Dangerous engineering dynamic acceptance auxiliary method based on fusion of unmanned aerial vehicle oblique photography and BIM

Abstract

The invention discloses an unmanned aerial vehicle oblique photography and BIM fused dangerous and large project dynamic acceptance auxiliary method, and particularly relates to the technical field of intelligent construction informationization. Multi-view images and laser point cloud data of a dangerous and large project area are collected through an unmanned aerial vehicle carrying an RTK-PPK dual-mode positioning system, a five-lens oblique photography camera and a laser radar; synchronous acquisition of multi-source data is realized; according to the method, multi-source data is subjected to fusion processing, edge endpoint features of a BIM model and point cloud data are automatically extracted through an improved ICP-NDT algorithm for coarse registration, sub-pixel-level fine registration is realized through iterative optimization, a digital twin model is introduced, BIM design parameters and actually measured data are compared based on a fuzzy logic multi-criterion acceptance decision tree, and the precision of the BIM design parameters is improved. An acceptance report is automatically generated, key data is returned to a cloud platform through a 5G network, the problem of model missing of the occlusion area is solved through multi-source data fusion, and the integrity of the three-dimensional model is ensured; millisecond-level defect identification is realized through edge intelligent decision, and potential safety hazard discovery timeliness is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent construction information technology, and more specifically, to a dynamic acceptance auxiliary method for critical and major engineering projects that integrates UAV oblique photography and BIM. Background Technology

[0002] Currently, in the field of building engineering, the acceptance of critical and large-scale projects mainly relies on traditional manual measurement and BIM model comparison. With technological advancements, UAV oblique photogrammetry technology, due to its high efficiency and high realism, is gradually being applied to engineering surveying and 3D modeling. This technology, by mounting multiple sensors on the same flight platform, simultaneously acquires images from multiple angles, including vertical and oblique views, enabling the rapid creation of realistic 3D models reflecting ground features. Meanwhile, BIM technology, as an important tool for the digitization of building engineering, provides parametric models that can describe the properties of building components in detail. Existing technologies have attempted to combine the realistic models generated by oblique photogrammetry with the design BIM models to assist in planning and acceptance. However, after in-depth analysis, the existing technology has the following significant drawbacks: Traditional oblique photogrammetry mainly relies on optical images to generate point clouds through multi-view geometric matching. In densely built-up areas or near-ground tree-covered areas, the model's geometric accuracy is easily poor due to the limited number of image matching connection points, resulting in distortion, blurring, and holes. Existing methods mostly use manual selection of feature points or placement of physical targets for coarse registration, resulting in low automation and registration accuracy that is easily affected by human factors, making it difficult to achieve sub-centimeter level high-precision registration.

[0003] Therefore, a dynamic acceptance auxiliary method for critical and major engineering projects that integrates UAV oblique photography and BIM is proposed to address the above problems. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a dynamic acceptance auxiliary method for critical and major engineering projects that integrates UAV oblique photography and BIM, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a dynamic acceptance auxiliary method for critical and major engineering projects integrating UAV oblique photography and BIM, comprising the following steps: S1. Collect multi-source data of the critical engineering area by using a drone equipped with an RTK-PPK dual-mode positioning system, a five-lens oblique photography camera and a lidar. The multi-source data includes multi-view images and laser point cloud data, thereby achieving synchronous acquisition of ground data. S2. After fusing the multi-source data, a dynamically updated digital twin model is constructed; S3. The UAV is also equipped with an airborne edge computing unit, which is equipped with a lightweight CNN model to perform real-time defect identification on the registered multi-source data. The defect identification includes concrete cracks, steel bar spacing deviations and formwork deformation. S4. Based on a fuzzy logic-based multi-criteria acceptance decision tree, the system compares BIM design parameters with measured data, automatically generates an acceptance report, and transmits key data back to the cloud platform via a 5G network.

[0006] Preferably, in step S2, the fusion process is as follows: A1. The multi-source data is coarsely registered by automatically extracting the edge endpoint features of the BIM model and point cloud data through the improved ICP-NDT algorithm, and then sub-pixel level fine registration is achieved through iterative optimization. A2. Kalman filtering is introduced to fuse GNSS, IMU, lidar and image data to construct the digital twin model.

[0007] Preferably, in step S1, the specific method of multi-scale data acquisition includes: simultaneously acquiring images from one vertical perspective and four oblique perspectives through a five-lens oblique photography system, fusing lidar point cloud data, and supplementing missing data in image occlusion areas.

