Low-altitude unmanned aerial vehicle flight building height rapid measurement method

Abstract

The present application relates to the technical field of geographic information data processing, and more particularly to a low-altitude unmanned aerial vehicle flight building height rapid measurement method, comprising obtaining satellite stereo image data and ground control point data in a task area, performing aerial triangulation based on the data to obtain measurement results; then generating dense point cloud data of the work area according to the measurement results; then extracting features from the satellite stereo image data and the dense point cloud data to obtain feature results, and automatically classifying the point cloud data by using a predefined deep learning model; then extracting ground point layer height and building point layer height information from the point cloud data according to the building contour and roof structure characteristics; and finally determining the building height according to the building point layer height information and the ground point layer height information. The present application combines multi-source data and intelligent algorithms to achieve efficient and accurate building height measurement and provide key data support for low-altitude unmanned aerial vehicle flight.

Description

Technical Field

[0001] This invention relates to the field of geographic information data processing technology, and in particular to a method for rapid measurement of building height using a low-altitude unmanned aerial vehicle (UAV). Background Technology

[0002] With the continuous development of drone technology, computer technology, space technology, geographic information technology, and society, the low-altitude economy has become an important driving force for future economic development, providing efficient and convenient services for people's lives. The rapid, accurate, and efficient acquisition of building height information data provides crucial foundational support for drone low-altitude flight safety and flight planning.

[0003] Traditional methods for obtaining building height suffer from various problems such as high workload, low efficiency, and low accuracy. A fast, efficient, and accurate method for obtaining building height can greatly reduce production costs and increase acquisition speed, so it needs to be explored and improved. Summary of the Invention

[0004] Therefore, the present invention provides a method for rapid measurement of building height by low-altitude unmanned aerial vehicle (UAV) flight, in order to solve the aforementioned problems existing in the prior art.

[0005] To achieve the above objectives, the present invention provides a method for rapid measurement of building height during low-altitude unmanned aerial vehicle (UAV) flight, comprising:

[0006] Step S1: Obtain satellite stereo image data and ground control point data within the mission area; perform aerial triangulation based on the satellite stereo image data and the ground control point data to obtain the measurement results.

[0007] Step S2: Generate dense point cloud data of the work area based on the measurement results;

[0008] Step S3: Extract features from the satellite stereo image data and the dense point cloud data to obtain feature results; fuse the feature results according to a predefined deep learning model and automatically classify them into point cloud data.

[0009] Step S4: Extract ground point layer height information and building point layer height information from the point cloud data based on the building outline and roof structure characteristics;

[0010] Step S5: Determine the building height based on the building point floor height information and the ground point floor height information.

[0011] Furthermore, the process of step S1 includes:

[0012] Acquire satellite stereo image data within the mission area, including stereo image pairs from multiple angles, and collect ground control point data for the mission area;

[0013] Feature points are extracted from each satellite image and matched to establish geometric relationships between adjacent images;

[0014] The extracted image features are combined with ground control point data to perform aerial triangulation to obtain the measurement results.

[0015] Furthermore, the process of combining the extracted image features with ground control point data to perform aerial triangulation to obtain the measurement results includes:

[0016] The correspondence between ground control point data and feature points extracted from satellite stereo imagery is established by matching corresponding points;

[0017] Based on the correspondence, the initial values ​​of the interior and exterior orientation elements of the image are constructed using the collinearity equation, and a regional network adjustment model is established to form an initial triangulation model.

[0018] The bundle adjustment algorithm is used to iteratively solve the image exterior orientation elements and the three-dimensional coordinates of ground points by combining the orbital parameters, attitude parameters and ground control point coordinates of the satellite stereo image. This optimizes the initial triangulation model and obtains the optimized result.

[0019] The optimization results are verified by weighted adjustment of the error equation to obtain the measurement results.

[0020] Furthermore, the process of step S2 includes:

[0021] Preprocessing of satellite imagery, including radiometric and geometric correction, is performed to correct image distortion.

[0022] Construct a multi-angle image matching volume based on the preprocessed image;

[0023] An initial sparse point cloud is generated from the matched cost volume using a disparity estimation algorithm;

[0024] The initial sparse point cloud is optimized and interpolated to generate a high-density point cloud;

[0025] The high-density point cloud is filtered to generate the dense point cloud data.

[0026] Furthermore, the process of generating an initial sparse point cloud from the matched cost volume using a disparity estimation algorithm includes:

[0027] A reasonable range of parallax values ​​is defined based on the baseline distance and imaging geometry of multi-angle images;

[0028] For each pixel and its candidate disparity values, the cost is calculated using the squared difference cost function, and a cost volume is constructed.

