Construction and application verification of a freestyle land-based training action assessment system based on multimodal information fusion

The lack of objective and quantitative feedback in freestyle land-based imitation training is a significant issue. This study constructed and validated a low-cost assessment system based on multi-modal information fusion to address this issue. The system integrates a standard USB camera and inertial measurement units (IMUs). An attention-based data fusion algorithm synergizes global information from vision with detailed kinematic data from IMUs. A temporal convolutional network (TCN) is then employed to quantitatively evaluate key technical indicators, such as elbow flexion angle. Experimental tests on 15 swimmers of varying skill levels showed that the system's mean absolute error (MAE) for predicting key angles was less than 5 degrees, and the concordance correlation coefficient (CCC) with expert evaluations exceeded 0.87. The results demonstrate that the proposed system can provide an objective and low-cost assessment solution for freestyle land-based training, offering a feasible path to support scientific training.