Application of construction site management based on digital twin and point cloud semantic segmentation technology

Unmanned Aerial Vehicle (UAV) photogrammetry and laser scanning can directly collect data in a variety of large and complex environments for inspection, precise navigation, and object recognition [43], for example, to identify structural elements such as walls, floors, and ceilings. The 3D point clouds, surface models, and orthogonal images generated from UAV images contain a wealth of information and are commonly used in outdoor environments. Meanwhile, with the development of mobile laser scanning technology, laser point clouds are widely used for their high accuracy and accessibility [20, 21]. Handheld Simultaneous Localization and Mapping (SLAM) is a LiDAR scanner based on SLAM algorithms to acquire point clouds through mobile scanning quickly; this 3D laser scanning technology has been widely used in architecture to provide a large amount of information and points [41; 42].The point cloud acquisition device used in this research is the GeoSLAM ZEB Horizon multifunctional Light Detection and Ranging (LiDAR) SLAM scanner, which can be handheld or backpack-mounted. It has a range of 100-m, can acquire 300,000 points per second, and achieves a relative accuracy of six mm.LiDAR SLAM uses laser sensors to create a map of its environment. LiDAR stands for Light Detection and Ranging and works by sending light pulses to features and measuring the time it takes for them to reflect. This gives the exact distance of the object or feature. The output is typically 2D (x, y) or 3D (x, y, z) point cloud data and can be effectively used for 3D reconstruction.3.2 RandLA-Net point cloud semantic segmentationDue to the complexity of the construction environment, this study establishes a construction site point cloud segmentation model based on RandLA-Net. It develops and establishes a construction site point cloud recognition system using the Python programming environment. Training parameters: 100 epochs, batch size 8, cross-entropy loss function, Adam optimizer, and NVIDIA RTX 3080 GPU computational setup.