Video anomaly behavior detection method based on attention-enhanced graph convolution and normalizing flows

Video anomaly behavior detection play a critical role in intelligent security, healthcare, and other domains, driven by the rapid advancement of video surveillance technologies. However, existing anomaly detection methods often focus primarily on local features when learning the strength of skeletal connections, neglecting global connectivity and feature channel information. This limitation hampers improvements in detection accuracy, making it difficult to enhance overall performance. To address this issue, we proposed a Video Anomaly Behavior Detection Method based on Attention-Enhanced Graph Convolution and Normalizing Flows. Initially, the spatial attention graph convolution technique is employed to acquire the spatial global characteristics of skeletons. Simultaneously, the channel attention graph convolution method is applied to capture the channel information of the spatial features. Subsequently, the spatial attention graph convolution and channel attention graph convolution are combined simultaneously within a spatiotemporal graph convolutional network, creating an attention-enhanced graph convolutional network. Additionally, we propose a detection framework that incorporates normalizing flows, which maps feature information to a Gaussian model serving as a prior distribution. This enables the calculation of the probability distribution for normal behaviors, using the likelihood of the Gaussian model to assess new samples. Experimental results demonstrate that the proposed method achieves superior detection performance with reduced false alarm rates on the ShanghaiTech and HR-ShanghaiTech datasets.