Real-time visual off-road path detection

Abstract

In this paper, we propose a fast and real-time capable system for visual off-road path detection. We equipped our robot AMOR with a single monocular camera and explored unstructured environments like woods. In these areas, it is almost harder to identify and classify drivable and non-drivable parts in an image. In urban regions, roads can be detected by lane markers or delimitations whereas the boundaries of a forest path blend into the environment almost seamlessly. In our work, we developed a software system that is based on mostly simple and low computationally intensive algorithms. We developed and tested the functions with a large dataset of camera images and also generated a manually Ground Truth for the evaluation.