Automated Video Analysis of Crowd Behavior

Transcription

1 Automated Video Analysis of Crowd Behavior Robert Collins CSE Department Mar 30, 2009 Computational Science Seminar Series, Spring 2009.

2 We Are... Lab for Perception, Action and Cognition

3 Research Interest: Looking at People (one of many!) Biometrics Pose/Activity Recognition Multi-person tracking Analysis of social behavior

4 Crowd Analysis: Motivation Automated video analysis of crowd scenes using computer vision tools, to facilitate: Real-time monitoring situation awareness notification / alarms After-action review trend analysis analyze abnormal events

5 Historical Perspective Our work has roots in video surveillance algorithms that automatically detect and track moving objects. Main components of a video surveillance system: Foreground/background subtraction Connected components to get blobs Data association to get trajectories

6 Simple Background Subtraction Keep a background image (assumed to have no objects present). Subtract current frame from background image Label pixels as foreground (1) or background (0) based on thresholding I(t) abs T M(t) B B = I(0); loop time t I(t) = next frame; diff = abs[b I(t)]; M(t) = threshold(diff, ); end

7 Adaptive Background Subtraction Continually update background image with current frame Yields resilience to slow changes in illumination I(t) abs T M(t) B(t-1) delay B(t) I(t) + (1 )B(t-1) B(0) = I(0); loop time t I(t) = next frame; diff = abs[b(t-1) I(t)]; M(t) = threshold(diff, ); B(t) = I(t)+(1 )B(t-1); end

8 Grouping FG Pixels Into Blobs dilate Connected components AND Bounding box = smallest rectangle containing all pixels on the object.

9 Data Association Determining the correspondence of blobs across frames is based on feature similarity between blobs. Commonly used features: location, size / shape, velocity, appearance For example: location, size and shape similarity can be measured based on bounding box overlap: A B score = 2 * area(a and B) area(a) + area(b) A = bounding box at time t B = bounding box at time t+1

10 Simple Surveillance Results movie

11 Problem: Blob Merge/Split merge occlusion occlusion split When two objects pass close to each other, they are detected as a single blob. Often, one object will become occluded by the other one. One of the challenging problems is to maintain correct labeling of each object after they split again.

12 Problem with Crowd Scenes Merge/Split events are the rule, rather than the exception!

13 Our Approach Estimate number/configuration of people using Markov Chain Monte Carlo (a stochastic search method) Good for low-resolution / wide-angle views. Still relies on foreground/background segmentation. Not appropriate for very high crowd density.

14 Problem Statement Given a foreground image, and person-sized bounding box*, find a configuration (number and locations) of bounding boxes that cover a majority of foreground pixels while leaving a majority of background pixels uncovered. foreground image person-sized bounding box *note: height, width and orientation of the bounding box may depend on image location we determine these relationships beforehand through a calibration procedure.

15 Likelihood Score To measure how good a proposed configuration is, we generate a foreground image from it and compare with the observed foreground image to get a likelihood score. config = {{x 1,y 1,w 1,h 1,theta 1 },{x 2,y 2,w 2,h 2,theta 2 },{x 3,y 3,w 3,h 3,theta 3 }} generated foreground image observed foreground image compare Likelihood Score

16 Likelihood Score Bernoulli distribution model likelihood simplify, by assuming log likelihood Number of pixels that disagree!

17 Searching for the Max The space of configurations is very large. We can t exhaustively search for the max likelihood configuration. We can t even really uniformly sample the space to a reasonable degree of accuracy. config k = {{x 1,y 1,w 1,h 1,theta 1 },{x 2,y 2,w 2,h 2,theta 2 },,{x k,y k,w k,h k,theta k }} Let N = number of possible locations for (x i,y i ) in a k-person configuration. Size of config k = N k And we don t even know how many people there are... Size of config space = N 0 + N 1 + N 2 + N 3 + If we also wanted to search for width, height and orientation, this space would be even more huge.

