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

Area of Science:

• Computer Vision
• Sports Analytics
• Machine Learning

Background:

• Athlete pose estimation faces challenges in complex sports, including multi-person confusion, temporal instability, and occluded joint inaccuracies.
• Existing methods struggle with the high dynamics and occlusions common in athletic activities.

Purpose of the Study:

• To develop an advanced pose estimation network, STG-DB Net, specifically designed to overcome the limitations of current models in sports environments.
• To enhance the accuracy, temporal stability, and robustness of athlete pose estimation under challenging conditions like occlusion and high-speed motion.

Main Methods:

• A three-tier architecture: instance separation (STGCN-IS), scene modeling (DSTformer), and pose refinement (BCQ).
• STGCN-IS uses bounding box constraints and skeletal clustering for multi-person separation.
• DSTformer employs dual-branch attention with occlusion-aware spatial attention and dynamic temporal windows.
• BCQ module integrates Behavior Cloning and Q-learning with sports-specific action templates for pose refinement.

Main Results:

• STG-DB Net outperforms eight baseline models on COCO Keypoints, MPII Human Pose, and LSP datasets across key metrics.
• Significant improvements observed: 5.7-16.9% AP50 in multi-person scenarios, 5.2-9.2 pixels decrease in MPJPE for dynamic movements, and 4.8-9.3% Recall increase in occlusion scenarios.
• Ablation studies confirm the critical contribution of STGCN-IS and DSTformer modules, with their removal causing 7.4-10.3% performance decline.

Conclusions:

• STG-DB Net demonstrates superior performance in athlete pose estimation within complex sports scenarios, addressing key challenges of multi-person tracking, temporal stability, and occlusion.
• The proposed architecture effectively enhances pose accuracy and motion plausibility using publicly available datasets.
• Future work will focus on multi-modal fusion, extreme scenario augmentation, and adaptive template generation to improve specialized optimization and generalization for diverse athletic disciplines.