Visual Odometry in Challenging Environments: An Urban Underground Railway Scenario Case

Abstract

Localization is one of the most critical tasks for an autonomous vehicle, as position information is required to understand its surroundings and move accordingly. Visual Odometry (VO) has shown promising results in the last years. However, VO algorithms are usually evaluated in outdoor street scenarios and do not consider underground railway scenarios, with low lighting conditions in tunnels and significant lighting changes between tunnels and railway platforms. Besides, there is a lack of GPS, and it is not easy to access such infrastructures. This research proposes a method to create a ground truth of images and poses in underground railway scenarios. Second, the EnlightenGAN algorithm is proposed to face challenging lighting conditions, which can be coupled with any state-of-the-art VO techniques. Finally, the obtained ground truth and the EnlightenGAN have been tested in a real scenario. Two different VO approaches have been used: ORB-SLAM2 and DF-VO. The results show that the EnlightenGAN enhancement improves the performance of both approaches.