Design of Forward Collision Warning System using Relative ...

Design of Forward Collision Warning System

using Relative Acceleration Estimation and

Multi-Object Tracking

Bo-Chiuan ChenAssociate Professor, Department of Vehicle Engineering

Director, Vehicle Technology Research Center

National Taipei University of Technology, Taipei, Taiwan

• Education– Ph.D., Mechanical Engineering, U. of Michigan

– M.S., Electrical Engineering, U. Michigan

– M.S., Mechanical Engineering, U. Michigan

– B.S., Naval Architecture and Ocean Engineering, NTU

• Experience– Assistant Professor, Vehicle Eng. Dept., NTUT

– Strategy and Software Engineer, Powertrain Control Dept., Visteon, USA

– Research Assistant, Automotive Research Center, U. of Michigan

• Awards– Outstanding Young Engineer Award, SAE Taipei Section, 2008.

– Excellent Paper Awards, National Conference on Vehicle Engineering, Taiwan, 2008, 2009, 2010.

– Top 5 in Super Mileage Competition, SAE Taipei Section, 2004-2010.

• Research Areas– Active Safety, Vehicle Dynamics and Control, Hybrid Electric Vehicle,

Engine Control, Optimum Control

Self Introduction

Forward Collision Warning System (FCWS) 2

Research Areas

• Active Safety– Electronic Stability Control

– Rollover Prevention

– Forward Collision Warning

– Lane Departure Warning

– Side Collision Warning

– Auto Parking

• Vehicle Dynamic and Control– Light Weight Electric Vehicle

– Electric Differential

– Electric Power Steering

– ABS/TCS

– Semi-active Suspension

• Hybrid Electric Vehicle– Hybrid Electric Scooter

– Power Management System

• Engine Control– Idle Speed Control

– Engine Management System

4 3 2 1


Statistics

• In U.S., the percentage of rear end collision in all collisions

was about 31.5% in 2009.

Source: Traffic Safety Facts 2009 Forward Collision Warning System (FCWS) 4

Forward Collision Warning System (FCWS)

Human Error

• 80% of drivers attempted no action in rear end collisions.

Source: SAE Paper 1999-01-0817 Source: AVEC 9437953

5

Existing Technologies

Audi

Ford

Honda

Volvo

BMW

Mercedes-Benz

VW

Toyota

Infinity


Object Detection Systems

• Infrared Laser

– Transmit energy in the THz range (1 THz=1012 Hz).

– Superior angular resolution.

– Limited performance due to atmospheric effects, such as fog and

rain.

– Does not perform well on wet objects or targets whose surface

roughness is the order of the laser wavelength (10.6 microns).

• Microwave/Millimeter wave radar

– Transmit energy in the tens of GHz range (1 GHz = 109 Hz)

– Better adverse weather penetration than active laser systems.

• Camera

– Usable distance accuracy for short-range detection (less than 55 m)

– Poor accuracy for long-range detection due to the pixel resolution


Forward Collision Warning System

• Autonomous Solutions:

– Identify valid target and measure range, range rate, and vehicle

speed (10Hz or faster).

– Vehicle path prediction

– Issue warning based on

• Time-to-Collision

• Time-headway (time gap)

• Threshold distance

From vehicle speed, road friction,

and human delays such as “blank

time” and “judgment time,” a “safe

following distance” d can be constructed.

RTTC

R

host

RTHW

V

Source: SAE 98PC-417


Relative Acceleration Estimation

• Lee and Peng (2005) mentioned that the leading vehicle

acceleration is a critical step for developing practical

collision warning/avoidance systems.

• Good estimation of relative acceleration is the key to reduce

the false alarm of FCWS.

– Dagan et al. (2004) calculated TTC from the momentary TTC

defined by Hayward and its derivative, which is closely related to

relative acceleration and can be computed by the scale change in

the image.

– Araki et al. (1996) applied a 3-state Kalman filter to estimate relative

velocity and acceleration.

