OTRNet: Pixel-Level Terrain Roughness Estimation for Off-Road Environments

Regional Traversability assessment constitutes a foundational task in autonomous driving, with terrain roughness serving as a critical factor influencing road perception. Off-road environments present unique safety challenges for vehicles due to undulating topography and heterogeneous distributions of surface materials and geometries, which contribute to pronounced spatial complexity and physical diversity. In this paper, we present the first large-scale multi-modal Off-road Terrain Roughness dataset NUDT-OTR. Accordingly, we propose a terrain roughness ground-truth generation framework leveraging LiDAR point cloud elevation differentials, where a dynamic-patch module is introduced in ground segmentation dealing with highly undulating terrains. Based on this, we further present a visual-based model OTRNet, which achieved robust estimation of terrain roughness in the BEV perspective through the designed Occlusion-Aware Module. Experimental results demonstrate significant estimation reliability of our approach across the proposed benchmark dataset.