Fitness Action Counting Based on MediaPipe
Xiangying Li, Mingwei Zhang, Junnan Gu, Zhi Zhang
Abstract
AI fitness has become a new and practical way of fitness, but most of the mainstream fitness apps focus on guiding and planning fitness activities, ignoring the detection and evaluation of users' fitness movements. Aiming at this phenomenon, this paper proposed a method to classify and count basic fitness movements based on Google Mediapipe framework. The method consists of three steps: First, a single fitness action is divided into two detection states: up and down, and the corresponding picture samples are collected and trained. Secondly, based on the generated training set (csv file), KNN algorithm was used to identify and classify different fitness actions. Finally, the classification results are processed and the fitness actions are counted. The best recognition angle and threshold are obtained through the test accuracy. Compared with the mainstream human pose recognition frameworks such as Openpose and Alphapose, Mediapipe's Blazepose algorithm has lower performance requirements, faster recognition speed and a high level of accuracy, which is more suitable for personalized needs for smart fitness on mobile devices today.