An unmanned aerial vehicle laser radar fusion engineering real scene three-dimensional surveying and mapping system and method

Abstract

The application discloses an unmanned aerial vehicle laser radar fusion engineering real scene three-dimensional surveying and mapping system and method, the system comprises: a multi-temporal data acquisition unit, for performing multi-temporal, multi-view joint acquisition on an engineering area, obtaining a laser point cloud sequence, an image sequence and corresponding time identification information; a feature registration extraction unit is used for unified coordinate registration of the laser point cloud sequence and the image sequence, the application relates to the field of engineering surveying and mapping. The unmanned aerial vehicle laser radar fusion engineering real scene three-dimensional surveying and mapping system and method, through joint registration of laser point clouds and images, multi-feature extraction and echo-texture conflict determination, realize the boundary completion, false face suppression and geometric correction of the three-dimensional surface, improve the accuracy and reliability of the surveying and mapping result, through the modular design of the system, realize the automatic acquisition, feature extraction, model correction and fusion output of multi-temporal, multi-view data, reduce manual intervention, improve the surveying and mapping efficiency.

Description

Technical Field

[0001] This invention relates to the field of engineering surveying, and more specifically, to a three-dimensional engineering real-scene surveying system and method that integrates UAV lidar. Background Technology

[0002] In the field of traditional engineering surveying, 3D surveying methods based on a single data source have many limitations. For example, while relying solely on lidar data can acquire relatively accurate spatial geometric information, it is insufficient in capturing detailed information such as the texture and color of objects; while using image data alone results in significant errors in geometric accuracy and spatial positioning, and is easily affected by factors such as lighting and occlusion.

[0003] With the continuous expansion of engineering construction scale and increasing complexity, higher demands are placed on the accuracy and completeness of 3D mapping of engineering scenes. Existing mapping methods are insufficient to meet the mapping needs of multi-element, high-precision, and dynamically changing engineering scenarios. Especially in large-scale infrastructure construction projects, objects within the project area have different stability and change characteristics, including permanent buildings and structures, phased construction ancillary facilities, and transient interference objects. How to accurately distinguish these objects with different properties and perform high-quality 3D modeling for each is a key problem currently facing the field of engineering surveying.

[0004] Furthermore, during multi-temporal data acquisition, factors such as attitude disturbances of the UAV platform, environmental changes, and noise from the acquisition equipment itself can lead to registration errors between laser point clouds and image data, thus affecting the accuracy of the 3D model. Simultaneously, the fusion of data acquired from different perspectives may result in surface conflicts and geometric distortions. Effectively addressing these issues to improve the quality and reliability of the 3D model is a pressing technical challenge. Summary of the Invention

[0005] The purpose of this invention is to provide an engineering real-scene 3D mapping system and method that integrates UAV lidar, which solves the problem that existing mapping methods are difficult to meet the mapping needs of multi-element, high-precision, and dynamically changing engineering scenes. In the process of multi-temporal data acquisition, due to factors such as attitude disturbance of the UAV platform, environmental changes, and noise of the acquisition equipment itself, there will be registration errors between the laser point cloud and image data, which will affect the accuracy of the 3D model.

[0006] This invention achieves the above objectives through the following technical solution: a UAV-based lidar fusion engineering real-scene 3D mapping system, the system comprising:

[0007] The system includes a multi-temporal data acquisition unit, a feature registration and extraction unit, a stability model calculation unit, a spatial unit layering unit, a candidate surface feature extraction unit, an echo texture conflict determination unit, a three-dimensional surface correction unit, and a three-dimensional model fusion output unit.

[0008] The multi-temporal data acquisition unit is used to perform multi-temporal and multi-view joint acquisition of the engineering area to obtain laser point cloud sequences, image sequences and corresponding time marker information;

[0009] The feature registration and extraction unit is used to perform unified coordinate registration on the laser point cloud sequence and the image sequence, and to extract the geometric features, texture features, temporal continuity features and component semantic features of the spatial units.

[0010] The stability model calculation unit is used to calculate the component stage stability parameters of each spatial unit based on occupancy persistence, position drift amplitude, morphological preservation degree and semantic category.

[0011] The spatial unit layering unit is used to divide the spatial unit of the test area into a permanent modeling layer, a stage auxiliary layer and a transient interference layer according to the component stage stability parameters.

[0012] The candidate surface feature extraction unit is used to extract laser echo intensity anomaly features, neighborhood distance abrupt change features, boundary break features, texture continuity features, edge closure features, and cross-view reprojection error features from candidate surfaces in the permanent modeling layer and the stage auxiliary layer.

[0013] The echo-texture conflict determination unit is used to construct a surface-level echo-texture conflict field and output the type determination result of the candidate surface;

[0014] The three-dimensional surface correction unit is used to perform boundary completion, pseudo-face suppression, geometric correction and local reconstruction processing based on the surface type determination result.

[0015] The 3D model fusion output unit is used to fuse the corrected 3D surface and layering results, and output an engineering real-scene 3D mapping model with stage labels, material credibility labels and surface type labels.

[0016] Furthermore, the feature registration and extraction unit includes:

[0017] The module includes an exterior orientation acquisition module, a coordinate mapping module, a gross error removal module, and a joint feature generation module.

[0018] The external orientation acquisition module is used to acquire the pose parameters and imaging parameters of the UAV platform;

[0019] The coordinate mapping module is used to establish the projection mapping relationship between the three-dimensional coordinates of the laser point cloud and the image pixel coordinates;

[0020] The gross error removal module is used to remove out-of-registration points caused by attitude disturbances, occlusion changes, and acquisition noise.

[0021] The joint feature generation module is used to generate geometric contour descriptions, surface texture descriptions, temporal variation descriptions, and semantic attribute descriptions of spatial units under a unified coordinate reference.

[0022] The joint feature generation module is also used to calculate the geometric contour description, surface texture description, temporal change description, and semantic attribute description based on the boundary point set of the spatial unit, the projected image block, the multi-temporal center coordinates, and the component category recognition results.

[0023] Furthermore, the stability model calculation unit includes:

[0024] The system includes a stability modeling module, a sample training module, a parameter inference module, and a weight correction module.

[0025] The stability modeling module is used to establish a component stage stability calculation model with occupancy persistence, position drift amplitude, shape preservation degree and semantic category as inputs;

[0026] The sample training module is used to perform normalization training on fused samples from different engineering scenarios.

[0027] The parameter reasoning module is used to output the component stage stability parameters of each spatial unit;

[0028] The weight correction module is used to correct the contribution weight of each input factor based on the project type, component distribution density, and data acquisition sequence integrity.

[0029] Furthermore, the spatial unit layering unit includes:

[0030] Threshold setting module, hierarchy determination module, and interference handling module;

[0031] The threshold setting module is used to set a first stability threshold and a second stability threshold based on the stability statistics of typical permanent components and typical stage auxiliary components.

[0032] The hierarchical determination module is used to classify a spatial unit as a permanent modeling layer when the stability parameter of the component stage is higher than the first stability threshold, as a stage-attached layer when the stability parameter of the component stage is between the first stability threshold and the second stability threshold, and as a transient interference layer when the stability parameter of the component stage is lower than the second stability threshold.

