Spatiotemporal pose lifting for tennis motion analysis

Tennis techniques are characterised by high openness, rapid speed, and complex limb coordination. Traditional evaluation methods based on manual observation or laboratory motion capture struggle to support efficient, precise, and continuous monitoring in real training scenarios. To address challenges in pose estimation during high-dynamic movements – such as occlusion, complex backgrounds, and temporal noise – this paper proposes PoseCResNet-R, a 2D human pose estimation model based on multi-scale feature fusion and spatio-temporal consistency correction. PoseCResNet-R performs 2D pose estimation, while STSPose refines the 3D motion sequence with temporal upsampling, ensuring smooth motion reconstruction. By integrating a Transformer-based temporal structure, we further develop STSPose, a 3D pose refinement model that achieves high-precision reconstruction of continuous 3D motion sequences from sparse 2D keypoints. Experimental results demonstrate that this framework achieves superior keypoint localisation stability, motion sequence smoothness, and inference efficiency in open tennis training scenarios. Building upon this foundation, this paper reveals the coupling mechanisms between lower-body power generation, trunk rotation, and upper-body acceleration from a kinematic chain transmission perspective. It analyzes quantitative relationships between joint angle changes and racket velocity/ball strike height, providing structured motion feedback for tennis technique training.