Crowd Behavior Analysis Based On Visual Surveillance Data

As the interdisciplinary of Computer Vision, Pattern Recognition and Artificial Intelligence, visual surveillance based crowd behavior analysis is currently a highly focused research area of intelligent visual surveillance technology. The thesis focuses on crowd behavior based on the trajectory data from video surveillance. The work mainly includes:1. Learning motion patterns based on latent structural information. Considering the influence of the scene structure having on crowd analysis, the latent structural information is proposed for trajectory analysis in unstructured scene. Based on latent structural information learned by clustering algorithm, trajectories are grouped, and then analyzed using temporal and spatial property to build motion patterns. Anomalies of single tracks are detected based on the learned motion patterns.2. Crowd analysis based on Hierarchical Dirichlet Process (HDP). We introduce the Topic Modeling from Natural Language Processing into trajectory based crowd analysis, learn the regions of interest(ROI) for crowd behavior using HDP and describe the changing motion features of crowd with time.3. Multi-scale statistical analysis for crowd behavior. To quantize the statistical crowd behavior features, we gather motion statistics for crowd on both global and local levels, and build the global velocity-orientation distribution matrix and local speed histogram and crowd density histogram of the ROIs as the statistical features templates for quantitative description of crowd behavior.