Content uploaded by Hossein Tehrani
Author content
All content in this area was uploaded by Hossein Tehrani on Nov 24, 2017
Content may be subject to copyright.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
1
Abstract— Lane change maneuver is a complicated maneuver
that causes many severe expressway accidents. Automated lane
change has great potentials to reduce the number of accidents.
Previous research mostly tried to find an optimal trajectory and
ignore the behavior model. To understand the human lane change
model, we carried out experiments at Japan expressways.
Through analyzing the human-driver lane change data, we
proposed a two-segment lane change model to mimic the
human-driver. We categorize the environment based on the
observation grid and propose alternative behaviors to handle the
complicated scenario. We develop an intuitive method to select the
suitable lane change behavior including active
(accelerate/decelerate) and passive (wait) based on distance and
related velocity (dx/dv) graph. We also extracted the most
preferable and safe condition for doing lane change based on the
human driver preference data. We evaluated the proposed model
by doing lane change simulations in PreScan environment with
considering the vehicle motion/control model. The simulation
results show the proposed model is able to handle the complicated
lane change scenario and the performance is near the human
driver.
Index Terms— Automotive engineering, Navigation, Vehicle
safety.
I. INTRODUCTION
TATISCAL data of expressway traffic accidents shows
that human error is a major reason for nearly 90% of
accidents (Volvo 2013) [1]. Lane change maneuver is a cause
for many severe expressway accidents due to wrong estimation
of surrounding environment or wrong maneuvering. Currently,
ADAS or automated driving has demonstrated the potential to
reduce the impact of human errors. The literature in lane change
or overtaking maneuver-based ADAS can be divided into
rule-based [2], [3] or utility-based [4] approaches. In [5] and [6],
Kasper et al introduced an object-oriented Bayesian network
approach to model the traffic scene for the detection of lane
change maneuvers. Bayesian network was also used by
Schubert [7] for lane change situation assessment and decision
Q. H. Do and S. Mita are with the Research Center for Smart Vehicles,
Toyota Technological Institute, Nagoya, Japan (e-mail: huydq@toyota-ti.ac.jp;
smita@toyota-ti.ac.jp).
H. Tehrani, M. Egawa and K. Muto are with Corporate R&D Div.3, DENSO
CORPORATION, Kariya, Aichi, Japan (e-mail: {hossein_tehrani,
masumi_egawa, kenji_Muto}@ denso.co.jp) .
K. Yoneda was with the Research Center for Smart Vehicles,Toyota
Technological Institute, Nagoya, Japan. Currently, he is with Autonomous
Vehicle Research Unit of Institute for Frontier Science Initiative, Kanazawa
University. (e-mail: k.yoneda@staff.kanazawa-u.ac.jp).
making. In [8], the authors model the surrounding environment
into an occupancy grid, then apply dynamic programing to find
drivability cells and find proper decision for lane keeping or
lane changing. Other related researches focus on maneuver
prediction. In [9], [10] authors trained a neural network for the
prediction of future lane change trajectory. A dynamic
Bayesian network is also used by Gindele [11] and
Schlechtriemen [12]-[13] for the behavior and trajectory
prediction. Meanwhile, Kumar utilized the support vector
machine and Bayesian filter for the same purpose [14]. The
fuzzy logic was used by Naranjo et al [15] for modelling the
lane change decision making. They use fuzzy controllers that
mimic human behavior and reactions during overtaking
maneuvers. Bahram [16] proposed a decision making based on
a nonlinear model predictive approach. Mixed logical
dynamical system is also used for solving lane change decision
making as in [17]-[19]. Simon and Markus [20] applied an
online Partially Observable Markov Decision Process to solve
the decision making for lane change. Brechtel et al. [21] applied
probabilistic MDP-Behavior planning for cars. Ardelt et al.
[22] presented a probabilistic approach to build a lane change
framework for automated vehicle. A lane selection method was
proposed by Jin [23], however, this method requires all related
vehicles connected. In [24], authors presented a collision
avoidance method by analyzing the kinematic model of the
vehicle and then calculated the minimum safety space
requirement for performing the lane change. Other methods for
lane change based on safety space checking are presented in
[25], [26]. However, they provided a more general framework
rather than focusing on lane change decision situations in
particular.
The lane change trajectory is generated according to the
vehicle state, surrounding vehicles and road information. The
control laws are then designed to use on-board sensors to track
the generated trajectory. Most of discussed researches either try
to predict the behavior of surrounding vehicle or based on the
kinematic functions and try to find an optimal trajectory or a
safest path [27]-[29]. They ignore the ego vehicle lane change
behavior model for complicated scenarios or in the presence of
time/distance constraint for merging/exiting of the expressway.
To exactly understand the human lane change model, we have
conducted lane change experiments at expressways in Japan.
The results of the human lane change motion analysis were
already published by the authors [30]. By means of analyzing
human lane change data including neighboring vehicles and
motion/behavior data, we realized that the human model was
Human Drivers based Active-Passive Model for
Automated Lane Change
Quoc Huy Do, Hossein Tehrani, Seiichi Mita, Masumi Egawa, Kenji Muto, Keisuke Yoneda
S
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
2
not a single stage. This paper is extended from our previous
work [31]. In this paper, a multi-segment behavior and motion
model to mimic the human-driver lane change operation is
proposed. At the “behavior segment”, the ego vehicle tries to
adjust the longitudinal distance, position or relative velocity to
make or find suitable free space in front of it and in the
destination lane. Alternative behaviors for “behavior segment”
are generated to be able to handle the lane change for any
scenario. A method to select the suitable behavior based on the
dx/dv graph segmentation is proposed. This graph provides the
safety distance dx and safety speed dv for doing the lane change
based on the human-driver lane change patterns.
At “motion segment”, the ego vehicle starts lateral motion and
enters into the destination lane. An optimization function is
proposed which integrates both lateral and longitudinal
trajectories in the same function. Through comparison with
human-driver data, we found that the proposed evaluation
function was able to select a motion close to human-driver.
