Lightweight pose-based fitness action recognition and repetition counting method

The advancement of artificial intelligence and computer vision technology has made sports detection a popular research field, but traditional detection methods rely on complex models and a large amount of computing resources, which have limitations in practical applications. Therefore, a sports motion detection and recognition technology based on lightweight motion detection and counting algorithm was proposed. By optimizing the OpenPose model, a lightweight network architecture was used to achieve faster processing speed while maintaining high detection accuracy. Meanwhile, an algorithm framework for detecting and counting sports actions was designed, and the performance of the algorithm was evaluated on multiple datasets. This algorithm demonstrated good convergence characteristics, low loss values, and high accuracy on both the training and testing sets. It achieved competitive performance in terms of average recognition accuracy, reaching 0.93, while maintaining relatively low misjudgment rates for six types of actions (8.49%, 7.46%, and 6.46%, respectively). In terms of computational efficiency, the proposed method achieved a processing speed of 60.16 ms and an inference time of 75.26 ms, indicating efficient execution. Meanwhile, the GPU utilization rate was 78%, suggesting that the algorithm required relatively fewer computational resources while maintaining effective performance. Experimental results in different real-world environments showed that the average counting accuracy exceeded 90%, demonstrating the feasibility and practical effectiveness of the proposed method.