[0033] The interference handling module is used to remove or independently identify transient interference layers.

[0034] Furthermore, the candidate surface feature extraction unit includes:

[0035] The module includes: echo anomaly extraction module, distance abrupt change extraction module, boundary break extraction module, texture continuity extraction module, edge closure extraction module, and reprojection error extraction module.

[0036] The echo anomaly extraction module is used to determine the degree of echo anomaly based on the difference in echo intensity between the target laser point and neighboring laser points.

[0037] The distance mutation extraction module is used to characterize the discontinuous changes in the spacing between local points on the candidate surface;

[0038] The boundary fracture extraction module is used to characterize the interruption location and interruption intensity of the candidate surface profile;

[0039] The texture continuity extraction module is used to characterize the continuity of the texture of a candidate surface on both sides of the boundary based on the texture direction, grayscale gradient and texture similarity of the candidate surface in adjacent image blocks.

[0040] The edge closure extraction module is used to characterize the degree of closure integrity of the candidate surface edge contour based on the start-end distance of the candidate surface edge points, the change in contour curvature, and the closure ratio.

[0041] The reprojection error extraction module is used to characterize the degree of deviation between the theoretical projection position and the actual projection position of candidate surface feature points in multi-view images.

[0042] Furthermore, the echo texture conflict determination unit includes:

[0043] The system includes a conflict model construction module, a feature fusion module, a conflict value calculation module, and a surface classification module.

[0044] The conflict model construction module is used to establish an echo-texture conflict field determination model for candidate surfaces;

[0045] The feature fusion module is used to perform spatial mapping, dimensional unification, and correlation fusion of laser-related features and image features;

[0046] The conflict value calculation module is used to output the surface conflict value based on the normalization results and corresponding weights of each feature;

[0047] The surface classification module is used to combine the surface conflict value and the surface category probability distribution to determine the candidate surface as a reliable solid surface, a surface with missing penetration measurement, a mirror pseudo-surface, an edge undersampled surface, or a pseudo-continuous surface.

[0048] Furthermore, the three-dimensional surface correction unit includes:

[0049] Missing measurement completion module, pseudo-surface suppression module, boundary completion module, and geometric correction module;

[0050] The missing measurement completion module is used to perform point cloud completion for the penetrating missing measurement surface based on the spatial distribution of effective points in the neighborhood;

[0051] The pseudoface suppression module is used to remove pseudoface points for mirrored pseudofaces based on abnormal echo characteristics and normal consistency.

[0052] The boundary completion module is used to perform boundary restoration for undersampled surfaces based on contour extraction results and neighborhood geometric constraints.

[0053] The geometric correction module is used to correct the three-dimensional coordinates of the distorted region on the pseudo-continuous surface based on the cross-view reprojection deviation and the local surface fitting results.

[0054] Furthermore, the 3D model fusion output unit includes:

[0055] The module includes a hierarchical fusion module, a tag generation module, and a results output module.

[0056] The layered fusion module is used to associate and fuse the corrected 3D surface with the spatial units corresponding to the permanent modeling layer and the stage auxiliary layer;

[0057] The label generation module is used to generate stage labels, material confidence labels, and surface type labels for the fused model. The stage label is used to characterize the modeling level to which the spatial unit belongs, the material confidence label is used to characterize the confidence level of the candidate surface, and the surface type label is used to characterize the specific type of the candidate surface.

[0058] The output module is used to output a three-dimensional mapping model of the engineering scene with multi-level semantic relationships.

[0059] Furthermore, it also includes a surveying result verification unit:

[0060] The mapping result verification unit is used to perform component stage consistency verification, surface geometric authenticity verification, and label correspondence verification on the output model based on the unified coordinate benchmark, spatial unit layering results, surface type determination results, and three-dimensional surface correction results.

[0061] The mapping result verification unit is also used to verify the boundary continuity between the permanent modeling layer and the stage sub-layer, the matching relationship between the material credibility label and the surface type label, and the semantic consistency between multi-temporal mapping results, and outputs the verified model as the engineering real scene 3D mapping result.

[0062] A method for fusion of UAV LiDAR and AI in engineering real-scene 3D mapping, applied to the aforementioned UAV LiDAR fusion engineering real-scene 3D mapping system, the method comprising:

[0063] S1. Perform multi-temporal and multi-view joint acquisition on the engineering area to obtain laser point cloud sequences, image sequences and corresponding time marker information;

[0064] S2. Perform unified coordinate registration on the laser point cloud sequence and the image sequence, and extract the geometric features, texture features, temporal continuity features and component semantic features of the spatial units;

[0065] S3. Calculate the component stage stability parameters of each spatial unit based on occupancy persistence, positional drift amplitude, morphological preservation degree, and semantic category;

[0066] S4. Based on the component stage stability parameters, the spatial unit of the test area is divided into a permanent modeling layer, a stage auxiliary layer, and a transient interference layer.

[0067] S5. Extract laser echo intensity anomaly features, neighborhood distance abrupt change features, boundary break features, texture continuity features, edge closure features, and cross-view reprojection error features from candidate surfaces in the permanent modeling layer and stage-attached layer.

[0068] S6. Construct a surface-level echo-texture conflict field and output the type determination result of the candidate surface;

[0069] S7. Perform boundary completion, pseudo-facet suppression, geometric correction, and local reconstruction based on the surface type determination results;

[0070] S8. Integrate the corrected 3D surface and layered results to output a 3D engineering real-scene mapping model with stage labels, material confidence labels, and surface type labels.

[0071] The beneficial effects of this invention are as follows:

[0072] 1. By jointly registering laser point clouds and images, extracting multiple features, and determining echo-texture conflicts, we can achieve boundary completion, pseudo-facet suppression, and geometric correction of three-dimensional surfaces, thereby improving the accuracy and reliability of mapping results.

[0073] 2. Based on the component stage stability parameters, the spatial unit of the survey area is divided into a permanent modeling layer, a stage auxiliary layer, and a transient interference layer to realize the stage modeling of dynamic components and construction process.

[0074] 3. Generate stage labels, material credibility labels, and surface type labels for 3D models, providing semantic information and credibility assessment of components, supporting engineering management and digital twin applications.

[0075] 4. Through modular system design, the system enables automatic acquisition, feature extraction, model correction, and fusion output of multi-temporal and multi-view data, reducing manual intervention and improving surveying efficiency.

[0076] 5. It has wide applicability and is suitable for various engineering construction sites, urban infrastructure monitoring and complex industrial environments, realizing refined three-dimensional mapping and dynamic monitoring. Attached Figure Description

[0077] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0078] Figure 1 This is a system block diagram of the present invention;

[0079] Figure 2 This is a flowchart of the feature registration and extraction unit of the present invention;

[0080] Figure 3 This is a flowchart of the stability model calculation unit of the present invention;

[0081] Figure 4 This is a flowchart of the overall method of the present invention. Detailed Implementation

[0082] The present application will now be described in further detail with reference to the accompanying drawings. It should be noted that the following specific embodiments are only used to further illustrate the present application and should not be construed as limiting the scope of protection of the present application. Those skilled in the art can make some non-essential improvements and adjustments to the present application based on the above application content.

