Chaotic Invariants of Lagrangian Particle Trajectories for Anomaly Detection in Crowded Scenes Shandong Wu, Brian E. Moore, Mubarak Shah Computer Vision Lab, University of Central Florida, USA sdwu@eecs.ucf.edu, Crowd Scene Analysis Video clip • Challenges – Very high density of objects – Diverse level of coherency of motions • Traditional methods – Only suitable for sparse scenes – Suffer from severe occlusions, small object sizes, similar appearance bmoore@math.ucf.edu, Optical flow Particle trajectory Experiment 1: Anomaly Detection Particle advection Calculating optical flow Clustering particle trajectory Representative trajectory Chaotic description Scene model Anomaly shah@eecs.ucf.edu Learning model Detection &Localization Calculating chaotic invariants Query clip Particle trajectory Chaotic Description Particle Advection Video clip – The Algorithm X time series of a representative trajectory • To identify the dynamics of representative trajectories Embedding (m: embedding dimension; J: time delay) Chaotic dynamics by measurable chaotic invariants • Chaotic feature set Optical flow – – – Sub-pixel level 2D optical flow interpolation L D M Clustering Orbit : Largest Lyapunov exponent : Correlation dimension : Mean of representative trajectories • Only necessary for position-caused anomalies Normality model For , find its nearest neighbor Representative trajectory p: mean period of orbit Pairs of nearest neighbors Euler’s method • 2D trajectory embedding: in terms of two 1D time series d(t): average divergence Particle’s position Calculating L Calculate mean rate of Separation of the nearest neighbors Use the distances of nearest neighbors L D Correlation sum Calculating D Particle trajectory A set of approximately parallel lines each with a slope roughly proportional to L A bunch of adjacent particle trajectories may belong to a single sub-object L – – Step 1: Remove relatively motionless particles and trajectories that carry minor information Step 2: Cluster by k-means according to position information Output: Representative trajectories Likelihood of representative trajectories w.r.t normality model D Clip 29 Clip 30 (Anomaly source GT) Advantages of the Algorithms Method: clustering – : Heaviside step function : threshold distance : number of points (crucial for small and noisy data set) Observation: • Experiment 2: Anomaly Localization Fit the average line Cluster Particle Trajectories – Dataset: Unusual crowd activity dataset from University of Minnesota for jth pair of nearest neighbors : distance between the jth pair of nearest neighbors • ROC curve and comparison with previous work Detection example • • (a) All representative trajectories (b) Clustering of trajectories with low probability (c) Localized major anomaly sources The above algorithms proven to be insensitive to the changes in – time delay – embedding dimension – size of data set and – to some extent noise More importantly, ensure L>0 for the condition of chaotic analysis A more efficient and manageable crowd scene representation ** Anomaly Detection • • • Definition of anomaly – Spatiotemporal change of system dynamics (chaotic or/and positions) • Global anomaly: entire change of dynamics • Local anomaly: dynamic changes near particular spatial points Normality model – Multi-variate GMM: 4D ( [Lx , Ly , Dx , Dy ] ) or 6D ( plus [M x , M y ] ) – Learning by: EM + IPRA algorithm Experiment 3: Position-caused Anomaly Localization Chaotic features Position feature Synthesized anomaly Detected anomaly Anomaly Localization Novelties / Contributions Localize the anomaly in terms of position & size – Calculate the likelihoods contributed by each representative trajectory – Localize those trajectories with low likelihoods – Cluster them according to position information – Filter out the clusters with fewer number of trajectories – The remaining clusters reveal the major abnormal regions • A novel combination of Lagrangian particle dynamics approach together with chaotic modeling • Unique utilization of clustering of particle trajectories for modeling crowded scenes (novel modeling element: representative trajectory) • Chaotic dynamics are introduced into the crowd context with effective algorithms for calculating chaotic invariants • Our methods are able to deal with both coherent and incoherent flows