Indoor map construction algorithm based on RGBD semantic segmentation

Abstract

In the field of robotics and autonomous driving, vision or laser-based map construction has always been the main direction for solving mobile vehicle's perception and localization. The stereo camera is widely used in robot map construction because it can perceive both color information and depth information. Based on the RGBD semantic segmentation network, this paper proposes a map construction algorithm based on deep semantic segmentation. By using the pixel information of deep semantic segmentation, the missing part of the 3D point cloud is filled to construct an octomap. After experiments, on the datasets, the algorithm has achieved better results than only using the depth information, and after actual deployment, the algorithm has completed the construction of real-time indoor semantic maps on the robot.