Research on motion decomposition of aerobics programs based on deep learning

Addressing the limitations of traditional aerobics motion analysis, which relies heavily on manual intervention and is highly subjective, as well as the limitations of existing deep learning algorithms in continuous motion decomposition, antiinterference capabilities, and specific adaptability, this study proposes the TA-GC-ADN aerobics motion decomposition algorithm and constructs a dedicated dataset, CAD-2024. The dataset covers 10 core aerobics movements, containing 500 video clips from 50 participants of varying skill levels. 120,000 frames of skeletal data (17 key joints) were extracted and expanded to 360,000 frames after preprocessing, achieving a 98.3% consistency in annotation. TA-GC-ADN employs a temporal-spatial dual-dimensional fusion framework. Its core modules include Temporal Attention (TAM) for locating keyframes, Bone Graph Convolution (B-GCM) for capturing joint linkage features, and a dual-branch interaction mechanism for optimizing boundary detection and classification. Experiments show that the algorithm achieves a decomposition accuracy of 96.8% under standard conditions, which is 2.5-5.7 percentage points higher than comparable algorithms such as GCN and LSTM. With 30% occlusion, the anti-interference rate is 92.3%, the boundary detection error is only 2.1 frames, and the FPS reaches 32.5, meeting real-time requirements. Ablation experiments verify the key supporting role of modules such as TAM and B-GCM in performance. This research provides a reliable technical solution for quantitative analysis of aerobics training and referee-assisted decision-making.