Search

CN-121982207-A - Building live-action three-dimensional reconstruction system based on point cloud data

CN121982207ACN 121982207 ACN121982207 ACN 121982207ACN-121982207-A

Abstract

The invention discloses a building live-action three-dimensional reconstruction system based on point cloud data, which relates to the technical field of building live-action three-dimensional reconstruction and comprises a preprocessing module, a multi-stage self-adaptive noise filtering module, a semantic and geometric joint reasoning module and a three-dimensional model generation module which are sequentially connected, wherein the preprocessing module is used for receiving original building point cloud data, carrying out registration and initial space division, and outputting a registered point cloud data set. According to the building live-action three-dimensional reconstruction system based on the point cloud data, a mechanism of multi-stage self-adaptive noise filtering and semantic geometry joint reasoning cooperation is introduced, so that the contradiction that the existing method is inaccurate in noise rejection and easy to lose feature details in a complex building scene is effectively solved. The system can intelligently distinguish real discrete noise from inherent complex geometric characteristics of a building, ensures the denoising effect, and simultaneously furthest reserves the integrity of details such as decorative components, curved roofs and the like, thereby realizing high-fidelity point cloud data preprocessing.

Inventors

  • YIN JIANDONG

Assignees

  • 杭州中联筑境建筑设计有限公司

Dates

Publication Date
20260505
Application Date
20260125

Claims (10)

