Room position and size deviation automatic detection method based on handheld point cloud collection

Abstract

This invention discloses an automatic detection method for room position and size deviations based on handheld point cloud acquisition, belonging to the field of three-dimensional spatial information processing technology. It aims to solve the problems of low data acquisition efficiency, disconnected processing, and lack of real-time feedback in existing technologies. This method relies on a handheld scanning terminal, a mobile application, and a cloud server. Through BIM-driven augmented reality scanning path planning, real-time on-site data quality monitoring, and intelligent point cloud processing with edge-cloud collaboration, it achieves automated identification and evaluation of room position and size deviations, and intelligently judges and visualizes the results. By adopting the above technical solution, this invention can achieve full automation and intelligence of the detection process, significantly improving operational efficiency and ensuring data quality and the reliability of analysis results.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional spatial information processing technology, and more specifically, to an automatic detection method for room position and size deviation based on handheld point cloud acquisition. Background Technology

[0002] In the entire lifecycle management of building construction and subsequent operation and maintenance, accurately verifying the consistency between the completed entity and the design blueprints is a core aspect of ensuring project quality, controlling construction costs, and mitigating safety risks. Specifically, the basic content of building quality inspection includes the precise measurement of macroscopic dimensions such as room height, width, and depth, as well as key components such as walls, columns, beams, and door / window openings, including their geometric dimensions, surface flatness, and spatial verticality. For a long time, technicians in this field have primarily relied on traditional contact or single-point measuring tools, such as tape measures, laser rangefinders, straightedges, and total stations widely used in surveying. These techniques, in their specific historical periods, provided fundamental solutions for measuring and defining project quality; their core working principle lies in approximating the geometric shape of components through discrete point measurements. However, the inherent limitations of such methods are becoming increasingly apparent. Not only are they inefficient and require high labor intensity from operators, but the more critical problem lies in the point-like sparsity of their measurement results. This makes it difficult to comprehensively and objectively reflect the continuous planar geometric features of components such as walls and floors, and the measurement process and results are easily affected by human operational errors.

[0003] To overcome the shortcomings of traditional measurement methods in terms of data completeness, ground-based 3D laser scanning technology has emerged. This technology, by emitting laser beams from a fixed station and receiving reflected signals, can acquire millions of 3D coordinate points of a target scene in a short time, forming high-density point cloud data. This technological paradigm greatly improves the comprehensiveness of data acquisition, enabling the use of surfaces instead of points to provide an unprecedentedly refined digital image of the geometric shape of building components, thus providing a reliable data foundation for the analysis of planar features such as flatness and verticality. However, the inherent operational mode of ground-based scanning technology presents new challenges when dealing with the dynamic and complex environment of construction sites. The equipment is typically bulky, the relocation and deployment processes are cumbersome, and to ensure data integrity, multiple station scans are often required in the same space, followed by complex point cloud data registration. This significantly prolongs the data processing cycle, making it difficult to meet the rapid response and immediate feedback requirements of construction sites.

[0004] In recent years, with the rapid development and maturation of Simultaneous Localization and Mapping (SLAM) technology, handheld 3D laser scanning devices integrating this technology have attracted widespread attention in the industry due to their unparalleled portability and flexibility. Operators can move freely within space using the handheld device to quickly collect point cloud data of the environment, which theoretically perfectly meets the rapid and mobile inspection needs of construction sites, effectively solving the problem of low operational efficiency of ground-based scanners. However, the in-depth application of this technology reveals a more hidden and profound contradiction: the current application of handheld scanning technology has largely resulted in a serious disconnect between front-end data acquisition efficiency and back-end data processing and application efficiency. Specifically, existing application solutions generally position handheld devices as simple "data acquisition tools," and the raw point cloud data they generate is essentially a set of geometric coordinates lacking semantic information and with a low degree of structure. To transform this raw data into meaningful quality inspection conclusions, subsequent data processing—including point cloud noise reduction, segmentation, alignment with the Building Information Model (BIM) or CAD drawings used as the design basis, and subsequent complex operations such as component identification, dimension extraction, and deviation comparison—still requires manual or semi-manual work by professional technicians on high-performance personal computers (PCs) using specialized software. This fragmented workflow means that the time saved by handheld devices in the front-end acquisition stage is consumed exponentially in the back-end "data processing." More importantly, this non-integrated process lacks an effective real-time feedback loop. On-site personnel cannot immediately ascertain whether the collected data quality meets the requirements for subsequent analysis, such as the presence of blind spots or insufficient point cloud density in key feature areas. These flaws are often only discovered when the data is imported into a PC for processing, leading to costly re-testing and significantly diminishing the high efficiency touted by handheld scanning devices in practical applications. The reason for this is that existing technologies have failed to deeply couple the portability of handheld devices with the intelligence and automation of data processing, and have failed to use the rich semantic information contained in the BIM model as prior knowledge to guide and drive the processing and analysis of point cloud data. As a result, a difficult-to-overcome gap has formed between the two key stages of "data acquisition" and "information extraction" in the entire inspection process.

[0005] Therefore, how to break through the technical bottleneck of "fast front-end, slow back-end" in current handheld scanning applications, deeply integrate the convenience of handheld scanning with the inherent intelligence of building information models, and construct a closed-loop method for the entire process from on-site data collection, real-time quality monitoring, cloud-based intelligent processing to automated deviation analysis and report generation, in order to overcome the disconnect between data collection and data application in existing technologies and achieve the immediacy, accuracy, and intelligence of building quality inspection, has become a key challenge and an urgent technical problem for those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies, such as the severe disconnect between the data acquisition efficiency and the efficiency of backend data processing and application of handheld scanning devices, the technological gap between data acquisition and information extraction, and the lack of effective real-time feedback and closed-loop management. To this end, this invention provides an automatic detection method for room location and size deviations based on handheld point cloud acquisition. This method constructs a fully automated system integrating augmented reality guidance, real-time quality monitoring, edge-cloud collaborative intelligent processing, and deep BIM-driven processes, aiming to achieve immediate, accurate, and intelligent detection and evaluation of location and size deviations in building interior spaces and components.

[0007] To achieve the above-mentioned objectives, this invention provides an automatic detection method for room position and size deviation based on handheld point cloud acquisition. The method relies on a system architecture consisting of a handheld scanning terminal, a supporting mobile application (App), and a cloud server. The handheld scanning terminal integrates a 3D LiDAR sensor, an inertial measurement unit (IMU), and a visible light camera. The 3D LiDAR sensor has a center wavelength of 905nm, a ranging range of at least 60 meters, a ranging accuracy better than ±10mm, and a point cloud acquisition rate of at least 320,000 points / second. The inertial measurement unit is a six-axis MEMS sensor, including a three-axis gyroscope and a three-axis accelerometer. The cloud server is equipped with a high-performance computing unit and a distributed database. The method specifically includes the following steps:

[0008] First, the task loading and augmented reality scanning path planning steps are performed. The operator launches the mobile application on the handheld terminal, connects to the cloud server via Secure Sockets Layer (SSL), and loads the Building Information Model (BIM) of the target project. The BIM data format is the Industry Basic Class (IFC) standard format. The system automatically performs BIM model parsing and automatic division of inspection task units. Specifically, the BIM parsing engine embedded in the system traverses the IFC data structure, identifies and extracts entity information for IfcBuildingStorey (floors), IfcGrid (grids), IfcSpace (spaces), and custom property sets (PropertySets) related to the division of construction flow sections. Based on the aforementioned information, the system automatically decomposes the entire project model into a series of independent inspection task units with unique identifiers (UUIDs). The data structure of each inspection task unit includes: task unit UUID, floor and area identifier, globally unique identifier (GUID) of the associated IfcSpace object, three-dimensional geometric bounding box of the space, and a list of GUIDs of all BIM components (such as walls, slabs, columns, doors, and windows) within the space.

