Accurate assessment of squat exercise quality is important for rehabilitation exercises performed outside clinical settings, where incorrect movement patterns may reduce training effectiveness and increase physical strain. This study evaluates squat repetitions from 2D human pose sequences for movement-quality assessment.

The squat subset of the REHAB24-6 rehabilitation exercise dataset is used. Each repetition has 2D pose sequences, temporal segmentation, and correctness labels. Based on bodyweight squat screening criteria, the assessment focuses on movement-quality indicators such as squat depth, knee alignment, left-right symmetry, foot stability, trunk posture, and movement control. In an initial experiment, biomechanically meaningful features were extracted from each repetition, including hip displacement, knee angle characteristics, left-right knee differences, foot movement range, movement duration, and hip velocity. Conventional machine learning models were evaluated using a leave-one-video-out protocol to reduce video-level data leakage, and the results provide an interpretable baseline.

This study further compares this baseline with temporal convolutional models and attention-enhanced graph neural models for pose-sequence analysis. The comparison focuses on improving correctness classification performance while keeping the output interpretable for rehabilitation assessment. The framework can also be extended to estimate error types, overall movement quality scoring, and personalized feedback.