[0029] Local optimization of the cost volume yields a preliminary disparity map;

[0030] Based on the preliminary disparity map and the camera's interior and exterior orientation elements, the disparity values ​​of pixels are converted into three-dimensional coordinates using collinearity equations to generate an initial sparse point cloud.

[0031] Furthermore, the process of step S3 includes:

[0032] Extract spectral and texture features from satellite stereo imagery data, and extract geometric features from dense point cloud data;

[0033] Using a predefined deep learning model, the extracted spectral features, texture features, and geometric features are fused to obtain the fusion result;

[0034] The fusion results are automatically classified using a trained deep learning model to generate classified point cloud data.

[0035] Furthermore, the process of automatically classifying the fusion result using a trained deep learning model to generate classified point cloud data includes:

[0036] The fusion result is preprocessed and formatted to obtain a formatted result;

[0037] The pre-trained deep learning model is used to perform multi-layer feature extraction and encoding on the formatted input results;

[0038] In the output layer of the model, the extracted features are mapped to the classification category through a fully connected layer to generate the classification probability of each point cloud point.

[0039] The category label of each point cloud point is determined based on the classification probability, thereby generating classified point cloud data.

[0040] Furthermore, the process of step S4 includes:

[0041] The point cloud data of the building area and its surrounding buffer zone are filtered out from the point cloud data based on the building outline;

[0042] Cluster analysis and geometric feature extraction are performed on the filtered point clouds to determine the structural type of the building roof;

[0043] Select the appropriate height extraction method based on the roof structure type;

[0044] Based on the height information of the ground point layer and the building point layer, the height extraction method is used to extract reasonable ground point layer heights and building point layer heights respectively.

[0045] Furthermore, the process of performing cluster analysis and geometric feature extraction on the filtered point cloud to determine the structural type of the building roof includes:

[0046] The filtered point cloud data is then denoised and smoothed to obtain the processing results.

[0047] The processing results are initially clustered using the Euclidean clustering method, dividing the point cloud into multiple subsets that may contain building rooftops;

[0048] The clustering results are further refined in each subset based on the geometric similarity between points to ensure that each cluster corresponds to a continuous area of ​​the roof.

[0049] Geometric features are extracted for each cluster, and the extracted geometric features are input into the trained classification model to determine the roof structure type corresponding to the cluster.

[0050] Furthermore, the process of selecting the corresponding height extraction method based on the roof structure type includes:

[0051] Identify the roof structure type of a building, including flat roofs, pitched roofs, ridged roofs, or complex curved roofs;

[0052] For flat roofs, the average elevation of the point cloud within the building area is used as the building's point layer height.

[0053] For pitched roofs, fit the roof plane and extract the height difference between the highest and lowest points in the roof area as the building's floor height.

[0054] For a ridged roof, the direction of the roof ridge is fitted by principal component analysis, and the heights of the highest and lowest points on both sides of the ridge are extracted. The average or maximum value is taken as the building floor height.

[0055] For complex curved roofs, the main structure is identified based on the curvature characteristics of the point cloud, and the weighted average of the local elevations is calculated as the building's point floor height.

[0056] Compared with existing technologies, the advantages of this invention lie in its ability to generate high-precision dense point cloud data by fusing satellite stereo imagery data and ground control point data for aerial triangulation. This data is then combined with a deep learning model to automatically classify the point cloud data, accurately extracting the ground and building point layer height information, and ultimately determining the building height. Each step is tightly integrated, leveraging the multi-angle observation advantages of satellite imagery and the precise geographic information of ground control points to ensure the accuracy of the measurement results. The introduction of the deep learning model improves data processing efficiency and classification accuracy, providing a reliable basis for subsequent height extraction. The entire method is highly efficient and automated, suitable for large-scale building height measurement tasks, effectively reducing labor and time costs, and providing crucial building height data support for low-altitude, economical UAV flights.

[0057] In particular, by combining feature point matching and corresponding point fusion with collinearity equations and regional network adjustment models, and using bundle adjustment algorithms for iterative optimization, the accuracy and reliability of aerial triangulation are significantly improved. The automated processing reduces manual operations, increases measurement efficiency, and reduces human error. It is applicable to different types of buildings and complex roof structures, and can quickly adapt to various measurement task requirements.

[0058] In particular, the initial sparse point cloud is optimized by removing noise and smoothing; high-density point clouds are generated through interpolation to fill in missing data, improving point cloud density and detail, and enhancing the integrity of building outlines and structures. Filtering generates dense point cloud data for further noise reduction, improving point cloud uniformity and accuracy.

