Monocular Visual Odometry using a Planar Road Model to Solve Scale Ambiguity

摘要

Precise knowledge of a robots’s ego-motion is arncrucial requirement for higher level tasks like autonomousrnnavigation. Bundle adjustment based monocular visual odometryrnhas proven to successfully estimate the motion of a robot forrnshort sequences, but it suffers from an ambiguity in scale. Hence,rnapproaches that only optimize locally are prone to drift in scalernfor sequences that span hundreds of frames.rnIn this paper we present an approach to monocular visual odometryrnthat compensates for drift in scale by applying constraintsrnimposed by the known camera mounting and assumptions aboutrnthe environment. To this end, we employ a continuously updatedrnpoint cloud to estimate the camera poses based on 2d-3dcorrespondences.rnWithin this set of camera poses, we identifyrnkeyframes which are combined into a sliding window and refinedrnby bundle adjustment. Subsequently, we update the scale basedrnon robustly tracked features on the road surface. Results on realrndatasets demonstrate a significant increase in accuracy comparedrnto the non-scaled scheme.