This study primarily considers the nonlinearity, dynamics, and multi-parameter coupling issues in evaluating the effectiveness of athletes’ physical training. It proposes a quantitative evaluation method based on video motion segmentation and a genetic algorithm-optimized backpropagation neural network. The method first segments the training video motion using a multi-scale spatiotemporal pyramid network. Simultaneously, it extracts joint Laplacian coordinates and high-order skeleton graph features. These features aid in constructing a frequency-aware spatiotemporal motion representation. Furthermore, a hybrid optimized GABP neural network model incorporating time-series correction is designed, utilizing a genetic algorithm to globally search network weights and structure. A linear regression error correction mechanism is also employed to improve prediction robustness on small sample data. Experiments are conducted using a fused dataset constructed from OpenSim simulation data and real motion data. Results show that the proposed method achieves an accuracy of 94.5% in the training effect classification task, with precision, recall, and Formula: see text1-score of 92.1%, 93.8%, and 92.9%, respectively, representing improvements of 6.2%, 4.9%, and 3.3% compared to traditional BP neural networks, support vector machines, and long short-term memory networks. Ablation experiments verified the contribution of mechanical parameters (such as explosive force inertia) and spatiotemporal characteristics to model performance. Noise tests showed that the model accuracy remained above 90% even under 15% Gaussian noise. This result demonstrates its good anti-interference capability.
Kang Chen (Tue,) studied this question.