Chi Guo,∗ Kai Huang,∗∗ Yarong Luo,∗ Huyin Zhang,∗∗ and Wenwei Zuo∗∗∗
Semantic mapping, object detection, laser returns segmentation, global optimization
Simultaneous localization and mapping (SLAM) is an important skill of mobile robots. Pure SLAM without considering semantic information can only finish basic tasks such as navigation. This paper focuses on semantic mapping on a mobile robot equipped with both a two-dimensional (2D) laser range finder and a monocular camera. We established a semantic mapping system which leverages object detection method based on the deep learning and light detection and ranging (LIDAR) SLAM, building occupancy maps marked with semantic labels. Firstly, to solve the problem that semantic information of 2D laser range data is difficult to be extracted, a method is porposed to calculate the semantic laser returns corresponding to recognized objects by matching the bounding boxes and geometrical segments of laser range data. Secondly, we designed a semantic grid and proposed a mapping strategy that generates the occupancy probability and semantic probability distribution in grids. We dynamically optimize semantic maps by the global optimization of submaps and clustering of semantic grids. We also evaluate the system using datasets collected in two typical indoor scenes and prove that it can build global consistent semantic maps.
Important Links:
Go Back