The structure of this paper is as follows; Section II explains
the analysis of human lane change data and our proposed
automated lane change flowchart. Section III introduces our
two segments lane change model and behavior generation and
selection criteria. The lane change motion planning is also
described in section III. Section IV presents our simulation and
comparison results. Finally, the conclusion is given in section
V.
II. AUTOMATED LANE CHANGE MODEL
A. Lane Change Experiment
To evaluate the human driver motion during the lane change,
we conducted experiments at Isewangan Expressway at Aichi
prefecture. The experiment route with the length of 21.4 km is
shown in Fig. 1(a) and marked inside red rectangle area.
Different long-time-career expert drivers (whose driving skills
are higher than average drivers) are selected for doing lane
change experiments. Lane change maneuvers were frequently
performed between different driving lanes to cover most
possible scenarios in the expressway. The lane change
experiments process utilized to record different human driver
behavior and motion is shown in Fig. 1(b). When the driver
receives request for changing the lane from test engineer, the
driver adjusts the longitudinal speed (acceleration/deceleration)
and finds a safe space and time instant to start steering for lane
change. Lane change in expressway is generally a challenging
task, and the driver should adjust both the lateral and
longitudinal acceleration to perform safe and comfortable lane
change. All the surrounding vehicles’ information, including
their relative distance, and speed, was also extracted. These
data are used to analyze the human driver behavior during the
lane change. Figure 1(c) shows the diagram of the whole lane
change experiment to extract data and evaluate the proposed
lane change method.
B. Lane Change Model
Through lane change experiments, we learned that the
human-driver model does not begin at the instant of the steering
wheel turn, but rather several seconds before the steering wheel
turn. Figure 2 (a) shows an example of a recorded lane change
scenario. The ego vehicle (red car) wanted to change to the
right lane. However, there was one vehicle ahead of the
ego-vehicle and one vehicle coming from behind in the
intended target lane. As shown in sample in Fig. 2 (a), turning
the steering wheel starts at time 17, although the human-driver
has already started to decelerate and reduce the vehicle speed
from 23.85 m/s to 22.23 m/s. The driver had to decelerate to
keep safe space from the front vehicle and wait for the vehicle
coming from behind in the right lane to pass before starting the
lane change. In this paper, a two-segment behavior/motion lane
change model is proposed as shown in Fig. 2 (b). In segment 1,
“behavior segment”, the ego vehicle tries to adjust the
longitudinal distance, position or relative velocity to make or
find suitable free space in front of it and in the destination lane.
In segment 2, “motion segment”, the ego vehicle starts lateral
motion and enters into the destination lane. The driver behavior
in segment 1 is highly dependent on the number of surrounding
vehicles, distances and their relative velocities. When the
vehicles enter/exit from an adjacent lane to the mainstream
(a) Experimented Expressway.
(b) Lane change experiments process to extract the human driver behavior.
(c) Lane change experiment diagram.
Fig. 1. Lane Change experiment.
time
Test Engineer Expert Driver
Lane change request
Start tag
Driver Evaluate
Environment
accelerate /decelerate
to adjust safety distance
/space
Lane change impossible
Cancel
experiment
tag Start steering
End of lane change
tag
Behavior Model
Motion Plan
Algorithms
Comparison and
Evaluation
Human ⇆Computer
Lane Change Experiments
Expert
Driver
-Lateral & Longitudinal Motion
-Lane information
-Surrounding vehicle
Data Extraction
Simulation(PreScan)
Feedback
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
3
traffic, the time/distance constraint is also a critical factor to
select the suitable behavior at segment 1. This factor is based on
the specific length of the adjacent road. There are also other
important parameters such as road curvature, visibility
condition or behavior of the surrounding vehicles. By analyzing
different cases of human lane change data, the following
behaviors for the segment 1 are extracted.
(1)
Although it is difficult to develop a general behavior model, it
is possible to propose a standard model that guarantees the
safety and smoothness of the lane change operation. For
example, if there is enough space at the destination lane, the
vehicle can turn the steering wheel and do the lane change. If
the relative speed of the approaching vehicles in the destination
lane is relatively high, ego vehicle may prefer to wait (passive)
until finding enough free space. In this case, it may even do
deceleration (active) to reduce the time or traveled distance.
The deceleration behavior may be useful in the case of
time/distance constraint when ego vehicle has to change the
lane to exit from the expressway. In the other case, when the
relative speed of the neighboring vehicles in the destination
lane is relatively low, ego vehicle may accelerate (active) to
pass the neighboring vehicles for doing the lane change. If there
is a sudden change in the behavior of the surrounding vehicles
during the lane change (sudden acceleration, deceleration or
lane change), ego vehicle might need evasive (active) maneuver
to avoid accident. On the other hand, the road curvature, speed
a) Lane change scenario.
(b) Two-segment model.
Fig. 2. Two-segment lane change model.
Target
Lane
Motion
Direction
Lane Changing
Finish Lane Change
Lane Changing
21
22
23
24
25
26
27
0 5 10 15 20 25
Speed (m/s)
Ego vehicle
Deceleration to make
free space/time for lane
change
Time(s)
Turning the
steering wheel
Change lane
and accelerate
to adjust
speed
Longitudinal
acc
Lateral
acc
Segment 1 Segment 2
Fig. 3. Automated lane change flowchart.
Estimate speed/Position of
neighboring vehicles Estimate the behavior of
neighboring vehicles
Generate lateral/longitudinal
trajectory for neighboring vehicles
Lane information
Make
grid map
Behavior A
Segment 1: Do lane change
Behavior B (Passive)
Segment 1: Wait
Segment 2: Lane change
Behavior C (Active)
Segment 1: Accelerate
Segment 2: Lane change
Behavior D (Active)
Segment 1: Decelerate
Segment 2: Lane change
Image sensor Laser scanner / Radar
Set of alternative
behaviors
Evaluation of
different
behaviors
Behavior selection
criteria's
Generate
acceleration/
deceleration/w
ait patterns
for Segment 1
Generate
motion for
Segment 2
Behavior A
is selected ? yes
no
Time buffer for
re-evaluation
(every
timestamp:
milliseconds)
Execution
and control
Real time
environment
assessment
Real time control & execution
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
4
limits and kinematic constraints of the vehicles such as
maximum speed or acceleration also have impact on the
driver’s behavior.
