Mobile laser scanning (MLS) generates dense and detailed point clouds that present difficulties in effective processing as a result of uneven point distribution. The voxelization processes the data, maintaining the spatial structure and reducing the size of the data as the point clouds are converted to regular 3D grids. The main objective of this study is to understand how voxelization affects MLS semantic classification. Although voxel-based methods are widely preferred, current state-of-the-art do not explicitly assess voxel-point strategies in a systematic manner. This paper considers point-voxel feature strategies for semantic classification of street-level objects from MLS data, and four different approaches are evaluated on three benchmark datasets: TUM-MLS1, Toronto-3D and SynthCity. Our main contribution is in-depth assessment of these approaches including two novel strategies: a multi-scale voxelization that captures both fine and coarse spatial details and a hybrid method combining voxel-based and point-based features. The experimental results indicate that the strategies improve both classification accuracy and efficiency while maintaining robust performance across datasets with diverse point densities. Voxelization strategies achieved over 90% overall accuracy establishing strong handcrafted baselines for urban MLS classification. These results provide important insights useful for 3D object recognition in complicated environments and urban mapping.
Seyfeli et al. (Wed,) studied this question.