[0059] In particular, by integrating spectral, texture, and geometric features, the characteristics of building point clouds can be comprehensively captured, improving classification accuracy and reducing misclassification. The deep learning model, trained on extensive data, adapts to different types of buildings and complex scenes, exhibiting strong generalization capabilities. The entire process from feature extraction to classification is automated, reducing manual intervention and improving measurement efficiency.

[0060] In particular, by combining cluster analysis and geometric feature extraction with a classification model, the type of roof structure can be accurately determined, thereby selecting an appropriate method to extract the building height and improving the accuracy and reliability of the measurement. Attached Figure Description

[0061] Figure 1 A flowchart illustrating a method for rapid measurement of building height using a low-altitude unmanned aerial vehicle (UAV) provided by the present invention;

[0062] Figure 2 This is a flowchart illustrating step S1 in a method for rapid measurement of building height using a low-altitude unmanned aerial vehicle (UAV) provided by the present invention.

[0063] Figure 3 This is a flowchart illustrating step S2 in a method for rapid measurement of building height using a low-altitude unmanned aerial vehicle (UAV) provided by the present invention.

[0064] Figure 4 This is a flowchart illustrating step S3 in a method for rapid measurement of building height using a low-altitude unmanned aerial vehicle (UAV) provided by the present invention. Detailed Implementation

[0065] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0066] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0067] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0068] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0069] Please see Figure 1 As shown, the present invention provides a method for rapid measurement of building height during low-altitude unmanned aerial vehicle (UAV) flight, comprising:

[0070] Step S1: Obtain satellite stereo image data and ground control point data within the mission area; perform aerial triangulation based on the satellite stereo image data and the ground control point data to obtain the measurement results.

[0071] Specifically, such as Figure 2 As shown, the process of step S1 includes:

[0072] Step S11: Acquire satellite stereo image data within the mission area, including stereo image pairs from multiple angles, and collect ground control point data for the mission area;

[0073] Specifically, this involves acquiring satellite stereo imagery data within the mission area. This step requires obtaining high-resolution, multi-band satellite imagery from reliable data sources, ensuring that the imagery coverage and timestamps meet the measurement requirements. Simultaneously, ground control point (GCP) data is collected. These points, obtained through field measurements or extracted from high-precision maps, are evenly distributed throughout the mission area, providing crucial geographic references for subsequent measurements.

[0074] Step S12: Extract feature points from each satellite image and match them to establish geometric relationships between adjacent images;

[0075] Specifically, feature points are extracted from each preprocessed satellite image, such as key points detected using algorithms like SIFT and SURF, including corner points, edges, and blob features. These extracted feature points carry rich local information, providing a foundation for subsequent matching. Feature point matching is then performed between adjacent images to find point pairs with the same or similar features, establishing geometric relationships between the images.

[0076] Step S13: Combine the extracted image features with ground control point data to perform aerial triangulation to obtain the measurement results.

[0077] Specifically, the process of combining the extracted image features with ground control point data to perform aerial triangulation to obtain the measurement results includes:

[0078] The correspondence between ground control point data and feature points extracted from satellite stereo imagery is established by matching corresponding points;

[0079] Specifically, algorithms such as SIFT and SURF are used to extract feature points from the preprocessed images. These points are stable under changes in illumination and viewpoint. Feature point matching is performed between adjacent images to find corresponding point pairs and establish geometric relationships between the images. False matches are removed and reliable corresponding point pairs are retained by filtering based on distance ratios and geometric constraints.

[0080] Based on the correspondence, the initial values ​​of the interior and exterior orientation elements of the image are constructed using the collinearity equation, and a regional network adjustment model is established to form an initial triangulation model.

[0081] Specifically, aerial triangulation is performed by combining extracted image features with GCP data. Correspondence between control points and image points is established through homologous point matching, and initial values ​​for the interior and exterior orientation elements of the image are constructed using collinearity equations. Collinearity equations describe the geometric collinearity relationship between ground features, image points, and the photography center. A regional network adjustment model is established, and the collinearity equations of multiple images are combined to form a mathematical model that describes the regional geometric relationships and observation data constraints.

[0082] The bundle adjustment algorithm is used to iteratively solve the image exterior orientation elements and the three-dimensional coordinates of ground points by combining the orbital parameters, attitude parameters and ground control point coordinates of the satellite stereo image. This optimizes the initial triangulation model and obtains the optimized result.

