Understanding Collision Processes using Video Data Workshop on Comparison of Surrogate Measures of Safety Extracted from Video Data Nicolas Saunier
[email protected]
January 12th 2014
Outline
1
Motivation
2
Methodology
3
Experimental Results using Video Data
4
Conclusion
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
2 / 40
Motivation
Outline
1
Motivation
2
Methodology
3
Experimental Results using Video Data
4
Conclusion
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
3 / 40
Motivation
Traffic Conflicts
A traffic conflict is “an observational situation in which two or more road users approach each other in space and time to such an extent that a collision is imminent if their movements ´ 1977] remain unchanged” [Amundsen and Hyden,
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
4 / 40
Motivation
Traffic Conflicts
A traffic conflict is “an observational situation in which two or more road users approach each other in space and time to such an extent that a collision is imminent if their movements ´ 1977] remain unchanged” [Amundsen and Hyden, Several traffic conflict techniques and lack of comparison Issues caused by the (mostly) manual data collection process cost reliability and subjectivity: intra- and inter-observer variability
Mixed validation results
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
4 / 40
Motivation
The Safety/Severity Hierarchy
Accidents F I PD
Serious Conflicts Slight Conflicts Potential Conflicts Undisturbed passages
Various surrogate safety measures ´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
5 / 40
Motivation
Past research [Davis et al., 2008]
There is some evidence that evasive actions undertaken by road users involved in conflicts may be of a different nature than the ones attempted in collisions
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
6 / 40
Motivation
Past research: The Whole Hierarchy
´ 2006] [Svensson, 1998, Svensson and Hyden, ´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
7 / 40
Motivation
Past research: The Whole Hierarchy
Feedback and learning process: collisions with injuries occurred at the ´ 2006] signalized intersection [Svensson, 1998, Svensson and Hyden, ´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
7 / 40
Motivation
Objectives
Understand collision processes by studying the similarities of interactions with and without a collision to design better counter-measures develop better surrogate measures based on better-known relationships between interactions with and without a collision
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
8 / 40
Motivation
Objectives
Understand collision processes by studying the similarities of interactions with and without a collision to design better counter-measures develop better surrogate measures based on better-known relationships between interactions with and without a collision
Methods collect large amounts of interaction data, in particular using video sensors design suitable interaction descriptors and safety indicators (obtained through a robust probabilistic framework) design suitable interaction similarity measures use and adapt data mining techniques to cluster the interactions
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
8 / 40
Methodology
Outline
1
Motivation
2
Methodology
3
Experimental Results using Video Data
4
Conclusion
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
9 / 40
Methodology
Interaction Descriptors
v2
U2
distance
´ N. Saunier, Polytechnique Montreal
Δv
v1
1
U
θ
Understanding Collision Processes
January 12th 2014
10 / 40
Methodology
Rethinking the Collision Course
A traffic conflict is “an observational situation in which two or more road users approach each other in space and time to such an extent that a collision is imminent if their movements remain unchanged” For two interacting road users, many chains of events may lead to a collision It is possible to estimate the probability of collision if one can predict the road users’ future positions the motion prediction method must be specified
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
11 / 40
Methodology
Motion Prediction