C. Automated Lane Change Flowchart
Automated lane change is an integration of
sensing/perception, planning (behavior & motion) and control.
The behavior/motion flowchart for doing the lane change is
shown in Fig. 3. It starts with estimation of motion parameters/
position of neighboring vehicles to estimate their trajectories.
Based on the driving lane information and the behavior/motion
parameters of the neighboring vehicles, a trajectory for a
certain period [] can be estimated. By using the trajectory
data points (), the occupancy grid map of current
and future state of the surrounding environment is estimated. In
the following, the main components of the automated lane
change flowchart are briefly explained.
D. Situation Modelling
The current lane change situation is modelled into a state
occupancy grid as in the Fig. 4. This state grid is attached to the
vehicle position. The grid cell’s width and orientation are equal
to the lane width. The middle cell length is equal to the
ego vehicle length plus a safety distance. The grid’s size is
calculated based on the relative velocity of the ego vehicle and
surrounding vehicles, and the time to complete the lane change.
The front cell size’s is calculated based on the closest
vehicle in the front, and behind cell’s size is calculated
based on the corresponding closest vehicle approaching from
behind.
(2)
is the intended or required time to perform lane
change maneuver, and it takes 4–6 s for a typical driver to
complete lane change and based on human driver data [32].
is the velocity of the closest vehicle in the front, and
is the velocity of the closest vehicle behind the ego vehicle;
and are human defined minimum safety distances;
is the current ego vehicle’s velocity. In literature, the grid
cells are generally considered to be same size [5],[8].
Empirically through simulations and recorded human-driver
data, we realized that the center cells in the grid were just
dependent on the vehicle size and they were generally smaller
than the front and back cells in the grid. Considering large size
for center cell causes the ego vehicle does not do the lane
change due to occupation of the center cell in the left or right
side. We also realized that the back cells have larger size
compared to front cells to mimic the human behavior.
By discretizing the surrounding environment into a nine-cell
grid, it is possible to present different states to perform the lane
change. There are eight surrounding cells in the occupancy grid
and each cell can be either free (cell value = 0) or occupied (cell
value = 1). Thus there are totally 2^8 = 256 states. Since the
vehicle only perform either the left side or right side lane
change maneuver, the other side’s cell can be temporarily
ignored so that the number of state can be reduced to 2^5 = 32
states.
E. Trajectory Estimation of Neighboring Vehicles
Recorded past motion data, road information and
behavior (lane change/lane keeping) are utilized to estimate the
trajectory of the neighboring vehicles. Polynomial function is
used to estimate the trajectory of the neighboring vehicles [33].
As shown in Fig. 5, the trajectory of a neighboring vehicle is
estimated by recorded position data and lane center point (we
assume that this vehicle will follow the center of the lane) and a
polynomial curve is fitted to these data. Quintic polynomials
are used to generate the estimated trajectory. This is given as
(4)
is polynomial function’s factor which can be
calculated based on the past motion data point and the lane
center points.
Thist is recorded time and Topr is total operation/prediction
time. Here, “Frenet Frame” method is applied in order to
combine different lateral and longitudinal motions in one
equation [33]. In this case, the lateral and longitudinal motion
of each vehicle can be presented by an equation which is based
on the distance traveled along the center line.
There are effective collisions checking methods in the literature
to check the collision between two trajectories [34]. To check
the collision in the simulation platform, the trajectory of the
neighboring vehicle is sampled for a certain period [0~T].
Inevitable Collision States (ICS) method [35] is applied to
check the collision possibility between two trajectories.
III. BEHAVIOR AND MOTION GENERATION
A general two-segment model for doing lane change has
been already proposed in previous section. At segment one,
“behavior segment”, the ego vehicle tries to adjust the
longitudinal relative distance, velocity to make or find suitable
free space in front or in the destination lane. In this segment, the
ego vehicle may accelerate, decelerate or just wait to make free
space/time interval at the destination lane. At segment 2,
Fig. 4. Environment grid modelling.
Fig. 5. Neighboring vehicle’s trajectory estimation.
Historical data Estimated Trajectory
(5 degree polynomial)
d
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
5
“motion segment”, the ego vehicle starts lateral/longitudinal
motion and enters the destination lane.
This paper extends “behavior segment” introduced in [36]
and briefly review the lateral/longitudinal motion generation
method for “motion segment” presented in [29].
A. Behavior Segment
For every state, different alternative behaviors are
considered as shown in Fig. 6. For the situation in Fig. 6,
different following behaviors are available to do the lane
change;
Behavior A: The ego vehicle does the lane change with
current speed as there is enough space and the relative
velocities of the vehicles in the right lane are not high.
Behavior B (active): The ego vehicle decelerates and
enters to the right lane at the back of the vehicles. It is
preferable behavior when the lane change at limited
time/distance has to be done (for example exit point in
the expressway).
Behavior C (active): The ego vehicle accelerates and
enters to the right lane at the front of vehicles.
Behavior D (passive): If the relative speeds of right lane
vehicles are high, the ego vehicle just waits until the
right lane vehicles passes and the right lane becomes
free to do right lane change.
The proper behavior can be selected based on the relative
distance and between the ego vehicle and neighboring vehicles.
The large of difference of relative velocity are shown in the
graph in Fig. 6 which was extracted from recorded human-data.
To have exact understanding of different behaviors, we
categorized 32 (2^5) occupancy gird states (left or right lane
change) to the following four main categories. The behavior
alternatives are limited based on the categories to reduce the
calculation time. Different categories for occupancy grid states
are shown in Fig. 7. In Fig.7, the white cells are empty cells,
light gray cells are temporarily not considered cells due to the
direction of the lane change and black cells are occupied cells.
The categories and alternative behaviors are defined based on
our experiments from analyzing the human-driver data for
different lane change scenarios.
Category A: There are two alternative behaviors for
occupancy grid states. The ego vehicle either waits or
performs the lane change based on the relative speed and
distance to the neighboring vehicles.
Category B: There are more alternative behaviors for
states in this category. The ego vehicle may do the lane
change but sometimes acceleration/deceleration or wait is
Fig. 6. Different available behavior and motion for lane change and velocity
profile corresponding to each behavior.