  1. 1. The building live-action three-dimensional reconstruction system based on the point cloud data is characterized by comprising a preprocessing module, a multi-stage self-adaptive noise filtering module, a semantic and geometric joint reasoning module and a three-dimensional model generating module which are connected in sequence; The preprocessing module is used for receiving the original building point cloud data, registering and dividing an initial space, and outputting a registered point cloud data set; the multistage self-adaptive noise filtering module is connected with the output end of the preprocessing module and is used for executing a processing flow at least comprising two different filtering mechanisms on the point cloud data set, identifying and marking potential noise points, wherein the first filtering mechanism screens based on local geometric attributes of the point cloud, and the second filtering mechanism screens based on distribution consistency of the point cloud in a feature space; The semantic and geometric joint reasoning module is connected with the output end of the multi-stage self-adaptive noise filtering module and is used for receiving filtered point cloud data, synchronously carrying out multi-type geometric primitive fitting and semantic category prediction, and carrying out secondary verification and correction on the potential noise points according to the relevance between the fitting result and the semantic prediction result to generate denoised point cloud data with semantic labels; The three-dimensional model generation module is connected with the output end of the semantic and geometric joint reasoning module and is used for generating a building live-action three-dimensional model with a complete topological structure and semantic annotation based on the denoised point cloud data with semantic labels and fusion of multi-scale geometric features and semantic context information.
  2. 2. The three-dimensional reconstruction system of building live-action based on point cloud data of claim 1, wherein the multistage adaptive noise filtering module comprises a primary filtering unit and a secondary filtering unit; The primary filtering unit is used for executing the first filtering mechanism and is specifically configured to perform self-adaptive space partitioning on the point cloud data set, calculate the local point density and curvature value of each point in each partitioning, mark the points with the local point density lower than a first dynamic threshold and the curvature value higher than a second dynamic threshold as primary suspected noise points, and self-adaptively determine the first dynamic threshold and the second dynamic threshold according to the statistical distribution of point clouds in the partitioning; The secondary filtering unit is connected with the primary filtering unit and is used for executing the second filtering mechanism, and the secondary filtering unit is specifically configured to extract multidimensional feature vectors which are not marked as primary suspected noise points in the point cloud data set, wherein the multidimensional feature vectors at least comprise space coordinates, normal vectors and color information, perform cluster analysis on the multidimensional feature vectors in a feature space, mark points belonging to clusters with the number of cluster members being less than a preset number threshold and points with the center distance of a main cluster exceeding a dynamic distance threshold as secondary suspected noise points, and the union set of the primary suspected noise points and the secondary suspected noise points forms the potential noise point set.
  3. 3. The three-dimensional reconstruction system for building live-action based on point cloud data as claimed in claim 2, wherein the semantic and geometric joint reasoning module comprises a geometric primitive fitting unit, a semantic prediction unit and a joint decision unit; The geometric primitive fitting unit is used for parallelly executing three types of geometric primitives of planes, cylindrical surfaces and spherical surfaces for each local area in the filtered point cloud data, and calculating residual errors of each point relative to various fitting primitives; The semantic prediction unit is used for carrying out semantic category prediction on each point in the filtered point cloud data based on a neural network model and outputting probability distribution of each point belonging to a plurality of predefined building component categories, wherein the building component categories at least comprise walls, roofs, windows, doors and decoration components; The joint decision unit is respectively connected with the geometric primitive fitting unit and the semantic prediction unit and is used for making decisions according to the residual errors and the probability distribution, and is configured to search a plurality of adjacent points in a space neighborhood of a target point marked as a potential noise point, if the fitting residual errors of points belonging to the same semantic category in the adjacent points on a geometric primitive of a certain type are generally lower than a residual error threshold value, the fitting residual errors of the target point on the geometric primitive of the certain type are higher than the residual error threshold value, and the predicted semantic category probability distribution of the target point is inconsistent with the main semantic category of the adjacent points, the target point is confirmed to be a final noise point, otherwise, the target point is removed from the potential noise point set.
  4. 4. The three-dimensional reconstruction system for building live-action based on point cloud data as set forth in claim 3, wherein said system further comprises an iterative optimization module connected between said multistage adaptive noise filtering module and said semantic and geometric joint reasoning module to form a closed feedback loop; The iterative optimization module is configured to feed back a final noise point set to the primary filtering unit after the final noise point set is output by the joint decision unit, the primary filtering unit dynamically adjusts the determination strategies of the first dynamic threshold and the second dynamic threshold in the next round of processing according to the spatial distribution and the attribute characteristics of the final noise points confirmed in the current iteration, and the system executes the complete filtering and reasoning process including the feedback loop at least twice.
  5. 5. The three-dimensional reconstruction system for building live-action based on point cloud data as set forth in claim 3, wherein said three-dimensional model generation module comprises a feature fusion unit and an implicit reconstruction unit; The feature fusion unit is used for extracting multi-scale features of the denoised point cloud data with semantic tags, wherein the multi-scale features comprise geometric detail features obtained by local neighborhood calculation, structural features obtained by larger area context information aggregation and semantic features output by the semantic prediction unit; The implicit reconstruction unit is connected with the feature fusion unit and is used for taking the enhanced point cloud features as input and generating a continuous seamless three-dimensional grid surface model by solving an implicit field function.
  6. 6. The three-dimensional reconstruction system for building live-action based on point cloud data as claimed in claim 5, wherein said three-dimensional model generation module further comprises a topology verification and repair unit; The topology verification and restoration unit is connected with the implicit reconstruction unit and is used for carrying out automatic topology analysis on the generated three-dimensional grid surface model, detecting and identifying non-manifold edges, isolated vertexes or self-intersecting patches existing in the model, carrying out automatic restoration of category perception on the identified topology errors based on semantic tags output by the semantic and geometric joint reasoning module, wherein the restoration at least comprises bridging of broken edges belonging to the same semantic component and filling of small-scale holes belonging to unstructured semantic categories and generated by noise removal.
  7. 7. The three-dimensional reconstruction system for building live-action based on point cloud data as set forth in claim 1, wherein the initial space in the preprocessing module is divided into a plurality of three-dimensional blocks aligned with the building structure by adaptively dividing the point cloud data into a plurality of three-dimensional blocks based on the point density variation in the horizontal plane with reference to the identified floor height, wherein the obvious floor height aggregation interval is identified based on a height distribution histogram in the vertical direction of the building extracted from the registered point cloud.
  8. 8. The three-dimensional reconstruction system for building live-action based on point cloud data as set forth in claim 2, wherein the adaptive spatial partitioning in the primary filtering unit is a non-uniform partitioning strategy, wherein small partitioning sizes are adopted in areas with high point cloud density, and large partitioning sizes are adopted in areas with low point cloud density.
  9. 9. The building live-action three-dimensional reconstruction system based on the point cloud data, according to claim 3, is characterized in that a neural network model adopted by the semantic prediction unit uses a composite loss function integrating geometric loss and semantic loss in the training process, wherein a geometric loss term encourages network learning to have the characteristic of invariance to the local geometric structure transformation of the point cloud, and the semantic loss term supervises the semantic category of the accurate predicted point of the network.
  10. 10. The three-dimensional reconstruction system for building live-action based on point cloud data as set forth in claim 1, wherein the system further comprises an output module connected with the three-dimensional model generation module, and the output module is used for converting the generated three-dimensional model for building live-action with complete topological structure and semantic annotation into a file conforming to industrial basic class standard or general three-dimensional format, and associating semantic tags, geometric parameters and topological relation information of each component in the storage model.

