Improving robustness of monocular VT&R system with multiple hypothesis

Abstract

Visual Teach and Repeat (VT&R) has proven to be an important ingredient for mobile robot navigation. For VT&R, visual localization on a known map is a challenging task, especially in the case of motion jitter, feature-poor scenes and occlusion. State-of-the-art feature-based localization or SLAM algorithms sometimes may fail to overcome these challenges, and, as a result, suffer from tracking loss. To solve the problem of tracking loss in monocular-SLAM-based VT&R, we propose a particle filter (PF) based algorithm, which can provide robust location estimation even under challenging conditions. Our experiments verify the ability of our proposed PF-VT&R method.