Anomaly Detection Based On Video Target Analysis Under Surveillance Scenarios

In recent years,with the development of economy,science and technology,peo-ple's demand for safety precautions has a sharp increase.China has also continuously implemented various policies to promote the development of related intelligent moni-toring industries.As a core part of intelligent monitoring,anomaly detection has a wide application in the field of public safety.Anomaly detection means that the computer can automatically detect the abnormal behavior of the characters in the surveillance video by designing an algorithm.Anomaly detection based on video target analysis under surveillance scenarios has important theoretical significance and research value.In the video surveillance,anomaly detection still has many challenges needed to be solved:(1)Most of abnormal behaviors usually are highly related to motion in the real scenes.The existing motion features usually focus on the direction information of optical flow and ignore the amplitude information of optical flow.Thus,these features can not extract the sufficient motion information.(2)The abnormal behaviors may have different definitions in different scenes.The existing anomaly detection algorithms usually perform binary classification to the videos:normal or anomaly.Thus,these methods can not give the reason of anomaly.The thesis studies anomaly detection based on object analysis under video surveillance and mainly focuses on solving the problems as described above and improves the performance of anomaly detection.The main work and innovations are as follows:1.An anomaly detection algorithm based on HMOF feature and tracking is pro-posed.Considering the inadequacies of traditional motion features that cannot effec-tively extract motion information,the proposed algorithm presents a new HMOF mo-tion feature.Compared with the existing feature descriptors,the HMOF feature is more sensitive to the amplitude of the optical flow and more efficient to distinguish anomaly information.Firstly,the foreground patches of the video are extracted.Next,the HMOF features are computed for foreground patches,and are reconstructed by the auto-encoder for better clustering.Then,Gaussian Mixture Model(GMM)Classifiers are used to distinguish anomalies from normal activities.Finally,visual tracking module is used to track the abnormal patches,which ensures the continuity of anomalous events in temporal and spatial domain.Extensive experiments show that the proposed algorithm not only achieves better performance than other existing algorithms,but also can detect anomalies in real time.Although the speed is reduced after adding the tracking module,the performance of the algorithm is further improved.2.An anomaly detection algorithm based on multivariate fusion is proposed to detect anomalies by analyzing the visual concept of each target in the surveillance video.Among them,the proposed algorithm will extract the visual concept of the target through three branches:object,action and motion.The object branch focuses on the appearance information of the target,the action branch focuses on the action category of the target,and the motion branch focuses on the distribution of the motion features.Although the focus of these branches is different,they can complement each other and jointly detect abnormal behavior.The final abnormal score can then be obtained by combining the results of the three branches.In the action branch,an action recognition module using inter-frame information is proposed to solve the multi-target and multi-action recognition in the surveillance video,which is not utilized before in the anomaly detection field.Extensive experiments show that the proposed algorithm outperforms the state-of-the-art methods and also can explain the reasons of the anomalies from mul-tiple perspectives.