Time
Velocity
Behavior A,D (Passive)
Behavior B (Active)
Behavior C (Active)
Behavior A : Lane change
Behavior B: Lane change with deceleration
(Active Model)
Behavior C: Lane change with acceleration
(Active Model)
Behavior D: Wait (Passive Model)
Time
Lateral acc
Behavior A Behavior C Behavior B
Ego vehicle
(a) Category A
(b) Category B
(c) Category C
(d) Category D
Fig. 7. All 32 states models for right lane change and available
alternative behaviors for each of them.
1- accelerate
2- decelerate
3- lane change
4- wait
1- accelerate
2- lane change
3- wait
1- accelerate
2- wait
3- lane change
1- accelerate
2- wait
3- lane change
1- accelerate
2- decelerate
3- lane change
4- wait
1- decelerate
2- lane
change
1- decelerate
2- lane change
3- wait
1- decelerate
2- lane change
3- wait
1-decelerate
2- wait
3- lane change
1-decelerate
2- wait
3- lane change
1- accelerate
2- decelerate
3- wait
1- accelerate
2- wait 1- accelerate
2- decelerate
3- wait
1- accelerate
2- wait
1- decelerate
2- wait 1- decelerate
2- wait
1- accelerate
2- decelerate
3- wait
1- accelerate
2- wait
1- decelerate
2- wait 1- accelerate
2- wait 1- decelerate
2- wait 1- decelerate
2- wait
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
6
preferable to do safer/smoother lane change. In the case of
time/distance constraint, acceleration/deceleration may be
necessary to satisfy the lane change limitations.
Category C: It is related to complicated state during lane
change. In this category, the right cell of the ego vehicle is
occupied and suitable behavior should be selected to
provide free space at the destination lane. Ego vehicle may
accelerate, decelerate or wait for states in this category.
Category D: The ego vehicle has to wait for doing the
lane change and any other behavior may be dangerous.
The observation grid state can transit from one category to
another depending on the driver and neighboring vehicles. The
human driver often draw the current situation in his brain and
match it with the four categories and figures out which situation
is next if he selects a specific behavior. Eventually, category A
is selected by accelerating or decelerating the vehicle, and lane
change maneuver is initiated.
B. Behavior Selection
To select the suitable behavior for lane changing, there are
many researches in the literature that are mainly based on
HMM [37] or Bayesian network [27]. In this section, an
intuitive method to extract the behavior patterns from the
human-driver recorded data is proposed. The graph is
drawn for all the recorded lane change data from start until
end of the lane change at . The is calculated when the
ego vehicle driver decides to do the lane change in the presence
of the vehicle in the destination lane. This graph corresponds to
the “behavior segment” in proposed lane change model. Each
draw includes the (dx, dv) status for the “behavior segment”
while the driver adjusts the longitudinal distance/speed to make
the safe space and time before the steering at . An example of
for human driver regarding to a typical right lane change
scenario is shown in Fig. 8. The horizontal axis is the related
distance between vehicle in the target lane and the ego
vehicle, and vertical axis is the velocity difference between
two vehicles. In Fig. 8, the human driver slowly accelerates
(active) until the neighboring vehicle in the right lane passes.
(dx, dv) points are shown in Fig 8 and there are noises due to
the sensor noise and detection/tracking uncertainties, the (dx,
dv) points are not smooth or linearly continuous. To handle the
sensor uncertainty and noise, we consider a margin for the (dx,
dv) data and interpolate a line to model the human driver
behavior.
More than 90 cases of human driver lane change were
analyzed, and graphs for all these cases were drawn.
The graph for right lane change cases is shown in Fig. 9.
The horizontal axis is the distance between vehicle in the target
lane and the ego vehicle, and vertical axis is the relative
velocity between two vehicles. The lines show different
recorded right lane change data that start from
(human-driver decide to do the right lane change) until the
(red star) that driver steers to enter the right lane and continue.
Based on the different samples of human-driver that are shown
in Fig. 9, the graph’s area is able to be divided into three
areas corresponding to three different behaviors including
and
based on the initial value for . At the start of the right
lane change, if the corresponding point is in “Lane
Change” area, the ego vehicle is able to enter to right lane. If the
is in “Wait” (passive) area, ego vehicle waits until
enters the “Lane Change” area (the right vehicle pass)
and then enters to the right lane. If the is in
“Accelerate” (active) area, ego vehicle will accelerate and wait
until value enters the “Lane Change” and then do the
lane change. The acceleration behavior is usual because the ego
vehicle enters the higher speed lane and it should adjust speed
with current lane. This graph also shows the safety relative
distance or relative speed for doing the safe and
comfortable right lane change. For sample in Fig. 9, the
driver accelerates and reduces its relative speed before enter to
the right lane. For sample, the driver waits until the right
lane vehicle passes it and then starts to enter to the lane change.
The graph in Fig.9, is only prepared for the positive
values of because all the human lane change data lie in this
area and vehicles in the right lane have generally higher speed.
To complete the graph and cover all the values of , it is
necessary to provide the human driver lane change data for the
negative values of .
The graph for the typical left lane change cases in
expressway was also prepared. As shown in Fig. 10, the driver
behavior is different from the right lane change because in this
case, the vehicle enters to the lower speed lane. Similarly, the
graph area can be divided into three areas
Fig 8. graph for human right lane change.
Lane Change Start
ego vehicle
Behavior Segment
Steering Start
-30
-25
-20
-15
-10
-5
0
5
10
15
20
25
0123456
dv: 4.95
dx: -12.2
Start Lane Change
dv: 2.54
dx: 8.06
End Lane Change
Behavior Segment
Motion
Segment
Acceleration
(active)
dv (m/s)
dx (m)
margin
Interpolated
line
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
7
corresponding to three different behaviors including
and
based on the initial value of .
The graphs in Fig. 9 and 10 provide an intuitive
model for performing safe and comfortable lane change. Based
on the initial value of , we are able to select the
suitable active/passive behaviors and generate speed and timing
profile for doing the automated lane change. As shown in Fig.