[0009] After the detection task units are divided, the system automatically plans the optimal scanning path for the selected task units. The goal of this path planning algorithm is to generate the shortest movement trajectory covering all key surfaces within the space, while meeting the preset point cloud density (unit: points / square meter). Specifically, the system first voxels the three-dimensional geometric space of the task unit, with a voxel resolution of 5cm. Based on the BIM model, the voxels occupied by solid components such as walls, floors, and columns within the space are marked as obstacles. Subsequently, in non-obstacle areas, an improved Rapid Expanding Random Tree Star (RRT*) algorithm is used for path search. The cost function of this algorithm considers not only the path length but also the visibility of path points to the BIM component surfaces and the expected scan coverage. The algorithm finally outputs a spatial three-dimensional path composed of a series of six-degree-of-freedom (6-DoF) pose points. The mobile application projects this three-dimensional path onto the camera preview screen of the handheld terminal in real time, and by rendering a continuous three-dimensional arrow model, forms augmented reality (AR) visual guidance, instructing the operator to move along this optimized path.

[0010] Secondly, the on-site data acquisition and real-time quality monitoring steps are performed. The operator, holding the scanning terminal, moves within the room to be tested following the AR path guidance. The handheld terminal's built-in Simultaneous Localization and Mapping (SLAM) system, by tightly coupling the point cloud data from the 3D LiDAR sensor with the readings from the inertial measurement unit, calculates the device's six-degree-of-freedom pose in the world coordinate system in real time and simultaneously generates raw point cloud data. During the scanning process, the mobile application performs real-time registration and quality analysis of the point cloud data and the BIM model in an independent computing thread. Specifically, the system uses the device pose output by the SLAM system as the initial transformation matrix, downsamples a portion of the point cloud data generated per second to a voxel grid to a resolution of 5cm, and then performs lightweight registration with a pre-loaded, similarly voxelized BIM model using a one-iteration Generalized Iterative Closest Point (G-ICP) algorithm. After registration, the system evaluates in real time the coverage and density of the acquired point cloud on the surface of the BIM component. The implementation method is as follows: the triangular mesh model (mesh) of each BIM component is parametrically unfolded into a two-dimensional UV texture coordinate space. For each newly acquired and registered point cloud point, its nearest projection point on the surface of the BIM component is calculated, and the UV coordinates corresponding to the projection point are recorded in a two-dimensional rasterized cumulative matrix. The mobile application renders the surface of the BIM model in real time with three colors based on the count value in the cumulative matrix: when the raster count value corresponding to a certain area is greater than a preset density threshold ρ... target When the count value is between 0 and ρ, it is rendered as green; when the count value is between 0 and ρ, it target When the count is between 0 and 1, it is rendered as yellow; when the count is 0, it is rendered as gray. Based on this visual feedback, the operator performs additional scans on the areas displayed as yellow until most areas turn green, thereby ensuring the integrity and uniformity of the original data collection.

[0011] Next, the point cloud preprocessing steps are performed in a collaborative manner between the edge and cloud. After a detection task unit completes scanning, the operator confirms the submission on the mobile application. The handheld terminal locally performs feature-sensitive adaptive downsampling processing on the point cloud data. This downsampling processing aims to reduce the data volume to the greatest extent possible without losing key geometric features. The specific algorithm is curvature-sensitive adaptive voxel raster downsampling. First, an octree index is constructed from the collected complete raw point cloud data. Second, the covariance matrix of the three-dimensional coordinates of the subset of point clouds contained in each leaf node of the octree is calculated. Third, the covariance matrix is ​​decomposed into eigenvalues ​​to obtain three eigenvalues ​​λ1≥λ2≥λ3. Then, a normalized curvature metric C=λ3 / (λ1+λ2+λ3) is calculated based on the eigenvalues. This metric C approaches 0 when the point cloud distribution is linear or planar, and has a higher value when the distribution is at corners or scattered points. Finally, the retention strategy of the point cloud within the leaf node is determined based on the metric C: retain all points within the leaf node, or retain only its centroid. The probability of this decision is positively correlated with the C value, ensuring that point clouds in high-curvature areas such as corners, opening edges, and structural edges are preserved with a high probability, while point clouds on flat surfaces are significantly downsampled. After processing using this method, the amount of point cloud data is compressed to 15% to 20% of the original size.

[0012] The downsampled feature-enhanced point cloud data, along with its corresponding detection task unit UUID, is packaged and uploaded to the cloud server via an API interface based on the HTTPS protocol. Upon receiving the data, the cloud server first performs deep noise reduction processing. This processing employs a pre-trained point cloud noise classifier model based on a graph convolutional network. This model can identify and remove complex discrete noise points caused by material reflection properties, airborne particles, etc., and its performance is superior to traditional statistical outlier removal (SOR) or radius filtering algorithms.

[0013] After noise reduction, the server performs the core BIM-driven point cloud unit identification and semantic segmentation. This step utilizes the BIM model as strong prior knowledge to guide the point cloud segmentation process. Specifically, based on the received detection task unit UUID, the server queries the associated BIM database to retrieve the GUID list of all BIM components within that task unit, along with their precise geometric definitions and spatial location information. For each BIM component in the list, the server performs the following operations: extracting the boundary polygons and extrusion vectors of the wall from the BIM definition to generate its precise digital geometric surface. Then, in the received point cloud data, a search space is defined with the geometry of the BIM wall as the center and extending outwards by 30cm. Within this defined search space, a Directed Random Sample Consensus (RANSAC) plane fitting algorithm is executed. The sampling process of this algorithm is constrained within the search space, ensuring the robustness and efficiency of the fitting process. All points within the fitted plane constitute the subset of the point cloud belonging to that wall. The server separates this subset of point cloud from the main point cloud and assigns it a semantic label, which is directly derived from the GUID of the BIM component. This process is performed sequentially on all BIM components within the task unit until the entire point cloud data is precisely segmented into individual component point clouds with semantic labels that correspond one-to-one with the BIM components.