2

ˆˆ ˆˆ ˆ ˆ ˆ( 1) ( ) ( ( ) ( )), ( )

1 0.5 0

0 1 , 0 , 1 0 0

0 0 1 1

T

k k y k k k R V A

T T

T

x Ax L Cx x

A G C


Multi-object Tracking

• Moon et al. (2009) proposed primary target selection

– A yaw-rate based subject vehicle’s lane detection, a motion based

analysis, and an integration process.

– Primary target might be changed quite often during transient yaw

motion.


Proposed Approach

• Recursive Least Squares

R̂ t

t kT

true TTC

estimated TTC

0

R t

rela

tive

dis

tan

ce (

m)

time (sec)

2 ˆˆ T

R R R RR k a k t b k t c k k k φ θ

2 ˆˆ T

D D D DD k a k t b k t c k k k φ θ

ˆ T

R R R Ra k b k c k θ

ˆ T

D D D Da k b k c k θ

2 1 ,T

k t t φ t kT

1

1 1Tk k k k k k

K P φ I φ P φ

ˆ ˆ ˆ1 1Tk k k y k k k θ θ K φ θ

1Tk k k k P I K φ P

1

0 0

3 3

ˆ 0 T

k k

k k k y k

θ φ φ φ

0

X

Y

(0,0)

ˆ 1R k

ˆ 1D k

ˆpredictR k

ˆpredictD k

( )R k

D k

r

overflow when t is large

reset to zero during safe maneuvers


Proposed Approach

• Variable forgetting factor

– Large forgetting factor is suitable for small relative acceleration.

However, the estimation performance deteriorates with large relative

accelerations.

– Small forgetting factor is suitable for large relative accelerations.

However, it might produce noisier estimations for small relative

acceleration.

– Adjust the forgetting factor according to the estimated relative

acceleration, i.e. variable forgetting factor, might be a good solution.

ˆR R Rd m A ˆ

D D Dd m

2 ˆˆ ˆˆR R Ra b cR k k t k t k

ˆˆ ˆ2 R RV k a k t b

ˆ ˆ2 RA k a k

2 ˆˆ ˆ ˆ2

ˆ

V k V k A k R kTTC k

A k


Optimization

• Kalman Smoothing

– the forward filtered data contains undesirable time delays,

– a backward Kalman filter is constructed to cancel its effect

– The averaged data is then used as the ground truth.

– The same smoothing procedure is applied to the relative orientation.

( ) ( ) ( )

( ) ( ) ( )

( ) ( ) ( ) ( ) ( ) ( )

0 1 0 0 0 0 0

0 0 0 0 0 1 01 0 0 0 0 1 0

, , , 1 0 0 1 0 0 00 0 1 0 0 0 1

0 0 0 0 1 0 0

0 0 0 0 0 0 1

KF f KF

KF

T

KF F F L L

f

t t t

t t t

t V t a t R t V t a t

x A x Gw

y Cx Nv

x

A G C N

ˆ ˆ ˆ( 1) ( ) [ ( ) ( )]KF df KF s KFk k k k x A x L y Cx

ˆ ˆ ˆ( 1) ( ) [ ( ) ( )],db sk k k k z A z L y Cz 1[ ]T T

db df df df

A A A A


Proposed Approach

• Multi-object Tracking

F

(k)

F

(k-1)

H

min

max

0

H

min

F

(k)

0

F

(k-1)

max

min max

if <

if and

possible collision, warning on

else

safe, warning off

else

safe, warning off

end

th

th

TTC TTC

count count 　

v

r

4secthTTC 3.5secthTTC

2thcount 3thcount

possible collision safe

1 ˆ ˆ ˆˆtan ,R V D


Experimental Setup

CAN analyzer

VBOX IMU

ibeo LUX laserscanner

Steering angle sensor

Mototron

Data

acquisition

USB

cameraData logging

Analog

CAN Bus

USB


ISO 15623

Longitudinal target discrimination ability test

Straight road lateral target discrimination ability test

Curved test track and target

discrimination ability test


ISO Test -1


2 4 6 8 10 12 14 16 18 20 220

10

20

30

40

rela

tive

dis

tan

ce

(m

)

