Abstract:Airborne LiDAR point cloud segmentation (PCS) is often employed as a preprocessing step for the subsequent object recognition for scene interpretation. Current segmentation methods often aim at single-wavelength LiDAR data by fully exploiting the spatial information, which makes them unsuitable for multispectral point cloud (MPC) data due to ignoring the use of spectral signatures. In this article, a normalized spatial–spectral supervoxel segmentation method is proposed for MPC data. Specifically, a normalized spectral–spatial metric is developed to construct the ${k}$ -dimensional tree (KD tree) for MPC data clustering. Considering the uneven density distribution of MPC, an adaptive energy minimization principle based on the sum of the distance is devised to accurately select the seed points of voxels, solving the problem of undersegmentation. To reduce the cross-boundary points, the normalized spectral–spatial metric with the concave–convex judgment is extended to further optimize the edges between adjacent voxels. An important asset of our method is to segment MPC without the need for any manual annotation. Experiments on two MPC datasets show that the proposed method yields better performance compared to several comparative methods.

A Normalized Spatial–Spectral Supervoxel Segmentation Method for Multispectral Point Cloud Data