Research on real-time athlete movement and pose estimation and error correction system based on deep learning

Research on Real-time Athlete Movement and Pose Estimation and Manual sports movement correction is often limited by high costs and the subjectivity of human coaches. To address these challenges, this paper proposes a real-time athlete pose estimation and error correction system based on knowledge distillation and biomechanical constraints. The system utilizes a modular architecture comprising four specialized components: data acquisition, pose estimation, biomechanical error analysis, and an interactive feedback interface. To achieve a balance between high precision and computational efficiency on edge devices, a knowledge distillation framework is employed, where a lightweight MobileNet-V3 student model is trained under the supervision of a high-performance HRNet teacher model. The student model is further optimized using 8-bit integer quantization and the TensorRT inference engine. The error correction logic relies on rigorous biomechanical modeling, calculating joint angles via vector dot products to identify deviations from professional standards in real-time. To ensure high throughput, a multi-threaded asynchronous pipeline is implemented to manage data flow. Experimental results demonstrate that the system achieves a mean Average Precision (mAP) of 92.4% and maintains a processing speed of 45 frames per second (FPS) on mobile platforms. Furthermore, a controlled user study confirms the system’s training effectiveness with statistical significance ( p < 0.05). This research provides an objective, low-latency solution for automated athletic coaching with quantitative posture feedback.