Research And Implementation Of Anomaly Detection Algorithm For Surveillance Video

In recent years,with the improvement of urban security system,intelligent monitoring technology has been widely used in all aspects of urban life,serving to maintain public security and stability and curb illegal crimes.Video anomaly detection,as a key technical basis for the construction of intelligent monitoring system,has gradually become another exciting research topic in the field of computer vision.It has farreaching research significance and great social value in social security,smart home,patient monitoring and many other fields.Driven by the upsurge of deep learning research,abnormal event recognition in surveillance video has attracted extensive attention from industry and academia.However,there are still some problems in the current video anomaly detection algorithm based on deep learning:Firstly,dual-flow anomaly detection algorithms based on appearance and motion usually use optical flow technology to learn the motion characteristics between adjacent frames of video,but optical flow only represents the pixel displacement relationship between adjacent frames and ignores the potential association problem of multiple consecutive frames.Secondly,the video anomaly detection framework based on future frame prediction describes the model of normal data and quantifies whether the current frame is abnormal according to the reconstruction quality training of the predicted single frame.However,using the reconstructed quality information of a single frame to realize the abnormal detection decision is not robust.The event is a continuous process including spatio-temporal factors.The single frame image only contains spatial dimension information,and the time dimension information is missing,which will inevitably affect the final model training and abnormal decision.Finally,there are few data sets in the research field of video anomaly detection,and frame/pixel level annotation is time-consuming and expensive.How to implement weakly supervised video anomaly detection algorithm based on video level tags is an urgent problem to be solved.In this regard,this paper carried out research and implementation of anomaly detection algorithm for surveillance video,mainly including the following three points:1.A video anomaly detection algorithm IP-VAD based on incremental prediction is proposed and implemented.First,the encoder is used to encode each frame independently to obtain the corresponding lowdimensional feature representation.Secondly,the differences represented by low-dimensional features of adjacent frames are aggregated in chronological order.The last frame is then reconstructed using the appearance decoder,the motion decoder predicts future deltas,and aggregates the last frame and future deltas for future frames.Finally,anomaly detection is realized according to the reconstruction quality quantization of future frames.Experimental results on two public data sets show that the incremental prediction method is better than the traditional integrated optical flow method.2.A fast anomaly detection algorithm VSPF based on video clip prediction is proposed and implemented.The VSPF algorithm designed video clip prediction to replace the traditional single frame prediction,and the temporal feature fusion method was designed to retain the temporal feature information of the predicted multi-frame sequence to the maximum extent,and obtain a more robust normal score according to the overall reconstruction quality of the predicted sequence.In addition,the algorithm can meet the high real-time requirements.3.Proposed and implemented IRE-MIL,a multi-instance learning weakly supervised video anomaly detection algorithm based on instance reconstruction enhancement.The premise of IRE-MIL algorithm execution is to use negative samples(normal video)to train models that only describe normal data.Firstly,a pre-trained autoencoder model is used to reconstruct all the examples in the positive and negative packets,and then features are extracted by deep neural network and input into fully-connected neural network to obtain the anomaly score of the examples.Finally,MIL sequencing model was used to train the positive sample score as much as possible.Experimental results show that the effect of the model is improved obviously.