Automated Proactive Road Safety Analysis - Freen.saunier.free.fr/saunier/stock/10-11-25-tram.pdf ·...

Automated Proactive Road Safety AnalysisTransportation Research At McGill Seminar

Nicolas [email protected]

November 25th 2010

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation

ProbabilisticFramework

ExperimentalResults

InvestigatingCollisionFactors

Conclusion

Outline

1 Motivation

2 Probabilistic Framework for Automated Road SafetyAnalysis

3 Experimental Results using Video Data

4 Investigating Collision Factors Using Microscopic Data

5 Conclusion

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

A World Health Issue

Over 1.2 million people die each year on theworld’s roads, and between 20 and 50 million suffernon-fatal injuries. In most regions of the world thisepidemic of road traffic injuries is still increasing.(Global status report on road safety, World HealthOrganization, 2009)

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

A World Health Issue

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Road Safety Analysis

• Limits of the traditional approach based on historicalcollision data:

• problems of availability and quality• insufficient data to understand the processes that lead

to collisions• reactive approach• pedestrians: issues are made more accute by the rarity

of collisions and the lack of data (exposure)

• Need for proactive approaches and surrogate safetymeasures that do not depend on the occurrence ofcollisions

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Surrogate Safety Measures

• Research on surrogate safety measures that• bring complementary information• are related to traffic events that are more frequent than

collisions and can be observed in the field• are correlated to collisions, logically and statistically

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Traffic Conflicts

A traffic conflict is “an observational situation inwhich two or more road users approach each otherin space and time to such an extent that a collisionis imminent if their movements remain unchanged”[Amundsen and Hyden, 1977]

• Traffic Conflict Techniques• Limits caused by the data collection process (human

observers in the field)• cost• intra- and inter-observer variability

• Mixed validation results

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

The Safety/Severity Hierarchy

FI

PD

Undisturbed passages

Potential ConflictsSlight Conflicts

Serious ConflictsAccidents

Various severity measures

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

MotivationNeed for automated tools to address the shortcomings ofreactive diagnosis methods and traffic conflict techniques

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

The Collision Course

A traffic conflict is “an observational situation inwhich two or more road users approach each otherin space and time to such an extent that a collisionis imminent if their movements remain unchanged”[Amundsen and Hyden, 1977]

The extrapolation hypotheses must be specified.

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Rethinking the Collision Course

• For two interacting road users, many chains of eventsmay lead to a collision

• It is possible to estimate the probability of collision ifone can predict the road users’ future positions

• learn road users’ motion patterns (includingfrequencies), represented by actual trajectories calledprototypes

• match observed trajectories to prototypes andextrapolate

[Saunier et al., 2007, Saunier and Sayed, 2008]

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

A Simple Example

0.4

0.7

0.6

0.3

1

2

t1

t2

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Collision Points

• Using of a finite set of extrapolation hypotheses, thecollision points CPn are enumerated

• The probability of collision si computed by summing theprobabilities of reaching each potential collision point

P(Collision(Ui ,Uj)) =∑

n

P(Collision(CPn))

• The expected Time To Collision is also computed (ifthere is at least one collision point, i.e.P(Collision(Ui ,Uj)) > 0)

TTC(Ui ,Uj , t0) =

∑n P(Collision(CPn)) tnP(Collision(Ui ,Uj))

[Saunier et al., 2010]

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Video Sensors

Video sensors have distinct advantages:• they are easy to install (or can be already installed)• they are inexpensive• they can provide rich traffic description (e.g. road user

tracking)• they can cover large areas• their recording allows verification at a later stage

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Video-based System

Motion patterns, volume, origin-destination counts,driver behavior

Road User Trajectories Interactions

Traffic conflicts, exposure and severity measures, interacting behavior

Image Sequence+

ApplicationsCamera Calibration

Labeled Images for Road User Type

+

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Feature-based Road UserTracking in Video Data

Good enough for safety analysis and other applications,including the study of pedestrians and pedestrian-vehicleinteractions [Saunier and Sayed, 2006]

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Motion Pattern Learning

Traffic Conflict Dataset, Vancouver Reggio Calabria, Italy58 prototype trajectories 58 prototype trajectories

(2941 trajectories) (138009 trajectories)

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

The Kentucky Dataset

• Video recordings kept for a few seconds before andafter the sound-based automatic detection of aninteraction of interest

• 229 traffic conflicts• 101 collisions• The existence of an interaction or its severity is not

always obvious• The interactions recorded in this dataset involve only

motorized vehicles• Limited quality of the video data: resolution,

compression, weather and lighting conditions

• Calibration done using the tool developed by KarimIsmail at UBC [Ismail et al., 2010b]

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Severity Indicators

3.0 3.5 4.0 4.5 5.0 5.5Time (second)

