Large Scale Crowd Local Motion Detection By Weighted Network Approach

Intelligent video surveillance is the research emphasis and difficulties in the field of artificial intelligence.The main content of intelligent monitoring system is detecting and analyzing crowd behavior in the video surveillance.We analyze crowd movement state from the consecutive frames in the image by extracting the crowd movement characteristics.When abnormal behavior occurs,alarm prompt warning in a timely manner,so we can reduce casualties and property losses in the largest extent.How to detect the local retrograde motion and the local abnormal movement area of large-scale crowd scenes is the main focus of this paper.Firstly,using the improved pyramid optical flow algorithm based on Lucas-Kanade to deal with crowd scene video sequence frame,getting the crowd velocity vector field,and regarding every velocity vector in two dimensional velocity vector field as a node in crowd weighted network;Secondly,the interrelation between velocity vectors is determined by the included angle between the two velocity vectors obtained by the velocity vector dot product formula,and using this value to evaluate the correlation degree between every two velocity vectors,Thus,the crowd weighted network is constructed,consequently,getting the adjacency matrix that represents the movement information of the crowd;Finally,analyzing the crowd weighted network model of local retrograde scene and local motion anomaly scene of large-scale crowd scenes,extracting node strength in crowd weighted network as characteristic parameter to construct two dimensional matrix,then the local retrograde and local abnormal motion areas are detected.In this paper,we conduct experiments on multiple groups of video in database,the results show that node strength is larger in the abnormal movement area and then detecting it.With respect to the local motion anomaly scene,we compare with other method which is based on a framework in which Lagrangian Particle Dynamics is used for detection of flow instabilities,the results obtained by this algorithm are nearer to the real value.