Towards Learning Obstacles to Avoid Collisions in Autonomous Robot Navigation

Abstract

Avoiding obstacles is one of the main tasks in robotic navigation. In this paper, robot navigation using monocular vision is presented. Therefore, an accuracy in the segmentation of obstacles is necessary to avoid collisions by estimating the Time-to-Contact. Our proposal in this research process is based on using YOLO so that through a training process, the robot identifies which regions of the image are potentially obstacles. The experimentation was performed in a real environment, with low daylight and without controlling lighting parameters. The first results of this approach are satisfactory although this project will continue with the learning of other obstacles.