关键词: 履带车辆, 越野环境, 可通行性分析, 高程地图, 语义分割

Abstract: To address the limitations of existing traversability analysis methods, which often suffer from incomplete obstacle recognition and poor generalization in complex off-road environments, this paper proposes an obstacle recognition algorithm based on open-vocabulary semantic segmentation. The method extracts semantic labels of obstacles and terrain around the vehicle, enabling effective identification of previously unseen obstacles in unstructured environments. The algorithm is validated on both datasets and real-world experiments, demonstrating its stability and comprehensive recognition capability. Building upon this, a multi-layer 2.5D map is constructed by integrating semantic labels with 3D point clouds. First, terrain is preliminarily classified into passability levels according to semantic labels. Second, terrain smoothness is quantified based on ground elevation, while geometric parameters of special environmental features (e.g., vertical walls) are measured. Furthermore, the vehicle’s driving posture is predicted by incorporating the geometric configuration of a tracked vehicle, thereby quantifying the coupling relationship among static slope stability, semantic terrain categories, and geometric attributes. A cost function is then designed to jointly assess traversal risk and cost, ultimately generating a vehicle-centric traversability map. Comparative experiments against state-of-the-art methods verify the effectiveness and reliability of the proposed approach, which enhances data support for decision-making, planning, and control of unmanned tracked platforms.

Key words: tracked vehicle, off-road environment, traversability analysis, elevation mapping, semantic segmentation