Abnormal Behavior Recognition Method Based on Fine-Grained Human Dynamic Skeleton Features

To address the insufficiency of feature extraction in existing video anomaly recognition methods, this paper proposes an improved graph-embedded pose clustering-based abnormal behavior recognition method. By employing fine-grained feature extraction, it enhances information representation capabilities and improves algorithm robustness. The framework is structured as follows: First, spatial and channel reconstruction convolutions are integrated into a Deep Residual Network backbone, effectively eliminating spatial-channel redundancy in extracted features while reducing computational complexity, thereby enhancing operational efficiency. Second, a hierarchical decomposed graph convolutional network (modified from spatio-temporal graph convolutional networks) is implemented for dynamic skeleton feature extraction, coupled with an attention-guided hierarchical aggregation module for multi-level feature fusion. Evaluations on ShanghaiTech and NTU-RGB+D datasets demonstrate detection accuracies of 76.4% and 74.6% respectively, validating the method's effectiveness.