[0014] Subsequently, an automated multi-dimensional data quality compliance assessment step is executed. After completing the individual component segmentation, the system performs a quantitative quality assessment on each component point cloud subset to determine whether it meets the accuracy requirements of subsequent deviation analysis. This assessment system includes three core indicators: coverage completeness S... cov Data Validity den and feature completeness S feat Coverage completeness S cov The calculation formula is: the area covered by the point cloud of the segmented component projected onto the geometric surface of its corresponding BIM component, divided by the total surface area of ​​the BIM component. Data Validity S den The calculation formula is: the average point density of the component point cloud divided by the preset target density threshold ρ. target The upper limit of the result is 1.0. Feature completeness S feat The calculation method is as follows: First, on the geometric model of the BIM component, the 3D Harris corner detection algorithm is used to extract all its geometric corners and sharp edge segments as a set of key feature points; then, for each key feature point, it is checked whether there are at least 10 points from the corresponding component point cloud subset within a spherical neighborhood with a radius of 5cm in 3D space. Feature completeness S feat This represents the percentage of successfully validated key feature points out of the total number of key feature points. Ultimately, a comprehensive quality score Q is calculated. scoreQ is calculated using a weighted average. score = 0.5 * S cov + 0.3 * S den + 0.2 * S feat .

[0015] The system will calculate Q score Compare with a preset acceptable threshold. If Q score If the data is greater than or equal to this threshold, the data quality is considered acceptable, and the process continues. If Q... score If the data quality falls below this threshold, it is deemed unqualified. In this case, the system will generate a command and send a push notification via the mobile application to the on-site operators requesting a retest. This request will clearly specify the GUID of the component with unqualified data quality and the reason for the unqualified result (e.g., "coverage completeness is only 70%" or "Northeast corner feature point is missing"), thus providing precise guidance for the retest.

[0016] After the data quality is deemed acceptable, the system initiates overall spatial position deviation detection and component-level geometric dimension deviation detection in parallel.

[0017] For overall spatial position deviation detection, the system treats all successfully segmented and labeled component point clouds as a rigid whole and performs global optimal registration with the corresponding component geometry set in the BIM model. This registration employs the more robust Levenberg-Marquardt Iterative Closest Point Algorithm (LM-ICP) to obtain an optimal 4×4 homogeneous transformation matrix T that transforms the entire point cloud from its current measured position to its BIM design position. deviation The matrix T deviation This represents the overall spatial deviation of the entire room relative to the design blueprint in its completed state. The system decomposes the translation vectors (ΔX, ΔY, ΔZ) and rotation Euler angles (ΔRx, ΔRy, ΔRz) from this matrix as a quantitative result of the room's overall position and orientation deviation.

[0018] For component-level geometric dimensional deviation detection, the system executes specific analysis algorithms for different types of components. Taking door and window openings as an example, for their individual point cloud subsets, the system first determines their principal plane through principal component analysis (PCA) and projects the point cloud onto this plane. On the two-dimensional projection, the Hough Transform algorithm is used to detect the four straight lines constituting the boundary of the opening. By calculating the intersection of these four straight lines, the coordinates of the four corner points of the opening are accurately obtained. Then, the measured width and height of the opening are calculated. The system then queries the Width and Height parameters of the IfcOpeningElement object of the opening from the BIM database as design values, and the difference between the two values ​​gives the dimensional deviation.

[0019] Taking wall components as an example, the system performs two types of geometric tolerance analyses on its individual point cloud subsets. The first is a flatness check: the system uses the least squares method to perform optimal plane fitting on the wall's point cloud, obtaining a reference plane equation Ax + By + Cz + D = 0. Then, it calculates the orthogonal distance from each point in this point cloud subset to this reference plane; the maximum absolute value of all distances is the flatness deviation of the wall surface. The second is a verticality check: the system extracts the unit normal vector n = (A, B, C) of the aforementioned fitted reference plane and calculates the angle θ = arccos(n·v) between this normal vector and the Z-axis unit vector v = (0, 0, 1) representing the absolute vertical direction. This angle θ is the verticality deviation of the wall.

[0020] Finally, the system performs logical composite judgment and delivers multi-dimensional data. After calculating the deviation values ​​for all inspection items, the system launches a configurable standard rule engine. This engine pre-loads a structured database that stores national building construction quality acceptance standards or project-defined tolerance limits. Each record includes the component type, inspection item name, and allowable upper and lower tolerance limits. For example, a record might be (Component type: interior wall, Inspection item: flatness, Tolerance limit: 4mm). The system automatically compares each calculated deviation value with the corresponding entry in the rule engine.

[0021] If the deviation values ​​of all test items are within their respective tolerance limits, the result is deemed "qualified". If any one or more items exceed the tolerance, the result is deemed "unqualified". For unqualified items, the system marks and records them. The process finally summarizes all qualified and unqualified results.

[0022] The system automatically generates a structured inspection report and delivers it via the mobile application and web portal. The report presents the results in a multi-dimensional, visual manner. In the mobile application on the handheld terminal, the system directly overlays the inspection results onto the BIM 3D model, using color coding (e.g., green for acceptable, red for out-of-tolerance) and numerical labels to visually display the deviation status of each component. Simultaneously, the system generates a detailed inspection report in a standard PDF or Excel document format with a single click. This report includes: a project overview, inspection scope, a summary table of deviation data for all inspection items, a detailed list of out-of-tolerance items and their specific out-of-tolerance values, and, for key non-compliant components (such as walls with out-of-tolerance flatness), a deviation visualization analysis diagram. This flatness deviation visualization analysis diagram is a heatmap where each point in the wall point cloud is assigned a color value based on its distance to the fitted reference plane. A continuous color spectrum from blue (concave) to green (flat) to red (convex) clearly shows the distribution and severity of uneven areas on the wall surface. This concludes the entire inspection process.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] (1) The entire inspection process has been automated and intelligent, significantly improving work efficiency. This invention integrates the traditionally separate data collection, processing and analysis processes by introducing BIM-driven automated task planning, AR path guidance, intelligent processing with end-to-cloud collaboration and rule-based automatic evaluation. This allows the entire process from on-site scanning to generating a detailed report to be completed in a very short time, achieving "instant analysis and immediate report generation", thereby shortening the inspection cycle that traditionally takes several hours or even days to the level of several minutes.

[0025] (2) The quality and reliability of data collection and analysis results are ensured. The real-time quality monitoring mechanism proposed in this invention ensures the integrity and validity of the original data from the source by providing visual feedback on the coverage and density of the point cloud during the scanning process, thus avoiding invalid rework due to data quality issues. In addition, the BIM-driven semantic segmentation method utilizes the BIM model as strong prior knowledge, which greatly improves the accuracy and robustness of building component identification in the point cloud, laying a solid data foundation for subsequent accurate deviation calculation.

[0026] (3) It deepens the application value of Building Information Modeling (BIM) in engineering quality inspection. This invention no longer treats the BIM model as a simple geometric reference background, but elevates it to an intelligent engine that drives and guides the entire inspection process. From the automatic decomposition of tasks and the planning of scanning paths to the semantic segmentation of point cloud data, and then to the automatic extraction and comparison of design values, the inherent logic and rich information of BIM are integrated throughout, realizing the deep integration of data collection, processing and design intent.

