Deep Transfer Learning Based Real Time Fitness Movement Identification

Sports are full of people’s lives, and regular exercise has become an indicator of people’s health. Due to the high price, most people who exercise at home will not hire fitness trainers, but learn about fitness through media communities. This is likely to lead to the wrong posture of fitness, which can lead to injury. A cheap, simple, and accurate fitness recognition system could increase fitness awareness. This paper proposes a deep transfer learning method that uses Yolov4 to classify fitness movements, which can instantly recognize fitness movements with only one network camera. We built a database, which contains 20 users and online fitness photos, a total of 16302 images, including 12 kinds of fitness movements. 10 user and online photos are used to train Yolov4, and another 10 user photos are used for testing. In the experiment based on Yolov4 to detect fitness, mAP is 99.71%, Precision is 97.9%, Recall is 98.56%, and F1-score is 98.23%. The results show that fitness movements can be detected accurately and quickly using this method.