Rastreamento e detecção de jogadores de futebol em tempo real usando YOLO v8

Abstract

This study tested the performance of the YOLOv8 model to detect and track football players in real time using match videos. The starting point of the research was the challenge of running computer vision algorithms in highly dynamic environments, marked by player occlusions, intense movement, changes in player scale on the screen, lighting variations, and the need for instant processing. To address this, I built a code structure capable of preprocessing the images, loading the data into the models, and running detection integrated with the ByteTrack and BoT-SORT trackers. The tests used clips from the SoccerNet dataset, divided into smaller files to avoid overloading the computer used for execution. In practice, the system ran the generic YOLOv8 “person” class, generated bounding boxes on the video frames, and traced the paths covered by the athletes on the field over time. The system performed well and was able to find and visually track almost all players in the tested scenes. The trackers processed the entire test video, generating thousands of detections and making it possible to draw the movement map of the athletes. In the excerpts analyzed more closely, the average efficiency reached 79,03%, which confirms that YOLOv8 has technical feasibility for live sports monitoring. This combination with ByteTrack and BoT-SORT proved effective for following player movement and opens possibilities for tactical and positional analysis of teams on the field. On the other hand, the project also ran into some important limitations. Since the model was pre-trained to identify any person, the system made some occasional mistakes, showed fluctuations in performance from one frame to another, and ended up confusing the referee with a player. Another key finding was that evaluating the system only by counting the total number of detections and the model’s confidence score is too superficial to measure the real quality of tracking. YOLOv8 is an excellent tool with strong potential for football applications, but the next steps require the use of specific video-tracking metrics, such as MOTA, IDF1, and HOTA, as well as training the artificial intelligence with images focused exclusively on match environments and player uniforms.