Research On Key Technologies Of Building High-Precision Surface Reconstruction Based On 3D Point Cloud

Lightweight 3D models of buildings are the key foundation for digital cities,digital twins and other application scenarios.Currently,indoor and outdoor reconstruction methods based on 3D point clouds are the current research hotspots in the fields of computer vision and remote sensing mapping.The current main technical challenges are focused on the following two points: firstly,due to the restricted acquisition viewpoint and scenery occlusion,the outdoor acquired single building 3D point cloud often has local data missing and is more seriously affected by noise and other factors;secondly,indoor 3D point cloud scenes,such as walls and other structural information are easily disturbed by indoor furniture and other movable targets.Therefore,this thesis adopts spatial slicing and graphical element connection strategies,and deep learning strategies based on indoor target detection,respectively,to carry out the exploration of indoor and outdoor 3D surface model reconstruction techniques for buildings,and the main research works are as follows.1.For the problem of missing data caused by uneven density and occlusion of 3D point clouds in large outdoor scenes,this thesis proposes a connection-aware map-based surface reconstruction algorithm for 3D point clouds of buildings.Firstly,a planar primitive growth structure and a multiconfidence hierarchy are designed based on the dynamical structure to solve the problem of primitive initialization failure caused by missing data.Secondly,a connected graph structure with different connection relationships is constructed for accelerating the 3D reconstruction solving process,based on which the building scaffolding structure is constructed through the graph structure for the point cloud data.Finally,the reconstruction energy equation is constructed based on the mapping relationship between the point cloud data and the triangular surface piece of the reconstruction model,and the minimization solution of the energy equation can effectively solve the problem of poor fitting of the algorithm to the data plane primitives and significantly improve the reconstruction robustness under the missing data.The experimental results for SUM Helsinki3 D,a large outdoor scene dataset,show that the proposed algorithm can effectively extract watertight models of buildings with different styles in different scenes,and can obtain Lo D2-level building models with more geometric details than the mainstream algorithms,with an average reduction of 23.7% in reconstruction errors and a model compression ratio of less than 3% in evaluation indexes.2.A natural language-driven reconstruction algorithm for building interior surfaces is proposed for the problem of movable occlusions in building interior 3D point cloud scenes.Firstly,a natural language verification module is designed to incorporate the strong a priori information in the scene description text language into the 3D reconstruction process.Secondly,an iterative natural language guidance module is constructed for the complementation and information fusion between different instance objects.This strategy not only improves the accuracy of the instance segmentation network and the 3D reconstruction network in both directions,but also effectively removes the redundant information of the movable objects in the indoor scenes so as to preserve the structural information in the room.Experimental results for indoor datasets Scan Net and Scan2 CAD show that,for indoor scene data,the natural language-driven instance reconstruction framework proposed in this thesis can fully decouple indoor movable objects from structural scenes,realize functions such as rearrangement of indoor scenes and removal of movable objects,and effectively improve several reconstruction accuracy metrics on the datasets.In summary,this thesis proposes a planar element-based connection-aware map and its growth strategy,as well as a natural language-driven reconstruction algorithm for indoor and outdoor building surfaces,respectively,targeting the challenges of missing data of outdoor point clouds and interferences of indoor movable targets.The experimental results for mainstream datasets show that the above reconstruction strategies effectively improve the accuracy of building reconstruction and can realize the integrated reconstruction of indoor and outdoor models.