SEMANTIC MAPPING OF INCREMENTAL 3D POINT CLOUDS BASED ON MULTI-HOP GRAPH ATTENTION NETWORK

Mapping and Semantic mapping are the important research areas for the autonomous navigation of mobile robots. However, realising point cloud information extraction in a dynamic environment is still a challenge in semantic mapping. To solve this problem, We propose a method called PMGAT-SM(semantic mapping of 3D point clouds based on multi-hop graph attention network) for achieving semantic mapping in 3D point cloud environments. In PMGAT-SM, we designed a point cloud classifier PMGAT, which extracts semantic information of unordered point clouds by constructing a graph. We combined PMGAT with dynamic growing clustering to achieve instance segmentation in complex environments. Our extensive experiments in the KITTI scene show the effectiveness of the semantic mapping model. Meanwhile, compared with popular semantic information extraction models, the PMGAT performs better in fine-grained feature extraction of point cloud segments.