LDTA-Pose: Advancing 2D Human Pose Estimation with Lightweight Dynamic Task Alignment based on modified YOLOv8

2D Human Pose Estimation plays a crucial role in analyzing performance in fitness activities. Current single-stage methods suffer from the lack of interaction between classification and regression branches, large network parameter sizes, and poor detection accuracy due to strong short-distance dependencies between keypoints. To address these issues in fitness, a lightweight dynamic task alignment framework based on Yolov8-pose is proposed. In order to enhance classification and regression alignment in single-stage networks, a dynamic task alignment detection head is proposed by leveraging label assignment strategies and learning task interaction features. To mitigate the issue of information loss caused by the unidirectional propagation in Yolov8-pose, the backbone is replaced with RevCol to enhance feature retention. Additionally, efficient self-adjusting weighted downsampling module is designed to retain more useful information. Furthermore, the C2f module in downsampling is enhanced with Context-Guided Blocks, integrating local and global feature fusion. Experimental results on a self-created fitness action dataset show that, compared to Yolov8n-pose, our proposed algorithm reduces parameters by 60.8%, decreases computational cost by 33.7%, and improves average detection accuracy by 1.89%.