PoseTrackNet: Integrating advanced techniques for accurate and robust human pose estimation in dynamic environments
Yuze He, Ke Chen, J. Hu · Alexandria Engineering Journal · 2025
Real-time pose estimation and tracking are essential to various applications like sports analytics, surveillance, and human–computer interaction. However, the current models suffer from failures in maintaining high accuracy with temporal consistency for different poses and complex movements, especially in dynamic environments. We propose a new model, namely PoseTrackNet, which is primarily designed for single-person pose estimation, featuring advanced components such as the Pose Correction Module, Deep SORT Tracking, and Keypoint Refinement. PoseTrackNet focuses more on enhancing accuracy and the robustness of keypoint prediction via refinement and ensuring temporal coherency across frames of a video. Specifically, PoseTrackNet is designed for single-person pose estimation, optimizing tracking accuracy in dynamic environments. Experimental results demonstrate that PoseTrackNet achieves superior performance compared to ten state-of-the-art models, with a rank-1 accuracy of 92.17% and a mAP of 88.73% in single-query scenarios. These results evidence its strength in dealing with challenging poses while maintaining precision across diverse conditions and, therefore, being suitable for real-time applications. This research will help further the field of pose estimation by providing a model that is more accurate and thus reliable for real-world applications. Our future work will concentrate on optimizing its computational efficiency and evaluating its performance on more diverse datasets.