Adaptive Block Elevation Mapping for Large-scale Scene.
Yang Zhou,Shiqiang Zhu,Huang Huang,Yuehua Li,Jason Gu
DOI: https://doi.org/10.1109/ROBIO55434.2022.10011710
2022-01-01
Abstract:Dense map that contains the surrounding geometry and vision information of a robot is widely used for path planning, navigation, obstacle avoidance and other applications. Considering the performance of the processing unit mounted on the robot is limited, mapping algorithm has to make compromise by sacrificing speed and precision. It will be more challenging when the dense mapping scene is very large because the memory consumption will be greatly increased and the map is difficult to be extended if beyonding the initial map. To suppress the negative impact from the increased map scale, we proposed a novel block mapping approach to generate the dense map in large scale of scene. In this work, the elevation map is selected as the base dense map. The entire elevation map is segmented into numerous block maps of which size is much smaller than that of the entire map. The present moment of lidar and vision measurements are used to generate the local elevation map. The local elevation map is used to update block maps which are adaptively generated along the motion trajectory. A memory-disk interaction mechanism, which the block maps will be loaded to memory or saved to local disk when needed, is introduced. Our block mapping approach is tested on the KITTI datasets, and the results demonstrate that the mapping approach can stably operate in a large scale of scene with a much smaller consumption of memory.