배드민턴 단식 선수 위치 추적 시스템 구현 및 검증

초록

The purpose of this study is to implement, apply, and evaluate a badminton singles player position tracking system using a deep learning model. Therefore, for the practical application of the developed model, a study was conducted on a total of 6 game videos that were not used for learning during the system implementation process. As a result, the total number of frames used was 334,654, with 321,235 frames that were successfully recognized and 13,419 frames that failed to recognize, and when the actual game video was applied, the player recognition accuracy was 95.86%. As a result of the content analysis of the factors of recognition failure, most of them occurred at the end of the rally, and the progress of the rally occurred only from a single factor due to overlapping players near the net. These results confirm that the deep learning model proposed in this study performs player recognition well in the actual game video rather than the test image.

키워드