Method, system, device and medium for constructing a three-dimensional model based on unmanned aerial vehicle surveying

Abstract

The present application relates to the technical field of building area three-dimensional model construction, in particular to a three-dimensional model construction method, system, device and medium based on unmanned aerial vehicle surveying and mapping; the method of the present application firstly performs pixel-level segmentation on the building area images obtained by the unmanned aerial vehicle to obtain ground object categories, and further segmentation is performed in combination with terrain information to obtain scene units; then, the scene unit point cloud data obtained by each unmanned aerial vehicle is registered within the unit, and combined with an absolute positioning reference, inter-unit registration processing is performed, and then through curved surface reconstruction and boundary fusion, a three-dimensional point cloud model is obtained; finally, through obtaining a multi-angle RGB image set of each scene unit, visible image extraction is performed, occlusion interference is removed, the best view image is selected to reduce texture deformation, and finally the texture of the best view image is mapped to the three-dimensional point cloud model to obtain a surveying and mapping area three-dimensional model, which has accurate geometric structure and high-fidelity visual features, and provides high-precision and real three-dimensional basic data for surveying and mapping visualization, engineering design and the like.

Description

Technical Field

[0001] This invention relates to the field of 3D model construction technology for building areas, specifically to a method, system, equipment, and medium for 3D model construction based on UAV mapping. Background Technology

[0002] 3D building models, as a core carrier of urban digital construction, play a crucial role in urban planning, cultural relic protection, and smart city management. In urban planning, they can visually represent the spatial layout and sunlight / ventilation relationships of building complexes, aiding in optimizing land use efficiency and traffic flow design. In cultural relic protection, high-precision models enable the digital archiving of structural features of ancient buildings, providing millimeter-level data support for restoration plan development. In smart city management, the integration of models with IoT sensing data allows for real-time monitoring of building energy consumption and pipeline operation status, promoting refined urban governance. Furthermore, in emergency rescue and disaster relief, 3D models can quickly locate structural weaknesses in buildings, providing decision-making basis for planning rescue routes in disasters such as earthquakes and fires.

[0003] Existing methods for 3D modeling of building areas based on UAV mapping still have some shortcomings. For example, in the reconstruction of complex structures, point cloud data is often sparse, resulting in low reconstruction accuracy and loss of building details. At the same time, in the texture mapping stage, ignoring the geometric characteristics of ground features and using a uniform orientation texture mapping method leads to texture mapping distortion. These problems significantly reduce the detail and realism of the 3D model, making it difficult to meet the needs of engineering applications and visualization. Summary of the Invention

[0004] The purpose of this invention is to provide a method, system, device, and medium for constructing three-dimensional models based on UAV mapping.

[0005] The technical solution of this invention is as follows:

[0006] A method for constructing a 3D model based on UAV mapping includes the following operations:

[0007] S1. Perform pixel-level segmentation on the image of the area to be surveyed to obtain the land feature categories; combine the land feature categories with the terrain information of the area to be surveyed to perform scene unit segmentation on the land feature categories to obtain several scene units;

[0008] S2. Based on the point cloud data acquisition method corresponding to the scene unit, acquire the point cloud data of each scene unit to obtain several scene unit point clouds; after preprocessing, perform intra-unit registration on point clouds at different angles of the same scene to obtain several intra-unit registered point clouds; based on the preset absolute positioning reference of the area to be surveyed, align all intra-unit registered point clouds to a unified coordinate system, and then perform inter-unit registration processing to obtain an initial point cloud model; after surface reconstruction of the point cloud corresponding to each scene unit in the initial point cloud model, and then perform point cloud boundary fusion, obtain a three-dimensional point cloud model;

[0009] S3. Acquire RGB images of each scene unit from multiple angles to form their respective RGB image sets; for each triangular facet of the scene unit in the 3D point cloud model, obtain the corresponding visible image from the corresponding RGB image set; select the best viewpoint image from the visible images, and map the texture information of the best viewpoint image to the corresponding triangular facet to obtain the 3D model of the survey area.

[0010] The intra-cell registration operation in S2 is as follows: After downsampling, the point clouds of different angles of the current scene cell are converted into multiple view depth images; feature points are extracted from the multiple view depth images respectively, and after matching, initial feature matching pairs are obtained; the initial feature matching pairs are purified to obtain filtered matching pairs; the filtered matching pairs are processed by spatial transformation matrix to obtain intra-cell registration transformation parameters.

[0011] A point cloud at a random angle is selected as the reference angle point cloud, and the rest are used as non-reference angle point clouds. The non-reference angle point clouds and their corresponding registration transformation parameters are subjected to coordinate transformation to obtain aligned point clouds. The reference point cloud and the aligned point cloud are fused to obtain the registered point cloud within the current scene unit.