Description

Building live-action three-dimensional reconstruction system based on point cloud data Technical Field The invention relates to the field of building live-action three-dimensional reconstruction, in particular to a building live-action three-dimensional reconstruction system based on point cloud data. Background The building live-action three-dimensional reconstruction technology is used for converting a building in the real world into a high-precision visual three-dimensional model through various digital means, and is widely applied to the fields of urban planning, building protection, intelligent operation and maintenance, virtual display and the like. The point cloud data acquired by the technologies of laser scanning, oblique photography and the like can truly reflect the geometric structure and the surface characteristics of the building, and provide a reliable data base for subsequent analysis, simulation and display. However, in the actual acquisition process, the point cloud data is often affected by multiple noises such as environmental interference, equipment errors, motion distortion and the like, so that the problems of burrs, holes, blurred edges and the like of the reconstructed model occur, and the accuracy and usability of the three-dimensional model are seriously affected. In order to improve the accuracy and robustness of three-dimensional reconstruction of a building, various point cloud denoising and reconstruction methods have been proposed in the prior art. For example, the grant publication number CN118918244a describes a method and a system for three-dimensional reconstruction of a building based on point cloud, and the method screens non-noise points by recursive nearest neighbor search, and determines discrete points by combining plane fitting, thereby finally realizing noise rejection and three-dimensional model generation. The technology can identify and remove isolated noise points to a certain extent, and improves the smoothness of the reconstructed surface. However, the method still has the following defects that firstly, recursive nearest neighbor searching depends on random selected points and fixed thresholds, when point cloud distribution is uneven or a complex structure is formed, effective points are prone to being deleted by mistake or noise point residues, secondly, the method for judging discrete points by plane fitting is poor in adaptability to a non-plane area and is prone to causing characteristic loss, wherein the non-plane area comprises a curved surface, an eave angle and a decoration member, thirdly, fusion of multi-source and multi-scale point clouds in building reality and semantic information association are not fully considered, structural hierarchy and reality are lacked in reconstruction results, and fine modeling and subsequent application are difficult to support. Therefore, a point cloud reconstruction system which can adapt to complex building scenes, maintain the integrity of geometric features and have high robustness is still lacking in the prior art, and a more intelligent, precise and practical three-dimensional reconstruction scheme for building reality is needed to be proposed. Disclosure of Invention The invention aims to provide a building live-action three-dimensional reconstruction system based on point cloud data, which aims to solve the problems in the background technology. In order to solve the technical problems, the invention provides the technical scheme that the building live-action three-dimensional reconstruction system based on the point cloud data comprises a preprocessing module, a multi-stage self-adaptive noise filtering module, a semantic and geometric joint reasoning module and a three-dimensional model generating module which are connected in sequence; The preprocessing module is used for receiving the original building point cloud data, registering and dividing an initial space, and outputting a registered point cloud data set; the multistage self-adaptive noise filtering module is connected with the output end of the preprocessing module and is used for executing a processing flow at least comprising two different filtering mechanisms on the point cloud data set, identifying and marking potential noise points, wherein the first filtering mechanism screens based on local geometric attributes of the point cloud, and the second filtering mechanism screens based on distribution consistency of the point cloud in a feature space; The semantic and geometric joint reasoning module is connected with the output end of the multi-stage self-adaptive noise filtering module and is used for receiving filtered point cloud data, synchronously carrying out multi-type geometric primitive fitting and semantic category prediction, and carrying out secondary verification and correction on the potential noise points according to the relevance between the fitting result and the semantic prediction result to generate denoised point cloud data