Towards using artificial intelligence as tool in artistic gymnastics coaching : case backward giant circle

Abstract

The objective of this thesis was to study whether it is possible to create a system that estimates artistic gymnast’s body joint angles based on a low-budget 2-dimensional single RGB video recording. To meet the objective, 54 video files were collected on gymnasts performing backward giant circle skill, together with assessments of the performances by two professional coaches. The video files contained total of 233 repetitions of the skill. A pilot system of computer vision algorithms was developed, using an open source human body pose recognition algorithm. An algorithm based on pixel grayscale value was developed and used to recognize starting and ending moment of a repetition and to sample each repetition at 7 key phases. Body joint angle estimates were calculated based on the body part location estimates of the 1631 samples. The work proved that it is possible to develop a system that estimates body joint angles of an artistic gymnast. It was found that rotation and cropping of the frames improved probability of yielding correct estimates. The angle estimate for knees had highest, up to 66%, correlation with coach evaluations. Hips and shoulders had weak but significant correlation with coach evaluations. The results indicate that it is possible to develop a low-budget system that could work as augmented tool in artistic gymnastics coaching. In addition, human body pose recognition provides a new method to biomechanical research of artistic gymnastics.