Semantic OctoMap SLAM Based on Improved HRNet

In this paper, a semantic OctoMap SLAM system is proposed by combining image semantic segmentation with SLAM technology. The system is divided into two parts. In the semantic segmentation part, the improved HRNet combining the Multi-Layer Serial Dilated Convolution (MLSDC) and the Multi-Layer DUpsampling (MLDU) is used for semantic segmentation of RGB data. In the visual SLAM part, the camera pose estimation and optimization are carried out on the input RGBD data, and then the point cloud is preliminarily generated. Next, the Bayesian semantic fusion algorithm is used to fuse the semantic segmentation results and project them into the generated point cloud. Then, the point cloud is filtered to remove outliers and noise points. Finally, the octree map is constructed. Experiments show that the improved HRNet algorithm can retain more detailed features compared with the traditional algorithm, and has a better segmentation effect on the fuzzy edge, and the overall accuracy is improved. The final octree semantic map has intuitive mapping effect.