Evaluating the Effectiveness of Projection Techniques for the Semantic Segmentation of LIDAR-Captured Point Clouds

This research evaluates the effectiveness of orthogonal, multi-view, and spherical projection techniques in the semantic segmentation of LIDAR-captured point clouds. Using U-Net, a deep learning architecture originally designed for biomedical image segmentation, the study processed 2D projections of point clouds to generate segmented images. Post-segmentation, label spreading was employed to map these labels back to the 3D point cloud. The orthogonal projection emerged as the most proficient, achieving an overall accuracy of 92% and a mean IoU of 49.69%. Multi-view projection, capturing occluded surfaces, registered slightly lower metrics, while spherical projection lagged with 82% accuracy due to inherent data warping. The findings underscore the pivotal role of projection techniques and U-Net in point cloud segmentation, suggesting orthogonal projection as the optimal approach for the most applications.