Abstract:
Identification and tracking of players in sports broadcasts has been an enduring challenge within the Deep Learning and Computer Vision domain. In the field of identification and tracking of players are important for team tactics, activity analysis of players, entertainment in broadcast sports videos. We present an approach based on YOLO and Deep SORT for identification and tracking of players. Our proposed research addresses the problem of identifying and tracking of players in broadcast videos of Soccer. Tracking and identification of Players in this scenario is a challenging task due to abrupt movements of players and sometimes visuals are not visible clearly. Player identification and tracking frameworks have been affected by neural networks. It’s been observed that the incorporation of CNN’s improves the performance of architecture. In this thesis, we present an approach to track and identify players, ball and referee in soccer broadcast videos, our approach is based on YOLO and Deep Sort. Furthermore, we introduced an additional feature of identification and tracking of referee and ball. We achieved mean average precision (MAP) of 95% percent in training and 94% in testing of our model on identification of objects. For tracking of objects we used Multi Object Tracking Accuracy (MOTA), our model achieved 95% on training and 91% on testing on tracking of objects.