11, there is a vehicle in the right lane and we are initiating the
right lane change with initial condition at , There
are many alternatives as shown by different lines from No. 1 to
No. 5 for doing the lane change. For example, case No.1, we
just wait in the (passive mode) until the back vehicle in the right
lane passes. For other cases (active model), we accelerate to
reduce our relative speed.
If we carefully look at the human lane change data in Fig. 9,
there is an area, where most of the human lane change begins.
This area is named as the . In
the , the relative velocities is
between [0 ~ 3 m/s] and relative distance is between [5 ~ 20
m]. To select the human like behavior for doing the right lane
change, we draw the shortest line from initial condition at
to and generate speed
profile and lane change timing. We track the and
when we enter into the ; we start to do
the lane change. One advantage of our approach in comparison
with other related method is the calculation time. Since all the
data are recorded and processed off-line to generate the dv-dx
graph. The behavior selection is simplified into matching the
current (dv-dv) point into the graph so that the calculation time
is small (less than 10 ms).
C. Velocity Profile Planning (Behavior Segment)
According to the active/passive behavior, the corresponding
desired vehicle’s velocity profile is calculated. The
acceleration/deceleration patterns are presented in the form of a
function of ego vehicle position, velocity and the target
leading/behind vehicle.
Figure 12 illustrates the vehicle acceleration behavior when a
leading vehicle exists in the neighboring lane. In this scenario,
ego vehicle accelerates and passes the leading vehicle. The
acceleration is continued till sufficient safe distance is achieved,
before performing the lane change maneuver. The ego vehicle’s
acceleration is calculated based on (5):
where vehicle position at time;
: velocity at time;
(): leading vehicle position at time ;
: leading vehicle position at time ;
: operation time horizon;
: safety reserving distance;
is calculated through leading vehicle velocity and time to
react
(6)
To generate smooth and comfort acceleration/deceleration
motion, the following cost function is minimized;
Fig. 12. Vehicle acceleration for passing.
Fig. 9. graph related to human driver behavior for doing the
right lane change.
Fig. 10. graph related to human driver behavior for doing the
left lane change scenario.
Fig. 11. Selection of suitable behavior pattern based on the initial value of
( for right lane change
-35
-30
-25
-20
-15
-10
-5
0
5
10
15
20
25
30
35
40
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10
Active
(Accelerate)
Lane Change
Passive
(Wait)
t0
ts
t0
ts
(m/s)
-35
-30
-25
-20
-15
-10
-5
0
5
10
15
20
25
30
35
40
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10
Active
(Decelerate)
Lane Change
Passive
(Wait)
-35
-30
-25
-20
-15
-10
-5
0
5
10
15
20
25
30
35
40
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10
Active
(Accelerate)
Lane Change
Passive
(Wait)
⑤
①
②
③
④
Safety
Margin
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
8
where and are human given weighted factor to balance
between the jerk and safety distance (). This cost function is
similar to the method presented in [20] though the operation
time is fixed in [20]. In our approach, the operation time is
not fixed to have more degree of freedom. The error in the
safety distance () is calculated by the following;
(8)
There is direct method to solve the above problem using
Lagrange multiplier and Gradient Descent. Though finding the
exact solution for optimization problem in (7) is difficult and
time consuming. Here, we turn our attention to approximations
of the minimizer through a simplification and sampling from
search space. Instead of calculating the best trajectory explicitly
and modifying the coefficients to get a valid alternative, we
generate in a first step, such as in [33, 38], alternative
trajectories for both . Later we can pick the valid and
safe motion which is safe and has the lowest cost value.
The quartic polynomial function is utilized to generate
acceleration/deceleration motion [34].
(9)
The coefficients are estimated by considering
ego vehicle constraints (maximum acceleration/deceleration
and speed), operation time and desired velocity at the
end. Alternative longitudinal trajectories are generated by
sampling a valid range of operation time and final velocity
(it can be extracted from graphs) while
considering the boundary conditions including maximum and
minimum acceleration . We generate alternative
speed profiles by sampling and select the best speed profile
which has the lowest value for cost function in (7).
D. Motion Segment
To model a lateral path during a lane change, many
approaches in literature use 5th degree polynomials as it
provides minimum jerk for steering [27, 37]. Polynomial
function provides a geometric modelling of the vehicle
trajectory that responds to the realistic demands of the lane
change maneuver.
(10)
The equation coefficients
are calculated
considering dynamic constraints (boundary conditions for
lateral acceleration) and values of the position, velocity and
acceleration at initial and endpoint. The initial velocity and
acceleration of vehicle can be obtained from the CAN and
generate alternative lateral trajectories by changing operation
time . is sampled from a valid range of operation time and
ending conditions to generate alternative lateral trajectories
while considering the boundary conditions including
and
to avoid slip. The best lateral
motion is selected which minimize the following cost function
that includes lateral jerk, heading error and smoothness as
shown in Fig. 13;
where , and are human given
weighted factors.
IV. SIMULATION
To evaluate and test the proposed model, a simulation
platform on the Prescan was developed. It includes different
modules for sensing, behavior/motion planning, trajectory
estimation of neighboring vehicles and ego vehicle’s control.
The behavior/motion generation module is developed under
C++ to increase the efficiency of the simulation platform.
A. Lane Change Scenario
Figure 14 shows simulation results for the right lane
change scenario with different relative speed for
neighboring vehicles (ego vehicle is red one and it is going to
do right lane change). In Fig. 14, the ego vehicle (red vehicle in
the central lane) wants to perform the lane change maneuver to
the right-side lane with initial
). Based on the proposed method, the ego vehicle
makes a speed and timing profile to the “most preferable area”
as shown by dotted red line in Fig. 15. It slowly accelerates
(active model) and starts the lane change when enter to the safe
area as shown in Fig. 14 simulation. We have tested the
Category C Category B Category A Lane Change
Fig. 14. Automated right lane change simulation.
Fig. 13. Lateral trajectory optimization.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
9
proposed method with different initial and the
results are shown by dotted lines in Fig. 15. The tested results
show that the proposed model works properly, without any
collision during the lane change simulation.