[0083] Specifically, a bundle adjustment algorithm is used for iterative optimization. This algorithm combines the orbital and attitude parameters of satellite imagery with the coordinate information of ground points (GCPs), and iteratively adjusts the exterior orientation elements of the imagery and the three-dimensional coordinates of ground points through calculations. This process continuously reduces the residual between the model's predicted values ​​and the observed values, thereby optimizing the initial triangulation model and obtaining more accurate parameter estimates.

[0084] The optimization results are verified by weighted adjustment of the error equation to obtain the measurement results.

[0085] Specifically, the optimization results are verified through weighted adjustment using the error equation. The error equation describes the difference between the observed and theoretical values, and the accuracy and reliability of the results are evaluated through weighted adjustment calculations. If the accuracy requirements are met, the measurement results are reliable; otherwise, the adjustment model needs to be adjusted and the calculation re-iterated until the accuracy standard is met.

[0086] Specifically, by using feature point matching and corresponding point fusion, combined with collinearity equations and regional network adjustment models, and employing bundle adjustment algorithms for iterative optimization, the accuracy and reliability of aerial triangulation are significantly improved. The automated processing reduces manual operations, increases measurement efficiency, and reduces human error. It is applicable to different types of buildings and complex roof structures, and can quickly adapt to various measurement task requirements.

[0087] Step S2: Generate dense point cloud data of the work area based on the measurement results;

[0088] Specifically, such as Figure 3 As shown, the process of step S2 includes:

[0089] Step S21: Preprocess the satellite imagery, including radiometric and geometric correction, to correct image distortion.

[0090] Specifically, by utilizing the sensor's radiometric calibration parameters and atmospheric radiative transfer model, radiometric distortion in the image is eliminated, ensuring that the image grayscale values ​​accurately reflect the radiative energy reflected or emitted by ground features. Through ground control points and the sensor's geometric model, the image is projected onto a unified geographic coordinate system, ensuring that the geometric position of the image pixels corresponds to the actual geographic coordinates.

[0091] Step S22: Construct a matching volume of multi-angle images based on the preprocessed images;

[0092] Specifically, features such as edges and corners are extracted from the image to provide a basis for subsequent matching. Based on feature similarity or other matching metrics, a 3D matching volume is constructed, quantifying the probability of pixel matching under different parallaxes.

[0093] Step S23: Generate an initial sparse point cloud from the matched cost volume using a disparity estimation algorithm;

[0094] Specifically, the process of generating an initial sparse point cloud from the matched cost volume using a disparity estimation algorithm includes:

[0095] A reasonable range of parallax values ​​is defined based on the baseline distance and imaging geometry of multi-angle images;

[0096] Specifically, the baseline distance and imaging geometry of multi-angle images are analyzed to determine a reasonable range for parallax values. Baseline distance refers to the distance between camera positions when shooting the same scene from different viewpoints. Imaging geometry includes the camera's interior orientation elements (such as focal length and principal point coordinates) and exterior orientation elements (such as camera position and attitude). Based on these parameters, the theoretical range of parallax can be calculated, providing a reference for subsequent parallax estimation. Considering the impact of building height and terrain variations on parallax, the parallax range is appropriately expanded to ensure coverage of all possible parallax values.

[0097] For each pixel and its candidate disparity values, the cost is calculated using the squared difference cost function, and a cost volume is constructed.

[0098] Specifically, for each pixel and its candidate disparity values, the squared difference cost for that pixel under different disparities is calculated. The squared difference cost function (SSD) is a commonly used similarity metric that evaluates the matching degree between two pixels by calculating the sum of the squares of the differences in grayscale values ​​within the two pixel windows. A three-dimensional cost volume is constructed, where two dimensions correspond to the rows and columns of the image, and the third dimension corresponds to the disparity value. Each element in the cost volume represents the matching cost of the corresponding pixel under a specific disparity.

[0099] Local optimization of the cost volume yields a preliminary disparity map;

[0100] Specifically, local optimization algorithms, such as dynamic programming or semi-global matching, are employed to process the cost volume. These algorithms improve the accuracy and consistency of the disparity map by finding the minimum-cost path within local regions or by utilizing contextual information to optimize disparity estimation. A preliminary disparity map is generated, where the disparity value of each pixel corresponds to the minimum-cost position of that pixel within the cost volume.

[0101] Based on the preliminary disparity map and the camera's interior and exterior orientation elements, the disparity values ​​of pixels are converted into three-dimensional coordinates using collinearity equations to generate an initial sparse point cloud.

