Prediction of valid jumping in long jump sport

Long Jumping, a branch of athletics, is a well-known competition activity in our country and around the world. Athletes try to reach the longest distance by performing their jumps within the framework of the determined rules. The jumps are evaluated by the referees and the distance of the jump is measured by deciding whether the jump is valid or not. Due to the nature of the competition, it is inevitable that the athletes must be fast. These controls made with the human eye put pressure on the referees and athletes, even if they are for a short time. Today, developing software and hardware resources have accelerated Artificial Intelligence studies. Artificial Intelligence solutions have been successfully applied in all areas of life. In this study, the ResNet-50 + FPN based Foot Pose Estimation Model, which estimates the foot positions of the athletes as a solution to the Prediction of Valid Jumping problem in the Long Jump Sport, and the Video Classifier Model with the GRU + fully connected layer, which makes the prediction whether the jump is valid by tracking the predicted foot poses in the video stream. Foot Pose Estimation and Video Classifier models were used in combination to solve the problem. In the training part, they were trained and tested individually. On the test dataset, the Foot Pose Estimation model achieved 0.992283 in the AP@IoU=0.5 metric, the Video Classifier model achieved 0.958333 in accuracy metric, and the Combined model achieved 0.971831 in accuracy metric.