[0083] Example 1:

[0084] Please see Figure 1-3 This invention provides a technical solution: a UAV-based LiDAR fusion engineering real-scene 3D mapping system, the system comprising:

[0085] Multi-temporal data acquisition unit, feature registration and extraction unit, stability model calculation unit, spatial unit layering unit, candidate surface feature extraction unit, echo texture conflict determination unit, three-dimensional surface correction unit, and three-dimensional model fusion output unit;

[0086] The multi-temporal data acquisition unit is used to acquire multi-temporal and multi-view data of the engineering area using a drone equipped with lidar and imaging module, and to obtain laser point cloud sequences, image sequences and corresponding timestamp information;

[0087] Among them, the laser point cloud sequence is a series of data sets composed of laser points obtained at different times by scanning the engineering area with lidar. These points contain the spatial location information of the target object; the image sequence is a series of image data captured by the imaging module on the UAV, such as a camera, at different times; and the timestamp information records the identification information of the acquisition time of the laser point cloud sequence and the image sequence, which is used to determine the specific time of data acquisition and provide a time reference for subsequent multi-temporal data analysis.

[0088] The feature registration and extraction unit is used to perform unified coordinate registration between the laser point cloud sequence and the image sequence, and to extract the geometric features, texture features, temporal continuity features and semantic features of the spatial units.

[0089] Among these features, unified coordinate registration transforms the laser point cloud sequence and the image sequence into the same coordinate system, ensuring accurate spatial correspondence for subsequent joint analysis; geometric features describe the spatial unit's shape, size, orientation, and other characteristics, such as the object's side length, angle, and curvature, used to describe the object's spatial morphology; texture features depict the details and patterns on the object's surface in the image, reflecting information such as the object's material and color distribution, which can be extracted using image processing techniques; temporal continuity features consider the changes in spatial units across different time phases, reflecting their continuity or trend over time; and component semantic features assign semantic information to various components in the engineering area, such as buildings and roads, including their type and function, which helps in understanding the structure and composition of the engineering area.

[0090] Furthermore, the specific implementation process of the joint feature generation module includes: dividing the point cloud after unified coordinate registration into spatial units according to the preset voxel side length or component boundary; extracting the boundary point set, normal vector, mean curvature value and circumscribed contour for each spatial unit to form a geometric contour description; projecting the spatial unit onto the corresponding image block, extracting the color histogram, gray-level co-occurrence matrix statistics and texture direction consistency to form a surface texture description; forming a temporal variation description based on the center point displacement, occupancy frequency and contour change rate of the same spatial unit in continuous temporal phases; inputting the point cloud projection area into the component semantic recognition model trained by labeled samples to obtain the probability of categories such as building body, road, scaffolding, formwork support, construction vehicle or personnel, and forming a semantic attribute description based on the category corresponding to the maximum category probability and its confidence level.

[0091] The stability model calculation unit is used to construct and train the component stage stability calculation model and the echo-texture conflict field determination model. Then, based on the occupancy persistence, positional drift amplitude, morphological preservation degree and semantic category of each spatial unit in multiple temporal phases, the corresponding component stage stability parameters are calculated through the trained model.

[0092] Among them, the component stage stability calculation model is a mathematical model used to calculate the stability of various components in an engineering area at different stages. It evaluates the stability of components by considering multiple factors. The echo-texture conflict field determination model is used to determine whether there is a conflict between laser echo features and image texture features, and to determine the conflict situation. After training, it can accurately identify the conflict field state under different feature combinations. Occupation persistence is the degree to which a spatial unit persists in multi-temporal data, reflecting the unit's temporal stability. The longer the persistence, the higher the occupation persistence. Position drift amplitude is the range of position changes of a spatial unit in different temporal data. The smaller the position drift amplitude, the more stable the unit's position. Morphological preservation is the degree to which the shape, size, and other morphological features of a spatial unit remain unchanged in different temporal phases. The better the morphological preservation, the more stable the unit. Component stage stability parameters are calculated by the trained component stage stability calculation model based on the above factors and are used to quantitatively describe the stability of components at different stages.

[0093] The spatial unit is a layered unit used to divide the spatial unit of the test area into a permanent modeling layer, a stage-attached layer and a transient interference layer according to the component stage stability parameters. The transient interference layer is removed or identified separately, while the permanent modeling layer and the stage-attached layer are included in the subsequent surface reliability determination process.

[0094] The permanent modeling layer, defined based on the component stage stability parameters, contains spatial unit layers of components that are stable and remain unchanged over a long period within the engineering area. These components are fundamental and stable in the project and are suitable for permanent modeling. The stage-related layer contains spatial unit layers of components that exist within a certain stage of the engineering area, are associated with major engineering activities, but may change over time, such as temporary facilities during construction. The transient interference layer contains spatial unit layers that appear briefly in the engineering area and interfere with the 3D mapping of the engineering scene, such as birds and vehicles. This layer is usually removed or separately identified to reduce its impact on the mapping results.

[0095] The candidate surface feature extraction unit is used to extract laser echo intensity anomaly features, neighborhood distance abrupt change features, boundary break features, texture continuity features, edge closure features, and cross-view reprojection error features from candidate surfaces in the permanent modeling layer and stage-attached layer.

[0096] Among them, the laser echo intensity anomaly feature refers to the significant difference between the echo intensity in the laser point cloud and the surrounding normal echo intensity, which may reflect special properties or anomalies of the object's surface; the neighborhood distance abrupt change feature refers to the sudden change in the distance between a point on the candidate surface and its neighboring points, which may indicate the presence of defects or special structures on the surface; the boundary breakage feature refers to the discontinuity or breakage of the candidate surface boundary, indicating that the surface boundary may have problems or be disturbed; the texture continuity feature refers to the degree of continuity of the texture of the candidate surface in the image. Surfaces with good texture continuity are usually uniform in material, while texture discontinuity may indicate changes or defects on the surface; the edge closure feature refers to whether the edges of the candidate surface are completely closed. Surfaces with closed edges are more geometrically complete, while unclosed edges may affect the accurate representation of the surface; the cross-view reprojection error feature refers to the error characteristics generated when reprojecting the candidate surface from images of different viewpoints to the same coordinate system. This feature can be used to evaluate the consistency and accuracy of the surface representation under different viewpoints.