[0102] Specifically, using the camera's interior and exterior orientation elements and collinearity equations, the disparity values ​​in the initial disparity map are converted into three-dimensional coordinates. The collinearity equation is a fundamental mathematical model in photogrammetry that describes the geometric collinearity between object points, image points, and the camera center on the ground. Through the collinearity equation, disparity values ​​can be converted into actual three-dimensional spatial coordinates. An initial sparse point cloud is then generated, where each point represents the three-dimensional spatial position corresponding to a pixel in the image.

[0103] Step S24: Optimize and interpolate the initial sparse point cloud to generate a high-density point cloud;

[0104] Specifically, sparse point clouds are filtered and smoothed to remove noise points and improve point cloud quality. A 3D model is then constructed using the sparse point cloud, and interpolation algorithms are used to fill in missing regions to generate a high-density point cloud.

[0105] Step S25: Filter the high-density point cloud to generate the dense point cloud data.

[0106] Specifically, algorithms such as voxel filtering and statistical filtering are applied to further remove noise from the point cloud. The final output dense point cloud data is more spatially dense and uniform, providing high-precision three-dimensional geometric information for subsequent building height measurements.

[0107] Specifically, the initial sparse point cloud is optimized by removing noise and smoothing it; high-density point clouds are generated through interpolation to fill in missing data, improving point cloud density and detail, and enhancing the integrity of building outlines and structures. Filtering generates dense point cloud data for further noise reduction, improving point cloud uniformity and accuracy.

[0108] Step S3: Extract features from the satellite stereo image data and the dense point cloud data to obtain feature results; fuse the feature results according to a predefined deep learning model and automatically classify them into point cloud data.

[0109] Specifically, such as Figure 4 As shown, the process of step S3 includes:

[0110] Step S31: Extract spectral and texture features from satellite stereo image data, and extract geometric features from dense point cloud data;

[0111] Specifically, multi-band data from satellite imagery is used, employing band operations (such as band ratios and differences) to capture the spectral characteristics of ground features and distinguish different types of land cover. Simultaneously, dimensionality reduction techniques such as Principal Component Analysis (PCA) are used to extract the principal components of spectral features, preserving most of the spectral information. The Gray-Level Co-occurrence Matrix (GLCM) algorithm is used to analyze the distribution and arrangement of pixel gray values ​​in the imagery, calculating parameters describing texture features, such as second moment, contrast, correlation, and entropy, to reflect the texture details of the ground cover surface. Geometric attributes such as height, density, and curvature of the point cloud are calculated based on the point cloud data. Height information reflects the relative height of buildings; density information helps identify densely populated areas of buildings; and curvature can be used to distinguish between building edges and flat areas.

[0112] Step S32: Using a predefined deep learning model, the extracted spectral features, texture features, and geometric features are fused to obtain a fusion result;

[0113] Specifically, the extracted spectral, texture, and geometric features are fused to form a more comprehensive feature representation. Deep learning-based feature fusion methods are employed, such as utilizing the multi-input branch structure of a convolutional neural network (CNN). Different features are input into different branches for initial feature extraction, and then these features are fused at a deeper level. During feature-level fusion, feature concatenation can be used, directly concatenating feature vectors from different modalities into a single long vector, which serves as input to the subsequent deep learning model. This simple and direct fusion method preserves the complete information of the original features, allowing the model to learn the correlations and complementary information between different features during training.

[0114] Step S33: The fusion result is automatically classified using a trained deep learning model to generate classified point cloud data.

[0115] Specifically, the process of automatically classifying the fusion result using a trained deep learning model to generate classified point cloud data includes:

[0116] The fusion result is preprocessed and formatted to obtain a formatted result;

[0117] Specifically, the fused feature data is preprocessed, including normalization or standardization, to make it conform to the input requirements of the deep learning model, and the data is formatted into a tensor form acceptable to the model.

[0118] The pre-trained deep learning model is used to perform multi-layer feature extraction and encoding on the formatted input results;

[0119] Specifically, a pre-trained deep learning model, such as a Transformer-based model, is selected. The formatted fused features are input into the model, and the model's multi-layered structure is used for feature extraction and encoding. Each layer performs a non-linear transformation on the features, progressively extracting higher-level abstract features.

[0120] In the output layer of the model, the extracted features are mapped to the classification category through a fully connected layer to generate the classification probability of each point cloud point.

[0121] Specifically, in the output layer of the model, the extracted high-level features are mapped to classification categories through a fully connected layer, and the output is converted into the probability that each point cloud point belongs to a different classification category using functions such as softmax.

[0122] The category label of each point cloud point is determined based on the classification probability, thereby generating classified point cloud data.

