Research On Video Anomaly Event Analysis Based On Spatio-Temporal Feature Enhancement

Video abnormal event detection is to effectively discover the situation and time status of abnormal events in video,which is a current research hotspot in the field of computer vision and multimedia.Video abnormal event detection analysis can be widely used in public places,traffic,security and other aspects of video surveillance,which can promote the development of a harmonious society and a safe society,with important economic and social benefits.At present,many researchers have made some progress in detecting video anomalous events from the segment or frame level in the video,and studied spatio-temporal network models for video anomalous event detection.However,most of these models are limited to local spatio-temporal feature expression,and there are also problems such as anomalous event detection generating anomalous fragmentation and inaccurate anomalous start and end positions,as well as the video anomalous event detection rate still needs to be improved.To address these problems,this thesis investigates video anomalous event detection methods based on significant sequence features and spatio-temporal graph networks and video anomalous event detection methods based on spatial consensus-driven graph networks to enhance the spatio-temporal feature representation of video anomalous events and improve the detection and classification performance of video anomalous events.The proposed methods are also used to design and implement a prototype system for video anomaly event analysis.The main contributions of this thesis are as follows:(1)A video anomaly event detection method based on salient features and spatiotemporal graph networks is proposed.Aiming at the current need of anomalous event classification and detection,and the problem that existing anomaly detection models produce anomaly fragmentation and spread of anomalous starting and ending positions,which lead to inaccurate classification of anomalous events,a video anomalous event detection and classification method using sequence salient features and spatio-temporal graph network is proposed,which establishes an anomalous event salient feature sequence screening and This method establishes an abnormal event sequence screening and refinement model based on the spatio-temporal fusion network abnormality detection model,eliminates the problem of fragmentation and spreading of abnormal segments,and achieves more accurate detection and localization of abnormal events.Comparing with the cutting-edge methods,the experimental results show that the proposed method reaches the cutting-edge level in video abnormal event localization performance and abnormal event classification accuracy.(2)A video anomaly event detection method based on spatial consensus-driven graph network is proposed.To address the spatial feature enhancement problem of video sequences,a spatial consensus-driven graph network-based video anomaly event detection and classification method is proposed,which adopts multiple spatial feature graphs with spatial heterogeneity for consensus-driven construction of spatial graph network structure,enhances the credibility of the graph structure,learns to express the spatio-temporal features of videos more reasonably,and improves the performance of anomaly event localization and classification.In the classification stage,in order to suppress the problem of category imbalance in training and normal categories being judged as abnormal,attentional consistency loss is proposed to improve the ability of abnormal classification more effectively.Experiments are conducted on two commonly used datasets,and the experimental results show that the proposed method improves the graph construction compared with the current cutting-edge research methods,and the model can locate the abnormal intervals more accurately,which can further improve the comprehensive performance of video abnormal event detection and classification.(3)A prototype system for video abnormal event analysis is designed and implemented.It aims to automatically detect abnormal events in the video,such as the sudden appearance of objects,abnormal behavior of people,etc.,and promptly notify relevant personnel for processing through alarms and reports.This system can present the abnormal detection results in a visualized form,which is convenient for users to make data analysis and decision.Besides,the operation interface is easy to understand,the process is simple and clear,and the practicality is strong.Finally,the detection results can be verified and modified through manual interaction,which increases the credibility of the detection results.