Predict trajectories according to various hypotheses iterate the positions based on the driver input (acceleration and steering) learn the road users’ motion patterns (including frequencies), represented by actual trajectories called prototypes, then match observed trajectories to prototypes and resample
Advantage: generic method to detect a collision course and measure severity indicators, as opposed to several cases and formulas (e.g. in [Gettman and Head, 2003]) [Saunier et al., 2007, Saunier and Sayed, 2008, Mohamed and Saunier, 2013, St-Aubin et al., 2014]
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
12 / 40
Methodology
A Simple Example
t1
0.7 0.3
t2 0.4
2
0.6
1 ´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
13 / 40
Methodology
Collision Points and Crossing Zones Using of a finite set of predicted trajectories, enumerate the collision points CPn and the crossing zones CZm . Severity indicators can then be computed: X P(Collision(Ui , Uj )) = P(Collision(CPn )) Pn
P(Collision(CPn )) tn P(Collision(Ui , Uj )) P P(Reaching(CZm )) |ti,m − tj,m | pPET (Ui , Uj , t0 ) = m P m P(Reaching(CZm )) TTC(Ui , Uj , t0 ) =
n
[Saunier et al., 2010, Mohamed and Saunier, 2013, Saunier and Mohamed, 2014]
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
14 / 40
Methodology
Similarity Measures
Traditional measures rely on fixed length descriptor vectors: extract agregated values from continuous time series indicator data considerable loss of information
Some measures naturally accomodate variable length vectors: Longest Common Sub-sequence
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
15 / 40
Methodology
Need for Improved LCSS 20
20
15
15
10
10
5
5
00
5
10
15
20
00
5
10
15
20
The series in each plot have maximum similarity if using δ = +∞: this is desired in the plot on the left since it is an exact sub-sequence, but not on the right if the rate of change is taken into account ´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
16 / 40
Methodology
The Aligned LCSS 1.6 1.4
Time to Collision (s)
1.2 1.0 0.8 0.6 0.4 0.2 0.02.0
2.5
3.0
Time (s)
3.5
4.0
4.5
Example of alignment of two very similar real TTC indicators: LCSS2,d0.2 s = 0.2 and ALCSS2,d0.2 s = 1 ´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
17 / 40
Methodology
The Aligned LCSS 16
4.5
14
4.0 3.5 Time to Collision (s)
Distance (m)
12 10 8 6
2.5 2.0 1.5
4 22
3.0
1.0 3
4
Time (s)
5
6
LCSS+∞ LCSS2 ALCSS2
7
0.52.0
Distance 0.87 0.35 0.42
2.5
3.0
3.5
4.5 4.0 Time (s)
5.0
5.5
6.0
6.5
TTC 0.64 0.12 0.42
These real profiles are more similar using LCSS with infinite δ than using ALCSS and a finite δ ´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
17 / 40
Methodology
Clustering Algorithm
All algorithms operating on a similarity matrix may be used Custom algorithm with cluster prototypes [Saunier et al., 2007] 1 2
3
Indicators are sorted by length For each indicator, if its maximum similarity is lower than a threshold, create a new cluster with indicator as prototypes Otherwise, assign it to the most similar prototype
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
18 / 40
Experimental Results using Video Data
Outline
1
Motivation
2
Methodology
3
Experimental Results using Video Data
4
Conclusion
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
19 / 40
Experimental Results using Video Data
The Kentucky Dataset
Video recordings kept for a few seconds before and after the sound-based automatic detection of an interaction of interest 213 traffic conflicts and 82 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
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
20 / 40
Experimental Results using Video Data
Road User Tracking
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
21 / 40
Experimental Results using Video Data
Motion Prediction
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
22 / 40
Experimental Results using Video Data
Motion Prediction
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
23 / 40
Experimental Results using Video Data
Motion Prediction
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
24 / 40
Experimental Results using Video Data
Collision Probability
Severity Indicators 0.40 0.35 0.30 0.25 0.20 0.15 0.10 0.05 0.00
2
3
4 Time (second)
5
6
2
3
4 Time (second)
5
6
TTC (second)
3.0 2.5 2.0 1.5 1.0 0.5
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
25 / 40
Experimental Results using Video Data
Description of Interactions [?]
