Human-vehicle Abnormal Detection Method In Surveillance Video Based On Temporal CNN And Sparse Optical Flow

With the rapid development of computer technology,surveillance video has become more and more widely used in social public places and has become an important part of intelligent monitoring systems.Since most anomaly detection methods in surveillance video rely on complex feature extraction,there is still room for improvement in the detection performance and efficiency of video anomaly detection.In this paper,CNN is combined with the change in time-series to detect abnormal regions,and optical flow is used to further detect anomalies to improve detection performance and efficiency.The main research work of this paper is as follows:Firstly,the human and vehicle abnormal detection method in surveillance video based on temporal CNN and sparse optical flow is presented.After clarifying the main problems and data source on this research,the technical route of merging temporal CNN and robust sparse optical flow is determined,and the fusion method is expounded.Secondly,the abnormal detection method in surveillance video based on temporal CNN is proposed.By inserting a binary quantization layer into the fully convolutional neural network to form a binary convolutional network,the spatial relationship between the input image and the output features is determined;the temporal CNN pattern(TCP)is introduced,and a binary map is performed by the video block overlap calculation.And a TCP map and is obtained by up sampling for the initially detection on a video anomaly sequence.Then,the video anomaly detection method based on robust sparse optical flow is proposed.The robust sparse optical flow is extracted by foreground mask,finding good feature points,spatial sampling and forward-backward filtering;the optical flow histogram which can represent effective motion features is calculated;The robust sparse optical flow obtained above and the optical flow histogram are featureaggregated to further detect video sequence anomalies.Finally,the experiments and the evaluations were carried out on the actual collected surveillance video and standard data sets.The popular anomaly detection methods on video in recent years are compared with our method on the standard dataset.The results show that the error rate of the method is reduced and the accuracy is kept high.The actual effect of the method is tested on the actual collected surveillance video dataset,and it verified the effectiveness of the method.The research shows that the proposed method improves the efficiency of traditional optical flow,reduces the cost of clustering,and can quickly and efficiently detect human and vehicle anomalies in surveillance video.