Transcript Slide 1

May 3, 2011
Safety-Related Developments in Advanced Driver
Assistance
Environmental Perception & Cooperative Driving
Jeroen Ploeg
TNO Technical Sciences
May 3, 2011
2
Outline
Introduction
Trends in Advanced Driver Assistance
Collision mitigation & avoidance
Probabilistic Risk Estimation for Vulnerable Road Users
Cooperative Driving
Cooperative Adaptive Cruise Control
Conclusions
May 3, 2011
3
Societal Trends
A growing need for mobility, individuality, freedom
Expected growth in mobility 20 – 30% in the coming decade (in the
Netherlands)
Consequence 1: current safety level will be hard to maintain
Consequence 2: “vehicle loss hours” will significantly increase
Consequence 3: emission level/fuel consumption increases
Significantly more road space in the living environment is not
acceptable
Advanced Driver Assistance (ADA) offers possibilities ...
May 3, 2011
4
Advanced Driver Assistance (ADA)
Advanced Driver Assistance (ADA) systems
“systems that support the driver in his driving task, primarily based
on information regarding the local traffic situation”
Vehicle dynamics systems excluded, such as
ABS
ESP
...
May 3, 2011
5
ADA trends
controllability: autonomous driving
Cooperative Driving results in2:
50% less traffic congestion
8% less traffic accidents
1
VRU safety 5%
: less CO2 emission
43% road fatalities are
pedestrians
5% cyclists
mobility: cooperative driving
safety: collision warning  mitigation 
avoidance
comfort: cruise control, advanced cruise control
1
2TETSC
PIN 2008-D-R0996/A:
annual report 2009
TNO report
“Smarter
and better – the benefits of intelligent traffic”
May 3, 2011
6
ADA trend 1: Vulnerable Road Users
Number of road fatalities in the Netherlands
1000
Other
Motorcycle
900
Pedestrian
800
mopeds
700
Bicycle
vans/truck
600
Car
500
400
300
200
Total: decreasing 100
number of fatalities
VRUs: 318 fatalities0 out of 750 total in 2008 (> 42%)
2004
2005
2006
2007
2008
Year
Rijkswaterstaat, Kerncijfers Verkeersveiligheid, www.rijkswaterstaat.nl/dvs, 2009
May 3, 2011
7
ADA trend 2: Cooperative Driving
Cooperative Driving
Influencing the individual vehicles, either through advisory or
autonomous actions, so as to optimize the collective behavior with
respect to:
Safety (also affects throughput)
Throughput (highways + urban roads)
Emission/fuel consumption (trucks)
Main enabler: wireless communications
May 3, 2011
8
Risk estimation for VRUs
Injury reduction
Driver warning
Autonomous braking
Airbag deployment to reduce impact
Environmental perception for VRU
Sensor
Sensor
Image
processing
Sensor fusion
Clustering,
assignment
Sensor
fusion
Risk estimation
Object
identification
Driver
warning
Autonomous
Risk
action
estimation
Airbag
deployment
eCall
Assessment (driving & pre-crash & crash)
May 3, 2011
9
Vehicle risk estimation
Predict trajectories of detected objects (vehicles) and host vehicle
across a certain time-horizon
Quantify optimal trajectory by use of a cost function
Collision-free -> keep safe distance to obstructing objects
Feasible trajectory -> low accelerations
Minimize the cost function by choosing optimal host vehicle trajectory
May 3, 2011
10
Vehicle risk estimation (cnt’d)
no collision
object
safety
distance
predicted
object
trajectory
host
predicted host
trajectory
minimum
distance
collision
May 3, 2011
11
Risk estimation for VRUs – main principle
VRUs behave rather non-deterministic
Probabilistic approach is proposed to cover the resulting uncertainty
in the path prediction of VRUs
Probability Density Functions (PDF)
P
P
Probabilistic
Risk
Estimation
v0
0
Assumption: object classification is known to choose the correct PDF
May 3, 2011
12
Implementation
1. At time t = t0 the position, orientation, velocity, (rotational velocity,
acceleration) are known of the detected object(s) and own vehicle
2. Determine position probability of object over a certain time horizon
3. Determine maximum `overlap’ of host and object probability density
function  collision probability
y
x
May 3, 2011
13
Simulation
x0  18 m
y 0  9 m
 0  0.65 rad
v0  5 m/s
a0  0 m/s2
k0  0 m
x0  0 m
y0  0 m
 0  0 rad
Here, normal distributions are chosen for:
v0  5 m/s
a0  0 m/s
2
k0  0 m
forward velocity
heading
May 3, 2011
14
MIO index
Time-to-collision [s]
Collision probability [%]
Experiments
Time [s]
MIO index 
collision probabilit y
time - to - collision
May 3, 2011
15
Summary risk estimation for VRUs
Probabilistic Risk Estimation (PRE) provides an estimation of the
collision probability in the presence of large uncertainties with respect
to future object behavior (such as with VRUs)
Modular, generic approach, serving multiple ADAS applications
object detection & classification
prediction & risk estimation
Liability issue: will the driver remain responsible?
May 3, 2011
16
Cooperative Driving
Two types of systems, roughly
Warning/advisory systems  not time-critical  event-triggered
Automatic systems  time-critical  time-triggered  real-time
closed loop control
 Cooperative Adaptive Cruise Control (CACC)
Basis: Adaptive Cruise Control (ACC)
+ wireless communication
Vehicle-following control objective
Increase safety by automatically
“smoothing” traffic through
shockwave mitigation
May 3, 2011
19
String stability – human driving behavior
Sugiyama, Y.; Fukui, M.; Kikuchi, M.; Hasebe, K.; Nakayama, A.; Nishinari, K.; Tadaki, S.
& Yukawa, S., Traffic Jams without Bottlenecks - Experimental Evidence for the Physical
Mechanism of the Formation of a Jam. New Journal of Physics, 2008, 10 (033001), 7
May 3, 2011
20
String stability – ACC
Infinite string
ACC, with time headway
h = 0.5 s
Initial velocity 72 km/h
Initial condition error of
one vehicle of 2 m
String unstable 
with linear controller,
a collision occurs
May 3, 2011
21
String stability
Take 2nd-order systems in series connection
yˆ i ( s )
1

