Computer vision techniques for autonomous vehicles applied to urban underground railway

Abstract

Autonomous vehicles’ presence is becoming a reality in everyday life, with autonomous driving cars on the road, GOA3-GOA4 trains in the railway domain, or automated guided vehicles in the industrial domain. These autonomous systems must execute complex tasks to perceive the environment and make decisions with limited human interaction or even without human interaction. In that way, localization and motion estimation are critical tasks for the operations an autonomous vehicle must accomplish. Position information is essential to identify the vehicle context and surroundings and move or act accordingly. Computer Vision-based approaches have shown promising results in mobile robotics, drones, or autonomous cars. However, the application and evaluation of CV-based solutions are more limited in the railway domain, especially in challenging environments. In this research, a state of the art of Visual Odometry (VO) and Visual SLAM (vSLAM) algorithms is carried out. In the SOTA, the analyzed VO/vSLAM algorithms are usually evaluated in outdoor street scenarios and do not consider the challenging perception conditions that can be found in urban underground railway scenarios, with low lighting conditions and texture-less areas in tunnels and significant lighting changes between tunnels and railway platforms. Moreover, there is no reference dataset in the VO/vSLAM community with such characteristics, raising the need to generate a proprietary dataset. Considering the lack of GPS signals in underground scenarios, a method is proposed to generate a ground truth of images and poses in underground railway scenarios. The generation process is based on synchronizing geodetic coordinates, train ATP data recorded from the radar and encoder sensors, and a railway gradient map provided by the railway infrastructure manager. Two state-of-the-art and recently proposed VO/vSLAM approaches (ORB-SLAM2 and DF-VO) have been tested in the generated proprietary datasets. These algorithms have achieved good performance in standard benchmarks such as KITTI and represent two distinct VO/vSLAM algorithm types: geometric and learning-based. However, the results show that the scenario lighting characteristics significantly affect the VO/vSLAM algorithms’ performance. In order to afford the challenging lighting conditions of the underground railway domain, the application of a data enhancement technique has been considered (EnlightenGAN). As calibration is critical for geometric VO/vSLAM algorithms, the impact of EnlightenGAN on the camera calibration parameters is also analyzed. The results demonstrate that EnlightenGAN does not considerably affect those parameters. Besides, it improves the performance of both VO/vSLAM approaches in challenging scenarios.