[0008] Preferably, in step S2, the specific implementation of the improved ICP-NDT algorithm includes: automatically extracting the beam-column intersection points and wall corner edge endpoint features of the BIM model and the point cloud for coarse registration, and performing iterative optimization by combining the ICP algorithm and normal distribution transformation.

[0009] Preferably, in step S3, the edge intelligent decision-making achieves millisecond-level defect identification through a lightweight CNN model, and automatically determines whether the verticality and spacing deviation of the formwork support exceed the allowable range based on a fuzzy logic decision tree, and triggers a graded warning. The lightweight CNN model employs an attention-based feature fusion module, which can process texture features from oblique photographic images and geometric features from laser point clouds in parallel. The module calculates channel attention weights and adaptively fuses multi-source features to improve the robustness of crack and deformation recognition under varying lighting conditions and partial occlusion, thereby reducing the false negative rate.

[0010] Preferably, in step S4, the acceptance report generation automatically outputs a dynamic report containing deviation color charts, data tables, and acceptance conclusions according to built-in standards, and then realizes multi-party visual collaborative management through the cloud platform, with early warning information pushed to relevant responsible persons in real time.

[0011] Preferably, step S1 further includes using a high-precision terrestrial laser scanner to supplement data acquisition in complex occlusion areas, thereby ensuring the integrity of the point cloud.

[0012] Preferably, in step S2, the digital twin model supports dynamic comparison and analysis of multi-time series point cloud data. The dynamic comparison and analysis automatically overlays and compares the real point cloud data collected this time with the point cloud model in the historical period. By calculating the change in Euclidean distance of the point cloud coordinates, the deformation trend of key parts of critical engineering projects is automatically identified and quantified. When the cumulative displacement exceeds the preset threshold, a trend warning chart is automatically generated in the acceptance report.

[0013] Preferably, the improved ICP-NDT algorithm described in step S2 integrates an adaptive parameter adjustment strategy based on point cloud density. This adaptive parameter adjustment strategy dynamically adjusts the nearest point search radius and NDT voxel grid size in ICP matching according to the distribution density of the point cloud in different regions, ensuring registration accuracy while optimizing computational efficiency.

[0014] The technical effects and advantages of this invention are as follows: 1. Compared with existing technologies, this method for assisting the dynamic acceptance of critical engineering projects by integrating UAV oblique photogrammetry and BIM utilizes a UAV platform equipped with RTK-PPK, a five-lens oblique photogrammetry camera, and LiDAR, combined with supplementary data acquisition by a high-precision ground laser scanner in complex areas. This innovative multi-source data fusion scheme effectively solves the problem of missing models in occluded areas from a single data source, ensuring the full elements, high accuracy, and completeness of the 3D real-scene model, and providing a reliable data foundation for subsequent analysis.

[0015] 2. Compared with existing technologies, this dynamic acceptance assistance method for critical engineering projects, which integrates UAV oblique photography and BIM, deploys a lightweight CNN model through an airborne edge computing unit and employs an attention-based feature fusion module. This enables millisecond-level real-time defect identification of registered multi-source data. This design pushes computing power to the edge, achieving real-time on-site analysis of key defects such as concrete cracks and rebar spacing deviations. It overcomes the latency caused by data transmission back to the cloud for processing, significantly improving the timeliness of safety hazard detection. The feature fusion module enhances the model's robustness to recognition under varying lighting conditions and partial occlusion by adaptively weighting and fusing texture and geometric features, reducing the false negative rate.

[0016] 3. Compared with existing technologies, this method for dynamic acceptance assistance of critical engineering projects by integrating UAV oblique photography and BIM uses an improved ICP-NDT algorithm and Kalman filtering to fuse data, construct a dynamically updated digital twin model, and support automatic overlay and comparison of multi-time series point cloud data. This combination of technologies can achieve sub-pixel-level precise registration and dynamic fusion of BIM design model and real-world point cloud data. By calculating the change in Euclidean distance, the deformation trend of key parts is automatically quantified, transforming the model from a static reference into a dynamic mapping volume, providing a theoretical basis for early warning of structural safety trends.