18 Searching for the Max Local Search Approach Given a current configuration, propose a small change to it Compare likelihood of proposed config with likelihood of the current config Decide whether to accept the change

19 Proposals Add a rectangle (birth) add current configuration proposed configuration

20 Proposals Remove a rectangle (death) remove current configuration proposed configuration

21 Move a rectangle Proposals move current configuration proposed configuration

22 Searching for the Max Naïve Acceptance Accept proposed configuration if it has a larger likelihood score, i.e. Compute a = L(proposed) L(current) Accept if a > 1 Problem: leads to hill-climbing behavior that gets stuck in local maxima Brings us here But we really want to be over here! start Likelihood

23 Searching for the Max The MCMC approach Generate random configurations from a distribution proportional to the likelihood! Generates many high likelihood configurations Likelihood Generates few low likelihood ones.

24 Searching for the Max The MCMC approach Generate random configurations from a distribution proportional to the likelihood! This searches the space of configurations in an efficient way. Now just remember the generated configuration with the highest likelihood.

25 Sounds good, but how to do it? Think of configurations as nodes in a graph. Put a link between nodes if you can get from one config to the other in one step (birth, death, move) death birth config A birth death config C config B move move move move move move move move birth death config D move move config E death birth Note links come in pairs: birth/death; move/move

26 Detailed Balance Consider a pair of configuration nodes r,s Want to generate them with frequency relative to their likelihoods L(r) and L(s) Let q(r,s) be relative frequency of proposing configuration s when the current state is r (and vice versa) A sufficient condition to generate r,s with the desired frequency is L(r) r q(r,s) L(r) q(r,s) = L(s) q(s,r) detailed balance s L(s) q(s,r)

27 Detailed Balance Typically, your proposal frequencies do NOT satisfy detailed balance (unless you are extremely lucky). To fix this, we introduce a computational fudge factor a Detailed balance: a* L(r) q(r,s) = L(s) q(s,r) a * q(r,s) Solve for a: L(r) r a = L(s) q(s,r) L(r) q(r,s) s L(s) q(s,r)

28 MCMC Sampling Metropolis Hastings algorithm Propose a new configuration Compute a = L(proposed) q(proposed,current) L(current) q(current,proposed) Accept if a > 1 Else accept anyways with probability a Difference from Naïve algorithm

29 Trans-dimensional MCMC Green s reversible-jump approach (RJMCMC) gives a general template for exploring and comparing states of differing dimension (diff numbers of rectangles in our case). Proposals come in reversible pairs: birth/death and move/move. We should add another term to the acceptance ratio for pairs that jump across dimensions. However, that term is 1 for our simple proposals.

30 MCMC in Action Sequence of proposed configurations Sequence of accepted configurations movies

31 MCMC in Action Max likelihood configuration Looking good!

32 Examples

33 Adding Shape to the Estimation video frame estimated state library of candidate shapes

34 Adding Shape to the Estimation MCMC iterations movie MCMC proposes changes to current configuration add/remove a person shift location of person change their shape

35 Soccer Dataset Evaluation run on every 10th frame green contours: true positives red boxes: false negatives pink contours: false positives movie detailed view

36 Caviar Dataset evaluation on six sequences learned shapes results from sample frames

37 Detections, Nov 22, Curtin Road

38 Crowd Behavior In areas of bidirectional motion, people tend to follow others to minimize collisions (maximize throughput). Known as the fingering effect. movie Green: leftward moving. Red: rightward moving,

39 Fingering Effect Density by image row Green (leftward); Red (rightward)

40 Detections, Sep 6, Gate A

41 Crowd Flow/Density Keep in mind this scene structure (as depicted by red lines)

42 Crowd Flow/Density 30 minute period

43 Crowd Flow/Density movie Time Lapse. Integrated over spatial/temporal windows.

44 GateA Path Counts \ movie Maintain a running count of number of people whose trajectories cross a set of user-specified lines (color-coded).

45 Collective locomotion find small groups traveling in a crowd sociological hypothesis: validating that the majority of people in a crowd cluster in small groups public safety: improving situation awareness and realtime emergency response during public disturbances through real-time monitoring and modeling of crowd behavior

46 Agglomerative Hierarchical Clustering

47 Sample Results note: computer only sees this view! Evaluation reveals substantial agreement between computer-generated groupings and those found by human coders (ground truth) p <.001

48 More Results

49 Research Directions Validation; improve algorithm robustness. Detection of stationary people Tackle the HARD problems. Primarily high-density crowds.

50 Research Directions Idea: might be good to try other human shapes (activities) or other animals/objects.