Human Pose Estimation Algorithm for Complex Scenes Based on Improved YOLOv8-Pose
Bingmin Chen, Kailin Lei, Jinchan Liu, Liuqi Lang · 2025
In sports scenes, there are complex factors such as multi-target occlusion, fast motion blur, and sudden changes in illumination, among others, and the traditional pose estimation algorithm is difficult to maintain computational accuracy while remaining lightweight, making it unsuitable for sports processing. To address the aforementioned issues, this study provides an improved approach based on the YOLOv8-Pose model for dealing with human pose estimation in complicated sports settings. This paper makes the following enhancements to YOLOv8-Pose. First, in the Backbone section, all of the original standard convolutional modules are replaced with the lightweight GhostConv module.This module generates redundant features through efficient linear transformation, which significantly reduces the computational complexity and parametric count of the model while maintaining the feature expression capability, making the model able to ensure the real-time monitoring requirements of sports pose processing. The DWR attention technique is innovatively implemented in the Head section. This attention mechanism may dynamically modify the sensory field weights, allowing the model to more correctly focus on the feature information of the critical portions when dealing with the task of estimating human posture, thereby addressing the issue of declining model accuracy. The experimental results show that, when compared to YOLOv8-Pose, the improved model significantly reduces the number of parameters by 87 %, with only 0.4 % loss of AP50 and 0.8 % loss of AP accuracy, and improves detection speed by 5.1 %, with each image detected in only 4.61 ms, making it suitable for accurate real-time monitoring and analysis of athletes' poses in sports scenarios.