Anomaly Detection Using Motion Patterns Computed from Optical Flow

Abstract

A method is proposed for detecting anomalies in extremely crowded scenes using analysis of motion patterns. The optical flow is computed by initializing the video as a dynamical system. Optical flow is a vector field where each vector represents the direction and amount of motion. This generated model can be used to define trajectories. Then these trajectories are clustered hierarchically using spatial and temporal information for learning the motion patterns. Based on the learned statistical motion patterns, anomalies are detected using statistical methods.