Research On Video Anomaly Detection Technology Based On Visual Saliency

With the increasing demand for public security and the construction of security projects such as safe city and Xueliang project,intelligent video surveillance technology has been widely used in security fields such as intrusion and congestion detection.The deployment of a large number of surveillance cameras has generated a huge amount of video data.Relying solely on manual surveillance video has been unable to meet regulatory needs.Therefore,there is an urgent need for a technology that automatically identifies and records abnormal events and provides early warning.Based on this demand,this paper proposes a video summarization and abnormal event detection technology based on visual saliency.It systematically studies how to remove a large amount of redundant information in the video by predicting the saliency information,so as to automatically judge and detect abnormal events according to the spatio-temporal key information.The proposed method can not only filter the key information in the long video sequence and form a video summary to provide assistance for manual video supervision,but also automatically identify the abnormal events in the video and quickly locate the starting time of the abnormal events.The details are as follows.In video saliency detection,aiming at the problem of inaccurate prediction caused by rotation and scaling of salient targets in the process of motion,a video saliency detection method based on convolution trajectory gated cyclic unit network was proposed.A context-aware pyramidal attention network was used to extract features and improve feature representation;a trajectorygated recurrent unit network fused spatio-temporal features to maintain spatio-temporal consistency;Finally,the saliency graph was output.The experimental results on the public dataset showed that the improved method improved the accuracy and can predict accurate results in complex scenes.In video summarization,traditional video summarization was difficult to fully express semantic information and subtle differences between adjacent frames,resulting in low accuracy,resulting in low accuracy,a video summarization method based on visual saliency was proposed.The frame importance score was calculated based on the saliency map results and the original video frame,so as to guide the key frame extraction.Also,time filtering was used to optimize the synthetic videos.Experiments on public datasets showed that the improved method achieved the optimal sampling proportion and the highest sampling difference ratio in critical and non-critical periods,and the user questionnaire results also got high votes.In video anomaly detection,aiming at the problem of poor model generalization performance caused by uneven distribution of positive and negative samples in the training set of abnormal event detection,a video abnormal event detection method based on visual saliency and MultiInstance Learning was proposed.The spatio-temporal features were non-uniform segmented based on the frame importance score,and the segmentation results were used as examples in the positive and negative packets.The Multi-Instance Learning model was used to judge whether an abnormal event had occurred in the video.The experimental results on public dataset showed that the non-uniform segmentation feature can improve the accuracy,and the anomaly score curve showed that the algorithm can also detect anomalies in real time.