B. Merge and Exit
The proposed model was extended to cover more
complicated cases of lane change for merging into traffic
highway. In these cases, the time/ distance for doing lane
change is limited. The suitable behaviors for merge and exit are
active model (accelerate or decelerate) to do the lane change as
soon as possible. The proposed segments and suitable
behavior for each segment is shown in Fig. 16. Two simulations
to merge or exit (Fig. 17 and Fig.18) which consider two initial
statuses for are shown by in Fig. 16. For
, the value of falls in the accelerate area and the
acceleration behavior is selected to merge the expressway. For
, the value of falls in the deceleration area and
the suitable behavior is decelerate to merge the expressway. As
shown in Fig. 16, the changes are not linear like normal
lane changes and there are sharp changes in the (accelerate
and decelerate) to enter the lane change area in shorter time.
The simulation results for are shown in the Fig. 18
(a), (b) respectively.
Simulations to exit the expressway are also carried out. Two
examples which considering two initial statuses that are shown
Category C Category B Category A Lane Change
(a) , merge in expressway simulation -deceleration
(b) Velocity profile to merge into highway for
Fig. 17. Simulation results to merge into highway for .
t=0 , start
Segment 1:
decelerate
Segment 2:
lane change
Merge
Fig. 15. graph related to simulated right lane change scenario
Fig. 16. graph for right lane change simulation scenarios
-35
-30
-25
-20
-15
-10
-5
0
5
10
15
20
25
30
35
40
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10
Lane
Change
Lane Change
decelerate
(active)
accelerate
(active)
Lane
Change)
Category C Category B Category A Lane Change
(a) , merge in expressway simulation -deceleration behavior
(b) Velocity profile
Fig. 18. Lane change simulation to merge.
t=0 , start
Segment 1:
decelerate
Segment 2:
lane change
Merge
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
10
by in Fig. 16. In these case, there are also sharp
changes in the (accelerate and decelerate) to enter the lane
change area in shorter time. Based on the position in a prior
known (or recorded global map), the ego vehicle is able to
know if it is close to the merge or exit position or not. The
system is then able to decide between patterns in Fig.9 or
Fig.16.
C. Comparison between Human and Computer
In our method, we “mimic” the human driver in “behavior
segment”, however, in “motion segment”, we applied
polynomial functions to generate lane change trajectory. In this
section, we compare human-driver with computer motion
generator for “motion segment” during the expressway lane
change. To evaluate both lateral and longitudinal trajectories in
same function, the lateral jerk and smoothness are utilized as
Kanayama and Hartman proposed [39]. Evaluation function is
defined as integral over the square of arc-length derivative of
curvature along the path for a function with
curvature . The results of comparison between
human-driver and computer motion generation for motion
segment are shown in Table I. In most cases the computer can
generate close or even better motion sets compared to
human-driver (lower lateral jerk and lower smoothness-1).
The motion curvature of one lane change sample for
computer motion generated and human-driver is shown in Fig.
19. As shown in this sample, the computer generated smoother
motion (lower curvature) compared to human-driver.
V. CONCLUSIONS
In this paper, a behavior/motion model for automated lane
change at expressway has been proposed. The proposed model
is mainly inspired by human-driver lane change and behavior
data that can handle difficult lane change scenarios. The
occupancy grid states are categorized and alternative behaviors
are defined for each corresponding category. To select the
suitable behavior, dx/dv graph is segmented based on the
human-driver lane change patterns. Our experiments were
carried on at the expressway where the lanes have different
ranges of velocity. In this case, change to left or right lane
means change to slower or higher speed lane. Thus, the
segmented dx/dv will be dependent on the left or right lane
change. The proposed model is intuitive and able to handle
complicated lane change scenarios even in the presence of
disturbances or sudden changes in behavior of surrounding
vehicles. In future research, more human-driver lane change
patterns are going to be extracted for better segmentation of
dx/dv graph. Our current behavior model does not consider the
interaction between the vehicles so that in future work, our
model will consider it.
REFERENCES
[1] Volvo Trucks, European Accident Research and Safety
Report.[Online]http://www.volvotrucks.com/SiteCollectionDocuments/
VTC/Corporate/Values/ART%20Report%202013_150dpi.pdf, 2013.
[2] L. Fletcher et al. (2009), The DARPA Urban Challenge, volume 56 of
Springer Tracts in Advanced Robotics, pp. 509–548. Springer Berlin
Heidelberg.
[3] W. He, X. Wang, G. Chen, M. Guo, T. Zhang, P. Han and R. Zhang,”
Monocular based lane-change on scaled-down autonomous vehicles,” in
Proc. IEEE Intelligent Vehicles Symposium (IV), 2011, pp.144-149.
[4] W. Dolan and B. Litkouhi , “A prediction- and cost function-based
algorithm for robust autonomous freeway driving,” in Proc. IEEE
Intelligent Vehicles Symposium (IV), 2010, pp. 512–517.
[5] D. Kasper, G. Weidl, T .Dang, et al., “Object-Oriented Bayesian networks
for detection of lane change maneuvers,” Intelligent Transportation
Systems Magazine, Vol. 4, No. 1, 2014, pp.19-31.
