As aerobics becomes a more popular competitive sport, it is essential to have accurate, ef?cient, and automated movement analysis to improve judging fairness and training performance. Traditional manual methods are labor-intensive, subjective, and prone to inconsistencies. To address these issues, this study proposes a novel motion analysis model for aerobics competition video that combines a classi?cation neural network with target box ratio detection and non-maximum suppression (NMS). The model uses a graph convolutional neural network (GCN) and an independent recurrent neural network (IndRNN) to capture the spatial and temporal features of human motion. The model signi?cantly reduces redundant detections and improves localization precision across multi-scale movements by incorporating target box ratio detection and NMS. Experiments conducted on the UCF101 and Sports-1M datasets demonstrated that the proposed model achieved an accuracy of 99%, RMSE of 0.11, and F1-score of 0.98. Compared with baseline models, it offered superior detection performance with faster processing times, especially in complex, multi-target scenarios. The key innovation lies in the fusion of GCN and IndRNN within an optimized, multi-branch framework using NMS to suppress spatial redundancy. This makes it well-suited for real-time, high-precision analysis in aerobic competition scoring systems. This research introduces a robust, scalable approach to automated sports video analysis. This approach has the potential to be used for judging assistance, providing athlete performance feedback, and creating intelligent sports analytics.