Research on Multi-model Fusion Attitude Detection and Velocity-Based Power System

In the context of the close integration of artificial intelligence and sports, the precise measurement and intelligent coach-free scientific guidance remain one of the hotly debated topics. However, most current technologies are still limited to manual supervision and measurement of relevant indicators through hardware sensors. This approach is time-consuming and labor-intensive, not only prone to interfering with athletes’ actual performance but also requiring coaches’ real-time supervision to prevent irrecoverable injuries. Therefore, this paper proposes the LW-Pose model, which introduces Depth-Wise (DW) convolutions and channel attention mechanisms to optimize the backbone network. We replace the original decoupled head section with the Vision Transformer method and analyze the output vectors using Simcc technology to ultimately obtain accurate posture outputs. The proposed method achieves an AP50 of 92.6% and AP of 71.8%. Based on this technology, a velocity-based power training system has been developed with accuracy of 97%.