The use of computer vision and multi-camera systems has become unavoidable in modern times. In sports, such systems have become a standard for match analysis and improving the fairness of refereeing decisions, as they provide coverage of large playing surfaces such as a football field. Due to the increasing dynamics of the game, this information is highly desirable for analyzing the weaknesses of teams and referees who must make decisions in a split second. One of the initial challenges lies in the synchronization of the system used to determine player positions, as well as in handling overlapping between players. Therefore, we need methods that provide effective solutions for extracting such information.
This thesis presents a system that enables the detection of football players using one or more cameras. It also includes a method that improves object detection by dividing images into smaller overlapping segments.
In the practical part, we first recorded a match of the Slovenian Prva Liga Telemach, synchronized the video using a practical method based on identifying the moment of ball release, and cropped the frames to exclude irrelevant surroundings. In selected situations, we then divided the frames into four overlapping segments. Using the YOLOv8 model, we detected the players and obtained their bounding boxes, which we validated with manually annotated reference boxes. Finally, we evaluated the success rate of the detections and analyzed the precision using the validation results.
|
