Research Of Algorithm For Crowd Abnormal Behavior Detection Based On The Technology Of Approximate Nearest Neighbor Search

In recent years,crowd abnormal behavior detection has become a vital part in the field of intelligent video surveillance analysis since there is an increasing demand of safety and security assurance for the public.With the rapid development of the information technology and the increasingly mass media,the issue that how to accurately extract and conclude the information that people truly care about from the massive video data has been a great focus for researchers.As an effective way to assist the safety assurance department for population management and safety monitoring in the target areas,crowd abnormal behavior detection can help to discover the anomlies in time and then reduce the potential loss to the minimum.So it has been a hot area of research.Although there are already many existing anomaly detection algorithms,it is still a challenge to detect and locate the anomalies accurately.By researching the existing algorithms in this area,in this paper,the difficulties and problems of anomaly detection methods are analyzed.Further,the novel methods for anomaly detection,which are based on the technology of approximate nearest neighbor search,are proposed,including the anomaly detection method based on Locality Sensitive Hashing and the method based on Locality Sensitive B-tree.The main innovations of this paper are listed as follows:1.A novel feature called Dynamics of Pedestrian Behavior(Do PB)is proposed.This feature can effectively describe the behavior and motion of the pedestrians.By analyzing the existing methods,it can be found that almost all of methods only use the motion and appearance cues when describing the crowd behaviors,but toally ignore the information of the dynamic changes of the motion and appearance.Do PB focus on the dynamic change,including the changing rate of pedestrian's motion and appearance in a certain period of time.By analyzing the dynamics,it can help to detect many anomalies,like the drastic behavior and the abnormal translation motion.2.The method for pedestrian detection and tracking based on RPCA(Robust Principle Componet Analysis)is proposed.In this method,RPCA is applied to detect and extract the moving targets,then individual pedestrian is segmented and tracked by matching each component between the adjacent frams.Since there might be some problems that the videos may have low resolution,the moving targets may have the similar gray value with the background,many tranditional algorithms for object detection and tracking sometimes can not detect all of the targets successufully.The application of RPCA can help to effectively solve this problem,thus each pedestrian can be detected and tracked.3.A novel anomaly detection method based on Locality Sensitive Btree is proposed.On the basis of realizing the anomaly detection method based on Locality Sensitive Hashing,this paper further proposes an improved method,which is based on Locality Sensitive B-tree,to describe the pedestrian behavior and detect the anomalies.Locality Sensitive Hashing is a commonly used method in approximate nearest neighbor search technology,which can cluster the data into several clusters.The appromate nearest neighbors can be found in the cluster where the query belongs.Locality Sensitive B-tree can achieve fast search of approximate nearest neighbors by reflecting and encoding the data in the new feaure space.In the proposed method,a Locality Sensitive B-Forest is constructed with several Locality Sensitive B-trees in it and each feature vector is encoded then.By analyzing the similarity between the encoded data,the approximate neareast neighbors can be found.The anomaly saliency is calculated based on the distances between the query and its nearest neighbors,and then the anomaly is detected and located.