Anomaly Detection In Surveillance Scenes Based On Deep Learning Method

With the progress in science and technology and the promotion of "smart city" and"safe city",video surveillance technology is used widely.There are many problems in traditional video surveillance systems,such as low efficiency of monitoring by human beings,redundancy storage,can't analyze videos automatically and warning anomaly and so on.In this context,intelligent video surveillance has become an urgent need in the field of public security.Anomaly detection plays an important part in intelligent video surveillance.It can analyze and detect anomalies accurately and quickly,so that the human observers can handle them in time.It is important for the development of city security and has always been the concern of both academic and industry.For the actual urban video surveillance scene,there are some problems and chal-lenges.The main performances are as follows:1)Affected by illumination,occlusion and so on,missing detections are likely to occur in object detection,which affects the performance of subsequent anomaly detection;2)The surveillance scenes are complex and diverse,many methods used low-level features or single kind of features,which are not sufficient and not discriminative for video representation;3)Most of the present algorithms only detect abnormal event,and do not explain why the event is judged to be abnormal.Based on these problems,this paper studies the abnormal behavior detection algorithm under video surveillance scene as follows:1.An anomaly detection algorithm based on the improved fusion of appearance and motion information of objects is proposed.Most of the previous algorithms divided the video into blocks,which leads to the segmentation of the target,resulting in the loss of motion information and affecting the performance of anomaly detection.In this paper,we apply the object-centric method to handle the problem,and the deep object detec-tion algorithm SSD is used for object detection.Considering that the low-level motion features used in traditional algorithms are not sufficient to represent the motion infor-mation,the paper extracts the optical flow based on the detected object,and develops a new effective multi-scale histogram of optical flow with local energy(MHOFE)to cap-ture motion cue.Then,a two-stream autoencoder is used to learn the anomaly detection model utilizing the normal samples of RGB images and optical flows,respectively.The latent features obtained in this way are more discriminative for anomaly detection.Fi-nally,the post-processing module based on tracking is proposed to find the abnormal objects missed by occlusions in detection.The experiment results show that the pro-posed algorithm provides superior results of anomaly detection and location compared to state-of-the-art methods.2.An anomaly detection algorithm based on multiple visual concept analysis of target is proposed.In the actual intelligent monitoring system,the detection and re-counting of the abnormal event are crucial,which can help human observers quickly judge if they are false alarms or not.This paper focuses on the recounting of anomalies,learning the basic visual concepts of targets through multiple branches:object,action and motion states,then incorporating them into the unified framework of anomaly de-tection.These branches jointly learn semantic information about the appearance infor-mation and motion states of the target.This paper proposes an improved action recogni-tion module,which combines object tracking with inter-frame information to solve the recognition problem of multi-target and multi-action recognition in surveillance videos.In the motion states branch,a multi-class approach based on k-means and one versus rest SVM is employed to separate each normality cluster from the rest,which can well model the difference of normal and abnormal motion patterns.Finally,the anomaly is detected by the joint analysis and fusion results of each branch.Experimental re-sults show that the algorithm can effectively detect anomalies and can also describe the events in the human-understandable form for event recounting.