[0027] (4) An efficient quality management closed loop has been established. The real-time analysis results and multi-dimensional visualization reports provided by this invention enable construction managers to accurately identify and locate quality problems at the first time and on the first site. Real-time non-conformity alarms and precise supplementary measurement guidance, combined with detailed deviation analysis reports, provide timely and reliable basis for subsequent corrective decisions, forming a rapid and closed-loop quality control process of "measurement-analysis-feedback-correction", effectively avoiding increased costs and project delays caused by delayed discovery of quality problems.

[0028] (5) It lowers the barrier to entry for using advanced detection technologies. Due to the high degree of automation of the entire process, the professional skills required of operators are greatly reduced. Operators only need to follow AR guidance to complete simple scanning actions, and all complex data processing and analysis work is automatically completed by the background system, enabling high-precision three-dimensional laser scanning detection technology to be widely used in a wider range of engineering scenarios.

[0029] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, embodiments of the present invention are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a schematic diagram of the system architecture described in this invention;

[0032] Figure 2 This is an overall flowchart of the method described in this invention;

[0033] Figure 3 This is a flowchart of the task loading and augmented reality scanning path planning described in this invention;

[0034] Figure 4This is a flowchart of the on-site data acquisition and real-time quality monitoring described in this invention;

[0035] Figure 5 This is a flowchart of the point cloud preprocessing for edge-cloud collaboration described in this invention;

[0036] Figure 6 This is a flowchart of the automated judgment of multi-dimensional data quality compliance as described in this invention;

[0037] Figure 7 This is a flowchart of the deviation detection and result delivery described in this invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0039] Reference Figure 1 This demonstrates the system architecture upon which the invention is based. The system consists of three core components: a handheld scanning terminal, a companion mobile application running on the terminal, and a cloud server. The handheld scanning terminal, such as a GeoScanner Pro S device, integrates a series of sensors and processing units. Specifically, its core sensing component is a three-dimensional LiDAR sensor with a center wavelength set at 905nm, safe for the human eye, belonging to Class 1 laser safety level. Its effective ranging range is up to 60 meters, with a ranging accuracy better than ±10mm at a distance of 10 meters. It boasts a point cloud acquisition rate of up to 320,000 points per second and a 360-degree horizontal field of view and a 270-degree vertical field of view. To achieve accurate attitude estimation, the terminal also integrates a high-performance six-axis inertial measurement unit (IMU), specifically a Bosch BMI160 microelectromechanical system (MEMS) sensor. This unit synchronously outputs angular velocity data from a three-axis gyroscope and linear acceleration data from a three-axis accelerometer at a frequency of 200Hz. In addition, the device is equipped with a 12-megapixel visible light camera with a 120-degree wide-angle field of view, used to acquire texture information of the surrounding environment and support the realization of augmented reality functions. The device's local processing power is provided by a high-performance ARM-based system-on-a-chip (SoC), such as the Qualcomm Snapdragon 8 Gen 2 processor, and is supplemented by 12GB of RAM to ensure the smooth operation of on-device algorithms.

[0040] The accompanying mobile application is installed and runs on the handheld scanning terminal, serving as the main interface for human-computer interaction and the scheduling center for edge-side processing tasks. This application communicates stably and reliably with the cloud server via a Secure Sockets Layer (SSL) encrypted channel.

[0041] The cloud server is deployed in a high-performance data center, and its hardware configuration includes a cluster of graphics processing units (GPUs) for massively parallel computing and a distributed database system, such as a PostgreSQL-based database, for persistent storage of massive amounts of data, extended with a PostGIS plugin for efficient processing of geospatial data. This server hosts the computationally intensive data processing and analysis tasks described in this invention.

[0042] Combination Figures 2-7 As shown below, the specific implementation steps of the automatic detection method for room position and size deviation based on handheld point cloud acquisition provided by the present invention will be described in detail.

[0043] like Figure 3 As shown, the first step of the process is task loading and augmented reality scanning path planning. The operator first launches the accompanying mobile application on the handheld scanning terminal. The application initiates a connection request to the cloud server through the aforementioned SSL encrypted channel and performs user authentication. After successful authentication, the operator selects and loads the Building Information Model (BIM) of the target project from the server's project list. The BIM model processed in this invention adopts the Industry Foundation Class (IFC) standard format, specifically IFC4. After the model is loaded, the system automatically launches an embedded BIM parsing engine, which can be built based on an open-source IFC processing library (such as IfcOpenShell). This engine is responsible for traversing the hierarchical structure of the entire IFC data file, accurately identifying and extracting information from key entities such as IfcBuildingStorey (floors), IfcGrid (grid lines), IfcSpace (spaces), and custom property sets (PropertySets) that record construction flow segment division information.

[0044] Based on this parsed structured information, the system automatically divides the inspection task units. This process logically and automatically decomposes a large building project model into a series of independent and manageable inspection task units. Each inspection task unit is created as an independent record in the database and assigned a globally unique identifier (UUID). The data structure of this record is designed to include the following fields: the UUID of the task unit itself, its floor and area identifier (e.g., "Building A - 4th Floor - East Zone"), the globally unique identifier (GUID) of the core IfcSpace object it is associated with, the specific coordinate range of the three-dimensional geometric bounding box of the space in the project coordinate system, and a list containing the GUIDs of all BIM components within the space (such as walls IfcWall, floors IfcSlab, columns IfcColumn, doors IfcDoor, windows IfcWindow, etc.). This division method decomposes the complex inspection work into atomic tasks based on individual rooms or spaces, greatly simplifying the subsequent data management and processing flow.

[0045] Once a detection task unit is selected, for example, when an operator selects the task unit representing "Room 401" on the application interface, the system automatically plans the optimal scanning path for that unit. The core objective of the path planning algorithm is to generate a walking trajectory with the shortest distance for the operator, while ensuring that the average density of the point cloud on the surfaces of all key components is not lower than a preset threshold (e.g., 5000 points / square meter). Specifically, the planning algorithm first voxels the three-dimensional geometric space of the selected task unit, with the voxel size resolution set to 5 centimeters. Subsequently, based on the BIM model, the system marks the voxels occupied by all solid components (walls, floors, columns, etc.) in the space as impassable obstacle voxels. On this basis, the system uses an improved Fast Expanding Random Tree Star (RRT*) algorithm to search for paths within the free space composed of all non-obstacle voxels. The improvement here is reflected in the design of its cost function. Traditional RRT* algorithms mainly optimize path length, while the cost function in this invention... c(x) Defined as a weighted sum:

[0046] c(x) = αL(x) + β(1-V s (x))+γ(1-C r (x))

[0047] in, L(x) It is the path length. V s (x) It is the visibility score of each point on the path to the key surfaces of the BIM component. Cr (x) This refers to the coverage of the BIM surface by the point cloud generated based on the expected path points. Weighting factor. α,β,γ The introduction of values ​​(e.g., 0.4, 0.3, 0.3) allows the algorithm to actively gravitate towards locations where all building surfaces can be better "seen" and scanned while seeking the shortest path. The algorithm's final output is a spatial 3D path composed of a series of sequentially arranged six-degree-of-freedom (6-DoF) pose points. Upon receiving this path data, the mobile application projects it in real-time onto the handheld terminal's camera preview via a graphics rendering engine, manifesting as a continuous, prominent 3D arrow model. This forms an intuitive augmented reality (AR) visual guide, instructing the operator to move and scan along this optimized path.

