Deep Learning Technique for Gymnastics Movements Evaluation Based on Pose Estimation

Gymnastics is characterized by a sequence of movements that the athlete must perform on different apparatuses. The player’s precision in executing these movements determines the points granted. Therefore, to mitigate the possibility of errors, it was proposed to implement a system in addition to, rather than exclusively depending on, human observation for evaluating the movements. Computer vision is employed to evaluate the robustness and steadiness of movements on still rings in gymnastics competitions. Considering the absence of an online dataset for these movements, a dataset was generated by compiling a series of video clips sourced from YouTube. The collection has ten distinct categories, each containing 2000 photos. The photos were partitioned, allocating 80% for training and 20% for testing. The video clips employed the YOLOV7 algorithm. The motion detection reached a 95% accuracy rate. To improve accuracy in distinguishing motion, a technique is used to place a group of points on the player’s body, and a set of mathematical equations is applied to measure the distance and angle between these positions. These elements all contribute to a degree of accuracy that reaches 99%.