Grid Point Serialized Transformer for LiDAR Point Cloud Semantic Segmentation in Various Densities and Heights Scenes

Point cloud semantic segmentation is among the important tasks to achieve comprehensive perception of 3D environments. However, current segmentation methods suffer from limited local receptive fields, poor extraction of global information, and insufficient scene generalizability. To alleviate these problems, we propose the grid point serialized transformer (GridPSFormer), a semantic segmentation method based on space-filling curves. GridPSFormer maps gridded points to 1D serialized orders via the 3D spatial serialized order module, which shows superior locality-preserving and provides a comprehensive understanding of the 3D space. Then, by combining the local serialized attention mechanism and the global serialized mamba module, GridPSFormer effectively captures the local and global features of serialized orders and improves the modeling ability of points at long distances. In addition, the classes-refined serialization module complements the semantic context information to enhance the generalizability in various densities and heights scenes. GridPSFormer achieved SOTA performance on three datasets, HRHD_HK, WHU_ALS, and DALES, with 64.89%, 68.17%, and 85.50% of mIoU as well as 91.41%, 82.99%, and 98.24% of OA, respectively. Experimental results demonstrated that the various spatial serialized orders can efficiently explore 3D spatial information and achieve accurate semantic segmentation of large-scale scenes with lower computational complexity.