Robust egomotion for large-scale trajectories

Abstract

This paper presents an effective egomotion solution based on high curvature image features described using local intensity histograms for stereo matching and tracking. To robustify the visual processing system, we propose feature extraction over moment image representation to overcome the adverse effects of illumination changes. A bundle adjustment optimisation technique, thoroughly analysed for different reprojection strategies, is developed for motion estimation of an autonomous platform. The quality of results is shown to be on par with high quality GPS-corrected-INS systems, even for long-range trajectories.