Optimized design of 3D human pose estimation algorithm based on improved adaptive attention mechanism for sports training

Accurate three-dimensional (3D) human pose estimation is fundamental to sports training analysis, athlete performance monitoring, and rehabilitation guidance. Existing deep learning-based methods are primarily limited by two key deficiencies: (1) fixed receptive fields of convolutional networks hinder the effective learning of long-range joint dependencies, and (2) standard residual backbones fail to differentiate the discriminative importance of individual feature channels, which is particularly detrimental to depth coordinate regression accuracy in complex athletic scenarios. In this paper, we propose a novel 3D human pose estimation network enhanced by an Improved Adaptive Attention Mechanism, specifically optimized for sports training scenarios. Building upon a weakly supervised transfer learning baseline, we introduce a Residual Attention Module (RAM) that integrates ECANet into the ResNet50 residual blocks to reinforce channel-wise feature discrimination critical for depth estimation. Additionally, we design an Adaptive Attention Mechanism based on Batch Normalization initialization to ensure seamless compatibility with pre-trained weights, preventing the disruption of learned semantic information during early training. Extensive experiments on the Human3.6M benchmark demonstrate that our method achieves significant performance gains, reducing the Mean Per Joint Position Error (MPJPE) from 64.9 mm to 58.6 mm–a 9.7% relative improvement. To explicitly validate the model’s utility in real-world sports training, we extend our evaluation to the AthleticsPose dataset, featuring authentic high-speed athletic motions. Furthermore, comprehensive evaluations on 2D sports datasets yield state-of-the-art results, including a 94.3% PCK@0.2 on the LSP dataset, with notable accuracy gains in challenging distal joints such as wrists and ankles. Ablation studies validate the distinct and cumulative contributions of the channel and spatial recalibration modules. Finally, qualitative assessments and inter-joint cosine similarity matrix analyses confirm that the proposed architecture markedly improves the modeling of structural biomechanical relationships and provides robust depth estimation under severe self-occlusions and complex asymmetric postures–capabilities directly beneficial for precise technique analysis in athletic environments.