Abstract: Video anomaly detection is widely applied in fields such as public security, transportation, and healthcare. However, human pose anomaly detection faces issues including susceptibility to environmental influences, difficulty in handling skeleton timelines, high computational complexity, and easy neglect of local important features in motion regions. To address these problems, a novel model  based on human skeleton, named spatiotemporal graph normalizing flow mixed attention state space model (STG-FAM), is proposed. This model effectively captures temporal dynamic features in skeleton timelines by introducing a selective state space model and normalizing flow into the spatiotemporal graph convolutional network. It utilizes a mixed attention mechanism to learn attention weights across channels and spatial domains, thereby enhancing the model’s focus on key nodes and spatiotemporal edges in the temporal skeleton and improving the model’s representational capacity and anomaly detection performance. The effectiveness of the proposed model is demonstrated  through experiments on two video anomaly detection datasets: the ShanghaiTech Campus dataset and the UBnormal dataset.

Key words: video anomaly detection, human skeleton, graph neural network, state space model, attention mechanism