A Wushu Posture Recognition System Based on MediaPipe
Jiewei Ma, Lianzhen Ma, Wenpian Ruan, Haidong Chen, Jinyong Feng
Abstract
With the development of information technology such as computer vision and human-computer interaction, online physical education has become an active field of current physical education research. Especially with COVID-19, a significant number of people are restricted to learning and exercising motor skills in small spaces, but most of the movement cannot be carried out in narrow environment. Wushu is exempt from this restriction, so it is often used in online physical education in China. In this context, we propose a Wushu posture recognition system based on camera and MediaPipe for tracking hand, head and body movements of users. According to the Landmarks returned by Mediapipe, we designed recognition algorithms for Fist, Palm, Hook, Tiger Talon, Forward Lunge, Horse Stance and Empty Step. By testing 400 photos, the experimental results show that these algorithms can effectively identify these movements. From there, we built a system using Python to help users perform Wushu training independently, safely, and efficiently without a teacher.