Research On Video Abnormal Behavior Detection Algorithm Based On Deep Learning

Video anomaly detection is a highly challenging task in unsupervised video analysis.Surveillance video anomaly detection has gained widespread attention owing to its applications in public security,social security management,and the rising trends in deep learning and computer vision.However,the surveillance video in real life has the problems of illumination change and chaotic background,and it is difficult to distinguish the influence of similar behavior,resulting in unsatisfactory detection of human abnormal behaviors.This paper studies the detection algorithm of human abnormal behavior in surveillance video based on deep learning.Deep learning technology is applied to the detection task of abnormal behavior related to human,and a new abnormal behavior detection network is constructed to realize the effective detection of human abnormal behavior in public areas.The main research contents of this paper are as follows:1.Aiming at the problem that the spatiotemporal graph convolutional network(ST-GCN)only acts on local neighbor nodes and lacks global information and needs to overcome the problems of illumination changes and chaotic backgrounds in surveillance videos,this paper proposes a self-attention augment graph convolution algorithm based on skeleton features to detect abnormal human behaviors.Skeleton data has been shown to be robust to complex backgrounds,illumination variations,and dynamic camera scenes,and are naturally constructed as a graph in non-Euclidean space.In order to solve the problem that spatiotemporal Graph convolutional network(ST-GCN)only acts on local neighborhood nodes and thus lacks global information,a new spatiotemporal self-attention augmented Graph convolutional network(SAAgraph)is proposed in this paper.The spatial graph convolution operator and the improved Transformer self-attention operator are combined to capture the local and global information of joints.Extensive experiments on two large open standard datasets,namely the Shanghai Tech Campus and CUHK Avenue datasets,reveal the latest performance of the proposed method compared to existing skeleton-based and graph convolution methods.2.Aiming at the problem that it is difficult to distinguish the behaviors of partial similar motion patterns due to the single skeleton mode and the lack of global time information,a skeleton-guided multi-modal abnormal behavior detection algorithm is proposed.In order to make full use of the advantages of RGB video mode and skeleton mode for abnormal behavior detection under similar behaviors,the action behavior features extracted from skeleton mode are used as guidance,and a new spatial embedding is used to strengthen the corresponding relationship between RGB video and skeleton pose.At the same time,the inter-frame relationship between the same nodes is extracted by using temporal self-attention to capture the global information of time and effectively extract discriminative abnormal behavior features.In two largescale public standard datasets,the results show that the proposed method can effectively strengthen the temporal relationship and modal correspondence of skeletonguided multimodal features,enabling more accurate detection of abnormal behaviors with similar motion patterns.This paper mainly considers the use of skeleton data to eliminate the interference of complex backgrounds,and conducts research on abnormal behavior detection methods based on skeleton-based spatiotemporal self-attention augmented graph convolution to extract the complex spatiotemporal relationships of human video skeletons.And by utilizing the complementarity of skeletal modal data and RGB modal data,a skeleton-guided multi-modal abnormal behavior detection method is proposed to extract the spatiotemporal robust representation of the interaction relationship in the video for abnormal behavior detection under similar behaviors.Its significance is to improve the detection performance of abnormal human-related behaviors,and to reduce false detections to a certain extent in practical applications in the future.