[0123] Specifically, based on the classification probability, the category label of each point cloud point is determined according to the category corresponding to the maximum probability, thereby realizing the automatic classification of point cloud data.

[0124] Specifically, by fusing spectral, texture, and geometric features, the characteristics of building point clouds can be comprehensively captured, improving classification accuracy and reducing misclassification. The deep learning model, trained on extensive data, adapts to different types of buildings and complex scenes, exhibiting strong generalization capabilities. The entire process from feature extraction to classification is automated, reducing manual intervention and improving measurement efficiency.

[0125] Step S4: Extract ground point layer height information and building point layer height information from the point cloud data based on the building outline and roof structure characteristics;

[0126] Specifically, step S4 includes the following process:

[0127] The point cloud data of the building area and its surrounding buffer zone are filtered out from the point cloud data based on the building outline;

[0128] Specifically, the building outlines, whether vector data or manually drawn, are converted into a coordinate system that matches the point cloud data. Through spatial querying, the point cloud within the building outline is selected from the point cloud data, and a certain range is extended as a buffer to ensure complete coverage of the building and its potentially affected surrounding areas.

[0129] Cluster analysis and geometric feature extraction are performed on the filtered point clouds to determine the structural type of the building roof;

[0130] Specifically, the process of performing cluster analysis and geometric feature extraction on the filtered point cloud to determine the structural type of the building roof includes:

[0131] The filtered point cloud data is then denoised and smoothed to obtain the processing results.

[0132] Specifically, the filtered point cloud data is denoised and smoothed by using a radius-based filtering method. An appropriate radius threshold is set to remove areas with low point density, reduce noise interference, and at the same time preserve the corner features of buildings to obtain the processing results.

[0133] The processing results are initially clustered using the Euclidean clustering method, dividing the point cloud into multiple subsets that may contain building rooftops;

[0134] Specifically, the Euclidean clustering method is used to perform preliminary clustering of the processed results, dividing the point cloud into multiple subsets based on the Euclidean distance between points. A distance threshold (e.g., 0.5 meters) is set to ensure that the clusters roughly coincide with the edges of actual building roofs, resulting in multiple subsets that may contain building roofs.

[0135] The clustering results are further refined in each subset based on the geometric similarity between points to ensure that each cluster corresponds to a continuous area of ​​the roof.

[0136] Specifically, within each subset, the clustering results are further refined based on the similarity of geometric features between points (such as height and curvature). Secondary clustering is performed using the height distribution and curvature value range of the points to ensure that each cluster corresponds to a continuous area of ​​the roof.

[0137] Geometric features are extracted for each cluster, and the extracted geometric features are input into the trained classification model to determine the roof structure type corresponding to the cluster.

[0138] Specifically, geometric features are extracted for each cluster, including mean height, mean curvature, and density. After normalizing these features, they are input into a trained classification model (such as a random forest or support vector machine). The decision boundary or classification rules obtained from model training are used to determine the roof structure type corresponding to the cluster.

[0139] Select the appropriate height extraction method based on the roof structure type;

[0140] Specifically, the process of selecting the corresponding height extraction method based on the roof structure type includes:

[0141] Identify the roof structure type of a building, including flat roofs, pitched roofs, ridged roofs, or complex curved roofs;

[0142] Specifically, it is important to clearly distinguish between flat roofs, pitched roofs, ridged roofs, and complex curved roofs. For example, flat roofs are characterized by small elevation changes and low curvature; pitched roofs have obvious sloping features and large elevation changes; ridged roofs have obvious linear ridge features; and complex curved roofs are characterized by irregular height and curvature distribution.

[0143] For flat roofs, the average elevation of the point cloud within the building area is used as the building's point layer height.

[0144] Specifically, for all point cloud points within a cluster, the arithmetic mean of their elevation values ​​is calculated. By traversing each point in the cluster, summing their elevation values, and then dividing by the total number of points, the building's floor height is obtained. Abnormal elevation points that significantly deviate from the average are removed to reduce measurement errors and the impact of noise.

[0145] For pitched roofs, fit the roof plane and extract the height difference between the highest and lowest points in the roof area as the building's floor height.

[0146] Specifically, the least squares method is used to fit a plane to the point cloud within the cluster, resulting in a fitted plane equation. The highest and lowest points are searched on the fitted plane, their elevation values ​​are recorded, and the difference between the two is calculated as the building's floor height. The error of the fitted plane is evaluated to ensure that the fitting result meets the preset accuracy requirements.

[0147] For a ridged roof, the direction of the roof ridge is fitted by principal component analysis, and the heights of the highest and lowest points on both sides of the ridge are extracted. The average or maximum value is taken as the building floor height.

