Resolving scale ambiguity for monocular visual odometry

Abstract

Scale ambiguity is an inherent problem in monocular visual odometry and SLAM. Our approach is based on common assumptions such that the ground is locally planar and its distance to a camera is constant. The assumptions are usually valid in mobile robots and vehicles moving in indoor and on-road environments. Based on the assumptions, the scale factors are derived by finding the ground in locally reconstructed 3D points. Previously, kernel density estimation with a Gaussian kernel was applied to detect the ground plane, but it generated biased scale factors. This paper proposes an asymmetric Gaussian kernel to estimate unknown scale factors accurately. The asymmetric kernel is inspired from a probabilistic modeling of inliers and outliers, that is, 3D point can comes from the ground and also other objects such as buildings and trees. We experimentally verified that our asymmetric kernel had almost twice higher accuracy than the previous Gaussian kernel. Our experiments was based on an open-source visual odometry and two kinds of public datasets.