Robust visual odometry using stereo reconstruction error model

Abstract

The fundamental problem of localization remains one of the key areas of improvement concerning any type of mobile vehicle. Lately great interest in the research community was presented for vision based localization methods. In this article a solution is proposed for radically improving precision when a stereo camera is used for image acquisition. While providing great number of advantages, stereo cameras have one major disadvantage: that of very noisy coordinates of the reconstructed 3D coordinates. Here we present a method to correspondingly take into account the noise in the case of consecutive spatial movements of a stereo camera. The solution is based on an elegant error model that is used for weighting each measurement correspondingly to its modeled error. Computationally the burden of such modifications to the direct visual odometry model does not affect the on-line speed.