Synthesizing fixed point of views from a spinning omnidirectional ball camera

Abstract

Recent years have seen the emergence of a new way to watch sports, that is, watching them from the ball's point of view (POV) taken with cameras embedded in the ball. The problem, however, is that the ball's POV is unstable since the ball spins. In this paper, we describe how we developed a prototype of the ball camera with an omnidirectional camera and propose an algorithm to fix the ball's POV to an arbitrary direction. Feature points are extracted from an entire omnidirectional image and the pose change between frames is estimated by feature matching and the five-point algorithm. The POV of the current frame is modified to the one of the initial frame by applying the product of estimated rotation matrices between all adjacent frames up to the current one. Our algorithm is not limited to our hardware and can be applied to any omnidirectional videos taken with different camera systems.