Detection and Tracking of Players in Sport Video

Praćenje igrača u sportskom videu korisno je za analizu strategije i kretanja istih, te služi kao svojevrsni izazov znanstvenoj zajednici koji se odluče baviti njim. Uzeli smo pristup praćenja po detekciji zato što je intuitivan i učinkovit. Za algoritam praćenja odabrali smo DeepSORT, a za detektore YOLO v4 i njegovu manju varijantu, YOLO v4-tiny, koja je pogodnija za izvedbu u stvarnom vremenu. Naučili smo mreže na podacima nogometnih utakmica i prate 3 klase: igrača, vratara i suca. Na našoj lokalnoj konfiguraciji, praćenje s prilagođenom YOLO v4 mrežom je prosječne brzine 5 okvira po sekundi, a MOTA i MOTP mjere na ispitnom videu su 62.9% i 32.3% respektivno. Modificirana YOLO v4-tiny varijanta izvršava se brzinom od 10 okvira po sekundi i na ispitnom je videu MOTA 46.5%, a MOTP 33.3%. Uzevši ubrzanje objekata u obzir prilikom računanja, uspjeli smo malo poboljšati rezultate DeepSORT-a. YOLO v4 mreža pokazuje robusnost na smetnje poput sakrivanja zbog svoje visoke preciznosti, dok tiny varijanta pruža bolje sučelje za pokretne sustave kod kojih je preciznost manje važna. Tracking players in sports videos is useful for motion and strategy analysis and is a challenge on its own. For this paper, we chose the tracking-by-detection approach as it is intuitive and effective. For tracking, we chose DeepSORT, one of the most prominent trackers today. For detection, we chose two detectors: YOLO v4, which is more complex and robust; and YOLO v4-tiny, which is better for detection in real-time. We trained the networks to discern between 3 classes: player, goalkeeper, and referee. At 5 FPS on our local machine, tracking with our custom YOLO v4 network gets a MOTA score of 62.9% and a MOTP score of 32.3%. Our custom YOLOv4-tiny, while performing tracking at 10 FPS, gets a MOTA score of 46.5% and a MOTP of 33.3%. When taking object acceleration into consideration, we have managed to slightly improve DeepSORT results. Each of these networks is suitable for its intended purpose: YOLO v4 is shown to be noise-resistant due to its high precision and recall, whereas YOLO v4-tiny is fast and would be a viable application for mobile systems that are more forgiving of detection errors.