Public Crowd Analysis Based On Particle Video

The crowd motion analysis in intelligent video surveillance systemis an important research sub-field. It has a large number of applications such ascrowd management, virtual environments, intelligent surveillance, public spacedesign, intelligent environments simulate, etc. The main purpose of crowd motionanalysis is by detecting, tracking, identifying the movement of the crowd, to theunderstanding and describing the crowd behavior. In this dissertation, we make useof particle video algorithm to solve high-density crowd dominant motion detectionand crowd motion segmentation and anomaly detection. The highlights of thedissertation include:1. To address the situation of complex scene geometry, multiple types ofocclusion, region with low texture, current existing crowd motion estimationalgorithms such as optical flow and feature tracking exist drift problems or trackinglost phenomenon, this thesis describes a new approach to video motion estimationcalled particle video and use the particle video to obtain accurate long-range featuretrajectories in complex crowd scene.2. Implement a high-density motion trajectory clustering based on particle videoalgorithm. For most particle trajectories obtained by particle video algorithm arescattered, in order to analyze the crowd motion behavior, we adopt a clusteringmethod based on longest common subsequences to identify dominant motions incrowd scene. Then using the dominant motion, we can analysis the crowd motionbehavior and anomaly detection.3. A crowd motion segmentation algorithm based on particle video algorithmand Finite Time Lyapunov Exponent (FTLE) field is proposed. Firstly Flow Map isobtained using particle video, whose spatial gradients are subsequently constructed aFinite Time Lyapunov Exponent (FTLE) field. The LCS divide flow into regions ofqualitatively different dynamics and are used to abnormal motion detection in anormalized cuts framework. 4. Combine with above two algorithms; we implement two main approaches forcrowd behavior analysis using object-based method and holistic-based method. Inobject-based methods, crowd behavior understanding is performed through somekind of segmentation or detection of individuals to analyze group behaviors, whileholistic approaches treat the crowd as a single entity, without the need of segmentingeach individual. This thesis adopt different algorithms to realize the above twomethods.