is improved using
the heuristic
16
CA 3065617 2019-12-19

search algorithm, and based on the fuzzy inference model, computing the safety-
and-
comfort-based optimal solution of the following-car acceleration range.
Meanwhile, the
model is tested and validated using the Apollo simulation platform to ensure
accuracy and
utility of the model.
[0076] The method for predicting a car-following behavior under the Apollo
platform as
disclosed is tested using the Apollo simulation platform. After the Apollo
software
environment is configured, the output interface of the Apollo platform is
docked with the
method. After successfully predicting information such as the following-car
acceleration
using the fuzzy inference behavior control policy model method, the method is
docked with
the decision planning module Planning of Apollo; finally, the Apollo software
implements
testing and validation of the fuzzy inference behavior control policy model;
during multiple
times of simulation process, the parameters are constantly adjusted, and the
algorithm is
optimized over the Apollo visual platform, specifically:
[0077] Step a: deploying the environment (e.g., Docker environment), and
pulling the
container mirror of Apollo;
[0078] Step b: entering the Apollo container, and compiling the simulation
environment (e.g.,
Dreamview simulation environment);
[0079] Step c: running the simulation environment after successful
compilation;
[0080] Step d: testing and validating the efficacy of the model using the
corresponding
simulation environment; the testing and validating interface refers to the
simulation
environment interface, wherein the interface is shown in Fig. 6, specifically:
[0081] First, docking the traffic flow and environment information outputted
by Apollo with
the input of the model; then, converting the predicted acceleration value
obtained using the
model into the Planning input simulation platform, wherein the specific
docking path is
shown in Table 3; next, a vehicle under the simulation platform may adjust the
velocity based
on the input plan, and the final predicted acceleration value and the actual
value are used to
optimize the driver's desired car-following distance through feedback
optimization.
[0082] Testing and validating the model in the Apollo simulation environment:
[0083] Docking the traffic flow and environment information outputted by
Apollo with the
input of the model; then, converting the predicted acceleration value obtained
through the
model into the Planning input simulation platform, wherein the specific
docking path is
17
CA 3065617 2019-12-19

shown in Fig. 6.
[0084] A vehicle under the simulation platform may adjust the velocity based
on the input
plan, and the final predicted acceleration value and the actual value are used
to optimize the
driver's desired car-following distance through feedback optimization.
.. [0085] The above embodiment the present invention are expressed only one of
several
embodiments, a more specific and detailed description thereof, but it is not
is thus able to be
construed as limiting the scope of the present invention patent. It should be
noted, one of
ordinary skill in the art, without departing from the concept of the present
invention,
numerous variations and modifications may be made, which are within the scope
of the
present invention. Thus, the scope of which shall be defined by the present
invention as the
standard.
18
CA 3065617 2019-12-19

Representative Drawing