IoT-enhanced multi-attention and lightweight feature integration for human pose estimation in motion training systems

Abstract

Human pose estimation is widely used in intelligent sports training, rehabilitation assistance, and human–computer interaction, providing precise motion feedback and training guidance. However, existing methods suffer from keypoint localization errors and insufficient global coherence in complex backgrounds, occlusions, and high-dynamic motion scenarios. To address these challenges, this paper proposes a hybrid IoT-vision deep learning model for human pose estimation and motion training feedback. IoT-based motion sensors are integrated with vision-based keypoint detection to enhance pose estimation accuracy, particularly in occluded or high-speed movement scenarios. The model employs the LSFE stacked feature extraction module to enhance multi-scale feature adaptability, incorporates the LFAM local attention mechanism (SPAM + CARM) to improve key joint modeling, and introduces the GEAM global enhancement module to ensure keypoint stability and consistency. Additionally, an ECA-based lightweight channel attention mechanism reduces computational complexity while enhancing key feature responses. Experimental results show that the proposed model achieves Mean Accuracy of 0.946 and 0.949 on the LSP and MPII datasets, respectively, with PCK scores of 0.97 and 0.95. This model demonstrates significant improvements over existing methods in real-time performance and robustness, particularly in complex scenarios such as sports training, rehabilitation, and monitoring.