Automated scoring in climbing competitions

Abstract

With all the technological advances and research in sports tracking and human movement analysis in climbing due to the increasing popularity of the sport, no evidence of the use of an automated scoring system for climbing competitions is found in literature. The lack thereof motivated the evaluation of state of the art technology regarding its fitness for such task and its practical viability. Therefore, in this thesis a framework for an automated scoring system for climbing competitions with a focus on lead climbing is built. A novel one-camera approach is taken and a low-cost system is proposed where a single perspective from behind the climber is used to provide video data. The system utilizes the video data from the camera in conjunction with object tracking, color masking and image processing techniques in Python to derive scores post-competition. With YOLOv3, state of the art computer vision technology is used for climber detection and tracking, while the climbing holds are manually marked and further delineated with HSV color masks in a simple user interface. The declaration of holds as "reached" utilizes difference calculation with the structural similarity (SSIM) index and thresholds it against empirically derived values. Two video resolutions, 4k and 1080p, are used for building and testing the system. The results from the testing are manually reviewed and provide insights for the system integrity (the system not failing) with mean integrity values for the tested video resolutions of 82.6% and 76.6% respectively. The lower performance values (correctly declared holds) of 48.7% and 40.4% respectively show potential for improvement. The workflow established in this thesis yields a usable scoring system for post-competition reviews and serves as a proof of concept for possible future developments in regards of automated scoring systems for climbing competitions.