Sparse frame selection for graph-based deadlift form assessment

Automated movement analysis from monocular RGB video can enhance injury prevention and performance assessment in strength training. Yet, existing approaches struggle to accurately evaluate the deadlift-one of the most widely practiced exercises with a high risk of injury when performed incorrectly. Assessing form from monocular RGB video remains challenging due to subtle biomechanical errors, redundant frames, and the computational demands of video analysis. This study introduces a Real-Time framework that integrates the modified Search-Map-Search frame-selection method with graph-based models, trained on a custom dataset that was created, labeled as good or bad form, and annotated by a certified expert for deadlift assessment. By selecting six representative frames per repetition, the proposed frame-selection method reduces redundancy by 89.8% while preserving critical biomechanical information. Benchmark results show that the proposed Temporal Graph Convolutional Network (proposed T-GCN) achieved 89.5% accuracy with high precision and recall; its graph-classification stage ran in 0.4 s, while the full video-to-decision pipeline ran in approximately 0.8 s on a desktop GPU. The lightweight proposed Spatio-Temporal Graph Convolutional Network (proposed ST-GCN) variant, with only 4,514 parameters, achieved comparable performance (85.3% accuracy). These findings demonstrate that the proposed framework achieves competitive accuracy while substantially reducing the number of processed frames and computational cost, enabling practical deployment on commodity RGB hardware.