Direct RGB-D Visual Odometry Based on Hybrid Strategy

Abstract

Edge-based Direct Visual Odometry (E-DVO) plays a crucial role in robot navigation across the low-texture and rich-texture scenes, however, two essential factors are overlooked in the traditional E-DVO methods. (1) Traditional E-DVO methods seriously rely on photometric or geometric cost, thereby generating the non-robust performance under the light changing or structure-less conditions; (2) EDVO methods generally suffer drift issue mainly derived from inaccurate rotation estimation for the long term visual odometry task. In this article, a novel hybrid cost function leveraging the photometric and geometric cost within a bi-direction framework is proposed to facilitate the addressed the former issue. While the latter issue is approached through hybridization of a simple yet effective switching strategy which can guarantee both robustness and accuracy by combining the global Manhattan model and direct edge alignment. We carry out various experiments on TUM RGB-D and ICL-NUIM datasets for performance evaluation. Results show that our method has the advantage of strong robustness and high accuracy compared with state-of-the-art methods, e.g., Canny-VO and ORB-SLAM2.