[0017] 4. Compared with existing technologies, this dynamic acceptance assistance method for critical and major engineering projects, which integrates UAV oblique photography and BIM, automatically compares BIM parameters with measured data through a multi-criteria acceptance decision tree based on fuzzy logic. It generates a dynamic acceptance report that includes deviation color maps, data tables, and conclusions. With the help of 5G and cloud platforms, it achieves collaborative management and early warning push. This design automates and standardizes the complex acceptance decision process, reduces human interference, improves the efficiency and objectivity of acceptance work, and realizes closed-loop management from data collection to decision output, making dynamic acceptance assistance possible. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0019] 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 only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1 As attached Figure 1 The method for assisting dynamic acceptance of critical and major engineering projects by integrating UAV oblique photography and BIM, as shown, is characterized by including the following steps: S1. Collect multi-source data of the critical engineering area by using a drone equipped with an RTK-PPK dual-mode positioning system, a five-lens oblique photography camera and a lidar. The multi-source data includes multi-view images and laser point cloud data, thereby achieving synchronous acquisition of ground data. S2. After fusing the multi-source data, a dynamically updated digital twin model is constructed; S3. The UAV is also equipped with an airborne edge computing unit, which is equipped with a lightweight CNN model to perform real-time defect identification on the registered multi-source data. The defect identification includes concrete cracks, steel bar spacing deviations and formwork deformation. S4. Based on a fuzzy logic-based multi-criteria acceptance decision tree, the system compares BIM design parameters with measured data, automatically generates an acceptance report, and transmits key data back to the cloud platform via a 5G network.

[0021] In a preferred embodiment, the fusion process in step S2 is as follows: A1. The multi-source data is coarsely registered by automatically extracting the edge endpoint features of the BIM model and point cloud data through the improved ICP-NDT algorithm, and then sub-pixel level fine registration is achieved through iterative optimization. A2. Kalman filtering is introduced to fuse GNSS, IMU, lidar and image data to construct the digital twin model.

[0022] Specifically, the UAV platform is equipped with an RTK-PPK dual-mode positioning system to obtain high-precision absolute coordinates, and integrates a five-lens oblique photography camera and a LiDAR as core data acquisition units. During flight operations, the five-lens camera simultaneously acquires image data from one vertical perspective and four oblique perspectives, while the LiDAR simultaneously acquires high-precision 3D point cloud data. These two types of data are strictly synchronized in time and space, and together constitute a multi-source dataset for critical engineering areas for complementary verification. The images provide rich texture information, while the point clouds provide precise geometric information. Especially in areas where the images may be occluded, the point cloud data can effectively supplement the missing 3D information. The collected multi-source data is first transmitted to the edge computing unit on the UAV. This unit has a built-in lightweight CNN model to process the registered multi-source data in parallel. The lightweight CNN model adopts a feature fusion module based on an attention mechanism. The working principle of this module is to extract texture features from the oblique image and geometric features from the laser point cloud, respectively. Then, by calculating the channel attention weights, different feature channels are adaptively assigned different importance weights and fused, thereby improving the ability to identify defects such as concrete cracks, steel bar spacing deviations and formwork deformation, especially enhancing robustness under conditions of lighting changes or partial occlusion. The results after edge intelligent recognition, together with the original multi-source data, are fused with the preset BIM design model through the improved ICP-NDT algorithm. The implementation principle of this algorithm is to first automatically extract the common significant feature points in the BIM model and the real point cloud, such as beam-column intersections and wall corner endpoints, for initial coarse registration. Then, iterative optimization is performed by combining the iterative nearest point algorithm and normal distribution transformation to achieve sub-pixel level fine registration. In this process, an adaptive parameter adjustment strategy based on point cloud density is integrated. This strategy dynamically adjusts the nearest point search radius and voxel grid size according to the density of the current point cloud area to balance registration accuracy and computational efficiency. After registration, the multi-source data is further processed using a Kalman filter algorithm, which integrates pose information from GNSS, IMU, LiDAR, and image data to construct a dynamically updated digital twin model that matches the actual engineering status. Finally, a multi-criteria acceptance decision tree based on fuzzy logic is activated. This decision tree automatically compares and analyzes the design parameters in the BIM model with the measured data in the digital twin model. Based on a preset fuzzy rule set, it comprehensively evaluates indicators such as the verticality and spacing deviation of the formwork support, automatically generates a report containing deviation color maps, data tables, and clear acceptance conclusions, and transmits key data and early warning information back to the cloud platform in real time via a 5G network to achieve multi-party visualized collaborative management.

