Traversability Analysis by Semantic Terrain Segmentation for Mobile Robots

Abstract

Mobile robots have the potential to be used in many outdoor tasks, such as search and rescue, patrolling, and delivery services. To enable robots to safely navigate through outdoor environments, it is important to analyse the terrain. We present a novel approach using Semantic Terrain Segmentation (STS), which relies on an adaptation of Deeplabv3+. Our goal is to segment the terrain according to different height thresholds, using only aerial RGB images. We collected and labelled a dataset, Vale, consisting of four categories based on the traversability constraints of three types of mobile robots: wheeled, tracked, legged and non-traversable. We trained our deep convolutional neural network (DCNN) on Vale, with transfer learning using a network pre-trained on Cityscapes. Our findings suggest that the proposed DCNN approach can identify and differentiate different height thresholds in the terrain, i.e. can segment based on mobile robot traversability.