Large Scale Point Cloud Semantic Segmentation for Indoor Digital Twin Generation

Automated processing of large amounts of sensor data poses a significant challenge in many fields, particularly with LiDAR point clouds. One field of application for LiDAR is the creation of high-definition (HD) maps, which are utilized in various domains, such as mobile indoor navigation. For visually impaired individuals, indoor navigation is crucial for enhancing their quality of life and independence. In this work, we focus on two advanced methods for semantic segmentation of point clouds generated using LiDAR and 360° camera sensors. These methods generate digital twins to create accurate and detailed representations of physical environments. The digital twins can be used to produce HD maps for indoor navigation. The first method involves converting the point cloud into a graph structure known as a superpoint graph (SPG). The second method, RandLA-Net, is based on efficient random sampling of points within the point cloud. Both methods are evaluated with our dataset, achieving an overall accuracy of 89%. This reproduced performance is consistent with their results on the public point cloud benchmark from Stanford University, demonstrating the efficiency of state-of-the-art semantic segmentation methods.