[0023] As a preferred implementation, coarse registration is first performed. The principle of the improved ICP-NDT algorithm is to automatically identify and extract key edge endpoints such as beam-column intersections and wall corners in the BIM model as reference features, and at the same time, extract the corresponding edge endpoint features from the real point cloud data obtained by LiDAR through curvature change detection. The expression for the improved ICP-NDT algorithm is:

[0024] in, This represents the optimal rigid body transformation matrix to be solved. It includes rotation and translation transformations and is the final output target of the algorithm. Its function is to accurately register the source point cloud into the target reference frame; This represents the optimal transformation matrix obtained through optimization. and The subscripts i and k represent points in the source point cloud to be registered, which are real-world 3D laser point cloud data collected by UAVs. The subscripts i and k represent the indexes of the points. They belong to the same point cloud set, but may be processed by different parts of the algorithm. Indicates the relationship between the target point cloud and the source point. The nearest point. This is the core concept of the ICP algorithm, found through nearest neighbor search; It represents the average three-dimensional coordinates of all points within a certain voxel grid into which the target point cloud is divided, and represents the center position of that grid; Let T represent the objective function of the ICP algorithm. Its goal is to find a transformation T such that, after transformation, the source point cloud corresponds to its closest point in the target point cloud. Minimize the sum of squared Euclidean distances between the two point clouds. Minimizing this directly means geometrically aligning the two point clouds. This represents the objective function of the NDT algorithm (with a negative sign). The NDT algorithm does not directly search for point-to-point correspondences, but instead calculates the transformed source point. The probability of falling into a voxel grid of the target point cloud. The exponential term is proportional to this probability density. Therefore, maximizing the sum of the probability densities of all source points is equivalent to minimizing the sum of the negative probability densities. Minimizing this term means distributing the points of the source point cloud as much as possible in the high-probability regions of the target point cloud, achieving statistical alignment and making it more robust to noise and outliers; and These represent weighting coefficients used to balance the contributions of the ICP and NDT terms to the overall objective function. The improved algorithm, by adjusting these two parameters, combines the high accuracy of ICP with the robustness of NDT. This represents the regularization or constraint term, which signifies the adaptive parameter adjustment strategy based on point cloud density. Here, ρ represents the density of the local point cloud. This term does not directly minimize the distance or maximize the probability, but rather imposes constraints on the transformation T, ensuring that parameters such as the search radius and voxel size of the algorithm can be adaptively adjusted in sparse or dense regions of the point cloud, thereby optimizing computational efficiency while maintaining accuracy. In implementation, the algorithm first calculates the centroids of the two sets of feature points and centers their coordinates. Then, it solves for the optimal rotation matrix and translation vector through singular value decomposition, completing the initial spatial alignment, i.e., coarse registration. Based on coarse registration, the algorithm enters the fine registration stage. Its principle combines the advantages of the Iterative Closest Point (ICP) algorithm and the Normal Distribution Transform (NDT). The ICP algorithm iteratively finds the closest point and minimizes the mean square error of the distance between corresponding point pairs, while the NDT algorithm divides the reference point cloud into a voxel grid and uses a multidimensional normal distribution to describe the statistical characteristics of points within each voxel. It solves the problem by optimizing the matching degree between the current point cloud and these distribution models. The transformation parameters are implemented by taking the coarsely registered point cloud as input, using ICP for initial iterative optimization to quickly approximate the model, and then switching to the NDT algorithm for fine adjustment. This combined iterative approach ultimately achieves sub-pixel level high-precision registration. After completing the precise spatial alignment of the point cloud and the BIM model, Kalman filtering is introduced for data fusion. The principle is to establish a system state equation and an observation equation, using the absolute position provided by GNSS, the inertial attitude measured by IMU, the high-frequency 3D point cloud generated by lidar, and the relative pose calculated by visual odometry from the oblique photogrammetric image as the observation input. In practice, the prediction step of the Kalman filter is based on the IMU data to deduce the system state, i.e. the current pose of the digital twin model. The update step uses observation data from GNSS, lidar and imagery to correct the prediction value. Through this recursive cycle of prediction and update, the noise of each sensor is effectively suppressed, multi-source heterogeneous data is dynamically fused, and finally a digital twin model with accurate position, fine geometry and dynamic updates over time is constructed.