[0097] Furthermore, the specific implementation process of the candidate surface feature extraction unit includes: establishing a K-nearest neighbor set centered on each laser point within the candidate surface, calculating the average point spacing from the center point to each neighboring point, and comparing it with the global average point spacing of the same candidate surface. When the local average point spacing exceeds the sum of the global average point spacing and a preset multiple standard deviation, it is recorded as a neighborhood distance abrupt change location, and the neighborhood distance abrupt change feature is formed by the proportion of abrupt change points and the abrupt change amplitude; projecting the candidate surface boundary points onto the main direction plane and sorting them according to the contour order, calculating the interval distance, fracture length, and fracture direction between adjacent boundary points, determining the location where the interval distance exceeds a preset multiple of the average distance between boundary points as the boundary interruption location, and determining the interruption intensity by the ratio of the fracture length to the perimeter of the candidate surface; extracting grayscale gradient direction, local binary pattern features, and texture descriptor similarity from the corresponding image blocks of the candidate surface, calculating the mean texture similarity of adjacent image blocks on both sides of the boundary, and forming texture continuity features; performing statistics on the first-to-last distance, closed curve length, and number of gaps for the boundary point set, and forming edge closure features by the ratio of the first-to-last distance to the contour perimeter and the proportion of effective closed edges.

[0098] The echo-texture conflict determination unit is used to spatially map the above-mentioned laser features and image features through the trained echo-texture conflict field determination model, construct a surface-level echo-texture conflict field, and determine whether the candidate surface belongs to at least one of the following: a credible solid surface, a penetration missing surface, a mirror pseudo-surface, an edge undersampled surface, or a pseudo-continuous surface.

[0099] Among them, spatial mapping maps and correlates laser features with image features in space, enabling analysis and comparison within the same spatial framework; surface-level echo-texture conflict field, constructed through spatial mapping, reflects the conflict between laser echo features and image texture features of candidate surfaces, visually demonstrating the distribution and degree of conflict; reliable solid surface, candidate surfaces whose laser echo features and image texture features are consistent and conform to actual engineering conditions, are reliable and can serve as the basis for 3D modeling; penetration-deficient surface, surfaces where lasers penetrate certain objects, such as transparent or thin objects, resulting in missing echoes in some areas, while corresponding features may exist in the image, are identified as penetration-deficient surfaces; mirror pseudo-surface, false surfaces generated by laser or image reflection, where there may be an abnormal correspondence between laser echo features and image texture features, are identified as mirror pseudo-surfaces; undersampled edge surface, candidate surface edge information is incomplete due to insufficient sampling during data acquisition, is identified as an undersampled edge surface; pseudo-continuous surface, surfaces that appear continuous but are actually not truly continuous due to abnormal laser echo or image features, are identified as such.

[0100] The 3D surface correction unit is used to perform boundary completion, pseudoface suppression, geometric correction and local reconstruction processing according to the surface type to generate a corrected 3D surface.

[0101] Among these, boundary completion addresses incomplete boundary issues such as undersampled edges by using algorithms and techniques to supplement missing boundary information, ensuring complete surface boundaries; pseudo-surface suppression reduces or eliminates the interference of false surfaces like mirror pseudo-surfaces, improving the accuracy of 3D modeling; geometric correction restores surfaces with positional deviations or morphological distortions to their correct spatial position and shape, ensuring the geometric accuracy of the 3D surface; and local reconstruction supplements missing information on surfaces with incomplete penetration or other missing information by reconstructing local areas, making the surface complete and consistent with reality.

[0102] The 3D model fusion output unit is used to fuse the corrected 3D surface with the layered modeling results, and outputs an engineering real-scene 3D mapping model with stage labels, material credibility labels and surface type labels.

[0103] Among them, the stage label is used to identify the stage information of different components in the engineering area, such as the construction stage and the operation stage, which helps to understand the development process and current status of the project; the material reliability label is a label assigned to the three-dimensional surface based on the analysis and judgment of the surface material, which reflects the reliability and accuracy of the surface material information; the surface type label is a label that identifies the type of the three-dimensional surface, such as a reliable solid surface, a penetrating missing surface, etc., which facilitates the classification, management and analysis of different surfaces in the model; the engineering real scene three-dimensional mapping model is a three-dimensional model that can realistically and accurately reflect the actual situation of the engineering area after the above series of processing and fusion, containing rich spatial information, semantic information and feature information.

[0104] It should be noted that during use, multi-temporal data acquisition can comprehensively acquire information about the engineering area at different times, providing a rich data foundation for subsequent analysis. Feature registration and extraction realize the fusion of laser and image data. Multi-feature extraction helps to accurately describe spatial units. Stability model calculation can quantify the stability of components and provide a scientific basis for spatial unit layering. Layered processing can specifically handle different types of spatial units, eliminate interference data, and improve modeling efficiency and accuracy. Candidate surface feature extraction comprehensively considers multiple surface characteristics. Echo texture conflict judgment accurately identifies surface types and provides a precise direction for 3D surface correction. 3D surface correction handles different types of surface problems to ensure surface quality. Finally, the 3D model fusion outputs a model with multiple labels, which facilitates the classification, management, analysis, and application of engineering real-world scenes, providing reliable 3D data support for engineering planning, construction, monitoring, etc.

[0105] In one embodiment, unified coordinate registration of the laser point cloud sequence and the image sequence includes: using high-precision exterior orientation elements acquired by the UAV-borne GNSS / IMU as the registration reference, spatial coordinate mapping registration of the three-dimensional geodetic coordinates of the laser point cloud sequence and the two-dimensional pixel coordinates of the image sequence is performed using the collinearity conditional projection transformation formula. The registration formula is as follows:

[0106]

[0107]

[0108] in,( , , ) represents the three-dimensional geodetic coordinates of the laser point cloud. , ) represents the corresponding pixel coordinates on the image, ( , , ( ) represents the three-dimensional coordinates of the camera station at the moment of UAV imaging. For the effective focal length of the imaging module, ( , The coordinates of the principal point pixel of the imaging module are shown below. After registration, the Random Sampling Consistency (RANSAC) algorithm is used to remove gross errors from the registration results, eliminating out-of-registration points caused by equipment errors and environmental interference, thus unifying the spatial coordinates of the laser point cloud and the image and ensuring the spatial consistency of the two types of data.

[0109] This design uses high-precision exterior orientation elements as a reference, employs the collinear conditional projection transformation formula to achieve coordinate mapping between the two types of data, and then uses the RANSAC algorithm to eliminate gross errors. The high-precision exterior orientation elements ensure the accuracy of the registration reference, the collinear conditional projection transformation formula can accurately realize coordinate transformation, making the laser point cloud and the image highly consistent in spatial position, and the RANSAC algorithm can effectively remove registration outliers caused by equipment errors and environmental interference, avoiding the impact of these abnormal data on subsequent analysis, improving data quality, and laying a solid foundation for subsequent accurate feature extraction, model construction and other operations, ensuring that the entire 3D mapping system can work based on accurate and consistent data.

[0110] In one embodiment, the construction, training, and parameter calculation of the component stage stability calculation model include:

[0111] Model Construction: A lightweight fully connected neural network is used to construct a component stage stability calculation model. The model network structure is: input layer - hidden layer 1 - hidden layer 2 - output layer. The number of neurons in the input layer matches the feature dimension, which is 4-dimensional, corresponding to occupancy persistence, positional drift amplitude, morphological preservation degree, and semantic category features. Hidden layer 1 has 64 neurons and uses the ReLU activation function; hidden layer 2 has 32 neurons and uses the ReLU activation function; the output layer has 1 neuron and uses the Sigmoid activation function, and the output value is the component stage stability parameter. ;

[0112] The training dataset was constructed by collecting laser point cloud and image fusion data of different engineering types and dividing them into training set, validation set and test set in a ratio of 8:1:1. The dataset was normalized to eliminate differences in feature dimensions, and Gaussian noise was added for data augmentation with a noise standard deviation of 0.02.

