Deep Learning-based Object Tracking in Soccer Data

Abstract

Along with the advent of deep learning, soccer video has been widely used for soccer analytics. The position of soccer ball or players is the most important data that enables a lot of analyses, but they are still being collected manually due to the difficulties of object tracking on a soccer video. In this thesis, the tracking of ball and players using the soccer video and GPS data is proposed. YOLO v5, as a deep learning-based object detection model, was trained on soccer video and resulted in an accurate performance. However, the soccer ball or players are too small for the field, so the image is divided to feed the model, which increased the computational cost. We propose a tracking algorithm that reduced inference time by 91% while maintaining accuracy to solve this problem. The tracking algorithm can also determine the presence of the ball with an accuracy of 98.5%. Furthermore, a CRNN model is used to handle the occlusion of the ball. The CRNN combines CNN and RNN to estimate the ball position when the occlusion occurred. The trained model predicted the bounding box with an error of a few pixels and handled the occlusion. The performance of the proposed ball tracking algorithm is compared with the Kalman filter and whose position error was 2.9 pixels on average, while the position error of the tracking algorithm was 1.1 pixels on average. For the player tracking, both GPS data and bounding boxes were used. The GPS data was used to identify each bounding box, and the boxes and GPS data were combined using data fusion method. As a result, the unstable object detection results became stable, while the position error is maintained.