Semantic Terrain Traversability Analysis Based on Deep Learning Aimed at Planetary Rover Navigation

Abstract

Autonomous navigation is becoming an increasing area of interest, especially in the space sector. This is evident in the development of planetary exploration rovers, which rely heavily on terrain traversability capabilities to explore extraterrestrial terrains. This technology assesses the ability of a rover to identify potential obstacles of a given terrain based on its physical characteristics and visual appearance. In this paper, we present an innovative architecture tailored for such analysis. Our approach combines stereo visual SLAM for trajectory reconstruction with supervised learning from labeled ground-truth data, specifically utilizing DeepLabv3+, for precise pixel labeling of terrain types. The segmented point cloud generated from stereo vision is then converted into an occupancy grid map, facilitating comprehensive terrain characterization. Implementing our method within the Robot Operating System (ROS) framework enables seamless integration with rover systems. The proposed system was deployed and tested on a prototype rover, successfully demonstrating its ability to map the surrounding environment and identify potential obstacles. Finally, the realized mapping is compared with satellite images, and the results prove the effectiveness of the studied algorithm to carry out terrain traversability analysis.