Athlete Motion Recognition and Biomechanical Analysis Based on a Multimodal CNN Framework

Abstract

Aiming at the problem of insufficient accuracy and low efficiency of traditional athlete action recognition methods, this paper proposes a multimodal fusion action recognition system and biomechanical quantitative analysis method based on an improved convolutional neural network (CNN). The system integrates high-speed cameras (video), inertial sensors (IMUs), and force measurement platforms to construct a multimodal data acquisition framework. The backbone adopts an improved ResNet-50 network embedded with a squeeze-and-excitation (SE) module to enhance channel attention. A spatiotemporal feature fusion module and dynamic time warping (DTW) algorithm are introduced to capture temporal continuity and synchronize multi-source data. The system achieves a recognition accuracy of 97.8% with an average processing time of 288.7 ms. The dataset includes 10,200 video segments (3–5 seconds each) and synchronized biomechanical data (e.g., GRF, joint angles, EMG) from 50 professional athletes across 6 sports (e.g., sprinting, long jump, tennis serve). The results demonstrate the effectiveness of the proposed method for intelligent sports analysis and injury prevention. This work provides a design paradigm for multimodal CNN-based action recognition and biomechanical evaluation.