Player identification with scoring detection

Sports video analysis is one of the most popular areas for image processing. These analyses are done on professional sports matches or training videos and generally employ traditional image processing methods instead of deep learning for scoring and player identification. The aim of this thesis is to obtain an algorithm to analyze a basketball training in real time. That is, the approach should be able to detect the correct players and scores for each player. For this purpose, we first propose a method to detect the shots and if they are makes or misses. Specifically, we use a deep learning model to detect the position of the hoop and background subtraction to detect the shot. Then, scoring is determined by another neural network using classification. After the determination of the shot, we go back in our buffered images to find the player who sent the shot. The player is detected by YOLOv3-tiny and identified by a deep unsupervised re-identification model. This unsupervised model is trained with Triplet Loss with the addition of a segmentation model to obtain high accuracy. To keep the real-time aim, all models employ mobile networks as base with different parameters and training methods. After each session, we run clustering to do the player re-identification again to improve on the result. The experiments were conducted on two training videos. The experiment results show the effectiveness of the current method with 95% accuracy on scoring identification and up to 92.5% overall accuracy on re-identification of 2 players in real-time, with up to 86.7% accuracy on 6 players with clustering.