[0148] Specifically, PCA analysis is performed on the point cloud within the cluster to determine the direction of the roof ridge. PCA calculates the covariance matrix of the point cloud data and solves for its eigenvalues ​​and eigenvectors to find the main direction of the data distribution, i.e., the ridge direction. Based on the ridge direction, the point cloud within the cluster is divided into regions on both sides of the ridge, and the point clouds on both sides are extracted separately. Elevation extreme points are extracted from the point clouds on both sides of the ridge, and the elevation difference between the highest and lowest points on both sides is calculated. The average or maximum value is taken as the building's floor height.

[0149] For complex curved roofs, the main structure is identified based on the curvature characteristics of the point cloud, and the weighted average of the local elevations is calculated as the building's point floor height.

[0150] Specifically, the main structural areas of the roof are identified based on the curvature characteristics of the point cloud. Curvature reflects the degree of bending of the roof surface, with different curvature values ​​corresponding to different geometric shapes. The roof is divided into multiple local regions according to the curvature characteristics, and each region is assigned a corresponding weight (e.g., based on region area or point density). The mean elevation of each local region is calculated, and then a weighted average is calculated based on the weights to serve as the building's point-level height.

[0151] Based on the height information of the ground point layer and the building point layer, the height extraction method is used to extract reasonable ground point layer heights and building point layer heights respectively.

[0152] Specifically, by combining cluster analysis and geometric feature extraction with a classification model, the type of roof structure can be accurately determined, thereby selecting an appropriate method to extract the building height and improving the accuracy and reliability of the measurement.

[0153] Step S5: Determine the building height based on the building point floor height information and the ground point floor height information.

[0154] Specifically, the building point-level height information and ground point-level height information are preprocessed, including outlier removal and smoothing, to improve the accuracy and reliability of the data. For each building area, the corresponding ground point-level height information is subtracted from the building point-level height information to obtain a preliminary building height. The preliminary calculated building height is then optimized, such as by using filtering algorithms to remove noise or by incorporating terrain information for correction, to improve the accuracy of the building height.

[0155] Specifically, this invention utilizes aerial triangulation by fusing satellite stereo imagery data and ground control point data to generate high-precision dense point cloud data. A deep learning model is then used to automatically classify the point cloud data, accurately extracting the ground and building point layer height information to ultimately determine the building height. Each step is tightly integrated, leveraging the multi-angle observation advantages of satellite imagery and the precise geographic information of ground control points to ensure the accuracy of the measurement results. The introduction of the deep learning model improves data processing efficiency and classification accuracy, providing a reliable basis for subsequent height extraction. The entire method is highly efficient and automated, suitable for large-scale building height measurement tasks, effectively reducing labor and time costs, and providing crucial building height data support for low-altitude, economical UAV flights.

[0156] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0157] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for rapid measurement of building height using a low-altitude unmanned aerial vehicle (UAV), characterized in that, include: Step S1: Obtain satellite stereo image data and ground control point data within the mission area; perform aerial triangulation based on the satellite stereo image data and the ground control point data to obtain the measurement results. Step S2: Generate dense point cloud data of the work area based on the measurement results; Step S3: Extract features from the satellite stereo image data and the dense point cloud data to obtain feature results; fuse the feature results according to a predefined deep learning model and automatically classify them into point cloud data. Step S4: Extract ground point layer height information and building point layer height information from the point cloud data based on the building outline and roof structure characteristics; Step S5: Determine the building height based on the building point floor height information and the ground point floor height information.

2. The method for rapid measurement of building height by low-altitude unmanned aerial vehicle (UAV) flight according to claim 1, characterized in that, The process of step S1 includes: Acquire satellite stereo image data within the mission area, including stereo image pairs from multiple angles, and collect ground control point data for the mission area; Feature points are extracted from each satellite image and matched to establish geometric relationships between adjacent images; The extracted image features are combined with ground control point data to perform aerial triangulation to obtain the measurement results.

3. The method for rapid measurement of building height by low-altitude unmanned aerial vehicle (UAV) flight according to claim 2, characterized in that, The process of combining the extracted image features with ground control point data to perform aerial triangulation to obtain the measurement results includes: The correspondence between ground control point data and feature points extracted from satellite stereo imagery is established by matching corresponding points; Based on the correspondence, the initial values ​​of the interior and exterior orientation elements of the image are constructed using the collinearity equation, and a regional network adjustment model is established to form an initial triangulation model. The bundle adjustment algorithm is used to iteratively solve the image exterior orientation elements and the three-dimensional coordinates of ground points by combining the orbital parameters, attitude parameters and ground control point coordinates of the satellite stereo image. This optimizes the initial triangulation model and obtains the optimized result. The optimization results are verified by weighted adjustment of the error equation to obtain the measurement results.

