Appearance-invariant Visual Localization for Long-term Navigation

Abstract

Long-term navigation refers to the ability of a mobile robot to navigate effectively over extended periods in environments that may undergo changes. Unlike short-term navigation, which focuses on immediate and local motion estimation, long-term navigation aims to identify environmental changes and maintain a sustainable map that is beneficial to the association between current observations and the visited location. In this paper, a Convolutional Neural Network (CNN) based discriminating model is proposed to quantify the repeatability and the discriminativity of visual features, aiming to filter out the unstable keypoints for the mapping process. Moreover, a novel self-adaptive mechanism for map maintenance is proposed to incrementally update the keypoint dataset in a memory-saving manner. Extensive experimental evaluations are conducted on the CMU Seasons dataset, which suggests that our method increases the success rate of localization from 50% to 86% on average.