Incremental large scale dense semantic mapping

In order to efficiently achieve accurate large-scale scene understanding result, A new large scale dense semantic mapping system was proposed. The system constructed a map by incrementally calculating with a conditional random field model. The method used stereo visual odometry to get the motion of the camera, and used the labeled image sequences to build semantic map. The key point was to incrementally build the semantic map which detected newly built voxels, over-segment the points within these voxels into supervoxels, labeled these supervoxels under the guidance of neighboring frames and used the rigid transformation matrix to fuse the newly labeled points with the already built map. A conditional random field model was constructed which took labeling results of sequential frames as the data term, took the coherent labeling constraint between neighboring supervoxels as the pairwise term and solved the model by graph cut. Experimental evaluations show that the approach can get an accurate large scale semantic map and decrease computational cost, The approach can improve the labeling results at image level. © 2016, Zhejiang University Press. All right reserved.