, 1  i  10
1

yˆ i 1 ( s)
2
s
2
s 1
2
n
n
yˆ 0 ( s)  uˆ ( s )
i ( s) 
 = 1.1
 = 0.73
 = 0.5
May 3, 2011
22
String stability – conditions
Define String Stability Complementary Sensitivity i(s) such that
yˆ i ( s)  i ( s) yˆ i 1 ( s)
with inverse Laplace transform i(t) (impulse response function)
Then, from linear system theory
yi (t )
L2
 i ( j)
H
yi 1 (t )
L2
known as the L2 gain, and
yi (t )
i.e., the L gain.
L
  i (t )
L1
yi 1 (t )
L
May 3, 2011
23
String stability – conditions (cnt’d)
Hence, in order to have disturbance attenuation in upstream direction,
we require
i ( j)
H
 1, 2  i  m
i.e., L2 string stability, or
 i (t )
i.e., L string stability
L1
 1, 2  i  m
May 3, 2011
24
String stability – conditions (cnt’d)
2nd-order systems in series connection
string stable
L2 string stable
L string unstable
string unstable
May 3, 2011
25
CACC design – communication topologies
Ad-hoc platooning: no designated platoon leader
Realistic solution for everyday traffic
Least demanding for communication
Unidirectional communication with directly preceding vehicle
May 3, 2011
26
CACC design – spacing policy
Spacing policy
d r ,i  r  h  vi
2im
h: time headway [s]
r: standstill distance [m]
Spacing policy improves string stability properties!
Controller acts on vehicle acceleration to realize the desired spacing
May 3, 2011
27
CACC design – controller
Spacing policy transfer function: H (s) 1  hs
Vehicle model G(s), communications time delay D(s), controller K(s)
May 3, 2011
28
CACC design – string stability
String stability complementary sensitivity
( s ) 
vˆi ( s)
1 D( s )  G ( s ) K ( s )


