Construction of Terrain Multidimensional Traversibility Feature Map for Off-Road Scenarios Based on Binocular Vision 2024-01-2046

Terrain Traversability Feature (TTF) map, which could be constructed by the images and point cloud data base on binocular vision, often using multi-frame fusion technology to expand the coverage area. However, the common challenges of off-road scenarios such as missing GPS data or single terrain features seriously hindered the alignment of adjacent frame data. Additionally, traditional TTF map depict the vehicle's surroundings only based on a few features such as terrain elevation or category. And it is insufficient for complex off-road scenarios navigation tasks.

This paper presents a method for constructing a Terrain Multidimensional Traversability Feature (TMTF) map for off-road scenarios based on binocular vision. First, we utilize the point cloud data from a binocular camera to construct a grid map model. Therefore, the geometric features of the terrain could be calculated with the grid as the basic unit, and a single-frame TMTF map of off-road scenarios is established. Subsequently, we propose an adhesion coefficient estimation method based on image semantics considering uncertainty, which successfully helps TMTF map to further describe the terrain category and mechanical characteristics. Then, an inter-frame pose transformation estimation method integrating wheel speed and direct method visual odometry is designed. It registers and fuses the historical and current TMTF maps to expands the vehicle's perception range, and effectively solves missing GPS data and single terrain features of off-road scenarios.

Finally, the test and verification are carried out based on the playback of the collected real vehicle data. The test results clearly demonstrate the constructed TTFM effectively represents multi-dimensional features like terrain type, geometry, and mechanics in off-road environments. At the same time, the traversability feature description of non-visible areas is successfully supplemented through fusion of multi-frame TMTF map.