Active Visual Localization using Predicted Images

摘要

Visual odometry (VO) is an important and intensively studied localization technique that uses camera images for navigating autonomous mobile robot systems. VO estimates the relative robot position from two consecutive camera images. However, the accuracy of VO often deteriorates in less feature-filled environments, such as near a white wall. In this paper, we propose a novel active visual localization method using machine learning technique to perform and maintain accurate VO while the robot is moving around in unknown environments. Our proposed method tries to estimate the VO accuracy without actually moving to the location and then deciding the next command to obtain better VO. The key idea of our proposed method is using predicted images at virtual target poses, instead of moving into its position. The effectiveness of our proposed method was verified by applying it for a free exploration task assigned to a wheeled mobile robot in a 3D visual simulator, and comparing the results with the conventional active visual localization. According to the results presented in this paper, our method comfortably outperformed the conventional technique.