A Proposal for Estimating Court Boundaries and Displaying Player Positions and Trajectories in Broadcast Handball Videos

Abstract

This study proposes a method for estimating court boundaries and displaying player positions and trajectories for broadcast handball videos. The videos typically cover only about one-third of the court, and the camera angle changes from left to right to follow the players. Initially, an object detection algorithm YOLOv5 is used to detect a goal object in the video, and the camera direction is determined from its position. Court boundaries are estimated in two ways: one using a bounding box around the goal object, and the other with a red area marked on the court. Once the court boundaries are estimated, DeepSort, an object tracking model, is employed to track the players, and their positions and movements are displayed on a top view of the court. The proposed methods were compared by conducting experiments on videos from different venues, and the results indicate that the estimation rates of the two methods were 80.7% and 85.3%, respectively.