[0048] like Figure 4 As shown, the second step of the process is on-site data acquisition and real-time quality monitoring. The operator holds a scanning terminal and moves smoothly within the room to be tested, guided by the AR arrows on the screen. During this process, the terminal's built-in Simultaneous Localization and Mapping (SLAM) system begins operation. This invention employs a tightly coupled LiDAR-Inertial Odometry (LIO-SAM) SLAM algorithm. This algorithm tightly fuses the raw point cloud data generated by the 3D LiDAR sensor at a rate of 320,000 points per second with the angular velocity and acceleration data output by the inertial measurement unit at a frequency of 200Hz within a factor graph optimization framework. This method can calculate the six-degree-of-freedom pose (3D position x, y, z and 3D attitude) of the scanning terminal in the world coordinate system (usually with the scanning start point as the origin) in real time and with high accuracy. roll, pitch, yaw ), and simultaneously construct the original point cloud map of the environment.

[0049] Meanwhile, to ensure data quality from the source, the mobile application performs real-time registration and quality analysis of point cloud data and the BIM model in parallel on a separate computing thread. This process is designed to be extremely lightweight to ensure that it does not affect the real-time performance of the SLAM system. Specifically, the system uses the device pose just output by the SLAM system as a high-precision initial transformation matrix to perform voxel raster downsampling on the newly generated portion of the point cloud data (approximately 320,000 points) every second, with the downsampling resolution set to 5 centimeters to reduce computational load. Subsequently, by executing one iteration of the Generalized Iterative Closest Point (G-ICP) algorithm, this batch of downsampled point clouds is quickly and lightweightly registered with the pre-loaded and similarly voxelized BIM model.

[0050] After registration, the system immediately performs real-time evaluation of the coverage and density of the acquired point cloud on the surface of the BIM component. Its implementation is innovative: the system first parametrically unfolds the triangular mesh model of each BIM component in the background, mapping it to a two-dimensional UV texture coordinate space. Then, the system maintains a two-dimensional rasterized cumulant matrix corresponding to this UV space. For each newly acquired and registered point cloud point on the component surface, the system calculates its nearest projection point on the component surface and maps the UV coordinates corresponding to that projection point to the corresponding raster in the cumulant matrix, incrementing the raster's count value by one. Based on the real-time count value in this cumulant matrix, the mobile application dynamically renders the BIM model surface in three different colors and displays them on the screen: when the raster count value corresponding to a certain area exceeds a preset density threshold ρ... target (For example, an equivalent value of 5000 points per square meter) When the area is rendered green on the screen, it indicates that sufficient data collection has been achieved; when the count value is between 0 and ρ target When the count is in between, it is rendered in yellow, indicating that the data is not yet sufficient; when the count is 0, it is rendered in gray, indicating that the area has not yet been scanned. Operators can use this intuitive visual feedback to selectively scan the yellow areas until most of the BIM model on the screen is green. This mechanism ensures the integrity and uniformity of the original data acquisition, fundamentally avoiding rework caused by data quality defects.

[0051] like Figure 5As shown, the third step in the process is edge-cloud collaborative point cloud preprocessing. Once the scanning of a detection task unit (e.g., "Room 401") is confirmed as complete according to the aforementioned quality monitoring mechanism, the operator clicks the "Complete and Submit" button on the mobile application. At this point, the handheld terminal first performs a feature-sensitive adaptive downsampling process on the collected complete raw point cloud data (which may contain tens of millions of points). The goal of this step is to reduce the data volume to the greatest extent possible without losing key geometric features (such as corners, opening edges, structural edges, etc.) to facilitate rapid uploading. Specifically, this downsampling algorithm is a curvature-based adaptive voxel raster downsampling method. First, the complete raw point cloud data is constructed into an octree data structure for spatial indexing. Second, for each leaf node in the octree containing a subset of point clouds, the covariance matrix of its three-dimensional coordinates is calculated. Third, the covariance matrix is ​​decomposed into eigenvalues, yielding three eigenvalues ​​λ1≥λ2≥λ3. Then, a normalized curvature metric C = λ3 / (λ1+λ2+λ3) is calculated based on these three eigenvalues. Theoretically, when the point cloud is linearly distributed (such as an edge line) or planar within a leaf node, λ3 will be very small, and the C value will approach 0; while when the point cloud is distributed as three-dimensional corner points or randomly scattered points, the three eigenvalues ​​are closer together, and the C value is higher. Finally, the system determines the retention strategy of the point cloud within the leaf node based on this curvature metric C, specifically a probability function positively correlated with the C value, for example, P(retain) = k * C + P base Where k is the scaling factor, P base This strategy ensures that point clouds in high-curvature regions are preserved with a high probability, while point clouds on flat surfaces are significantly downsampled, represented by their centroids. After processing using this method, the amount of point cloud data can typically be compressed to 15% to 20% of its original size, while retaining most of the geometric details.

[0052] The downsampled feature-enhanced point cloud data, along with its corresponding detection task unit UUID, is packaged and securely uploaded to the cloud server via a RESTful API interface based on HTTPS. Upon receiving the data packet, the cloud server first sends it to a deep noise reduction module. This module employs a pre-trained point cloud noise classifier model based on a Graph Convolutional Network (GCN), such as a model based on the DGCNN architecture. Trained on a large dataset of synthetic point clouds containing simulated noise, this model accurately learns and identifies patterns of complex discrete noise points caused by surface material reflection characteristics, airborne particles, and sensor electrical noise, effectively removing them. Its noise reduction performance significantly outperforms traditional statistical outlier removal (SOR) or radius-based filtering algorithms, especially showing greater robustness when processing point clouds with non-uniform density.

[0053] After noise reduction, the server executes one of the most crucial steps in the entire process: BIM-driven point cloud unit identification and semantic segmentation. This step aims to precisely segment a large, unordered point cloud dataset into independent point cloud subsets, each corresponding to a BIM component and bearing a clear semantic label. Based on the received detection task unit UUID, the server queries its associated BIM database to retrieve a list of GUIDs and their precise geometric definitions (such as boundary polygons, extrusion vectors, spatial locations, etc.) for all BIM components within that task unit. For each BIM component in the list, the server performs the following series of automated operations: First, it extracts the precise digital geometric surface model of the wall from the BIM definition. Then, within the newly received and noise-reduced point cloud data, it defines a virtual 3D search space centered on the BIM wall's geometric model and extending uniformly outwards by 30 centimeters. Within this defined search space, the server executes a Directed Random Sample Consensus (RANSAC) plane fitting algorithm. Because the sampling process is strictly constrained to a small area containing the target component, interference from distant, irrelevant point clouds is greatly eliminated, ensuring extremely high efficiency and robustness of the plane fitting process.

