Point Cloud Segmentation & 3D Model Construction

A two-pronged pipeline for segmenting and labelling indoor LiDAR scans of multistory buildings captured using an iPad Pro and constructing a simplified polygonal room-wise solid model is introduced. The proposed pipeline aims to offer a quick and user-friendly method, prioritizing speed and ease of use over accuracy and detail, to gain a preliminary understanding of the building and its potential. It involves building an inventory of the building’s contents and creating a simplified 3D model that can be used in existing environmental analysis tools. Implemented in C++ and exposed as a Python module utilizing pybind11 [14] the pipeline employs machine learning, specifically Yolo v8.1, trained on Coco to identify and label objects based on sequential RGB images captured with the iPad’s camera sensor. Simultaneously, an analytical approach constructs a simplified 3D model by performing planar region growing, ray tracing and clustering in conjunction with a linear integer solver for finding a water-tight solution for a 3D cell complex generated by intersecting a set of planes. The combined approaches yield a simplified solid room-wise polygonal 3D model containing segmented and labelled point clouds. The pipeline has been evaluated on scans of 2 to 3-story office buildings. The results demonstrate that the pipeline can segment and label scans with high accuracy, and the reconstructed 3D model accurately captures the overall shape of rooms. However, additional work is required to enhance the approach further. Future efforts should focus on robustness and increasing the number of supported objects that can be detected.