Omnidirectional Egomotion Estimation From Back-projection Flow

Abstract

The current state-of-the-art for egomotion estimation with omnidirectional cameras is to map the optical flow to the sphere and then apply egomotion algorithms for spherical projection. In this paper, we propose to back-project image points to a virtual curved retina that is intrinsic to the geometry of the central panoramic camera, and compute the optical flow on this retina: the so-called back-projection flow. We show that well-known egomotion algorithms can be easily adapted to work with the back-projection flow. We present extensive simulation results showing that in the presence of noise, egomotion algorithms perform better by using back-projection flow when the camera translation is in the X-Y plane. Thus, the proposed method is preferable in applications where there is no Z-axis translation, such as ground robot navigation.