Attention-guided multi-view video fusion for real-time sports injury risk assessment

Sports injuries impose significant burdens on athletes and healthcare systems, yet current assessment methods lack real-time intervention capabilities. This study presents an attention-guided multi-view video fusion framework that enables real-time injury risk evaluation during athletic training. The system employs cross-view attention mechanisms to adaptively integrate information from multiple camera perspectives, coupled with temporal attention modules to maintain assessment stability across movement sequences. We constructed a comprehensive dataset comprising 11,206 movement samples from 156 athletes across four sports disciplines. Experimental results demonstrate that the proposed model achieves 85.7% classification accuracy and a mean absolute error of 6.83 in continuous risk scoring, representing 28.4% improvement over single-view baselines. With an average processing latency of 87 milliseconds, the system meets real-time deployment requirements while maintaining strong agreement (Pearson correlation of 0.79) with expert clinical assessments. Field validation across three sports facilities confirmed practical feasibility for proactive injury prevention through immediate corrective feedback during training sessions.