4. The method for rapid measurement of building height by low-altitude unmanned aerial vehicle (UAV) flight according to claim 3, characterized in that, The process of step S2 includes: Preprocessing of satellite imagery, including radiometric and geometric correction, is performed to correct image distortion. Construct a multi-angle image matching volume based on the preprocessed image; An initial sparse point cloud is generated from the matched cost volume using a disparity estimation algorithm; The initial sparse point cloud is optimized and interpolated to generate a high-density point cloud; The high-density point cloud is filtered to generate the dense point cloud data.

5. A method for rapid measurement of building height using a low-altitude unmanned aerial vehicle (UAV) according to claim 4, characterized in that, The process of generating an initial sparse point cloud from the matched volume using a disparity estimation algorithm includes: A reasonable range of parallax values ​​is defined based on the baseline distance and imaging geometry of multi-angle images; For each pixel and its candidate disparity values, the cost is calculated using the squared difference cost function, and a cost volume is constructed. Local optimization of the cost volume yields a preliminary disparity map; Based on the preliminary disparity map and the camera's interior and exterior orientation elements, the disparity values ​​of pixels are converted into three-dimensional coordinates using collinearity equations to generate an initial sparse point cloud.

6. The method for rapid measurement of building height by low-altitude unmanned aerial vehicle (UAV) flight according to claim 5, characterized in that, The process of step S3 includes: Extract spectral and texture features from satellite stereo imagery data, and extract geometric features from dense point cloud data; Using a predefined deep learning model, the extracted spectral features, texture features, and geometric features are fused to obtain the fusion result; The fusion results are automatically classified using a trained deep learning model to generate classified point cloud data.

7. A method for rapid measurement of building height using a low-altitude unmanned aerial vehicle (UAV) according to claim 6, characterized in that, The process of automatically classifying the fusion result using a trained deep learning model to generate classified point cloud data includes: The fusion result is preprocessed and formatted to obtain a formatted result; The pre-trained deep learning model is used to perform multi-layer feature extraction and encoding on the formatted input results; In the output layer of the model, the extracted features are mapped to the classification category through a fully connected layer to generate the classification probability of each point cloud point. The category label of each point cloud point is determined based on the classification probability, thereby generating classified point cloud data.

8. A method for rapid measurement of building height using a low-altitude unmanned aerial vehicle (UAV) according to claim 7, characterized in that, The process of step S4 includes: The point cloud data of the building area and its surrounding buffer zone are filtered out from the point cloud data based on the building outline; Cluster analysis and geometric feature extraction are performed on the filtered point clouds to determine the structural type of the building roof; Select the appropriate height extraction method based on the roof structure type; Based on the height information of the ground point layer and the building point layer, the height extraction method is used to extract reasonable ground point layer heights and building point layer heights respectively.

9. A method for rapid measurement of building height using a low-altitude unmanned aerial vehicle (UAV) according to claim 8, characterized in that, The process of performing cluster analysis and geometric feature extraction on the filtered point cloud to determine the structural type of the building roof includes: The filtered point cloud data is then denoised and smoothed to obtain the processing results. The processing results are initially clustered using the Euclidean clustering method, dividing the point cloud into multiple subsets that may contain building rooftops; The clustering results are further refined in each subset based on the geometric similarity between points to ensure that each cluster corresponds to a continuous area of ​​the roof. Geometric features are extracted for each cluster, and the extracted geometric features are input into the trained classification model to determine the roof structure type corresponding to the cluster.

10. A method for rapid measurement of building height using a low-altitude unmanned aerial vehicle (UAV) according to claim 9, characterized in that, The process of selecting the corresponding height extraction method based on the roof structure type includes: Identify the roof structure type of a building, including flat roofs, pitched roofs, ridged roofs, or complex curved roofs; For flat roofs, the average elevation of the point cloud within the building area is used as the building's point layer height. For pitched roofs, fit the roof plane and extract the height difference between the highest and lowest points in the roof area as the building's floor height. For a ridged roof, the direction of the roof ridge is fitted by principal component analysis, and the heights of the highest and lowest points on both sides of the ridge are extracted. The average or maximum value is taken as the building floor height. For complex curved roofs, the main structure is identified based on the curvature characteristics of the point cloud, and the weighted average of the local elevations is calculated as the building's point floor height.

Images