[0012] The operation to obtain the initial point cloud model in S2 is as follows: All scene units' registered point clouds are compared with a preset absolute positioning reference, and after coordinate transformation, a unified point cloud for each scene unit is obtained; the unified point clouds of adjacent scene units are processed by extracting overlapping regions to obtain overlapping point sets; the overlapping point sets of adjacent units are processed by registration based on the iterative nearest point method to obtain inter-unit transformation parameters; one scene unit's unified point cloud is randomly selected as the reference unit point cloud, and the rest are used as non-reference unit point clouds; the non-reference unit point clouds and their corresponding unit transformation parameters are processed by coordinate correction to obtain aligned unit point clouds; the reference unit point cloud and the aligned unit point cloud are fused to obtain the initial point cloud model.

[0013] In S3, the method for obtaining the visible image of the current triangular facet is as follows: based on the vertex coordinates and camera parameters of the current triangular facet, multiple planar projection points are obtained; the images of all planar projection points in the corresponding RGB image set within the image field of view are retained to obtain the first candidate RGB image set; the images of the facet facing the camera direction in the first candidate RGB image set are retained to obtain the second candidate RGB image set; the images in the second candidate RGB image set that do not occlude the current triangular facet are taken as the visible images.

[0014] In S3, the method for selecting the best viewpoint image is as follows: obtain the angle value between the normal vector of the triangular facet and the optical axis direction vector of each visible image, and filter out the best viewpoint image based on the angle value and the preset angle threshold of the scene unit region corresponding to the triangular facet.

[0015] Before pixel-level segmentation in S1, image enhancement processing is performed on the image of the area to be surveyed to obtain an enhanced image of the area to be surveyed, which is then used to perform pixel-level segmentation. The image enhancement processing includes: obtaining the image quality assessment result of the image of the area to be surveyed based on pixel brightness values ​​and brightness histograms; if the image quality assessment result is that the image is overexposed, performing brightness suppression and detail restoration processing on the image of the area to be surveyed to obtain an enhanced image of the area to be surveyed; if the image quality assessment result is that the image is underexposed, performing shadow brightening and noise suppression processing on the image of the area to be surveyed to obtain an enhanced image of the area to be surveyed; if the image quality assessment result is that the image is backlit, performing dual-region adaptive enhancement processing on the image of the area to be surveyed to obtain an enhanced image of the area to be surveyed.

[0016] Dark area brightening and noise suppression can be achieved through Gamma correction and Zero-DCE network, respectively.

[0017] A 3D model construction system based on UAV mapping, used to implement the above-mentioned 3D model construction method based on UAV mapping, includes:

[0018] The scene unit generation module is used to perform pixel-level segmentation of the image of the area to be surveyed to obtain the land feature categories; and to perform scene unit segmentation of the land feature categories based on the land feature categories and the terrain information of the area to be surveyed to obtain several scene units.

[0019] The 3D point cloud model generation module is used to acquire point cloud data for each scene unit according to the point cloud data acquisition method corresponding to the scene unit, resulting in several scene unit point clouds. After preprocessing, the point clouds of several scene units are registered within the same unit for point clouds at different angles of the same scene, resulting in several registered point clouds within the scene units. Based on the preset absolute positioning reference of the area to be mapped, all registered point clouds within the scene units are aligned to a unified coordinate system, and then inter-unit registration is performed to obtain an initial point cloud model. After surface reconstruction of the point cloud corresponding to each scene unit in the initial point cloud model, the point cloud boundary is fused to obtain a 3D point cloud model.

[0020] The 3D model generation module for the survey area is used to acquire RGB images of each scene unit from multiple angles to form their respective RGB image sets. For each triangular facet of the scene unit in the 3D point cloud model, a visible image is obtained from the corresponding RGB image set. From the visible images, the best viewpoint image is selected, and the texture information of the best viewpoint image is mapped to the corresponding triangular facet to obtain the 3D model of the survey area.

[0021] A 3D model building device based on UAV mapping includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the above-mentioned 3D model building method based on UAV mapping.

[0022] A computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described method for constructing a 3D model based on UAV mapping.

