An Unstructured Terrain Traversability Mapping Method Fusing Semantic and Geometric Features

This paper presents a terrain traversability assessment method for autonomous navigation of mobile robots in unstructured outdoor environments. The method extracts terrain features from 3D point clouds and RGB images and merges them into a global terrain traversal cost map for safe trajectory planning. First, in view of the sparsity of raw point cloud data, Bayesian generalized kernel inference is applied to estimate the attributes of unknown grids, and Kalman filtering is used to fuse multi-frame information to generate dense local elevation maps in real time. Subsequently, the terrain semantic mapping takes into account the sensor noise and the semantic segmentation uncertainty and probabilistically updates the surface semantic information based on Bayesian filtering. Finally, geometric features are extracted from the elevation map and fused into the probabilistic semantic map to generate the terrain traversability map. The method is applied to a wheeled robot to perform traversability mapping tests in unstructured outdoor terrain. The experimental results show that the success rate of trajectory planning is improved by 7.7%-33.3% compared with other advanced terrain traversability analysis method.