Research On Video Anomaly Detection Algorithm Based On Enhanced Spatio-temporal Features

With the development of society,video surveillance cameras are spread all over the places where people live and study,ensuring social order and the safety of people's lives and property.Detecting abnormal events automatically from videos is the development requirement of modern video surveillance systems.According to the development requirement,video anomaly detection technology is derived,and video anomaly detection technology has become a current focus of research.Video is high-dimensional data that contains both timing information and image information.How to extract complete spatiotemporal features to express the basic events in the video has become a difficult point in current research,and the probability of occurrence of abnormal events in the video is low.When testing video anomaly detection,the dominant normal video clips could easily overwhelm the abnormal video clips,leading to failing to declare the abnormal events.In this article,the key research is enhancing video spatio-temporal features to express complete video basic events and enhancing the recognition of abnormal events to improve the accuracy of video anomaly detection.The main contents include:(1)Aiming at the problem that the traditional two-dimensional convolutional neural network in the image field cannot effectively use in the timing information on video,this paper proposes a 3DGRU video anomaly detection algorithm based on enhanced spatiotemporal features.The Chinese full name of the 3DGRU algorithm is that a multiinstance video anomaly detection algorithm basing on C3 D and GRU network to enhance spatio-temporal features.The algorithm is based on the three-dimensional convolutional neural network named C3 D to initially extract the temporal and spatial features of video blocks,and enhancing the long-term timing information and short-term timing information expression in the video features through the GRU structure,integrating the C3 D network and the GRU network to build a complete video basic event temporal and spatial expression Features,to detect video anomaly events by multi-instance ranking loss function.In experiments on the UCF-Crime data set,the 3DGRU algorithm proposed in this paper is compared with the MF algorithm in 2020,and the 3DGRU algorithm's AUC value has increased by 1.58%.(2)Aiming at the problem that abnormal video clips are easily affected by the predominant normal video clip bias,resulting in underreporting of abnormal events,this paper proposes an ITBM video anomaly detection algorithm based on enhanced spatiotemporal features,on the basis of enhancing the temporal and spatial characteristics,the recognition of abnormal events is improved.The Chinese full name of ITBM algorithm is a binary cross-entropy video anomaly detection algorithm based on I3 D and TCN network to enhance spatio-temporal features.In the feature extraction stage,the algorithm uses I3 D to extract the RGB and optical flow fusion features of the video frames,using the TCN time convolutional network to explicitly encode the feature values of the video segments,in order to enhance the multi-scale temporal correlation of adjacent video segments;In the anomaly detection stage,a joint loss function of cross entropy and mean square error is proposed.The center entropy loss function is used to expand the distance between normal video clips and abnormal video clips,and the mean square error loss function is used to reduce the intra-class distance of abnormal video clips.The joint loss function improves the separability of normal videos and abnormal videos,and improves the recognition of abnormal video clips,thereby reducing the false negative rate of abnormal events in video abnormal detection and improving the accuracy of abnormal event detection.In experiments on the Shanghai Tech data set,the ITBM algorithm proposed in this article is compared with the AR-Net algorithm in 2020,and the ITBM algorithm's AUC value has increased by 2.14%.