Efficient construction of topological semantic map with 3D information

Abstract

Path planning for human beings is based on scene understanding and semantic map as well as geometry features. The point cloud map can help robots to perform path planning. However, it is quite different from the natural way of human beings. In this paper, we propose to construct a novel framework for topological semantic map. Specifically, we construct a 2D semantic map by projecting 3D scene semantic information recognized by convolutional neural network onto a 2D plane. 3D reconstruction of the environment is achieved by RGB-D SLAM 3D space mapping algorithm. The intersections in the 2D map are recognized offline, and the semantic annotation of the intersection in the topological map is utilized to build up a complete object-based semantic map. Experimental results demonstrate the feasibility of the proposed approach.