Computer Vision System For Backflip Motion Recognition in Gymnastics Based On Deep Learning

Abstract

Reliance on computer vision systems in the sports field is one of the very important topics, which are of high importance, especially in the arbitration process or evaluating the accuracy of the player’s performance of the movement. It is better to rely on computer vision systems that are more accurate in the arbitration process. In this article, a method was presented to distinguish one of the important movements of the gymnastics player, by relying on deep learning techniques. The dataset was built based on high-quality video clips found on YouTube for tournaments held from the period 2018-2022, due to the absence of The dataset available. This data was divided into three sections: 70% for training, 10% for validation, and 20% for testing. Two models of the convolutional neural network yolov7 and yolov5 were trained, and the results obtained after testing the results of the models show that the seventh version was the best , Recall, Precision and Mean Average Precision criteria were adopted to evaluate the performance of these technologies.