[0025] As a preferred implementation, when the UAV flies along a preset route, its integrated five-lens oblique photography camera system consists of a vertically mounted front-view lens and four oblique-view lenses tilted at a certain angle in the forward, backward, left, and right directions, respectively. The five lenses are controlled by a unified synchronous triggering unit during flight and are strictly exposed simultaneously, thereby synchronously acquiring a vertical view image and four complementary oblique view images of the target area. This multi-view acquisition principle ensures the efficient acquisition of texture information of the side of the ground features. Simultaneously, the lidar on the same platform emits laser pulses and receives echoes, directly acquiring high-density 3D point cloud data of the ground surface and structures. During implementation, the lidar and the five-lens camera are spatiotemporally synchronized through a high-precision RTK-PPK positioning system and a high-frequency IMU unit, ensuring that each laser point cloud data point has precise absolute coordinates and attitude information, and each image has corresponding exterior orientation elements. The core principle of this design is to supplement the image occlusion areas that may exist in oblique photography using laser point clouds. Because lasers can penetrate vegetation gaps or obtain points under building obstruction from different angles, in specific implementation, the synchronously acquired point cloud data is initially registered with multi-view images. In the subsequent modeling process, for areas that cannot be clearly imaged in multiple view images due to occlusion, the 3D geometric information of the laser point cloud already acquired at that location is directly used for model construction and representation. This effectively solves the problem of model distortion or missing parts in occluded areas when relying solely on images for 3D reconstruction, ensuring the geometric integrity of the constructed digital twin model.

[0026] As a preferred embodiment, in step S2, the specific implementation of the improved ICP-NDT algorithm includes: firstly, automatic extraction of feature points is performed. The principle is to use the BIM model as a precise geometric reference. The intersection lines between its beams and columns and the edge lines of the walls have clear endpoint coordinates in three-dimensional space. These endpoints are predefined as feature point sets. For laser point clouds, the corresponding beam-column intersections and wall corner edge points in the actual scene are detected by calculating the curvature change or normal vector mutation of the local point cloud cluster, thereby obtaining the real scene feature point set. The coarse registration is based on these two sets of feature point sets. After calculating their centroids and centering their coordinates, the singular value decomposition method is used to solve for the optimal rotation matrix and translation vector between the two sets of point sets to complete the initial spatial alignment. Based on the good initial position provided by coarse registration, the algorithm enters the iterative optimization stage of fine registration. The principle is to alternately use the advantages of ICP algorithm and NDT algorithm. In implementation, ICP algorithm is first used for several iterations to quickly reduce the point-to-point distance error. Then, the algorithm is switched to NDT algorithm to divide the target point cloud into voxel grid and calculate the normal distribution parameters of the points in each grid. The pose is further adjusted by optimizing the probability that the source point cloud falls into these distributions after the transformation. This combined iterative optimization approach overcomes the shortcomings of the ICP algorithm, which is susceptible to the influence of initial position and outliers, and the NDT algorithm, which has slow convergence in the early stages. It gradually converges to sub-pixel level registration accuracy, laying the foundation for building a high-precision digital twin model.

[0027] As a preferred implementation, a lightweight CNN model mounted in the UAV's onboard edge computing unit is responsible for receiving multi-source data that has been spatially registered with the BIM model. This model performs millisecond-level forward inference calculations on the input registered oblique photographic images and laser point cloud data. The expression for this lightweight CNN model is:

