딥러닝 기반 배드민턴 선수 위치 추적시스템 개발

초록

The purpose of this study is to realize a system that can automatically recognize players and record location information within the video of a badminton singles match. To achieve the research purpose, 500 training data and 60 test data were used to develop a badminton single player tracking model. As a result, 100 models of training data showed a mAP score of 90.0%, and 300 models showed 95.83%. The final model, 500 training data models, showed high performance at 98.33%, and it was concluded that at least 500 training data were needed to apply deep learning in badminton. These results are considered to show satisfactory performance even if the amount of training data is not large because there was no camera viewpoint movement using the image obtained from a single fixed camera, and the badminton court, which is the background of the image, was kept constant. Therefore, it was confirmed that the amount of data used for learning is important for the development of high-performance deep learning models, but it is necessary to proceed with sufficient consideration of external environmental factors.

키워드