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

Colli

sion P

robabili

ty

3.0 3.5 4.0 4.5 5.0 5.5Time (second)

1.5

2.0

2.5

3.0

3.5

4.0

4.5

TTC

(se

cond)

Side conflict

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Severity Indicators

5.0 5.5 6.0 6.5Time (second)

0.00

0.05

0.10

0.15

0.20

0.25

0.30

Colli

sion P

robabili

ty

5.0 5.5 6.0 6.5Time (second)

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0

TTC

(se

cond)

Side conflict

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Severity Indicators

2 3 4 5 6Time (second)

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

Colli

sion P

robabili

ty

2 3 4 5 6Time (second)

0.5

1.0

1.5

2.0

2.5

3.0

TTC

(se

cond)

Parallel conflict

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Severity Indicators

3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0Time (second)

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

Colli

sion P

robabili

ty

3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0Time (second)

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0

4.5

TTC

(se

cond)

Side collision

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Severity Indicators

2 3 4 5Time (second)

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

Colli

sion P

robabili

ty

2 3 4 5Time (second)

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0

4.5

TTC

(se

cond)

Side collision

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Severity Indicators

3.0 3.5 4.0 4.5 5.0 5.5 6.0Time (second)

0.0

0.1

0.2

0.3

0.4

0.5

Colli

sion P

robabili

ty

3.0 3.5 4.0 4.5 5.0 5.5 6.0Time (second)

0.0

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0

4.5

TTC

(se

cond)

Parallel collision

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Distribution of Indicators

Maximum Collision Probability Minimum TTC

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0Collision Probability

0

20

40

60

80

100

120

140Traffic Conflicts

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0Collision Probability

0

10

20

30

40

50

60

70Collisions

0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5TTC (second)

0

10

20

30

40

50

60Traffic Conflicts

0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5TTC (second)

0

5

10

15

20

25

30Collisions

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Spatial Distribution of theCollision Points

Traffic Conflicts

0

8

16

24

32

40

48

56

64

72

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Spatial Distribution of theCollision Points

Collisions

0

6

12

18

24

30

36

42

48

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Study Before and After theIntroduction of a Scramble

Phase

Data collected in Oakland, CA [Ismail et al., 2010a]

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Distribution of SeverityIndicators

0 2 4 6 8 10 12 14 16 18 200

100

200

300

400

500

600Histogram of Before-and-After TTC

TTCmin

in seconds

Fre

quen

cy o

f tr

affic

eve

nts

-10 -8 -6 -4 -2 0 2 4 6 8 100

2000

4000

6000

8000

10000

12000

DSTmax

in seconds

Fre

quen

cy o

f tr

affic

eve

nts

Histogram of Before-and-After DST

-20 -15 -10 -5 0 5 10 15 200

500

1000

1500

2000

2500

3000

PET in seconds

Fre

quen

cy o

f tr

affic

eve

nts

Histogram of Before-and-After PET

0 2 4 6 8 10 12 14 16 18 200

1000

2000

3000

4000

5000

|PET| in seconds

Fre

quen

cy o

f tr

affic

eve

nts

Histogram of Before-and-After |PET|

-5 0 5 10 15 200

1000

2000

3000

4000

GTmin

in seconds

Fre

quen

cy o

f tr

affic

eve

nts

Histogram of Before-and-After GT

0 2 4 6 8 10 12 14 16 18 200

1000

2000

3000

4000

|GTmin

|in seconds

Fre

quen

cy o

f tr

affic

eve

nts

Histogram of Before-and-After |GT|

TTC Before

TTC After

DST Before

DST After

PET Before

PET After

|PET| Before

|PET| After

GT Before

GT After

|GT| Before

|GT| After

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Before and After Distribution ofthe Collision Points

a)

b)

c)

d)

Before Scramble After Scramble

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Objectives

• Understand collision processes by studying thesimilarities of interactions with and without a collision(conflicts)

• There is some evidence that evasive actionsundertaken by road users involved in conflicts may beof a different nature than the ones attempted incollisions [Davis et al., 2008]

• Importance for surrogate safety measures: whatinteractions without a collision may be used assurrogates for collisions?

