Abnormal Crowd Behavior Detection Based On Local Motion Clustering

Intelligent video surveillance is characterized by dispensing with human intervention,all-weather monitoring, high detecting rate and alarming real time, and has been thecurrent topic in computer vision. It has attracted more and more attentions of researchersin the human cost enhancing fast era. Abnormal crowd detection based on video sequencehas a very high social value and a broad application prospects for that it is able to analyzethe events in the scene automatically, and emit alarms about suspicious act to operators.Thus, how to detect the abnormal crowd behavior accurately is the main focus of thispaper.Among the current popular crowd behavior analysis method, the method based onmacro usually fails to locate the local abnormal behavior for that it ignores the individualmotion information. While the micro analysis method is restricted for the reason thataccurate detection and tracking of individuals is almost unavailable when pedestrians aredense. To solve the above problems, the paper starts with mesoscopic level to analyzecrowd behavior, and puts forward a crowd group segmentation and abnormal behaviordetection method by means of clustering local motion information. The main contentconsists of micro motion information extraction, mesoscopic group clustering andabnormal behavior detection and locating.Firstly, the optical flow between two consecutive frames is calculated and smoothedto provide reliable micro motion information for mesoscopic clique clustering. Then thealgorithm of Mean Shift is used to cluster micro local motion, meanwhile micro categoriesare merged into mesoscopic crowd group according to velocity and location. Finally, thevelocity space is divided into normal area and abnormal area, the class whose clustercenter falls in abnormal area and sustains a certain time is judged to be the abnormalbehavior. Map all the sample points of the abnormal class from velocity space to imagecoordinate to obtain the location of the abnormal behavior.The experimental results show that the proposed method is capable of detectingglobal exceptions and local exceptions, as well as the locating of exception positions. Therate of correct reaches99.23%when detecting global exceptions and86.25%when it comes to local exceptions.