Research On Abnormal Behavior Detection Technology Based On Sparse Structure Feature

The Video Surveillance Technology is constantly upgraded with the great support of the progressive science and technology,and more and more attentions are paid to how to realize the intelligent video surveillance.Anomaly detection,which mainly aims to intelligently recognize and analyze the behavior features of the crowd in scenes with large population mobility and high population density,is an inevitable problem to achieve the intelligent video surveillance.Once abnormal behaviors are found,alarm signal will be sent automatically so as to improve the emergency respondence ability of relevant departments and timely deal with abnormal problem.Therefore,it is necessary to intelligently recognize and analyze the behavior feature in the video surveillance scene.Moreover,anomaly detection has significant practical value in terms of public security,which means it worth further research.The key parts of abnormal detection lie in the feature representation of the object behavior and the construction of detection model.The main works in this dissertation about study of abnormal detection are as follows.Distinguish the moving objects block and background blocks based on the average optical flow amplitudes.If the average optical flow amplitude is larger than the threshold,extract features for the object block.This method helps to reduce the erroneous judgment of identifying the slight shake of trees nearby as a moving object when the global threshold extracts moving objects,so that the detrimental effect of the extracted features caused by the variable distances between the camera and the objects,can be improved.Next,study a co-sparse representation model for anomaly detection.Normal behavior features are co-sparse,while anomaly ones are not.Combined with the prior,an anomaly detection model based on co-sparse constraint is proposed.In the train stage,feature samples set are clustered into K clusters and an analysis dictionary is learned for each cluster.In the detection stage,extract the group features from object blocks and select the best dictionary for each test feature according to the nearest neighbor rule.Experimental results show that the proposed achieves,comparable performance.Eventually,study an effective model called Dirty.In order to make full use of the neighborhood information of the moving object,extract the group features for detection and judgment.In view of the group structure among feature,features are divided into groups.Generally,the spatial distribution of normal features tend to be more concentrated and shows similarities,but also exists differences.Considering the diversity of normal features,the normal sample can be clustered into different classes.Apply the Dirty model to the anomaly detection to discribe the relationship among all kinds of normal samples.l_1,?-norm represents similarity,l?1,1?-norm represents dissimilarity and then the Coordinate Descent method is used to solve the Dirty model.The best feature from the clustered is selected to build a dictionary.The experimental results show that the Dirty model has good experiment result as well as less time consuming.Experiments for local and global anomaly detection,operated on the UCSD Ped1 data set and the UMN data set,indicate that the two proposed have achieved good detection results.