[0054] All point cloud points (i.e., interior points) within the best-fit plane are identified as a subset of the point cloud belonging to that wall. The server separates this subset from the main point cloud and assigns it a semantic label, which is directly derived from the GUID of the BIM component. This process is performed sequentially on all BIM components (walls, slabs, columns, doorways, etc.) within the task unit until the point cloud data of the entire room is completely and accurately segmented into a series of individual component point clouds, each corresponding to a BIM component and possessing a unique semantic identity.

[0055] like Figure 6As shown, the fourth step in the process is the automated assessment of multi-dimensional data quality compliance. After completing the individual component segmentation, to ensure the accuracy and reliability of subsequent deviation analysis, the system performs a quantitative quality assessment on a subset of the point cloud for each component. This assessment system includes three core indicators: coverage completeness S... cov Data Validity den and feature completeness S feat Coverage completeness S cov The calculation method is as follows: project the point cloud of the segmented component onto the geometric surface of its corresponding BIM component, calculate the covered area, and then divide it by the total surface area of ​​the BIM component itself to obtain a percentage. Data Validity S den The calculation method is as follows: calculate the average point density of the component's point cloud and divide it by the preset target density threshold ρ. target The results were normalized and capped at 1.0. Feature completeness S feat The calculation method is more refined: First, the system uses the 3D Harris corner detection algorithm to extract all geometric corners and sharp edge segments on the precise geometric model of the BIM component, and uses the set of sampled points of these points and segments as the key feature point set. Then, for each key feature point, the system checks whether there are at least 10 points from the corresponding component's point cloud subset within a spherical neighborhood with a radius of 5 cm in its 3D space. Feature completeness S feat Ultimately, it is defined as the percentage of key feature points that are successfully validated (i.e., the number of points in the neighborhood meets the standard) out of the total number of key feature points.

[0056] Finally, a comprehensive quality score Q score The formula for calculating the weighted average of the above three indicators is as follows:

[0057] Q score = 0.5 * S cov + 0.3 * S den + 0.2 * S feat

[0058] The weighting coefficients reflect the degree of importance attached to different quality dimensions, with complete coverage considered paramount. The system will calculate Q... score Compare with a preset passing threshold (e.g., 85 points). If Q score If the data quality of the component is greater than or equal to the threshold, the data quality is considered acceptable, and the process continues. Otherwise, if Q... scoreIf the data falls below this threshold, it is deemed unqualified. In this case, the system automatically generates a command and sends a retest request to the scanning terminal of the on-site operator via push notification service from the mobile application. This request clearly specifies the GUID of the component with unqualified data quality and the specific reason for the failure (e.g., "coverage completeness is only 70%, data missing near the east corner" or "feature point missing at the upper edge of the doorway"), thus providing precise and closed-loop guidance for the retesting work.

[0059] like Figure 7 As shown, after the data quality of all relevant components is determined to be qualified, the system will start two core analysis tasks in parallel: overall spatial position deviation detection and component-level geometric dimension deviation detection.

[0060] For overall spatial position deviation detection, the system treats all successfully segmented and labeled component point clouds as a rigid whole and performs a global optimal registration with the corresponding component geometry set in the BIM model. This registration employs the more robust and convergent Levenberg-Marquardt Iterative Closest Point Algorithm (LM-ICP), and sets strict convergence criteria (e.g., the change in the transformation matrix between two iterations is less than 1e-6 or the maximum number of iterations is reached, 100). The algorithm ultimately obtains an optimal 4×4 homogeneous transformation matrix T that accurately transforms the entire point cloud from its current measured position to its BIM design position. deviation This matrix itself contains information about the overall spatial deviation of the room relative to the design blueprint in its completed state. The system further decomposes the translation vectors (ΔX, ΔY, ΔZ) and rotation Euler angles (ΔRx, ΔRy, ΔRz) from this matrix, and these values ​​are used as the final quantitative results of the room's overall position and orientation deviation.

[0061] For component-level geometric dimensional deviation detection, the system invokes specific analysis algorithms for different types of components. Taking door and window openings as an example, after extracting a subset of their individual point clouds, the system first determines their principal plane using Principal Component Analysis (PCA) and projects all point cloud points onto this plane, thus reducing the dimensionality of the three-dimensional problem to two dimensions. On the two-dimensional projected point set, the system employs the Hough Transform algorithm, which, due to its insensitivity to noise and missing data, robustly detects the four straight lines constituting the boundary of the opening. By calculating the intersection of these four straight lines, the system can accurately obtain the coordinates of the four corner points of the opening. Furthermore, by calculating the distance between the corner points, the measured width and height of the opening can be obtained. The system then queries the BIM database for the Width and Height parameters of the corresponding IfcOpeningElement object for the opening as design values, and subtracts the design values ​​from the measured values ​​to obtain the dimensional deviation of the opening.

[0062] Taking wall components as an example, the system performs two key geometric tolerance analyses on their individual point cloud subsets. The first is flatness checking: the system uses the least squares method to perform optimal plane fitting on the wall's point cloud, obtaining a reference plane equation Ax + By + Cz + D = 0 representing the overall trend of the wall surface. Then, the system calculates the orthogonal distance from each point in this point cloud subset to this reference plane, compares the absolute values ​​of all distances, and takes the maximum value, which is defined as the flatness deviation of the wall surface. The second is verticality checking: the system extracts the unit normal vector n = (A, B, C) of the aforementioned fitted reference plane and calculates the angle θ = arccos(n·v) between this normal vector and the Z-axis unit vector v = (0,0,1) representing the absolute vertical direction. This angle θ is the verticality deviation of the wall, which can be directly converted into deviations in millimeters per meter of height.

[0063] The final step in the process is the logical composite judgment of the results and the delivery of multi-dimensional data. After the deviation values ​​of all inspection items have been calculated, the system will launch a flexibly configurable standard rule engine. This engine is pre-loaded with a structured database that stores relevant national building construction quality acceptance standards (such as the "Unified Standard for Acceptance of Construction Quality of Building Engineering" GB50300) or tolerance limits defined by the project owner. Each record in the database contains the component type, the name of the inspection item, and the upper and lower limits of the allowable tolerance. For example, a typical record is (Component type: interior plastered wall, Inspection item: surface flatness, Tolerance limit: 4mm). The system will automatically compare each calculated deviation value with the corresponding entry in the rule engine.

[0064] If the deviation values ​​of all test items fall within their respective tolerance limits, the final result of the test task unit is judged as "qualified". If any one or more deviations exceed the limits, it is judged as "unqualified". For all unqualified items, the system will highlight them and record them in detail. The process finally summarizes all qualified and unqualified judgments.

[0065] Based on these summarized results, the system automatically generates a structured inspection report, which is delivered via a mobile application and a corresponding web portal. The report presents the inspection results in a multi-dimensional and highly visualized manner. In the mobile application on handheld devices, the system directly overlays the inspection results onto the BIM 3D model, using intuitive color coding (e.g., green for qualified components and red for components exceeding tolerances) and numerical labels attached to the components to visually demonstrate the deviation status of each component. Simultaneously, the system supports one-click generation of detailed inspection reports in standard PDF or Excel document formats. The report is rigorous and comprehensive, including a project overview, inspection scope (corresponding task unit UUID), a summary table of deviation data for all inspection items, a detailed list of exceeding tolerance items and their specific deviation values, and can include a deviation visualization analysis chart for critical non-compliant components (such as walls with severely excessive flatness). For example, in a visualization analysis chart of wall flatness deviation, each point in the wall point cloud is assigned a specific color value based on its orthogonal distance to the fitted reference plane. For instance, a continuous color spectrum from blue (representing concave areas) to green (representing flat areas) to red (representing convex areas) clearly and quantitatively displays the precise distribution and severity of uneven areas on the wall. This concludes a complete and automated process for detecting room position and dimensional deviations.

