Research Of Anomalous Crowd Behavior Detection Based On Video

Intelligent video surveillance has been become more and more important due to heightened security concerns.So as a branch of intelligent video surveillance technology,anomalous crowd behavior detection based on video has got the attention of researchers home and abroad.Thus,this paper mainly focuses on detecting of anomalous crowd behavior detection in the video surveillance system.The main work is as follows:Firstly,pre-processing and extracting the foreground.Mean filtering method is utilized to deal with the noise and the regular grid is used to divide the video sequences into a set of patches.The statistical averaging value of motion vectors is calculated in each patch to reduce the effect of noise and reduce the subsequent computation complexity,and the size of the patch is determined by the experiment;The k-means algorithm(k =2)is employ to extract the foreground area,and the threshold is set as the mean of the centroids of the two clusters.Secondly,the acceleration feature is extracted to detect the anomalous crowd behaviors in video surveillance systems.The acceleration is calculated by the image gray-scale invariance and the method of local optimization,and we evaluate the algorithm of the acceleration on the synthetic data which is known.When abnormal events happen,the velocity of the crowd will change a lot,so we employ the acceleration feature to reflect the change of the velocity.Anomalous crowd behaviors are detected by setting the acceleration feature to a suitable value.We appraise the proposed method on the UMN public dataset and the PETS2009 public dataset,and perform the comparative experiment between the proposed method and social force method.Finally,the divergent center is calculated to locate the anomalous crowd behavior.The graphical method and K-nearest neighbor search are employed to detect the one divergent center of the anomalous crowd event and the proposed method is analyzed by the simulation.Divergent centers imply possible places where abnormal events happen.When there are multiple divergent centers in the anomalous crowd event,the one divergent center method may not achieve accurate detection.Thus,the original algorithm is improved by employing the distance segmentation method to identify multiple different divergent centers.We analyze the performance by simulation and validate the proposed detection algorithm on public databases.