[0023] The beneficial effects of this invention are as follows:

[0024] This invention provides a method for constructing a 3D model based on UAV mapping. First, the image of the building area acquired by the UAV is segmented at the pixel level to obtain the land feature categories. Combined with terrain information, scene unit segmentation is performed to obtain scene units. Then, the point cloud of each scene unit is preprocessed and intra-unit matching is performed to ensure the fusion of details of the point cloud from multiple angles within the same scene. Combined with absolute positioning reference, inter-unit registration processing is performed to achieve accurate alignment of the point cloud across the entire domain. Then, surface reconstruction and boundary fusion are performed to obtain a 3D point cloud model that retains the detailed features of the land features. Finally, by acquiring multi-angle RGB image sets of each scene unit, the texture information is fully covered. Visible image extraction is performed to remove occlusion interference, and the best viewpoint image is selected to reduce texture deformation. Finally, the texture of the best viewpoint image is mapped onto triangular patches. The resulting 3D model of the surveyed area has both accurate geometric structure and high-fidelity visual features, providing high-precision and realistic 3D basic data for surveying visualization, engineering design, and other applications. Detailed Implementation

[0025] This embodiment provides a method for constructing a 3D model based on UAV mapping, including the following operations:

[0026] S1. Perform pixel-level segmentation on the image of the area to be surveyed to obtain the land feature categories; combine the land feature categories with the terrain information of the area to be surveyed to perform scene unit segmentation on the land feature categories to obtain several scene units;

[0027] S2. Based on the point cloud data acquisition method corresponding to the scene unit, acquire the point cloud data of each scene unit to obtain several scene unit point clouds; after preprocessing, perform intra-unit registration on point clouds at different angles of the same scene to obtain several intra-unit registered point clouds; based on the preset absolute positioning reference of the area to be surveyed, align all intra-unit registered point clouds to a unified coordinate system, and then perform inter-unit registration processing to obtain an initial point cloud model; after surface reconstruction of the point cloud corresponding to each scene unit in the initial point cloud model, and then perform point cloud boundary fusion, obtain a three-dimensional point cloud model;

[0028] S3. Acquire RGB images of each scene unit from multiple angles to form their respective RGB image sets; for each triangular facet of the scene unit in the 3D point cloud model, obtain the visible image from the corresponding RGB image set; select the best viewpoint image from the visible images, and map the texture information of the best viewpoint image to the corresponding triangular facet to obtain the 3D model of the survey area.

[0029] The specific steps and details are as follows.

[0030] S1. Perform pixel-level segmentation on the image of the area to be surveyed to obtain the land feature categories; combine the land feature categories with the terrain information of the area to be surveyed to perform scene unit segmentation on the land feature categories to obtain several scene units.

[0031] The image is first segmented at the pixel level to obtain the land feature categories. Then, combined with the terrain information, scene unit segmentation is performed to obtain scene units. This achieves a fine division of the area to be surveyed from micro-land features to macro-scenes. It not only clarifies the land feature attributes of each scene unit, but also aggregates similar land features into scene units. This provides accurate regional feature guidance for UAVs to acquire point cloud data, which is conducive to improving the spatial matching degree and detail integrity of point cloud data.

[0032] First, based on the planning information of the construction area, the area to be surveyed was delineated. To improve the accuracy and rationality of subsequent construction design, some areas surrounding the construction area were included in the area to be surveyed, and high-resolution orthophotos were acquired using a drone equipped with a high-resolution camera, serving as the image of the area to be surveyed.

[0033] Then, the image of the area to be surveyed is segmented at the pixel level (which can be achieved through machine learning methods and deep learning network models) to identify the land features in the area to be surveyed (such as buildings (including building roofs and facades), bare land, vegetation (vegetation canopy and vegetation surface), hillsides, etc.) and obtain the land feature categories.

[0034] To prevent poor image quality due to weather conditions or drone shooting angle and shooting time, which could affect the accuracy of pixel-level segmentation results, image enhancement processing is performed on the image of the area to be surveyed before pixel-level segmentation to obtain an enhanced image of the area to be surveyed, which is then used to perform pixel-level segmentation.

[0035] The specific operations for image enhancement processing are as follows.

[0036] Based on pixel brightness values ​​and brightness histograms, image quality assessment results of the area to be mapped are obtained.

[0037] Specifically, the color space of the image of the area to be measured is converted to obtain the luminance component. The luminance component is then statistically calculated using pixels to obtain the proportion of high-luminance pixels (high-luminance pixel value > 230), the proportion of low-luminance pixels (low-luminance pixel value < 25), the mean luminance, the standard deviation of luminance, and a luminance histogram. If the proportion of high-luminance pixels is greater than the overexposure threshold, and saturation accumulation occurs in the region of interest (bright area) on the right side of the luminance histogram, the image quality assessment result is overexposed. If the proportion of low-luminance pixels is less than the underexposure threshold, and the mean luminance is less than the mean luminance threshold, and accumulation occurs in the region of interest (dark area) on the left side of the luminance histogram, the image quality assessment result is underexposed. If the luminance histogram exhibits a bimodal distribution, and the standard deviation of luminance is greater than the standard deviation threshold, the image quality assessment result is backlit. If none of the above three criteria are met, the image quality assessment result is a qualified image.