[0113] Model training parameters were set as follows: the optimizer was Adam, the initial learning rate was set to 0.001, the learning rate was dynamically adjusted using a cosine annealing strategy, the learning rate decay coefficient was set to 0.95, the batch size was set to 32, the number of training epochs was set to 100, the early stopping strategy was set to 10, and training was stopped when the validation set loss did not decrease for 10 consecutive epochs. The loss function was the mean squared error (MSE).

[0114] Model training and validation involves inputting the training set into the model for iterative training, validating the model's performance using the validation set after each training round, and calculating the model's coefficient of determination using the test set. ,when The model is deemed to have completed training and met performance standards at that time.

[0115] The stability parameter calculation involves inputting the extracted spatial unit features into the trained component stage stability calculation model, and outputting the component stage stability parameters. Alternatively, a weighted linear fusion formula can be used to assist in the calculation. The formula is as follows:

[0116]

[0117] in, These are the weight coefficients for persistence, positional drift, shape preservation, and semantic category, respectively, satisfying... ,and ;

[0118] The rules for determining each weight coefficient are as follows: The basic weights are determined according to the engineering survey type; for building construction projects, the weights are... Municipal road and bridge projects Then, fine-tuning is performed based on the distribution characteristics of components in the survey area, with the fine-tuning range not exceeding the basic weight. ;

[0119] Among them, the persistent The calculation formula is:

[0120]

[0121] This refers to the actual duration of existence of a spatial unit within the multi-temporal acquisition time. This represents the total time spent by the drone collecting data on the engineering area. ;

[0122] Position drift amplitude The calculation formula is:

[0123]

[0124] The geometric center coordinates of the initial acquisition frame of the spatial unit. For the first The geometric center coordinates of the frame This represents the total number of frames captured. The geometric characteristic scale of a spatial unit. ;

[0125] Degree of shape retention The calculation formula is:

[0126]

[0127] The three-dimensional volume of the initial frame of the spatial unit. For the first The three-dimensional volume of a frame ;

[0128] The stability baseline value corresponding to the semantic category of spatial unit is pre-calibrated based on the attributes of engineering components, including permanent structures such as building main body and foundation components. Stage-specific auxiliary components such as scaffolding and formwork supports Instantaneous interference such as construction vehicles and on-site personnel .

[0129] This design constructs a lightweight, fully connected neural network model, rationally sets the number of neurons and activation functions in each layer, collects and divides the dataset into multiple data types, normalizes the data and adds noise enhancement, sets appropriate training parameters, calculates parameters after training and validation, and can also be used to assist with weighted linear fusion formulas. The advantages of this design are that the lightweight network structure is simple and efficient, suitable for processing this type of data; the reasonable dataset construction and processing methods can improve the model's generalization ability; appropriate training parameter settings help the model converge quickly and achieve good performance; and the combination of multiple parameter calculation methods can accurately evaluate the stability of the component stage from different perspectives, providing a reliable basis for subsequent spatial unit division, making the division results more consistent with the actual engineering situation.

[0130] In one embodiment, dividing the test area into spatial units based on the component stage stability parameters includes:

[0131] Based on the modeling requirements of the 3D mapping of the engineering scene and the characteristics of the components in the survey area, a first stability threshold is pre-set. Second stability threshold And satisfy ;

[0132] The threshold is determined by using the statistical mean of stability parameters of typical permanent components and stage-specific auxiliary components in the survey area as a benchmark. Take 0.8 times the statistical mean of the stability of permanent components. Take 0.5 times the statistical average of the stability of the auxiliary components in the stage, and it must meet the following requirements. The range of values ​​is , The range of values ​​is It can be fine-tuned according to the surveying needs of different project types, such as building construction, municipal engineering, and road and bridge construction, with the fine-tuning range not exceeding [a certain value]. ;

[0133] When the component stage stability parameters of the space unit When this happens, the spatial unit is determined to be the core modeling object of the project and is divided into a permanent modeling layer;

[0134] when When the space unit is determined to be a necessary ancillary object in the construction phase, it is divided into a phase ancillary layer.

[0135] when When the spatial unit is determined to be a temporary interference object at the engineering site, it is classified as an instantaneous interference layer.

[0136] This design pre-sets two stability thresholds and determines the rules based on the statistical mean of stability parameters of typical components, combined with the range of values. Spatial units are divided according to the relationship between stability parameters and thresholds. The thresholds are set based on the statistical mean of stability parameters of typical components, making the division standard scientific and reasonable. It can accurately distinguish components with different stability levels. The clear value range and fine-tuning rules can be flexibly adjusted according to the characteristics of different engineering types, enhancing the adaptability of the division method. Through this division method, the core modeling objects, necessary auxiliary objects, and temporary interference objects of the project can be clearly defined, which facilitates the subsequent targeted processing of different types of spatial units and improves the efficiency and accuracy of 3D mapping.

[0137] In one embodiment, extracting laser echo intensity anomaly features and cross-view reprojection error features includes: the laser echo intensity anomaly features are calculated based on the echo intensity difference between the target laser point and neighboring laser points, and the calculation formula is as follows:

[0138]

[0139] in, The actual echo intensity value of the target laser point. The mean echo intensity of all valid laser points in the neighborhood of the target laser point is used. The neighborhood range is adaptively set to 5-20 neighborhood points based on the point density of the laser point cloud.

[0140] Echo intensity anomaly threshold The rules for determining it are as follows:

[0141] Based on the coefficient of variation of laser echo intensity for common materials in the test area, such as concrete and brick, Take three times the coefficient of variation of echo intensity for conventional materials, and the value range is [missing value]. When the test area contains special materials, such as glass, metal, or coated templates, accounting for more than [a certain percentage], hour, Increase by 0.3-0.5;

[0142] when When the laser point is identified as an echo intensity anomaly, the corresponding surface area is found to have an echo anomaly.

[0143] The cross-view reprojection error feature is calculated based on the pixel distance between the actual and theoretical projections of candidate surface feature points in multi-view images. It reflects the spatial matching degree between 3D feature points and image texture. The calculation formula is as follows:

[0144]

[0145] in, To determine the effective number of viewing angles for reprojection, the overlap between viewing angles must be no less than [value missing]. , For candidate surface feature points in the th The actual pixel coordinates on the image from each viewpoint The feature points are projected from three-dimensional coordinates to the first feature point using the projection transformation formula. Theoretical pixel coordinates on a viewpoint image A larger value indicates a worse consistency of cross-view projection of feature points.

