RGB-D Vision Navigation via Motion Saliency Measurement and Twin Reprojection Optimization in Complex Dynamic Scenes

Abstract

Localization in an unexplored environment is a fundamental capability for robotic vision navigation. However, due to the static world assumption, it still suffers the impoverishment of robustness and accuracy in complex dynamic workspaces. Moving objects, with indeterminate motion status in dynamic scenarios, usually increase the difficulty and complexity to the localization of the robotic vehicles. To address this problem, a robust and real-time RGB-D vision navigation system based on motion saliency measurement (MSM) and twin reprojection (TR) optimization is proposed to allow accurate localization for the robotic vehicles under complex dynamic scenes. Firstly, a novel saliency-induced dense motion removal (SDMR) method is developed to detect and eliminate the dynamic regions in RGB-D inputs, which can effectively filter out the outlier data that are associated with the moving objects. Then, a robust matching strategy for edge drawing lines (EDLines) feature is devised to acquire fine line inliers by constructing keypoint correspondence. Furthermore, the TR error is built by depth measurement for the line features. It is incorporated into a new error optimization function to achieve optimal pose estimation. The experimental results demonstrate that the SDMR can accurately detect dynamic objects and eliminate movement regions in complex dynamic scenarios. The proposed navigation system proves to attain at least 26% improvement of localization accuracy over other advanced dynamic navigation solutions. Test code is available on https://github.com/SunIMLab/TL-REE .