3D scene recovery based on multiple objects tracking in sport videos

Abstract

This paper proposes a new method for estimating the player's and ball's 3D position information from monocular broadcast videos. For players, the homography between image and playfield is used to estimate their positions. By analyzing the geometry relation between the ball, its "virtual" shadow and camera position, we derive equations for estimating the flying ball's 3D position. Moreover, we propose a method to predict the flying plane if it cannot be determined from images. This method designs a new cost function, which arrives at the minimum when the predicted flying plane is reasonable. This method has at least two merits. One is that it can estimate the flying ball's position without referring to other objects with known height; the other is that only one assumption is made and the camera is in a fixed position. Experimental results are satisfying.