Combining sparse and dense methods in 6D Visual Odometry

Abstract

Visual Odometry is one of the most powerful, yet challenging, means of estimating robot ego-motion. By grounding perception to the static features in the environment, vision is able, in principle, to prevent the estimation bias rather common in other sensory modalities such as inertial measurement units or wheel odometers. We present a novel approach to ego-motion estimation of a mobile robot by using a 6D Visual Odometry Probabilistic Approach. Our approach exploits the complementarity of dense optical flow methods and sparse feature based methods to achieve 6D estimation of vehicle motion. A dense probabilistic method is used to robustly estimate the epipolar geometry between two consecutive stereo pairs; a sparse feature stereo approach to estimate feature depth; and an Absolute Orientation method like the Procrustes to estimate the global scale factor. We tested our proposed method on a known dataset and compared our 6D Visual Odometry Probabilistic Approach without filtering techniques against a implementation that uses the well known 5-point RANSAC algorithm. Moreover, comparison with an Inertial Measurement Unit (RTK-GPS) is also performed, for providing a more detailed evaluation of the method against ground-truth information.