Research on Deep Learning-Based Human Pose Recognition and Intelligent Scoring System for Aerobics Movements

Aerobics is characterized by continuous, rhythmsensitive full-body movements that are traditionally evaluated by expert judges. Manual scoring is time-consuming and suffers from subjectivity, while wearable sensors increase burden and may degrade natural performance. This paper proposes an end-to-end, computer-vision-driven aerobics scoring system that tightly couples deep learning-based human pose estimation, posetemporal modeling, and interpretable rule-constrained regression to deliver real-time, objective, and explainable scores. To advance computer technology beyond direct application, we co-design (i) a lightweight pose backbone that combines a tiny Vision Transformer for efficient global feature aggregation with (ii) a spatiotemporal graph refinement module that enforces kinematic consistency on joint trajectories, and (iii) a phase-aware temporal scoring encoder that attends to judge-critical phases (preparation, extension, peak, recovery). The system outputs a final score 0-100 and six sub-scores aligned with judging principles: posture accuracy, limb extension, balance stability, rhythm consistency, amplitude, and continuity. We further contribute an edge-oriented deployment pipeline including mixed-precision inference, structured pruning, and quantization-aware training to achieve high throughput on commodity GPUs and embedded devices. Extensive experiments show improved pose robustness, lower score prediction error, and useful per-joint error localization that supports coaching feedback.