From LiDAR point cloud towards digital twin city: Clustering city objects based on Gestalt principles

Recent advancement of remote sensing technologies has brought in accurate, dense, and inexpensive city-scale Light Detection And Ranging (LiDAR) point clouds, which can be utilized to model city objects (e.g., buildings, roads, and automobiles) for creating Digital Twin Cities (DTCs). However, processing such unstructured point clouds is very challenging, epitomized by high cost, movable objects, limited object classes, and high information inadequacy/redundancy. We noticed that many city objects are not in random shapes; rather, they have invariant cross-sections following the Gestalt design principles, including proximity, connectivity, symmetry, and similarity. In this paper, we present a novel unsupervised method, called Clustering Of Symmetric Cross-sections of Objects (COSCO), to process urban LiDAR point clouds to a hierarchy of objects based on their characteristic cross-sections. First, city objects are segmented as connected patches of proximate 3D points. Then, symmetric cross-sections are detected for symmetric city objects. Finally, the taxonomy and groups of city objects are recognized from a hierarchical clustering analysis of the dissimilarity matrix. Experimental results showed that COSCO detected the correct taxonomy and types of 12 cars from 24,126 LiDAR points in 8.28 s. Based on the cross-sections and taxonomy, a digital twin was created by registering online free 3D car models in 29.58 s. The contribution of this paper is twofold. First, it presents an effective unsupervised method for understanding and developing DTC objects in LiDAR point clouds by harnessing innate Gestalt design principles. Secondly, COSCO can be an efficient LiDAR pre-processing tool for recognizing symmetric city objects' cross-sections, positions, heading directions, dimensions, and possible types for smart city applications in GIScience, Architecture, Engineering, Construction and Operation (AECO), and autonomous vehicles.