Semantic–geometric hybrid terrain traversability mapping with risk assessment for UGV off-road navigation

Reliable terrain traversability assessment for autonomous navigation of unmanned ground vehicles (UGVs) in off-road environments requires considering both terrain physical properties and UGV mobility capabilities. To this end, this paper proposes a real-time terrain traversability mapping system based on semantic–geometric hybrid representations that incorporates explicit risk assessment into the mapping process. First, a lightweight semantic segmentation model is developed using knowledge distillation and model quantization to produce image semantic masks. These masks are associated with point clouds, and the resulting semantic point clouds are discretized into vehicle-centric local grids. Terrain height, surface normal, roughness, and the semantic probability distribution of terrain categories are then extracted from each grid cell to construct a single-frame local terrain traversability element map (TTEM). Grid-level temporal fusion integrates multi-frame observations into an updated local TTEM without retaining large amounts of raw observation data. Finally, the anisotropic risks of overturn failure, clearance failure, vibration failure, and traction failure are quantified considering both terrain properties and UGV mobility parameters. These risk indicators are then integrated into a traversability score to generate the local terrain traversability map (TTM). In addition to lightweight semantic segmentation and grid-level temporal fusion, a multi-module asynchronous parallel pipeline is introduced at the system level to further improve computational efficiency. Experiments on the custom StereoOffroad dataset show that the proposed system can effectively assess terrain traversability. Controlled simulations show that the risk indicators are associated with vehicle responses and can distinguish failure events across three simulated UGV platforms. The proposed method also outperforms representative existing methods in overall failure-event discrimination when the four test scenarios are evaluated jointly. Additionally, runtime evaluations show that the proposed system satisfies real-time mapping requirements under constrained computational resources.