Vision-based Terrain Segmentation and Roughness Estimation

Abstract

Intelligent terrain perception for search-and-rescue robotic applications, requires a high-level understanding of both the terrain type and its chief physical characteristics. Roughness is one such important terrain property, since it could play a key role in robot control/planning strategies, while navigating
in an unknown environment. In this paper, we present a single deep neural network architecture that predicts the pixel-wise terrain labels (i.e., sand, stone, wood, metal, road/sidewalk, and grass) and regresses their roughness from an input RGB image. Our approach, inspired by human analogy, leverages the basic image feature space from a pre-trained network (SegNet) to estimate the roughness. We experimentally validate our approach in real-world images, using RGB cameras. Moreover, we implement the algorithm on our four-legged centaur-like robot CENTAURO and demonstrate the use of our method in
assuring the stability of the robot in real-world scenarios, where the robot is traversing terrains of varying roughness.