Research On Modeling And Detection Of Abnormal Behavior Based On Spatio-temporal Features

With the development of computer technology and innovation of video monitor technology,people pay more and more attention to smart video monitor.Abnormal behavior detection in video is an important research topic in the intelligent monitoring system,and it is becoming a much attention application field in recent years.Meanwhile,abnormal behavior detection has great academic value and excellent business perspective in public safety assurance due to its characteristic with real-time,smart and economic.It can analyze the behaviors happened in video monitor and give the alarm once abnormal happens,and shorten the time to solve problem.The study of abnormal behavior detection mainly focuses on two parts: the features representation of target behavior and the structure of detection model.The details of the study are listed as follow:1.In the monitoring video sequence,different locations between the camera and objects will influence the accuracy of features extraction due to distance,which results in misjudging behaviors.Therefore,this dissertation will different treat the video sequence according to the position to eliminate the effect of the distance between the camera and objects2.3D-SIFT,better robustness than optical flow,is used to extract objects features and transform 3D-SIFT features into entropy attribute,And a multiple attribute based detection model for abnormal behavior is constructed using the local attributes,including entropy attribute,optical flow velocity of each node,and the global attribute-KL divergence.3.Study the co-sparse representation based anomaly detection model.Combined with the co-sparse priori of object features,an abnormal event detection model based on co-sparse regularization is proposed,use 1l norm to reconstruct the object features..4.Study the group sparse representation based anomaly detection model.In view of group structural characteristics between features,the dissertation group features into parts,and introduce the group sparse representation model into abnormal events detection.The dissertaion classifys the dictionary according to the position,and obtain sub-dictionaries using PCA.And then,object features are reconstructed using sparse coding algorithm based on 2,1l.In order to compare and test the validity of the proposed models,this dissertation studies the abnormal behavior detection over the UCSD Ped1 dataset,UMN dataset and WEB dataset.Extensive experimentsshow that the proposed algorithms have good performance in abnormal behavior detection in accuracy and detection time.