Abstract: To address the challenge of trajectory tracking in skeleton sledding under high-speed conditions, a dual-stage dynamic perception-prediction framework was proposed. Innovatively, a multi-level feature decoupling detector was proposed, achieving a 96% localization accuracy at speeds up to 130 km/h through an adaptive channel enhancement mechanism, effectively suppressing track interference. A spatiotemporal correlation feature tracker was designed, integrating geometric, texture, and temporal information to reduce trajectory prediction error to ±2.3 cm. Experiments on Skeleton-SOT-2022 dataset demonstrate that the system achieved an average IoU>0.85 and could output key parameters such as entry angle and center-of-mass offset in real time. This research breaks the technical bottleneck in high-speed motion scenarios, providing intelligent analysis tools for skeleton sledding training.