2 4 6 8 10 12 14 16 18 20 22

-5

0

5

rela

tive

ve

l. (

m/s

)

2 4 6 8 10 12 14 16 18 20 22-5

0

5

time (s)

rela

tive

acce

l. (

m/s

2)

3 state kalman

RLS (variable )

Kalman smooth

ISO Test -1

2 4 6 8 10 12 14 16 18 20 22250

260

270

280

290

an

gle

(d

eg

)

relative velocity vector

2 4 6 8 10 12 14 16 18 20 220

1

2

3

4

time (s)T

TC

(s)

conventional

3 state kalman

RLS(variable )

Kalman smooth

warning

8.5s 11s 17s 18s 8.5s 11s 17s 18s


ISO Test - 2


ISO Test - 2

2 4 6 8 10 12 14 16 18 20 220

10

20

30

40

50

rela

tive

dis

tan

ce

(m

)

2 4 6 8 10 12 14 16 18 20 22

-15

-10

-5

0

rela

tive

ve

l. (

m/s

)

2 4 6 8 10 12 14 16 18 20 22

-5

0

5

10

time (s)

rela

tive

acce

l. (

m/s

2)

3 state kalman

RLS (variable )

Kalman smooth

2 4 6 8 10 12 14 16 18 20 22250

260

270

280

290

relative velocity vector (main lane)

an

gle

(d

eg

)2 4 6 8 10 12 14 16 18 20 22

260

270

280

290

300

310

relative velocity vector (adjacent lane)

an

gle

(d

eg

)

2 4 6 8 10 12 14 16 18 20 220

1

2

3

4

time (s)

TT

C (

s)

conventional

3 state kalman

RLS (variable )

Kalman smooth

9s 12s 15s 16.5s 9s 12s 15s 16.5s

adjacent

adjacent

adjacent adjacent


ISO Test - 3


ISO Test - 3

2 4 6 8 10 12 14 16 18 20 22

10

20

30

40

50

rela

tive

dis

tan

ce

(m

)

2 4 6 8 10 12 14 16 18 20 22

-15

-10

-5

0

rela

tive

ve

l. (

m/s

)

2 4 6 8 10 12 14 16 18 20 22

-10

-5

0

5

time (s)

rela

tive

acce

l. (

m/s

2)

3 state kalman

RLS (variable )

Kalman smooth

2 4 6 8 10 12 14 16 18 20 22250

260

270

280

290

relative velocity vector (main lane)

an

gle

(d

eg

)2 4 6 8 10 12 14 16 18 20 22

260

270

280

290

300

310

relative velocity vector (adjacent lane)

an

gle

(d

eg

)

2 4 6 8 10 12 14 16 18 20 220

1

2

3

4

time (s)

TT

C (

s)

conventional

3 state kalman

RLS (variable )

Kalman smooth

6s 9.5s 13s 15.5s 6s 9.5s 13s 15.5s

adjacent

adjacent

adjacentadjacent


Expressway Tests

Merging TrafficLane Change

Cut-in Aggressive Lane Change


Conclusions

• Relative acceleration is considered to improve the accuracy

of TTC estimation.

– Recursive least square technique with variable forgetting factor is

used to estimate coefficients of two second order polynomials for the

relative distance and relative orientation, respectively.

• The region of interest for FCWS is extended from the main

lane to adjacent lanes.

– According to the measured relative distance and relative orientation,

a multi-object tracking algorithm is developed in this research.

– When the TTC is below the threshold value, relative velocity vector

is used to determine if there is an impending threat for collision.

• Experimental results show that the proposed algorithm

– can pass all 3 tests of ISO 15623

– and issue valid warnings to the driver without false alarms for the

expressway tests.


Future Works

• Automatic evasive maneuver for collision avoidance

Source: Continental Emergency Steer Assist 2010Forward Collision Warning System (FCWS) 25

Q & A


Date post:	30-Dec-2021
Category:	Documents
Upload:	others
View:	7 times
Download:	0 times

Design of Forward Collision Warning System using Relative ...

Documents