Categorical attributes Type of day Lighting condition Weather condition Interaction category
Interaction outcome
´ N. Saunier, Polytechnique Montreal
Values weekday, week end daytime, twilight, nighttime normal, rain, snow same direction (turning left and right, rear-end, lane change), opposite direction (turning left and right, head-on), side (turning left and right, straight) conflict, collision
Understanding Collision Processes
January 12th 2014
26 / 40
Experimental Results using Video Data
Description of Interactions Numerical attributes Road user type passenger car van, 4x4, SUV bus truck (all sizes) motorcycle Type of evasive action No evasive action Braking Swerving Acceleration 3 attributes from ∆v 6 values from s ´ N. Saunier, Polytechnique Montreal
Units number of road users number of road users number of road users number of road users number of road users number of evasive actions number of evasive actions number of evasive actions number of evasive actions km/h km/h
Understanding Collision Processes
January 12th 2014
27 / 40
Experimental Results using Video Data
Distribution of Speed Attributes
50
Norm of the velocity difference Collision Conflict
60 50
40
Speed Collision Conflict
Speed (km/h)
DeltaV (km/h)
40 30
20
20
10
0
30
10
DeltaVmin
DeltaV
DeltaVmax
´ N. Saunier, Polytechnique Montreal
0
Smin1
Smin2
S1
Understanding Collision Processes
S2
Smax1
January 12th 2014
Smax2
28 / 40
Experimental Results using Video Data
3 Clusters Interaction outcome
140
Collision Conflict
120
Number of interactions
100 80 60 40 20 0 ´ N. Saunier, Polytechnique Montreal
Cluster 1
Cluster 2
Understanding Collision Processes
Cluster 3 January 12th 2014
29 / 40
Experimental Results using Video Data
Clusters: Speed Attributes
70
70 60
50
50
40
40
Speed (km/h)
DeltaV (km/h)
60
Norm of the velocity difference Cluster 3 Cluster 2 Cluster 1
30
30
20
20
10
10
0
DeltaVmin
DeltaV
DeltaVmax
´ N. Saunier, Polytechnique Montreal
Speed Cluster 3 Cluster 2 Cluster 1
0
Smin1
Smin2
Understanding Collision Processes
S1
S2
Smax1
January 12th 2014
Smax2
30 / 40
Experimental Results using Video Data
Clusters: Interaction Category 140 120
Number of interactions
100
Interaction category Side straight Same dir. chang. Same dir. rear Same dir. turn R Same dir. turn L
Cluster 1: collisions, highest speeds, categories side straight and same direction turning right
80
Cluster 2: almost pure conflicts, lowest speeds
60
Cluster 3: collisions, medium speeds, categories same direction turning left and right and same direction changing lanes
40 20 0
Cluster 1
Cluster 2
´ N. Saunier, Polytechnique Montreal
Cluster 3 Understanding Collision Processes
January 12th 2014
31 / 40
Experimental Results using Video Data
Indicator Clustering using Aligned LCSS [Saunier and Mohamed, 2014]
Indicator
Threshold
Distance (Dist) Speed differential (SD) Velocity angle (VA) Time to collision (TTC) Probability of Collision (PoC)
1m 1.5 m/s 0.15 rad 0.2 s 0.1
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
Minimum Clustering Similarity 0.23 0.4 0.4 0.3 0.5
Number of Clusters 6 4 4 4 6
January 12th 2014
32 / 40
Experimental Results using Video Data
´ N. Saunier, Polytechnique Montreal
Dist (m)
Dist (m)
50 Cluster 1 - 23.1%(28/121) 50 Cluster 2 - 42.7%(35/82) 40 40 30 30 20 20 10 10 0 2 0 2 4 6 8 10 12 14 0 2 0 2 4 6 8 10 12 14 Time (s) Time (s) Cluster 3 0 . 0%(0 / 8) Cluster 4 - 42.1%(8/19) 50 50 40 40 30 30 20 20 10 10 0 2 0 2 4 6 8 10 12 14 0 2 0 2 4 6 8 10 12 14 Time (s) Time (s) Cluster 5 38 . 5%(5 / 13) Cluster 6 - 11.5%(6/52) 50 50 40 40 30 30 20 20 10 10 0 2 0 2 4 6 8 10 12 14 0 2 0 2 4 6 8 10 12 14 Time (s) Time (s) Dist (m)
Dist (m)
Dist (m)
Dist (m)
Indicator Clustering using Aligned LCSS
Understanding Collision Processes
January 12th 2014
32 / 40
Experimental Results using Video Data
Indicator Clustering using Aligned LCSS Cluster 1 - 27.6%(42/152)
20
20
15
15
10 5 02
10 5
4
6
8 10 12 Time (s) Cluster 3 - 37.5%(21/56)
14
0 25
20
20
15
15
SD (m/s)
SD (m/s)
25
10 5 02
Cluster 2 - 30.0%(18/60)
25
SD (m/s)
SD (m/s)
25
0
2
4 Time (s)
6
8
Cluster 4 - 3.8%(1/26)
10 5
0
2
´ N. Saunier, Polytechnique Montreal
4 6 Time (s)
8
10
00
2
Understanding Collision Processes
4 6 Time (s)
8
10
January 12th 2014
32 / 40
Experimental Results using Video Data
VA (deg.)