[0028] in, This represents the input multi-view oblique photographic imagery data. It is a four-dimensional tensor, whose dimensions are typically represented as [batch size, height, width, number of channels]. It contains RGB images captured from five lenses of a drone; This represents the input laser point cloud data. For efficient processing by CNNs, point clouds are typically preprocessed into regular data structures, such as voxel grids: dividing 3D space into a grid, with each grid encoding the features of its internal points; express; This represents two independent, lightweight feature extraction subnetworks, which are two structurally simplified convolutional neural networks, typically composed of lightweight operations such as depthwise separable convolutions and pooling layers. Specifically designed for use from image data Extract advanced texture features. Specifically designed to extract advanced geometric features from point cloud data, they output two sets of feature maps respectively; This represents a feature fusion module based on an attention mechanism; Indicates the characteristics after receiving and fusing. Typically, global average pooling and fully connected layers are used to ultimately output a probability vector. ; This represents the final output of the model. It is a vector where each element represents the probability that the input data belongs to a specific defect category. The category with the highest probability is taken as the recognition result. The principle lies in leveraging the ability of convolutional neural networks to extract image features hierarchically. First, texture features of concrete cracks are extracted from oblique images using convolutional and pooling layers. Simultaneously, geometric features of the formwork and reinforcing bars are extracted from depth or intensity maps generated from point cloud data. Then, a multi-criteria acceptance decision tree based on fuzzy logic is activated. The input to this decision tree is the quantified defect results identified by the CNN model, such as the pixel width of cracks and the image spacing of reinforcing bars. Its core principle is to transform precise numerical inputs into membership degrees of fuzzy sets. For example, the measured value of the vertical deviation of the formwork is input into a pre-defined fuzzification interface, which defines fuzzy sets such as "no deviation," "slight deviation," and "severe deviation," along with their membership functions. The fuzzy inference engine then performs inference based on a pre-defined fuzzy rule base established by domain experts. The rules are in the form of "if," for example, if the vertical deviation belongs to "slight deviation," then... If the deviation is minor and the spacing deviation is slight, the overall risk level is medium. Finally, the result of fuzzy inference is converted into a precise graded warning signal through the defuzzification interface. When the judgment deviation exceeds the allowable range, the system automatically triggers a graded warning. For example, green represents normal and no warning is needed, yellow represents attention and observation is needed, and red represents danger and immediate action is needed. This warning signal and key data are transmitted back through the 5G network. The lightweight CNN model adopts a feature fusion module based on an attention mechanism, which can process texture features from oblique photogrammetry and geometric features from laser point clouds in parallel. The module calculates channel attention weights and adaptively fuses multi-source features to improve the robustness of crack and deformation recognition under lighting changes and partial occlusion environments, and reduce the false negative rate. The calculation process of channel attention weights can be decomposed into the following key steps and expressions: First, the two feature maps from the image and point cloud are stitched together, and the global spatial information is compressed. , This represents the input feature map. It refers to the feature map that has been initially extracted by a lightweight CNN and is ready for fusion. This represents the total number of channels in the feature map; and This represents the spatial height and width of the feature map; This represents the global descriptor of the c-th channel. It is a scalar obtained by averaging the feature values ​​of all spatial locations (i,j) of that channel; then, using a fully connected layer and gating mechanism, the importance weight of each channel is calculated based on global information. ; This represents the weight matrix of the first fully connected layer. Its function is to reduce the dimensionality of the global descriptor z, thereby reducing computational cost and introducing non-linearity. This represents the weight matrix of the second fully connected layer. Its function is to restore the dimension to the original number of channels; This is typically represented by the ReLU function. Introducing nonlinear relationships enhances the model's expressive power. This represents the Sigmoid activation function. It compresses the output values ​​to the range of 0 to 1, so that the final weights s can be represented as a "gated" signal, i.e., an importance coefficient; This represents the final calculated channel attention weight vector; Next, the calculated weights are applied to the original feature map to complete adaptive fusion. , This represents the feature map of the c-th channel after reweighting. This represents channel-by-channel multiplication. The original feature map... Each value is multiplied by its corresponding channel weight. The principle is that if a certain channel is crucial for identifying the deformation of the formwork support under occlusion, then the network will learn to assign it a weight close to 1. To enhance this feature, a lower weight is assigned to texture channels that are greatly affected by lighting and have a lot of noise, thus suppressing their effect.

[0029] As a preferred implementation, the multi-criteria acceptance decision tree based on fuzzy logic automatically judges indicators such as the verticality and spacing deviation of the formwork support, and its output drives the report generation engine. The engine first calls the built-in industry acceptance standard library as the judgment benchmark, and quantifies the fuzzy logic results output by the decision tree into specific deviation values ​​and conclusions. Then, it automatically generates a dynamic acceptance report. The report contains three core elements. The first is a deviation color chart, which is obtained by visually rendering the difference between the measured value and the BIM design value on the fused digital twin model with different colors. For example, green represents the qualified range, yellow represents slight deviation, and red represents serious deviation. Secondly, there are data tables. The system automatically extracts comparative data from key parts and clearly lists the design values, measured values, and deviation values ​​in a structured table format. Thirdly, there are acceptance conclusions, which are automatically generated by the system according to preset rules. After the report is generated, the system transmits the complete dynamic report and key data necessary for collaborative management, such as warning levels, deviation color snapshots, and key data tables, back to the cloud platform through the high bandwidth and low latency link of the 5G network. After receiving the data, the cloud platform updates it to the shared digital twin model, realizing multi-party online visual collaborative management, and sends the warning information to the preset relevant personnel in real time through the platform's message push mechanism according to the warning level.