• Use of data mining techniques (k-means andhierarchical agglomerative clustering method) to clusterthe data

[Saunier et al., 2011]

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Description of Interactions

Categorical attributes ValuesType of day weekday, week endLighting condition daytime, twilight, nighttimeWeather condition normal, rain, snowInteraction category same direction (turning left

and right, rear-end, lanechange), opposite direction(turning left and right, head-on), side (turning left andright, straight)

Interaction outcome conflict, collision

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Description of InteractionsNumerical attributes UnitsRoad user typepassenger car number of road usersvan, 4x4, SUV number of road usersbus number of road userstruck (all sizes) number of road usersmotorcycle number of road usersbike number of road userspedestrian number of road usersType of evasive actionNo evasive action number of evasive actionsBraking number of evasive actionsSwerving number of evasive actionsAcceleration number of evasive actions3 attributes from ∆v km/h6 values from s km/h

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Distribution of Speed Attributes

DeltaVmin DeltaV DeltaVmax0

10

20

30

40

50

Del

taV

(km

/h)

Norm of the velocity difference

CollisionConflict

Smin1 Smin2 S1 S2 Smax1 Smax20

10

20

30

40

50

60

Spee

d (k

m/h

)

Speed

CollisionConflict

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

3 Clusters

Cluster 1 Cluster 2 Cluster 30

20

40

60

80

100

120

140

Num

ber o

f int

erac

tions

Interaction outcome

CollisionConflict

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Clusters: Speed Attributes

DeltaVmin DeltaV DeltaVmax0

10

20

30

40

50

60

70

Del

taV

(km

/h)

Norm of the velocity difference

Cluster 3Cluster 2Cluster 1

Smin1 Smin2 S1 S2 Smax1 Smax20

10

20

30

40

50

60

70

Spee

d (k

m/h

)

Speed

Cluster 3Cluster 2Cluster 1

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Clusters: Interaction Category

Cluster 1 Cluster 2 Cluster 30

20

40

60

80

100

120

140

Num

ber o

f int

erac

tions

Interaction category

Side straightSame dir. chang.Same dir. rearSame dir. turn RSame dir. turn L

• Cluster 1: collisions,highest speeds,categories side straightand same directionturning right

• Cluster 2: almost pureconflicts, lowest speeds

• Cluster 3: collisions,medium speeds,categories samedirection turning left andright and same directionchanging lanes

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Conclusion

• Tools and framework for automated road safetyanalysis using video sensors

• Large amounts of data: data mining and visualizationfor safety analysis

• Future work:• Still more work on data collection techniques (computer

vision)• Validation of proactive methods for road safety analysis• Understanding and modelling of collision processes:

collect more data

• Need for more open science: data and code sharinghttp://nicolas.saunier.confins.net

http://nicolas.saunier.confins.net

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Collaboration with• Clark Lim and Tarek Sayed (University of British

Columbia)• Karim Ismail (Carleton University)• Nadia Mourji, Bruno Agard (Ecole Polytechnique de

Montreal)

Questions?

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Amundsen, F. and Hyden, C., editors (1977).Proceedings of the first workshop on traffic conflicts,Oslo, Norway. Institute of Transport Economics.

Davis, G. A., Hourdos, J., and Xiong, H. (2008).Outline of causal theory of traffic conflicts and collisions.

In Transportation Research Board Annual MeetingCompendium of Papers.08-2431.

Ismail, K., Sayed, T., and Saunier, N. (2010a).Automated analysis of pedestrian-vehicle conflicts: Acontext for before-and-after studies.Transportation Research Record: Journal of theTransportation Research Board.In press.

Ismail, K., Sayed, T., and Saunier, N. (2010b).Camera calibration for urban traffic scenes: Practicalissues and a robust approach.

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

In Transportation Research Board Annual MeetingCompendium of Papers, Washington, D.C.10-2715.

Saunier, N., Mourji, N., and Agard, B. (2011).Investigating collision factors by mining microscopicdata of vehicle conflicts and collisions.In Transportation Research Board Annual MeetingCompendium of Papers, Washington, D.C.11-0117.

Saunier, N. and Sayed, T. (2006).A feature-based tracking algorithm for vehicles inintersections.In Third Canadian Conference on Computer and RobotVision, Quebec. IEEE.

Saunier, N. and Sayed, T. (2008).A Probabilistic Framework for Automated Analysis ofExposure to Road Collisions.

AutomatedProactive

Road SafetyAnalysis

N. Saunier

Motivation


ExperimentalResults


Conclusion

Transportation Research Record: Journal of theTransportation Research Board, 2083:96–104.

Saunier, N., Sayed, T., and Ismail, K. (2010).Large scale automated analysis of vehicle interactionsand collisions.Transportation Research Record: Journal of theTransportation Research Board, 2147:42–50.

Saunier, N., Sayed, T., and Lim, C. (2007).Probabilistic Collision Prediction for Vision-BasedAutomated Road Safety Analysis.In The 10th International IEEE Conference on IntelligentTransportation Systems, pages 872–878, Seattle. IEEE.

Automated Proactive Road Safety Analysis - Freen.saunier.free.fr/saunier/stock/10-11-25-tram.pdf ·...

Documents

Transcript of Automated Proactive Road Safety Analysis - Freen.saunier.free.fr/saunier/stock/10-11-25-tram.pdf ·...