Lightweight 3D Convolutional Networks for Biomechanical Motion Analysis in Martial Arts Training

To reduce the subjectivity of traditional martial arts training assessment and avoid the deployment requirements of wearable devices, the study develops an artificial intelligence (AI)-assisted analysis system based on a 3D Convolutional Neural Network (3D CNN). This study adopts a 3D CNN as the backbone network and develops optimization strategies from four aspects. Spatiotemporal decomposition convolution factorizes standard 3D convolution kernels into a combination of spatial convolution and temporal convolution. The temporal attention mechanism adaptively computes the importance weight of each frame, enabling the model to focus on key action frames. A dynamic inference mechanism determines whether computation should be terminated early based on the confidence of intermediate-layer outputs, thereby reducing redundant computation. Knowledge distillation uses only the class-similarity soft distribution of a large teacher model to supervise the lightweight student model. This reduces the loss of feature representation caused by network compression and helps maintain recognition accuracy with fewer parameters. Experimental results are obtained on the Chain-of-Thought Action Form Assessment (CoT-AFA) and Martial Arts, Dancing and Sports Dataset (MADS) public datasets. The results show that the proposed method achieves a Top-1 accuracy of 74.2% on the CoT-AFA dataset, which is close to the 75.3% achieved by Video Swin Transformer. Meanwhile, the proposed method reduces the number of parameters by approximately 52.3% (14.2M vs. 29.8M) and decreases Floating Point Operations per Second by approximately 43.7% (52.4G vs. 93.0G). In addition, the mean squared error (MSE) for quality evaluation is reduced to 1.48. On the MADS dataset, the Top-1 accuracy reaches 68.9%, while the mean per-joint position error is reduced to 60.5 mm. The results support the use of the system for video-based analysis in martial arts training.