180 Cluster 2 - 23.0%(20/87) 160 140 120 100 80 60 40 20 0 2 0 2 4 6 8 10 12 14 Time (s)
180 Cluster 3 - 17.5%(10/57) 160 140 120 100 80 60 40 20 0 2 0 2 4 6 8 10 12 14 Time (s)
180 Cluster 4 - 33.3%(16/48) 160 140 120 100 80 60 40 20 0 2 0 2 4 6 8 10 12 14 Time (s)
´ N. Saunier, Polytechnique Montreal
VA (deg.)
VA (deg.)
180 Cluster 1 - 35.6%(36/101) 160 140 120 100 80 60 40 20 0 2 0 2 4 6 8 10 12 14 Time (s)
VA (deg.)
Indicator Clustering using Aligned LCSS
Understanding Collision Processes
January 12th 2014
32 / 40
Experimental Results using Video Data
4.5 Cluster 2 - 38.2%(55/144) 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.01 2 3 4 5 6 7 8 9 Time (s)
Cluster 3 - 33.3%(3/9) 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.01 2 3 4 5 6 7 8 9 Time (s)
Cluster 4 - 5.0%(1/20) 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.01 2 3 4 5 6 7 8 9 Time (s)
´ N. Saunier, Polytechnique Montreal
TTC (s)
TTC (s)
TTC (s)
4.5 Cluster 1 - 19.4%(13/67) 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.01 2 3 4 5 6 7 8 9 Time (s)
TTC (s)
Indicator Clustering using Aligned LCSS
Understanding Collision Processes
January 12th 2014
32 / 40
Experimental Results using Video Data
1.0 0.8 0.6 0.4 0.2 0.0 2 1.0 0.8 0.6 0.4 0.2 0.0 2
1.0 0.8 0.6 0.4 0.2 10 0.0 2
Cluster 2 - 45.3%(34/75)
PoC
Cluster 1 - 18.6%(24/129)
4 6 8 Time (s) Cluster 3 - 18.2%(2/11) 0
2
2
1.0 0.8 0.6 0.4 0.2 10 0.0 2
2
1.0 0.8 0.6 0.4 0.2 10 0.0 2
4 6 8 Time (s) Cluster 4 - 0.0%(0/5)
10
4 6 8 Time (s) Cluster 6 - 20.0%(1/5)
10
0
2
PoC
1.0 0.8 0.6 0.4 0.2 0.0 2
4 6 8 Time (s) Cluster 5 - 18.2%(2/11) 0
0
2
0
2
PoC
PoC
PoC
PoC
Indicator Clustering using Aligned LCSS
0
´ N. Saunier, Polytechnique Montreal
4 6 Time (s)
8
Understanding Collision Processes
4 6 Time (s)
8
10
January 12th 2014
32 / 40
Conclusion
Outline
1
Motivation
2
Methodology
3
Experimental Results using Video Data
4
Conclusion
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
33 / 40
Conclusion
Conclusion
Better tools to measure interaction similarity (and general time series similarity)
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
34 / 40
Conclusion
Conclusion
Better tools to measure interaction similarity (and general time series similarity) Mounting evidence that not all interactions should be used for surrogate safety measure
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
34 / 40
Conclusion
Conclusion
Better tools to measure interaction similarity (and general time series similarity) Mounting evidence that not all interactions should be used for surrogate safety measure Future work: collect more data and compare methods
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
34 / 40
Conclusion
Need for Open Science Scientific principle of reproducibility to what extent are the mixed validation results reported in the literature related to a lack of comparisons and reproduciblity of the various methods proposed for surrogate safety analysis?
Need to share data and tools used to produce the results public datasets and benchmarks public / open source software
Traffic Intelligence open source project https: //bitbucket.org/Nicolas/trafficintelligence
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
35 / 40
Conclusion
Collaboration with Nadia Mourji, Bruno Agard, Mohamed Gomaa ´ Mohamed, Paul St-Aubin (Polytechnique Montreal) Funded in part by the Natural Sciences and Research Council of Canada (NSERC)
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
36 / 40
Conclusion
Collaboration with Nadia Mourji, Bruno Agard, Mohamed Gomaa ´ Mohamed, Paul St-Aubin (Polytechnique Montreal) Funded in part by the Natural Sciences and Research Council of Canada (NSERC)
Questions?
