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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
Conclusion
A World Health Issue
AutomatedProactive
Road SafetyAnalysis
N. Saunier
Motivation
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
Conclusion
The Safety/Severity Hierarchy
FI
PD
Undisturbed passages
Potential ConflictsSlight Conflicts
Serious ConflictsAccidents
Various severity measures
AutomatedProactive
Road SafetyAnalysis
N. Saunier
Motivation
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
Conclusion
MotivationNeed for automated tools to address the shortcomings ofreactive diagnosis methods and traffic conflict techniques
AutomatedProactive
Road SafetyAnalysis
N. Saunier
Motivation
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
Conclusion
A Simple Example
0.4
0.7
0.6
0.3
1
2
t1
t2
AutomatedProactive
Road SafetyAnalysis
N. Saunier
Motivation
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
Conclusion
Spatial Distribution of theCollision Points
Traffic Conflicts
0
8
16
24
32
40
48
56
64
72
AutomatedProactive
Road SafetyAnalysis
N. Saunier
Motivation
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
Conclusion
Spatial Distribution of theCollision Points
Collisions
0
6
12
18
24
30
36
42
48
AutomatedProactive
Road SafetyAnalysis
N. Saunier
Motivation
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
Conclusion
Before and After Distribution ofthe Collision Points
a)
b)
c)
d)
Before Scramble After Scramble
AutomatedProactive
Road SafetyAnalysis
N. Saunier
Motivation
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
AutomatedProactive
Road SafetyAnalysis
N. Saunier
Motivation
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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
ProbabilisticFramework
ExperimentalResults
InvestigatingCollisionFactors
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.