Fitness Action Counting Algorithm based on Pose estimation

We have developed a fitness action counting algorithm based on pose estimation to address the inefficiency and high error rate of wearable devices and manual supervision counting in fitness training. The algorithm utilizes an optimized MediaPipe Pose model for human pose estimation, obtaining keypoint positional information of users during fitness exercises and converting this information into feature vectors. Subsequently, a KNN algorithm is employed to design an action counting classifier, performing binary classification on the feature vectors to achieve action counting. Experimental results demonstrate that the algorithm achieves an inference speed of 30 FPS, with a counting accuracy of up to 99.5%, thus demonstrating its efficiency and practicality. This method exhibits significant innovation in effectively addressing the efficiency and accuracy issues present in existing methods, and is applicable to fitness facilities and online fitness platforms.