[0146] This design employs two methods: First, the laser echo intensity anomaly feature is calculated based on the difference in echo intensity between the target and neighboring points, thus determining the anomaly threshold. Second, the cross-view reprojection error feature is calculated based on the projection pixel distance of candidate surface feature points in multi-view images. For the laser echo intensity anomaly feature, the difference is calculated by comparing it with neighboring points, accurately capturing anomalies in the target point's echo intensity. The determined anomaly threshold considers both conventional and special materials, making the judgment more reasonable. The cross-view reprojection error feature reflects the spatial matching degree between 3D feature points and image texture. Through multi-view projection calculation, the projection consistency of feature points can be comprehensively evaluated, and the numerical value intuitively reflects the degree of matching. These two feature extraction methods reflect surface features from different angles, providing rich and accurate feature information for subsequent echo-texture conflict field determination.

[0147] In one embodiment, the construction, training, and surface type determination of the echo-texture conflict field determination model include:

[0148] The model is constructed using a convolutional neural network (CNN) + fully connected network (FC) to build an echo-texture conflict field determination model. The CNN part consists of three convolutional blocks used to extract spatial correlation information of features. Each convolutional block contains a convolutional layer, a BatchNorm layer, a ReLU activation function, and a max pooling layer. The kernel size is set to 3×3, the stride is set to 1, the padding method is Same, and the number of output channels is 16, 32, and 64 respectively. The FC part consists of hidden layer 1, hidden layer 2, and an output layer. Hidden layer 1 has 128 neurons, hidden layer 2 has 64 neurons, both using the ReLU activation function, and the output layer has 5 neurons using the Softmax activation function, corresponding to the probability distribution of 5 surface types.

[0149] Training dataset construction: Laser-image fusion feature data of different surface types in engineering scenarios were collected and labeled into five categories: reliable solid surfaces, penetration-deficient surfaces, mirror pseudo-surfaces, edge-undersampled surfaces, and pseudo-continuous surfaces. These were divided into training, validation, and test sets in a 7:2:1 ratio. The feature data were then normalized to... For each interval, random pruning and feature replacement were used to augment the data, and the sample size of the dataset was increased to twice the original sample size after augmentation.

[0150] Model training parameter settings: The optimizer used is AdamW, the weight decay coefficient is set to 0.0001, the initial learning rate is set to 0.0005, a step decay strategy is adopted, the learning rate is decayed to 0.1 every 20 rounds, the batch size is set to 16, the number of training rounds is set to 80, the early stopping strategy is set to 8, training is stopped when the validation set loss does not decrease for 8 consecutive rounds, and the loss function is cross-entropy loss.

[0151] Model training and validation involves inputting the training set into the model for iterative training. After each training round, the model performance is validated using a validation set. Precision, recall, and F1 score are calculated using a test set. When the F1 score for each class is... At that time, it is determined that the model training is complete and the performance meets the target;

[0152] Collision field construction and surface determination, normalized laser echo intensity anomaly characteristics Neighborhood distance abrupt change characteristics Boundary fracture characteristics Texture continuity features Edge closure features and cross-view reprojection error characteristics The trained model is first processed by a CNN layer to extract features related to the feature space, then by an FC layer to output the probabilities of each surface type, and simultaneously calculates the conflict values ​​of the conflict field. The formula is:

[0153]

[0154] in, For each feature, the normalized value is... These are the weighting coefficients for each feature. The weights of laser-related features are as follows: The image feature weights are, in order, satisfied... ;

[0155] The rules for determining each weighting coefficient are as follows:

[0156] Based on the quality of laser point cloud data and image data, if the laser point cloud point density And image resolution The total weight of laser-related features is set to 0.4, and the total weight of image-related features is set to 0.6. If the laser point cloud density... or image resolution The total weight for laser-related features is set at 0.6, and the total weight for image-related features is set at 0.4. Features within the same category are weighted evenly, and this weight can be fine-tuned based on the anomaly rate of the survey area, with the adjustment range for a single feature weight not exceeding [a certain value]. ;

[0157] Preset the four-level conflict threshold ,satisfy And all belong to ;

[0158] The threshold is determined as follows: The conflict values ​​of the calibration samples in the test area are sorted in ascending order, and then determined according to the 20%, 40%, 60%, and 80% quantiles. The initial value is then corrected according to the accuracy requirements of engineering surveying, achieving high-precision surveying and planar accuracy. Adjust the initial value down by 0.05-0.1 for standard accuracy surveying, including planar accuracy. Keep the initial value unchanged;

[0159] Combining the probability distribution of the model output with the conflict value The range of values ​​determines the surface type:

[0160] when When this occurs, it is determined to be a trusted entity surface;

[0161] when When this occurs, it is determined to be an undersampled edge surface;

[0162] when When the surface is in its pseudo-continuous state, it is determined to be a pseudo-continuous surface.

[0163] when At that time, it was determined to be a penetration of the missing measurement surface;

[0164] when When the time is right, it is determined to be a mirror pseudo-plane.

[0165] This design constructs a CNN+FC model, rationally sets the structure and parameters of each layer, collects data from multiple categories to build a dataset, normalizes and enhances the data, sets appropriate training parameters, and after training and validation, calculates conflict values ​​based on input features and combines them with the model output to determine the surface type. The CNN part can effectively extract the spatial correlation information of features, while the FC part further processes the output probability distribution. The combination of the two allows for comprehensive feature analysis. The reasonable dataset construction and processing method improves the model's ability to recognize different surface types. Appropriate training parameter settings help the model converge quickly and achieve good performance. By calculating conflict values ​​and combining them with the model output to determine the surface type, multiple factors are comprehensively considered, making the determination results more accurate and reliable. It can clearly distinguish different types of surfaces and provide an accurate basis for subsequent targeted correction processing.

[0166] In one embodiment, the correction process performed according to the surface type includes:

[0167] For different types of surface defects, differentiated correction and reconstruction processing is performed by combining the geometric features of the laser point cloud with the texture features of the image. Specifically:

[0168] For the missing measurement surface, the neighborhood geometric interpolation method is used to complete the point cloud of the missing area. The formula for calculating the three-dimensional coordinates of the completed points is as follows:

[0169]

[0170] in, To determine the number of effective laser points in the neighborhood of the missing measurement area, the following requirements are specified. To ensure interpolation accuracy, The three-dimensional coordinates of the effective points in the neighborhood. The unit normal vector of the plane fitted to the effective points in the neighborhood. To complete the perpendicular distance from the point to the fitted plane, The threshold value is determined by taking the average point spacing of the neighboring point cloud of the missing measurement area, and the value range is [value range missing]. Fine-tuning was performed based on the accuracy requirements of the engineering surveying.

[0171] For mirror pseudo-facets, high-abnormal echo points generated by virtual reflections are screened out by echo intensity threshold, and pseudo-facet point clouds are removed by combining geometric normal vector consistency analysis to achieve pseudo-facet suppression and ensure the geometric authenticity of the surface.

[0172] For undersampled surfaces at the edges, a contour extraction algorithm is used to extract the incomplete contour of the surface, and the contour is completed by cubic spline interpolation. Then, the edge missing point cloud is generated by combining the neighborhood geometric features to complete the boundary.