[0030] As a preferred implementation, after the UAV completes its predetermined flight path, the system automatically analyzes the data integrity based on the initially generated point cloud model and identifies sparse or missing areas of the point cloud caused by obstruction from buildings or the surrounding environment. For these identified complex obstructed areas, operators will use a high-precision ground laser scanner for supplementary data collection. This scanner is set up at multiple stations around the obstructed area, performing detailed scans of parts that are difficult for the UAV to detect from different ground angles to ensure no blind spots. During implementation, the high-precision ground laser scanner must use the same coordinate system reference as the UAV data collection system, by deploying at least [number missing] stations on the ground. Three control points, precisely measured in UAV aerial surveying, are used as common targets to achieve accurate alignment of point cloud data from different sources. The ground point cloud data acquired in addition has higher point accuracy and resolution and is then imported into the data processing system to be fused with the aerial point cloud data acquired by the UAV. During fusion, the improved ICP-NDT algorithm is used to accurately register the ground scan point cloud and the UAV point cloud in the overlapping area, thereby forming a complete, seamless, and unobstructed overall point cloud dataset. This step fundamentally solves the problem of model incompleteness that may exist when relying solely on aerial acquisition and provides a solid data foundation for building a high-precision digital twin model.

[0031] As a preferred implementation, after collecting and generating the current period's real-world point cloud model during each inspection, the system automatically retrieves the point cloud models of the same critical parts of the major project stored in the historical period from the cloud platform as a comparison benchmark. Its core principle is to use an improved ICP-NDT algorithm to accurately align the multi-time series point cloud data acquired at different time points in space, ensuring that they are in the same coordinate system to achieve effective dynamic comparison. In specific implementation, the system automatically overlays and compares the real-world point cloud data collected this time with the selected historical point cloud model. The core of the comparison analysis is to calculate the change in Euclidean distance of the point cloud coordinates of the same spatial location point at different times. This change represents the displacement or deformation of the point in the normal direction. The system generates a color-coded deformation cloud map by traversing and calculating the Euclidean distance of all point clouds of the critical parts, and automatically calculates the cumulative displacement within the calculated area. When the system detects that the cumulative displacement of a critical part, such as the top of the formwork, exceeds the preset safety threshold, it will automatically trigger an early warning mechanism in the current acceptance report, generating a time-series curve or deformation distribution chart that intuitively displays the deformation development trend, thereby realizing the automatic identification, quantification, and early warning of the deformation trend of critical parts of critical engineering projects.

[0032] As a preferred implementation, the algorithm first performs density analysis on the input point cloud data during runtime. The principle is to divide the point cloud space into temporary initial grids and count the number of points within each grid, thereby calculating a local point cloud density distribution map. During implementation, for densely distributed point cloud regions, the algorithm automatically assigns a smaller nearest-point search radius and a smaller voxel grid size for the registration calculation. This is because high-density point clouds provide abundant feature point pairs, and a small radius for fine-grained searching can effectively improve registration accuracy and detail reproduction. Conversely, for sparsely distributed point cloud regions, the algorithm automatically assigns a larger nearest-point search radius and a larger voxel grid size. This is because sparse point clouds have fewer feature points, and expanding the search range helps find effective corresponding point pairs, avoiding registration failure. At the same time, a larger voxel size ensures the stability of the statistical distribution and improves the robustness of the algorithm. The core principle of this adaptive adjustment strategy is to dynamically adapt the key parameters of the algorithm to the point cloud distribution characteristics of the current processing area. This ensures overall registration accuracy while avoiding ineffective fine calculations in sparse areas or incorrect matching in dense areas due to excessively large search ranges. Ultimately, it optimizes the computational efficiency of the entire registration process globally.

[0033] The above describes the working principle of the dynamic acceptance assistance method for critical and major engineering projects that integrates UAV oblique photography and BIM.