[6] D. Kasper, G. Weidl, T .Dang, et al., “Object-Oriented Bayesian networks
for detection of lane change maneuvers,” in Proc. IEEE Intelligent
Vehicles Symposium (IV), 2010, pp.673-678
TABLE I. COMPUTER AND HUMANI DRIVER ANALYSIS FOR LANE CHANGE
DATA
Case
Operation Time - (s)
Smoothness-1
Lateral Jerk
Human-
driver
Compu-
ter
Human-
driver
Compu-
ter
Human-d
river
Compu-
ter
1
4.92
5.075
1.23E-07
1.19E-07
0.746
0.761
2
7.135
7.175
1.12E-06
7.35E-07
0.810
0.802
3
4.195
4.06
8.35E-07
6.12E-07
1.280
1.145
4
4.265
4.22
4.00E-07
2.93E-07
2.275
2.102
5
4.15
4.17
7.34E-07
4.91E-07
0.432
0.303
6
3.66
3.705
3.10E-07
1.89E-07
1.306
1.022
7
3.02
3.075
9.70E-08
6.48E-08
0.229
0.199
8
3.995
4.04
3.19E-07
2.02E-07
1.049
0.863
9
5.605
5.6
8.94E-07
6.69E-07
1.859
1.463
10
4.285
4.315
3.94E-08
2.09E-08
0.351
0.205
11
3.195
3.18
2.37E-07
1.18E-07
0.334
0.242
12
3.025
3.02
3.09E-06
3.14E-06
6.611
8.112
13
4.83
4.965
3.22E-07
6.32E-08
0.573
0.130
14
3.85
4.1
1.46E-06
1.04E-06
4.226
3.823
15
4.585
4.665
2.62E-06
3.01E-06
10.156
12.169
16
5.155
4.96
1.23E-07
8.44E-08
0.760
0.597
17
4.335
4.545
1.79E-07
4.50E-08
0.473
0.205
18
4.425
4.375
1.01E-07
3.61E-08
0.475
0.186
Fig. 19. Human-driver and computer curvature over time for lane
change- blue curve shows smoothed IMU data from driver, red curve shows
the computer
Time (s)
Lane Change Curvature
Human Driver
Computer
Lane Change Operation
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
11
[7] R. Schubert, K. Schulze, and G. Wanielik. (2010). Situation assessment
for automated lane-change maneuvers. IEEE Transactions on Intelligent
Transportation Systems, vol. 11, no. 3, pp. 607-616.
[8] S. Sivaraman and M. M. Trivedi, "Dynamic probabilistic
drivability maps for lane change and merge driver
assistance,” IEEE Transactions on Intelligent
Transportation Systems, vol.15, pp. 2063– 2073, 2014.
[9] R.S. Tomar, S. Verma, G.S. Tomar, “Prediction of lane change
trajectories through neural network,” in Proc of International Conference
on Computational Intelligence and Communication Networks, 2010, pp.
249 – 253.
[10] A. Polychronopoulos, M. Tsogas, A. Amditis, et al, ”Dynamic situation
and threat assessment for collision warning systems: the euclide
approach,” in Proc. IEEE Intelligent Vehicles Symposium (IV), 2004, pp.
636-641.
[11] T. Gindele, S. Brechtel, R. Dillmann, “A probabilistic model for
estimating driver behaviors and vehicle trajectories in traffic
environments,” in Proc International IEEE Conference on Intelligent
Transportation Systems, 2010, pp.1625-1631.
[12] J. Schlechtriemen, A. Wedel, J. Hillenbrand, G. Breuel, K. Kuhnert,” A
lane change detection approach using feature ranking with maximized
predictive power”, in Proc. IEEE Intelligent Vehicles Symposium (IV),
2014, pp.108-114.
[13] J. Schlechtriemen, F. Wirthmueller, A. Wedel, G. Breuel, K. D. Kuhnert,”
When will it change the lane? a probabilistic regression approach for
rarely occurring events,” in Proc. IEEE Intelligent Vehicles Symposium
(IV), 2015,pp.1373-1379.
[14] P. Kumar, M. Perrollaz, S. Lef`evre, and C. Laugier,” Learning-based
approach for online lane change intention prediction,” in Proc. IEEE
Intelligent Vehicles Symposium (IV), 2013, pp.797-802.
[15] E. Naranjo, C. Gonzalez, R. Garcia, and T. de Pedro, (2008). Lane-change
fuzzy control in autonomous vehicles for the overtaking maneuver. IEEE
Transactions on Intelligent Transportation Systems, vol. 9, no. 3, pp.
438-450.
[16] M. Bahram, A. Wolf, M. Aeberhard and D. Wollherr,” A
prediction-based reactive driving strategy for highly automated driving
function on freeways ,” in Proc. IEEE Intelligent Vehicles Symposium
(IV), 2014, pp.400-406.
[17] J. Nilsson and J. Sj¨oberg,” Strategic decision making for automated
driving on two-lane, one way roads using model predictive control,” in
Proc. IEEE Intelligent Vehicles Symposium (IV), 2013,pp.1253-1258.
[18] Y. Du, Y. Wang and C. Chan,” Autonomous lane-change controller via
mixed logical dynamical”, in Proc . 17th International Conference on
Intelligent Transportation Systems, 2014, pp 1154 – 1159.
[19] Y. Du, Y. Wang and C. Chan,” Autonomous lane-change controller”, in
Proc. IEEE Intelligent Vehicles Symposium (IV), 2015, pp.386-393.
[20] U. Simon, and M. Markus, “Probabilistic online POMDP decision
making for lane changes in fully automated driving,” in Proc of IEEE
Conference on Intelligent Transportation Systems, 2013, pp.2063-2070.
[21] S. Brechtel, T. Gindele, and R. Dillmann, “Probabilistic MDP-behavior
planning for cars,” in Proc of 14th International IEEE Conference on
Intelligent Transportation Systems, 2011, pp.1537-1542.
[22] M. Ardelt, C. Coester, and N. Kaempchen. (2012). Highly automated
driving on freeways in real traffic using a probabilistic framework. IEEE
Transactions on Intelligent Transportation Systems, vol. 13, no. 4, pp.
1576-1585.
[23] Q. Jin, G. Wu, K. Boriboonsomsin, and M. Barth,” Improving traffic
operations using real-time optimal lane selection with connected vehicle
technology,” in Proc. IEEE Intelligent Vehicles Symposium (IV),
2014,pp.70-75.
[24] H. Jula, E. Kosmatopoulos and P. Ioannou. (2000). Collision avoidance
analysis for lane changing and merging. IEEE Transactions on Vehicle
Technology, vol. 49, pp.2295 -2308.
[25] C. Rodemerk, S. Habenicht, A. Weitzel, H. Winner and T.Schmitt,”
Development of a general criticality criterion for the risk estimation of
driving situations and its application to a maneuver-based lane change
assistance system,” in Proc. IEEE Intelligent Vehicles Symposium (IV),
2012, pp.264-269.
[26] G.Schildbach, F. Borrelli, “Scenario Model Predictive Control for Lane
Change Assistance on Highways”, in Proc of IEEE Intelligent Vehicles
Symposium, 2015, pp.611-616.
