BCST-GCN: a skeleton-based spatiotemporal graph convolutional network with bidirectional cross-attention for pig behavior recognition

To address the issues of weak inter-frame motion correlation and poor recognition robustness in video-based pig behavior recognition, as existing methods fail to fully exploit the spatiotemporal dynamic features of skeletons and can hardly capture fine behavioral details, this study proposes a skeleton-based spatiotemporal dynamic modeling method for pig behavior recognition. We use DeepLabCut (DLC) to accurately extract pig skeleton keypoints and construct the topological structure, streamline the ST-GCN by removing redundant network layers, and design an improved BCST-GCN model with a global-local self-attention BC module to dynamically reconstruct topological correlations, so as to effectively capture non-physical connections and complex spatiotemporal behavior characteristics. Experimental results show that the proposed framework can effectively recognize typical behaviors such as feeding, walking, lying, and dog-sitting posture, and the improved model yields 6.94%, 5.61%, and 6.88% increments in accuracy, precision, and recall respectively compared with the baseline model. The proposed method achieves accurate and efficient pig behavior recognition, solves the problems of weak temporal correlation and insufficient feature extraction in traditional models, and provides a reliable technical solution for intelligent monitoring in pig farming scenarios, supporting the intelligent upgrading of the breeding industry.