[0173] For pseudo-continuous surfaces, geometrically distorted regions are screened out based on the cross-view reprojection error threshold. The rule for determining this threshold is: take twice the image pixel resolution. When the reprojection error exceeds this threshold, it is judged as a distorted region. Combining the neighborhood geometric continuity of the laser point cloud with the texture consistency of the image, a local surface fitting method is used to perform geometric correction, correct the three-dimensional coordinates of the distorted region, and restore the true geometric relationship of the surface.

[0174] This design, addressing various surface defects, combines the geometric features of laser point clouds and image textures. For penetrating surfaces with missing measurements, neighborhood geometric interpolation is used for completion. Mirror-like pseudo-facets are suppressed through echo intensity thresholding and geometric normal vector analysis. Edge-undersampled surfaces are completed using contour extraction and interpolation. For pseudo-continuous surfaces, distorted areas are filtered and geometrically corrected based on reprojection errors. Specialized correction methods are employed for different surface defects, accurately resolving various problems. By combining the geometric features of laser point clouds and image textures, the advantages of both data types are fully utilized, making the correction and reconstruction more realistic. For example, penetrating surfaces with missing measurements are completed using neighborhood geometric information, ensuring geometric rationality; mirror-like pseudo-facets are suppressed through echo intensity and geometric angles, ensuring surface authenticity. These methods effectively improve surface quality, providing a guarantee for generating accurate 3D models.

[0175] In one embodiment, the output 3D mapping model of the engineering scene has multi-level semantic tags, which correspond one-to-one with the layering results and surface determination results in the mapping process, specifically including:

[0176] Stage labels, based on the layering results of spatial units, are divided into permanent modeling layers and stage-specific layers, which intuitively reflect the engineering construction stage attributes of model components.

[0177] Material trust labels are based on the conflict values ​​of the echo-texture conflict field. The credibility level is divided into credibility levels, and the credibility level threshold is the same as mentioned above. To remain consistent, among which Marked as high confidence, this indicates that the material in this area has no obvious laser echo or texture anomaly; The labeling as medium confidence indicates that there are slight feature anomalies in the area, which can be corrected to meet the mapping accuracy requirements; Marked as low confidence, this area is made of special material and has obvious penetration or virtual reflection phenomena, which need to be checked carefully;

[0178] Surface type labels are assigned based on the determination results of candidate surfaces, and are divided into reliable solid surfaces, penetrating missing surfaces, mirror pseudo-surfaces, edge undersampled surfaces, and pseudo-continuous surfaces, clearly reflecting the geometric and texture features of the model surface.

[0179] This design incorporates multi-level semantic tags for the model, including stage tags, material reliability tags, and surface type tags. These tags correspond to the hierarchical mapping process and surface assessment results. The multi-level semantic tags provide detailed annotations of the model from different dimensions, making the model information richer and more comprehensive. Stage tags intuitively reflect the attributes of the component's construction stage, facilitating the analysis and management of different stages of the project. Material reliability tags categorize reliability levels, allowing users to quickly understand the material conditions of each area and focus on low-reliability areas. Surface type tags clearly present the surface's geometric and textural features, aiding in understanding the model's surface condition. This tagging information facilitates the application of 3D mapping models for engineering real-world scenarios, enhancing the model's practicality and value in fields such as engineering planning, construction, and monitoring.

[0180] Example 2:

[0181] Please see Figure 4 A method for fusion of UAV LiDAR and AI in engineering real-world 3D mapping, applied to the aforementioned UAV LiDAR fusion engineering real-world 3D mapping system, includes the following steps:

[0182] S1. Perform multi-temporal and multi-view joint acquisition on the engineering area to obtain laser point cloud sequences, image sequences and corresponding time marker information;

[0183] S2. Perform unified coordinate registration on the laser point cloud sequence and the image sequence, and extract the geometric features, texture features, temporal continuity features and component semantic features of the spatial units;

[0184] S3. Calculate the component stage stability parameters of each spatial unit based on occupancy persistence, positional drift amplitude, morphological preservation degree, and semantic category;

[0185] S4. Based on the component stage stability parameters, the spatial unit of the test area is divided into a permanent modeling layer, a stage auxiliary layer, and a transient interference layer.

[0186] S5. Extract laser echo intensity anomaly features, neighborhood distance abrupt change features, boundary break features, texture continuity features, edge closure features, and cross-view reprojection error features from candidate surfaces in the permanent modeling layer and stage-attached layer.

[0187] S6. Construct a surface-level echo-texture conflict field and output the type determination result of the candidate surface;

[0188] S7. Perform boundary completion, pseudo-facet suppression, geometric correction, and local reconstruction based on the surface type determination results;

[0189] S8. Integrate the corrected 3D surface and layered results to output a 3D engineering real-scene mapping model with stage labels, material confidence labels, and surface type labels.

[0190] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0191] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A UAV-based LiDAR fusion engineering real-scene 3D mapping system, characterized in that, The system includes: The system includes a multi-temporal data acquisition unit, a feature registration and extraction unit, a stability model calculation unit, a spatial unit layering unit, a candidate surface feature extraction unit, an echo texture conflict determination unit, a three-dimensional surface correction unit, and a three-dimensional model fusion output unit. The multi-temporal data acquisition unit is used to perform multi-temporal and multi-view joint acquisition of the engineering area to obtain laser point cloud sequences, image sequences and corresponding time marker information; The feature registration and extraction unit is used to perform unified coordinate registration on the laser point cloud sequence and the image sequence, and to extract the geometric features, texture features, temporal continuity features and component semantic features of the spatial units. The stability model calculation unit is used to calculate the component stage stability parameters of each spatial unit based on occupancy persistence, position drift amplitude, morphological preservation degree and semantic category. The spatial unit layering unit is used to divide the spatial unit of the test area into a permanent modeling layer, a stage auxiliary layer and a transient interference layer according to the component stage stability parameters. The candidate surface feature extraction unit is used to extract laser echo intensity anomaly features, neighborhood distance abrupt change features, boundary break features, texture continuity features, edge closure features, and cross-view reprojection error features from candidate surfaces in the permanent modeling layer and the stage auxiliary layer. The echo-texture conflict determination unit is used to construct a surface-level echo-texture conflict field and output the type determination result of the candidate surface; the echo-texture conflict determination unit includes: a conflict model construction module, a feature fusion module, a conflict value calculation module, and a surface classification module; The conflict model construction module is used to establish an echo-texture conflict field determination model for candidate surfaces; The feature fusion module is used to perform spatial mapping, dimensional unification, and correlation fusion of laser-related features and image features; The conflict value calculation module is used to output the surface conflict value based on the normalization results and corresponding weights of each feature; The surface classification module is used to combine the surface conflict value and the surface category probability distribution to determine the candidate surface as a reliable solid surface, a surface with missing penetration measurement, a mirror pseudo-surface, an edge undersampled surface, or a pseudo-continuous surface. The three-dimensional surface correction unit is used to perform boundary completion, pseudo-face suppression, geometric correction and local reconstruction processing based on the surface type determination result. The 3D model fusion output unit is used to fuse the corrected 3D surface and layering results, and output an engineering real-scene 3D mapping model with stage labels, material credibility labels and surface type labels.