[27] J. Ziegler, P. Bender, T. Dang and C. Stiller, “Motion planning for Bertha
- a local, continuous method,” in Proc of Intelligent Vehicles Symposium,
Dearborn, Michigan, USA, 2014, pp. 450-457.
[28] Q. H. Do, L. Han, H. Tehrani and S. Mita, “Safe path planning among
multi obstacles,” in Proc of IEEE Intelligent Vehicles Symposium, 2011,
pp. 332 – 338.
[29] H. Tehrani, K. Muto, K. Yoneda and S. Mita, “Evaluating human &
computer for expressway lane changing,” in Proc of IEEE Intelligent
Vehicles Symposium, 2014, pp. 382 – 387.
[30] Q. H. Do, H. Tehrani, M. Egawa, K. Muto, K. Yoneda, and S. Mita,
“Distance constraint model for automated lane change to merge or exit,”
in Proc of the 3rd International Symposium on Future Active Safety
Technology Towards zero traffic accidents, 2015, pp. 17-24.
[31] California Driver Handbook-Safe Driving Practices: Merging and
Passing, California Department of Motor Vehicles, Sacramento, CA,
USA,2012.
[32] W. Yao, H. Zhao, F. Davoine, H. Zha,” Learning lane change trajectories
from on-road driving data,” in Proc. IEEE Intelligent Vehicles
Symposium (IV), 2012, pp.885-890.
[33] M. Werling, J. Zeigler, S. Kammel, S. Thrun, “Optimal trajectory
generation for dynamic street scenarios in a Frenet Frame,” in Proc of
IEEE International Conference on Robotic and Automation, 2010, pp.
987-993.
[34] D. Althoff, M. Buss, A. Lawitzky, M. Werling, D. Wollherr “On-line
trajectory generation for safe and optimal vehicle motion planning,” AMS
2012, pp. 99-107.
[35] T. Fraichard and H. Asama. (2004). Inevitable Collision States. A step
towards safer robots?. Advanced Robotics, vol. 18, pp. 1001–1024.
[36] H. Tehrani, Q. H. Do, M. Egawa, K. Muto, K. Yoneda, and S. Mita,
“General behavior and motion model for automated lane change," in Proc
of IEEE Intelligent Vehicles Symposium, 2015, pp. 1154-1159.
[37] L.Peng, A. Kurt, U. zgner ,“Trajectory prediction of lane changing
vehicle based on driver behavior estimation and classification,” in Proc of
Intelligent Transportation Systems, 2014 , pp. 942 – 947.
[38] A A. Geiger, M. Lauer, F. Moosmann, B. Ranf et al. (2012). Team
AnnieWAY's Entry to the 2011 Grand Cooperative Driving Challenge.
IEEE Transactions on Intelligent Transportation Systems, vol 13, no 3,
pp. 1008 – 1017.
[39] Y. Kanayama, G. R. De Haan,” Least Cost Paths With Algebraic Cost
Functions”, in Proc. IEEE International Workshop on Intelligent Robots,
1988, pp.341-346.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
12
Quoc Huy Do received his B.S. and M.S. degrees in
Information Technology from Hanoi University of
Technology, Vietnam, in 2006 and 2008, respectively.
He served on the faculty of Hanoi University of
Science and Technology in the academic years of
2006–2009. He received his Ph.D in information and
systems from Toyota Technological Institute (TTI) in
2013. Currently, he is a post-doctoral fellow at
Research Center for Smart Vehicles of TTI. His
research field is path planning, vehicle dynamic and
automated lane change for autonomous vehicles.
Hossein Tehrani Nik Nejad received his B.S. and
M.S. degrees in industrial engineering from Sharif
University of Technology, Iran, in 1996 and 1998,
respectively. He received his Ph. D degree in
mechanical engineering from Osaka Prefecture
University, Japan, in 2009. In 2009, he joined the
Toyota Technological Institute at Japan as
Post-doctoral Fellow in the Smart Vehicle Center. In
2012, he joined the DENSO CORPORATION. Dr.
Tehrani is an IEEE member and also a technical
committee member of ITS-Nagoya chapter. His
research interests are automated driving, sensor
processing, machine learning and deep neural
networks.
Seiichi Mita received his B.S., M.S., and Ph.D.
degrees in electrical engineering from Kyoto
University in 1969, 1971, and 1989, respectively. He
studied at Hitachi Central Research Laboratory,
Kokubunji, Japan, from 1971 to 1991, investigating
signal processing and coding methods. Currently, he
is a professor at Toyota Technological Institute (TTI)
in Nagoya (since 1999) and a director of the Research
Center for Smart Vehicles at TTI. Currently, he is
greatly interested in the research area of autonomous
vehicles and sensing systems. Dr. Mita is a member
of the Institute of Electronics, Information and
Communication Engineers and the Institute of Image
Information and Television Engineers in Japan. He is
also a member of IEEE. He received the best paper
award from the IEEE Consumer Electronics Society
in 1986 and the best paper and author awards from
the Institute of Television Engineers in Japan in 1987
and 1992, respectively.
Kenji Muto received the M.S. degree in
Electronic-Mechanical Engineering from Nagoya
University in 1997. He currently works for research
and development department in DENSO
CORPORATION. His research interests have been in
vehicular sensing and communication system and its
applications.
Masumi Egawa received his the B.E. and M.S.
degrees from Nagoya Institute of Technology, in
Nagoya, Japan, in 1996 and 1998 respectively. He
currently works for DENSO CORPORATION.
His research interests have been in information
security, wireless communication, and applications
of machine learning methods
Keisuke Yoneda received his B.S. degrees in
engineering from Toyohashi University of
Technology in 2007, and M.S., and Ph.D. degrees in
information science from Hokkaido University in
2009 and 2012, respectively. He is currently work as
Assistant Professor at Autonomous Vehicle Research
Unit of Institute for Frontier Science Initiative,
Kanazawa University. He is interested in the research
area of autonomous vehicles, artificial intelligence
and artificial life. He is a member of the Japan
Society for Precision Engineering and the Society of
Automotive Engineers of Japan, Inc