[0066] To further illustrate the technical effects of the present invention, a specific embodiment and comparative example are given below.

[0067] Example 1

[0068] This embodiment aims to inspect the interior space of room number "R-401" on the 4th floor of Building A in a commercial complex. The room has a design size of 5.0m × 4.0m and a floor height of 3.0m. It includes four interior walls, one floor slab, one ceiling, one doorway measuring 2.1m × 0.9m, and one window opening measuring 1.5m × 1.2m.

[0069] The operator uses the method of the present invention. After loading the project BIM model on the mobile application, the system automatically identifies "R-401" as an independent inspection task unit, and the UUID is {a1b2c3d4-e5f6-7890-1234-567890abcdef}. The automatically generated AR guidance path is 25.3 meters in total length and contains 12 key pose points. The operator follows the AR guidance and completes the data collection in 2 minutes and 15 seconds. During this period, the real-time quality monitoring interface shows that the coverage rates of all walls, floors, and ceilings reach the green standard. A total of 15.24 million original point clouds are generated by the scan. After clicking to submit, it takes 28 seconds locally on the terminal to complete the curvature adaptive downsampling, generating a feature-enhanced point cloud of 2.81 million points (data compression rate is 18.4%), and uploading it to the cloud server within 15 seconds through the 5G network.

[0070] After receiving the data, the cloud server takes 35 seconds to complete GCN noise reduction, BIM-driven semantic segmentation, and automated data quality assessment. The Q of all component point clouds score is higher than 85 points, and it is determined to be qualified. Subsequently, the system starts the deviation analysis in parallel and completes all calculations in 42 seconds. The results show that the overall room is translated 8 mm in the positive X-axis direction and rotated 0.05 degrees around the Z-axis. The flatness of the north wall is 3.2 mm, and the perpendicularity is 0.08 degrees. The measured width of the door opening is 902 mm (deviation +2 mm), and the height is 2095 mm (deviation -5 mm). All deviation values are within the specification limits preset for the project.

[0071] Finally, the system takes 5 seconds to generate a report. In total, from the start of scanning to obtaining the detailed PDF report, the whole process takes (135s + 28s + 15s + 35s + 42s + 5s) = 260 seconds, approximately 4.3 minutes. The report is pushed to the project manager's mobile phone in real time through the App and is updated synchronously on the Web side.

[0072] Comparative Example 1

[0073] This comparative example uses the method of traditional hand-held scanning combined with post-processing software to detect the same "R-401" room as in Example 1.

[0074] The operator uses the same hand-held scanner, but it does not have the functions of AR guidance and real-time quality monitoring. The operator walks and scans in the room based on experience, which takes about 3 minutes. After the scan is completed, the original point cloud data containing more than 15 million points is exported to the workstation computer through a data cable, which takes about 5 minutes.

[0075] Data processing engineers open the data on their workstations using third-party point cloud processing software (such as CloudCompare). First, they spend approximately 15 minutes manually denoising and roughly cropping the data. Then, by manually selecting areas, they segment the point clouds of walls, floors, ceilings, doorways, and window openings one by one. This process depends on the engineer's skill level and takes about 45 minutes, and there are subjective errors in the segmentation boundaries. After segmentation, they manually register the point clouds of each component with the corresponding STL models exported from the BIM software, performing ICP registration for each component, which takes about 20 minutes.

[0076] After registration, the software's built-in tools are used for deviation analysis. Wall flatness analysis takes approximately 5 minutes, and door and window opening size measurement takes approximately 10 minutes. Finally, screenshots of all results are taken, data is copied, and a test report is manually written, taking approximately 30 minutes.

[0077] During the data analysis phase, it was discovered that the point cloud was sparse in the northwest corner of the room, which was not noticed during the scanning process, resulting in low reliability of the flatness and verticality analysis results for that wall. To ensure quality, the operators need to return to the site to conduct supplementary measurements in that area, incurring an additional 30 minutes of communication, transportation, and work time.

[0078] Without considering rework, the total time from the start of scanning to obtaining the report was (3min+5min+15min+45min+20min+15min+30min) = 133 minutes, approximately 2.2 hours.

[0079] The key performance indicators of Example 1 and Comparative Example 1 are quantitatively compared below with Table 1:

[0080] Table 1 Comparison of Key Performance Indicators

[0081]

[0082] Through the detailed description and data comparison of the above embodiments and comparative examples, it can be clearly seen that the automatic detection method for room location and size deviation based on handheld point cloud acquisition described in this invention, by constructing a fully automated process driven by BIM and co-located with the cloud, demonstrates significant advantages over existing technologies in terms of detection efficiency, data quality, result reliability, and ease of operation.

[0083] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An automatic detection method for room position and size deviation based on handheld point cloud acquisition, relying on a system consisting of a handheld scanning terminal integrating a three-dimensional lidar sensor, an inertial measurement unit, and a visible light camera, a mobile application running on the scanning terminal, and a cloud server, characterized in that, The method includes: Task loading and augmented reality scanning path planning: The building information model of the target project is loaded through a mobile application. The building information model is automatically parsed and decomposed into detection task units with unique identifiers. A spatial three-dimensional path is automatically planned for the selected detection task unit. The spatial three-dimensional path is visualized on the display interface of the scanning terminal through augmented reality. On-site data acquisition and real-time quality monitoring: Operators use handheld scanning terminals and follow augmented reality visual guidance to move around. By using point cloud data from tightly coupled 3D LiDAR sensors and readings from inertial measurement units, the device pose is calculated in real time and raw point cloud data is generated synchronously. At the same time, the acquired point cloud data is registered with the building information model in real time in a lightweight manner, and the coverage quality of the point cloud on the surface of the building information model components is evaluated and visualized in real time. The edge-cloud collaborative point cloud preprocessing performs feature-sensitive adaptive downsampling processing on the raw point cloud data collected locally on the scanning terminal, uploads the downsampled point cloud data to the cloud server, performs noise reduction processing on the received point cloud data on the cloud server, and uses the building information model as prior knowledge to automatically segment the point cloud data into individual component point clouds with semantic tags that correspond one-to-one with the components in the building information model. The system automatically assesses the compliance of multi-dimensional data quality. For each individual component point cloud, it calculates a comprehensive quality score from three dimensions: coverage completeness, data validity, and feature completeness, and compares it with a preset pass threshold. If the score is not up to standard, a retest instruction is generated. Deviation detection and result delivery: For point clouds of individual components that meet quality standards, overall spatial position deviation detection and component-level geometric dimension deviation detection are performed. The detected deviation values ​​are logically combined with preset specification rules for judgment, and a structured inspection report containing a visual analysis diagram is generated.