2. The UAV-based LiDAR fusion engineering real-scene 3D mapping system according to claim 1, characterized in that, The feature registration and extraction unit includes: The module includes an exterior orientation acquisition module, a coordinate mapping module, a gross error removal module, and a joint feature generation module. The external orientation acquisition module is used to acquire the pose parameters and imaging parameters of the UAV platform; The coordinate mapping module is used to establish the projection mapping relationship between the three-dimensional coordinates of the laser point cloud and the image pixel coordinates; The gross error removal module is used to remove out-of-registration points caused by attitude disturbances, occlusion changes, and acquisition noise. The joint feature generation module is used to generate geometric contour descriptions, surface texture descriptions, temporal variation descriptions, and semantic attribute descriptions of spatial units under a unified coordinate reference. The joint feature generation module is also used to calculate the geometric contour description, surface texture description, temporal change description, and semantic attribute description based on the boundary point set of the spatial unit, the projected image block, the multi-temporal center coordinates, and the component category recognition results.

3. The UAV-based LiDAR fusion engineering real-scene 3D mapping system according to claim 1, characterized in that, The stability model calculation unit includes: The system includes a stability modeling module, a sample training module, a parameter inference module, and a weight correction module. The stability modeling module is used to establish a component stage stability calculation model with occupancy persistence, position drift amplitude, shape preservation degree and semantic category as inputs; The sample training module is used to perform normalization training on fused samples from different engineering scenarios. The parameter reasoning module is used to output the component stage stability parameters of each spatial unit; The weight correction module is used to correct the contribution weight of each input factor based on the project type, component distribution density, and data acquisition sequence integrity.

4. The UAV-based LiDAR fusion engineering real-scene 3D mapping system according to claim 2, characterized in that, The spatial unit layered unit includes: Threshold setting module, hierarchy determination module, and interference handling module; The threshold setting module is used to set a first stability threshold and a second stability threshold based on the stability statistics of the component. The hierarchical determination module is used to classify a spatial unit as a permanent modeling layer when the stability parameter of the component stage is higher than the first stability threshold, as a stage-attached layer when the stability parameter of the component stage is between the second stability threshold and the first stability threshold, and as a transient interference layer when the stability parameter of the component stage is lower than the second stability threshold. The interference handling module is used to remove or independently identify transient interference layers.

5. The UAV-based LiDAR fusion engineering real-scene 3D mapping system according to claim 3, characterized in that, The candidate surface feature extraction unit includes: The module includes: echo anomaly extraction module, distance abrupt change extraction module, boundary break extraction module, texture continuity extraction module, edge closure extraction module, and reprojection error extraction module. The echo anomaly extraction module is used to determine the degree of echo anomaly based on the difference in echo intensity between the target laser point and neighboring laser points. The distance mutation extraction module is used to characterize the discontinuous changes in the spacing between local points on the candidate surface; The boundary fracture extraction module is used to characterize the interruption location and interruption intensity of the candidate surface profile; The texture continuity extraction module is used to characterize the continuity of the texture of a candidate surface on both sides of the boundary based on the texture direction, grayscale gradient and texture similarity of the candidate surface in adjacent image blocks. The edge closure extraction module is used to characterize the degree of closure integrity of the candidate surface edge contour based on the start-end distance of the candidate surface edge points, the change in contour curvature, and the closure ratio. The reprojection error extraction module is used to characterize the degree of deviation between the theoretical projection position and the actual projection position of candidate surface feature points in multi-view images.

6. The UAV-based LiDAR fusion engineering real-scene 3D mapping system according to claim 5, characterized in that, The three-dimensional surface correction unit includes: Missing measurement completion module, pseudo-surface suppression module, boundary completion module, and geometric correction module; The missing measurement completion module is used to perform point cloud completion for the penetrating missing measurement surface based on the spatial distribution of effective points in the neighborhood; The pseudoface suppression module is used to remove pseudoface points for mirrored pseudofaces based on abnormal echo characteristics and normal consistency. The boundary completion module is used to perform boundary restoration for undersampled surfaces based on contour extraction results and neighborhood geometric constraints. The geometric correction module is used to correct the three-dimensional coordinates of the distorted region on the pseudo-continuous surface based on the cross-view reprojection deviation and the local surface fitting results.

7. The UAV-based LiDAR fusion engineering real-scene 3D mapping system according to claim 4, characterized in that, The 3D model fusion output unit includes: The module includes a hierarchical fusion module, a tag generation module, and a results output module. The layered fusion module is used to associate and fuse the corrected 3D surface with the spatial units corresponding to the permanent modeling layer and the stage auxiliary layer; The label generation module is used to generate stage labels, material confidence labels, and surface type labels for the fused model. The stage label is used to characterize the modeling level to which the spatial unit belongs, the material confidence label is used to characterize the confidence level of the candidate surface, and the surface type label is used to characterize the specific type of the candidate surface. The output module is used to output a three-dimensional mapping model of the engineering scene with multi-level semantic relationships.

8. The UAV-based LiDAR fusion engineering real-scene 3D mapping system according to claim 6, characterized in that, It also includes a surveying result verification unit: The mapping result verification unit is used to perform component stage consistency verification, surface geometric authenticity verification, and label correspondence verification on the output model based on the unified coordinate benchmark, spatial unit layering results, surface type determination results, and three-dimensional surface correction results. The mapping result verification unit is also used to verify the boundary continuity between the permanent modeling layer and the stage sub-layer, the matching relationship between the material credibility label and the surface type label, and the semantic consistency between multi-temporal mapping results, and outputs the verified model as the engineering real scene 3D mapping result.

9. A method for 3D engineering scene mapping using UAV lidar fusion, characterized in that, The method, applied to the engineering real-scene 3D mapping system for implementing UAV lidar fusion as described in any one of claims 1-8, comprises: S1. Perform multi-temporal and multi-view joint acquisition on the engineering area to obtain laser point cloud sequences, image sequences and corresponding time marker information; S2. Perform unified coordinate registration on the laser point cloud sequence and the image sequence, and extract the geometric features, texture features, temporal continuity features and component semantic features of the spatial units; S3. Calculate the component stage stability parameters of each spatial unit based on occupancy persistence, positional drift amplitude, morphological preservation degree, and semantic category; S4. Based on the component stage stability parameters, the spatial unit of the test area is divided into a permanent modeling layer, a stage auxiliary layer, and a transient interference layer. S5. Extract laser echo intensity anomaly features, neighborhood distance abrupt change features, boundary break features, texture continuity features, edge closure features, and cross-view reprojection error features from candidate surfaces in the permanent modeling layer and stage-attached layer. S6. Construct a surface-level echo-texture conflict field and output the type determination result of the candidate surface; S7. Perform boundary completion, pseudo-facet suppression, geometric correction, and local reconstruction based on the surface type determination results; S8. Integrate the corrected 3D surface and layered results to output a 3D engineering real-scene mapping model with stage labels, material confidence labels, and surface type labels.

Images