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
36 / 40
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
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
37 / 40
Conclusion
Video-based System
Image Sequence + Camera Calibration
Road User Trajectories
Interactions
Applications
+
Labeled Images for Road User Type
´ N. Saunier, Polytechnique Montreal
Motion patterns, volume, origin-destination counts, driver behavior
Understanding Collision Processes
Traffic conflicts, exposure and severity measures, interacting behavior January 12th 2014
38 / 40
Conclusion
Feature-based Road User Tracking in Video Data
Good enough for safety analysis and other applications, including the study of pedestrians and pedestrian-vehicle interactions [Saunier and Sayed, 2006] ´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
39 / 40
Conclusion
Motion Pattern Learning
Traffic Conflict Dataset, Vancouver 58 prototype trajectories (2941 trajectories)
´ N. Saunier, Polytechnique Montreal
Reggio Calabria, Italy 58 prototype trajectories (138009 trajectories)
Understanding Collision Processes
January 12th 2014
40 / 40
Conclusion
´ C., editors (1977). Amundsen, F. and Hyden, 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 Meeting Compendium of Papers. 08-2431. Gettman, D. and Head, L. (2003). Surrogate safety measures from traffic simulation models, final report. Technical Report FHWA-RD-03-050, Federal Highway Administration. Ismail, K., Sayed, T., and Saunier, N. (2010a). Automated analysis of pedestrian-vehicle conflicts: Context for before-and-after studies. ´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
40 / 40
Conclusion
Transportation Research Record: Journal of the Transportation Research Board, 2198:52–64. presented at the 2010 Transportation Research Board Annual Meeting. Ismail, K., Sayed, T., and Saunier, N. (2010b). Camera calibration for urban traffic scenes: Practical issues and a robust approach. In Transportation Research Board Annual Meeting Compendium of Papers, Washington, D.C. 10-2715. Mohamed, M. G. and Saunier, N. (2013). Motion prediction methods for surrogate safety analysis. In Transportation Research Board Annual Meeting Compendium of Papers. 13-4647. Accepted for publication in Transportation Research Record: Journal of the Transportation Research Board. Saunier, N. and Mohamed, M. G. (2014). ´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
40 / 40
Conclusion
Clustering surrogate safety indicators to understand collision processes. In Transportation Research Board Annual Meeting Compendium of Papers. 14-2380. Saunier, N. and Sayed, T. (2006). A feature-based tracking algorithm for vehicles in intersections. ´ In Canadian Conference on Computer and Robot Vision, Quebec. IEEE. Saunier, N. and Sayed, T. (2008). A Probabilistic Framework for Automated Analysis of Exposure to Road Collisions. Transportation Research Record: Journal of the Transportation Research Board, 2083:96–104. presented at the 2008 Transportation Research Board Annual Meeting. Saunier, N., Sayed, T., and Ismail, K. (2010). ´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
40 / 40
Conclusion
Large scale automated analysis of vehicle interactions and collisions. Transportation Research Record: Journal of the Transportation Research Board, 2147:42–50. presented at the 2010 Transportation Research Board Annual Meeting. Saunier, N., Sayed, T., and Lim, C. (2007). Probabilistic Collision Prediction for Vision-Based Automated Road Safety Analysis. In The 10th International IEEE Conference on Intelligent Transportation Systems, pages 872–878, Seattle. IEEE. St-Aubin, P., Saunier, N., and Miranda-Moreno, L. F. (2014). Road user collision prediction using motion patterns applied to surrogate safety analysis. In Transportation Research Board Annual Meeting Compendium of Papers. 14-5363. ´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
40 / 40
Conclusion
Svensson, A. (1998). A Method for Analyzing the Traffic Process in a Safety Perspective. PhD thesis, University of Lund. Bulletin 166. ´ C. (2006). Svensson, A. and Hyden, Estimating the severity of safety related behaviour. Accident Analysis & Prevention, 38(2):379–385.
´ N. Saunier, Polytechnique Montreal
Understanding Collision Processes
January 12th 2014
40 / 40