STG-DB net for athlete pose estimation and pose rational optimization in complex sports scenes

To address technical challenges inherent in athlete pose estimation within complex sports scenarios-specifically the confusion of multi-person skeletal instances, insufficient temporal stability during highly dynamic movements, and degraded localization accuracy for occluded joints-this paper proposes STG-DB Net, a pose estimation network tailored for sports environments. The model constructs a three-tier collaborative architecture comprising instance separation, scene modeling, and pose refinement. First, the STGCN-IS module is introduced to incorporate bounding box constraints and a skeletal distance clustering strategy, thereby achieving effective separation of multi-person skeletons and mitigating the aliasing of joint features across individuals. Second, a dual-branch attention structure-DSTformer-is designed to integrate occlusion-aware spatial attention with adaptive dynamic temporal windows, thereby enhancing the model's feature modeling capabilities under conditions of occlusion and high-speed motion. Finally, a BCQ pose refinement module is constructed based on a combination of Behavior Cloning and Q-learning; this module leverages statistical action prior templates derived from sports datasets to apply smoothing and motion plausibility constraints to pose sequences, thereby improving the temporal consistency and motion distribution fidelity of the output poses. Experimental results demonstrate that, across the COCO Keypoints, MPII Human Pose, and LSP datasets, STG-DB Net outperforms eight baseline models across six core metrics: in multi-person co-occurrence scenarios, AP50 improves by 5.7-16.9%; in highly dynamic scenarios, MPJPE decreases by 5.2-9.2 pixels; and in occlusion scenarios, Recall improves by 4.8-9.3%. Ablation studies further confirm that removing either the STGCN-IS or DSTformer module results in a decline of 7.4-10.3% in key performance metrics. This paper quantifies and validates pose estimation accuracy and motion plausibility solely based on publicly available datasets; specialized evaluations regarding professional sports technique correction, practical training guidance, and real-world deployment effectiveness have not yet been conducted. Consequently, there remains room for improvement in the model's specialized optimization capabilities for specific sports actions. Furthermore, the model currently faces performance bottlenecks in scenarios involving complete occlusion, ultra-high-speed extreme sports, and niche athletic disciplines. Future research will incorporate multi-modal information fusion, data augmentation for extreme scenarios, and strategies for the adaptive generation of dynamic templates to further enhance the model's generalization and adaptability within complex sports environments.