Semantic 3D mapping can be used for many applications such as robotnavigation and virtual interaction. In recent years, there has been greatprogress in semantic segmentation and geometric 3D mapping. However, it isstill challenging to combine these two tasks for accurate and large-scalesemantic mapping from images. In the paper, we propose an incremental and(near) real-time semantic mapping system. A 3D scrolling occupancy grid map isbuilt to represent the world, which is memory and computationally efficient andbounded for large scale environments. We utilize the CNN segmentation as priorprediction and further optimize 3D grid labels through a novel CRF model.Superpixels are utilized to enforce smoothness and form robust P N high orderpotential. An efficient mean field inference is developed for the graphoptimization. We evaluate our system on the KITTI dataset and improve thesegmentation accuracy by 10% over existing systems.
展开▼
