Abstract
This master’s thesis explores challenges and methodologies of accurately tracking players and the ball within football video sequences, employing advanced object detection and tracking techniques. The research is integrated into a larger project aimed at reconstructing football sequences in a virtual reality (VR) environment, thus enhancing the training and strategic analysis capabilities in sports environments.
The thesis primarily utilizes models based on convolutional neural networks, such as the YOLOv9 model for object detection to identify the positions of the players and the ball within different frames. Advanced methods such as the online tracking algorithm BoT-SORT and the minimum cost flow algorithm are employed for detection-based tracking, optimizing the accuracy of continuous player and ball trajectories in the complex, fast-moving setting of a football game. We also employ the TrackNet V3 model for predicting ball trajectories, as an alternative to detectionbased tracking. TrackNet utilizes multiple sequential frames to predict the trajectory of the ball, enhancing detection accuracy even during fast play or when the ball is momentarily occluded.
Results from these experiments indicate promising player tracking, though challenges persist when it comes to, e.g., id-switching. Ball tracking is shown to be more difficult due to its small size and high movement speed, which often leads to reduced detection reliability. TrackNet successfully predicts the ball’s trajectory when it is moving quickly and not occluded, but instead struggles when the ball is still or frequently occluded (such as during dribbling). The outcomes contribute towards the understanding of object tracking in sports and the development of applications for enhanced game analysis.