2. The method according to claim 1, characterized in that, In the task loading and augmented reality scanning path planning, the process of automatically parsing and decomposing the building information model into detection task units specifically includes: Traverse the industrial basic data structure of the building information model to identify and extract entity information of floors, grid lines, spaces, and custom attribute sets related to the division of construction flow sections; Based on the entity information, the project model is automatically decomposed into a series of independent detection task units, and a data structure containing the following information is created for each detection task unit: a unique identifier of the task unit, the floor and area identifier to which it belongs, a globally unique identifier of the spatial object associated with it, the three-dimensional geometric bounding box of the space, and a list containing globally unique identifiers of all building information model components within the space.

3. The method according to claim 1 or 2, characterized in that, In the task loading and augmented reality scanning path planning, the process of automatically planning the spatial three-dimensional path specifically includes: The three-dimensional geometric space of the detection task unit is voxelized, and the voxels occupied by the component entities are marked as obstacles according to the building information model. In non-obstacle areas, an improved fast expanding random tree star algorithm is used for path search, wherein the cost function of the algorithm is defined as a weighted sum that comprehensively considers path length, visibility of path points to the surface of building information model components, and expected scan coverage. The final output is a spatial three-dimensional path consisting of a series of sequentially arranged six-degree-of-freedom pose points.

4. The method according to claim 1, characterized in that, In the aforementioned on-site data acquisition and real-time quality monitoring, the process of real-time evaluation and visualization of point cloud coverage quality specifically includes: In a separate computing thread, the device pose output by the real-time localization and mapping system is used as the initial transformation matrix. The newly acquired point cloud data is then lightweightly registered with the preloaded building information model through a one-iteration generalized iterative nearest point algorithm. The triangular mesh model of each building information model component is parametrically unfolded into a two-dimensional UV coordinate space, and a two-dimensional rasterized cumulant matrix corresponding to this UV space is maintained. For each newly acquired and registered point cloud point, calculate its nearest projection point on the surface of the building information model component, and record the UV coordinates corresponding to the projection point in the cumulative matrix, thereby increasing the count value of the corresponding grid. Based on the count values ​​in the cumulative matrix, the surface of the building information model is rendered in real time with different colors for visualization: when the grid count value corresponding to a certain area is greater than a preset density threshold, it is rendered with the first color; when the count value is between 0 and the preset density threshold, it is rendered with the second color; when the count value is 0, it is rendered with the third color.

5. The method according to claim 1, characterized in that, In the point cloud preprocessing process of the aforementioned end-to-cloud collaboration, the feature-sensitive adaptive downsampling processing performed locally on the scanning terminal specifically includes: An octree index is constructed from the complete raw point cloud data collected. For each leaf node in the octree, calculate the covariance matrix of its three-dimensional coordinates for the subset of point clouds contained in the tree. The covariance matrix is ​​decomposed into eigenvalues ​​to obtain three eigenvalues ​​λ1, λ2, and λ3, where λ1 ≥ λ2 ≥ λ3. Calculate a normalized curvature metric C = λ3 / (λ1 + λ2 + λ3) based on the eigenvalues; The point cloud preservation strategy within the leaf node is determined based on the curvature metric C. This strategy ensures that the higher the curvature metric C, the higher the probability that the point cloud within the leaf node is completely preserved, while the lower the curvature metric C, the higher the probability that the point cloud within the leaf node is downsampled using its centroid as a representative.

6. The method according to claim 1 or 5, characterized in that, In the point cloud preprocessing process of the aforementioned end-to-cloud collaboration, the process of automatically segmenting the point cloud data into individual component point clouds using the building information model as prior knowledge on the cloud server specifically includes: Based on the received unique identifier of the detection task unit, retrieve the list of globally unique identifiers of all building information model components within the unit, along with their precise geometric definitions and spatial location information, from the database. For each building information model component in the list, perform the following operations: Extract its digital geometric surface from its building information model definition, and expand outward by a preset distance from this geometric surface to define a three-dimensional search space; Within the three-dimensional search space, a directional random sampling consensus algorithm is executed to fit a subset of point clouds that matches the geometric features of the component; The fitted subset of point cloud is separated from the main point cloud and assigned a globally unique identifier derived from the building information model component as a semantic label.

7. The method according to claim 1, characterized in that, The automated process for determining multi-dimensional data quality compliance specifically includes: For each individual component point cloud, calculate its coverage completeness, data validity, and feature completeness. The coverage integrity is calculated by dividing the area covered by the point cloud of the segmented component projected onto the geometric surface of its corresponding building information model component by the total surface area of ​​the building information model component. The data validity is calculated by dividing the average point density of the component point cloud by a preset target density threshold; the feature completeness is calculated by extracting key feature point sets from the geometric model of the building information model component and checking whether there are a sufficient number of points from the corresponding component point cloud subset in the neighborhood of each key feature point in three-dimensional space. A comprehensive quality score is calculated by weighting the coverage completeness, data validity, and feature completeness. The overall quality score is compared with a preset pass threshold. If it is lower than the threshold, it is determined to be unqualified, and a retest instruction containing the identifier of the unqualified component and the specific reason for the unqualification is generated and pushed to the operator through the mobile application.

8. The method according to claim 1, characterized in that, In the deviation detection and result delivery process, the overall spatial position deviation detection process specifically includes: All successfully segmented and labeled individual component point clouds are treated as a rigid whole and globally optimally registered with the corresponding component geometry set in the building information model. The global optimal registration adopts the Levenberg-Marquardt iterative nearest point algorithm to obtain an optimal 4×4 homogeneous transformation matrix that transforms the entire point cloud from the measured location to the design location of the building information model; The translation vector and rotation Euler angles are decomposed from the homogeneous transformation matrix to quantify the overall position and orientation deviation of the room.

9. The method according to claim 1, characterized in that, In the deviation detection and result delivery process, the procedure for performing component-level geometric dimension deviation detection includes checking the flatness and verticality of the wall components, wherein: The flatness inspection specifically involves: using the least squares method to perform optimal plane fitting on the individual point cloud subset of the wall component to obtain a reference plane equation, and calculating the orthogonal distance from each point in the point cloud subset to this reference plane, and taking the maximum value among all absolute distances as the flatness deviation of the wall surface. The verticality check specifically involves: extracting the unit normal vector of the reference plane, calculating the angle between the normal vector and the Z-axis unit vector representing the absolute vertical direction, and using this angle as the verticality deviation of the wall.

10. The method according to claim 1, characterized in that, In the deviation detection and result delivery process, the process of generating a structured detection report includes: The test results are overlaid on the 3D model of the building information model, and color coding and digital labels are used to visually display the qualified or out-of-tolerance status of each component. Furthermore, for wall components with excessive flatness deviation, a flatness deviation visualization analysis map is generated. This analysis map is a heat map in which each point of the wall surface point cloud is assigned a color value according to its distance from the fitting reference plane, and the distribution and severity of uneven areas of the wall surface are clearly displayed through a continuous color spectrum.

Images