Adaptive Graph Convolution Embedding Normalizing Flow for Video Anomaly Detection

Abstract Video anomaly detection is a challenging problem characterized by significant uncertainty, as it is influenced by various factors including appearance, posture, light intensity, scene changes, and etc. To alleviate the influence of the above factors, this article focuses more on the description of skeleton features. We use the Normalizing Flow framework to process human pose data to reduce the impact of strong generalization capabilities of other models. However, Normalizing Flow algorithms were unable to fully utilize the dependencies between non-directly connected nodes when convolving posture data, and ignored the spatial differences of nodes in the human body structure. Hence, we propose an improved video anomaly detection model called Adaptive Graph Convolution Normalizing Flow(AGC-NF). Our model can generate relationships between nodes that are not directly connected, and can adaptively update the connection relationship and connection strength between each node through training. Simultaneously, the incorporation of spatial distance constraints into node division amplifies local distinctions. After experimental evaluation, AGC-NF achieves improvement over mainstream Normalizing Flow methods on ShanghaiTech-HR datasets and UBnormal datasets.