Claims

1. A dynamic acceptance auxiliary method for critical and major engineering projects integrating UAV oblique photography and BIM, characterized in that, Includes the following steps: S1. Collect multi-source data of the critical engineering area by using a drone equipped with an RTK-PPK dual-mode positioning system, a five-lens oblique photography camera and a lidar. The multi-source data includes multi-view images and laser point cloud data, thereby achieving synchronous acquisition of ground data. S2. After fusing the multi-source data, a dynamically updated digital twin model is constructed; S3. The UAV is also equipped with an airborne edge computing unit, which is equipped with a lightweight CNN model to perform real-time defect identification on the registered multi-source data. The defect identification includes concrete cracks, steel bar spacing deviations and formwork deformation. S4. Based on a fuzzy logic-based multi-criteria acceptance decision tree, the system compares BIM design parameters with measured data, automatically generates an acceptance report, and transmits key data back to the cloud platform via a 5G network.

2. The method for assisting dynamic acceptance of critical and major engineering projects by integrating UAV oblique photography and BIM according to claim 1, characterized in that: In step S2, the fusion process consists of the following steps: A1. The multi-source data is coarsely registered by automatically extracting the edge endpoint features of the BIM model and point cloud data through the improved ICP-NDT algorithm, and then sub-pixel level fine registration is achieved through iterative optimization. A2. Kalman filtering is introduced to fuse GNSS, IMU, lidar and image data to construct the digital twin model.

3. The method for assisting dynamic acceptance of critical and major engineering projects by integrating UAV oblique photography and BIM according to claim 1, characterized in that: In step S1, the specific method of multi-scale data acquisition includes: simultaneously acquiring images from one vertical perspective and four oblique perspectives through a five-lens oblique photography system, fusing LiDAR point cloud data, and supplementing missing data in image occlusion areas.

4. The method for assisting dynamic acceptance of critical and major engineering projects by integrating UAV oblique photography and BIM according to claim 1, characterized in that: In step S2, the specific implementation of the improved ICP-NDT algorithm includes: automatically extracting the beam-column intersection points and wall corner edge endpoint features of the BIM model and the point cloud for coarse registration, and performing iterative optimization by combining the ICP algorithm and normal distribution transformation.

5. The method for assisting dynamic acceptance of critical and major engineering projects by integrating UAV oblique photography and BIM according to claim 1, characterized in that: In step S3, the edge intelligent decision-making achieves millisecond-level defect identification through a lightweight CNN model. The decision tree based on fuzzy logic automatically determines whether the verticality and spacing deviation of the formwork support exceed the allowable range and triggers a graded warning.

6. The method for assisting dynamic acceptance of critical and major engineering projects by integrating UAV oblique photography and BIM according to claim 5, characterized in that: The lightweight CNN model employs an attention-based feature fusion module, which can process texture features from oblique photogrammetry images and geometric features from laser point clouds in parallel. The module calculates channel attention weights and adaptively fuses multi-source features to improve the robustness of crack and deformation identification under varying lighting conditions and partial occlusion, thereby reducing the false negative rate.

7. The method for assisting dynamic acceptance of critical and major engineering projects by integrating UAV oblique photography and BIM according to claim 1, characterized in that: In step S4, the acceptance report is generated automatically based on built-in standards, and includes a color chart of deviations, data tables, and acceptance conclusions. Then, multi-party visual collaborative management is achieved through the cloud platform, and early warning information is pushed to relevant responsible persons in real time.

8. The method for assisting dynamic acceptance of critical and major engineering projects by integrating UAV oblique photography and BIM according to claim 1, characterized in that: Step S1 also includes using a high-precision terrestrial laser scanner to supplement data collection in complex occlusion areas, thereby ensuring the integrity of the point cloud.

9. The method for assisting dynamic acceptance of critical and major engineering projects by integrating UAV oblique photography and BIM according to claim 1, characterized in that: In step S2, the digital twin model supports dynamic comparison and analysis of multi-time series point cloud data. The dynamic comparison and analysis automatically overlays and compares the real point cloud data collected this time with the point cloud model in the historical period. By calculating the change in Euclidean distance of the point cloud coordinates, the deformation trend of key parts of critical engineering projects is automatically identified and quantified. When the cumulative displacement exceeds the preset threshold, a trend warning chart is automatically generated in the acceptance report.

10. The method for assisting dynamic acceptance of critical and major engineering projects by integrating UAV oblique photography and BIM according to claim 1, characterized in that: In step S2, the improved ICP-NDT algorithm integrates an adaptive parameter adjustment strategy based on point cloud density. The adaptive parameter adjustment strategy dynamically adjusts the nearest point search radius and NDT voxel grid size in ICP matching according to the distribution density of point cloud in different regions to ensure registration accuracy while optimizing computational efficiency.

Images