vˆi 1 ( s) H ( s) 1  G( s) K ( s)
Hence, without communications delay
( s ) 
1
1

H ( s) hs  1
Consequently, L2 string stable (L string stable as well).
May 3, 2011
29
CACC design – simulation results
Without communication (h = 0.5 s)
With communication (h = 0.5 s)
May 3, 2011
30
CACC design – simulation results (cnt’d)
“platoon” of 8 vehicles, 1st vehicle introduces speed variations
May 3, 2011
31
CACC experiments
Test fleet: 7x Toyota Prius, equipped with
Wireless communications (IEEE 802.11g)
GPS
CACC control computer
Low-level vehicle control
computer (interacts with
the vehicle CAN bus to
automatically accelerate/
decelerate)
May 3, 2011
32
CACC experiments (cnt’d)
Test fleet: 7x Toyota Prius (no. 7 is missing :-)
May 3, 2011
33
CACC experiments (cnt’d)
Lelystad, March 18, 2011
May 3, 2011
34
CACC experiments (cnt’d)
Velocity responses of test fleet
ACC (i.e., no WiFi)
CACC
May 3, 2011
35
CACC – object tracking
Objective
Determine relevant target vehicles based on multiple sensors, s.a.
wireless comm. (802.11p) and radar
Or, in other words:
match radar data with WiFi data
fuse data to get reliable object motion data
Packet-loss & inaccuracy of communicated GPS data biggest
challenge
May 3, 2011
37
CACC – object tracking (cnt’d)
Preprocessing
Feature
Filtering
Data
Clustering
Object
State
Estimation
Object
Classification
n = 6of
Kalman
groups
• “Raw”
• Coordinate
measurement
• Define
transformations
Region
data
• Cluster
Interest
data
to offilter
• different
(ROI)
Application
• Relevant
forsources
objects
specific object classification
• according
Each
group
m Kalman
filters
• make
Radardatahost
comparable
vehicle,
based
on:
tocontains
(expected)
• Motion
• Most
objects
data
Important
“as
good
Object(s)
as possible”
(MIO)
• Basic
• Range,
data acceptance/rejection
•bearing,
Host•vehicle
Each
Method:
rangeKalman
motion
rate
Quality
(kinematic)
•filter
Bidirectional
•Threshold
Not
suits
good
a specific
clustering
enough:
CACCdata
graceful
combination
degradation
•• Ignore
Relative
objects
• to
Application
hostdriving
vehicle
• •Based
WiFi,
specific
in radar,
on distance
WiFi
••Forward
to
+ be
radar
judged
MIO on controller level
m
=ROI
3 to Kalman
• Wireless
opposite
• Communication
Reject
direction
data
• Assign
outside
clusters
••Backward
to be filter
implemented
MIO
objects on controller level
Total
n·m = 18
filters
• Set-up
• Position,
objectvelocity,
data• matrix
•acceleration
Activate,
reset,
• Application
de-activatespecific!
filters
• max. n objects (n
= 6, currently)
• Absolute coordinates
• Reliability
measure
• Estimation error covariance per object
May 3, 2011
38
CACC – object tracking (cnt’d)
Results
Simulated scenario
6 vehicles
Wireless comm. + radar
Forward MIO & backward MIO
tracking
May 3, 2011
39
CACC – object tracking (cnt’d)
Results (cnt’d)
Measurements
Real-time implementation
Prius radar measurements
No wireless comm. yet
Forward MIO tracking
May 3, 2011
40
CACC – object tracking (cnt’d)
Results (cnt’d)
Measurements
Real-time
implementation
Prius radar
measurements
No wireless
comm. yet
Forward MIO
tracking
May 3, 2011
41
CACC – object tracking (cnt’d)
Results (cnt’d)
Measurements
Real-time
implementation
Prius radar
measurements
No wireless
comm. yet
Forward MIO
tracking
May 3, 2011
42
Summary CACC
CACC enables automatic smoothing of traffic through enforcement of
string stable behavior  increases safety by decreasing the number
of potentially dangerous events
Design focusing on implementation is feasible
CACC can be regarded as add-on to ACC
Standardization in wireless communications well under way (IEEE
802.11p, ETSI Geo-routing & message content)
Object tracking is a generic component (also used in VRU safety)
May 3, 2011
43
Conclusions
Advanced Driver Assistance:
Increased focus on VRU safety
Increased focus on Cooperative Driving (wireless communications)
Both types, although very different by nature, rely to a large extend on
detection, estimation & classification of road users
Both types are time- & safety-critical and even automatic
Changing the role of the driver from “real-time controller” to
“supervisor” opens up a whole new perspective with respect to
improving traffic safety.