[0038] If the image quality assessment result indicates that the image is overexposed, the image of the area to be mapped will undergo brightness suppression and detail restoration processing to obtain an enhanced image of the area to be mapped.

[0039] Specifically, the image of the area to be mapped is converted to a color space to obtain a luminance component image. The luminance component image is then convolved with multiple Gaussian kernels of different scales. This process suppresses the illumination components in the bright areas of the overexposed image through multi-scale Gaussian filtering, compressing the brightness to a reasonable range and resolving the overexposure problem, thus obtaining multi-scale illumination component estimates. The logarithmic domain of the luminance component image is then differenced with the logarithmic domain of each scale illumination component estimate, preserving the original lost textures of ground features in the bright areas, such as building edges and vegetation outlines, to obtain multi-scale reflection components. The multi-scale reflection components are then averaged and fused to integrate the details of ground features of different sizes, and color restoration is performed to obtain an enhanced image of the area to be mapped.

[0040] If the image quality assessment result indicates that the image is underexposed, the image of the area to be mapped will undergo shadow brightening and noise suppression processing to obtain an enhanced image of the area to be mapped. The aforementioned shadow brightening and noise suppression processing can be achieved through Gamma correction and low-light enhancement networks (such as Zero-DCE networks), respectively.

[0041] If the image quality assessment result indicates that the image is backlit, the image of the area to be mapped will undergo dual-region adaptive enhancement processing to obtain the enhanced image of the area to be mapped.

[0042] Specifically, the image of the area to be measured is converted to a color space to obtain a luminance component image. Based on the luminance component image and an adaptive threshold, the image of the area to be measured is segmented to obtain a dark area mask (low-brightness foreground area) and a bright area mask (high-brightness background area). The dark area mask and the bright area mask are respectively brightened by a low Gamma value (achieved by exponentiation of the dark area mask value with a preset low Gamma value, such as 0.5) and brightened by a high Gamma value (achieved by exponentiation of the bright area mask value with a preset high Gamma value, such as 1.5), so that... In backlit images, the dark textures (such as wall cracks and leaf veins) become clearly discernible from being invisible, and bright areas (such as the sky and reflective ground) are prevented from pixel saturation, preserving previously lost bright details such as cloud patterns and road markings, resulting in a brightness-corrected mask. The brightness-corrected mask and the guided filter (which can be obtained based on the image edge information of the area to be mapped) are fused to obtain a smoothly transitioned brightness enhancement component. The brightness enhancement component and the color component of the image of the area to be mapped are then subjected to inverse color space conversion to obtain an enhanced image of the area to be mapped with clear foreground and background features.

[0043] If the image quality assessment result is a qualified image, the image of the area to be mapped will be subjected to contrast enhancement processing to obtain the enhanced image of the area to be mapped.

[0044] Next, by combining the land feature categories (image semantic information) and the terrain information of the area to be surveyed, the land feature categories are segmented into scene units to obtain several scene units.

[0045] Specifically, for buildings, adjacent building areas with the same height range are merged into a single scene unit; for hills, based on the terrain information (slope and aspect) of the area to be mapped, the hills are subdivided into continuous blocks of gentle and steep slopes, each serving as an independent scene unit; for vegetation, continuous areas can be directly used as a scene unit; and for bare land (including roads, lawns, etc.), continuous areas can be used as a scene unit.

[0046] S2. Based on the point cloud data acquisition method corresponding to the scene unit, acquire the point cloud data of each scene unit to obtain several scene unit point clouds; after preprocessing, perform intra-unit registration on point clouds at different angles of the same scene to obtain several intra-unit registered point clouds; based on the preset absolute positioning reference of the area to be surveyed, align all intra-unit registered point clouds to a unified coordinate system, and then perform inter-unit registration processing to obtain an initial point cloud model; after surface reconstruction of the point cloud corresponding to each scene unit in the initial point cloud model, and then perform point cloud boundary fusion, obtain a three-dimensional point cloud model.

[0047] Each scene unit point cloud obtained according to the scene unit point cloud acquisition method is preprocessed and internally matched to ensure the fusion of details of point clouds from multiple angles in the same scene. Combined with the absolute positioning reference, inter-unit registration is performed to achieve accurate alignment of the global point cloud. Then, through surface reconstruction and boundary fusion, a three-dimensional point cloud model is obtained, which accurately preserves the fine features of different scene units such as building edges, steep slope textures, and vegetation layers, and eliminates the splicing gaps between units to ensure the spatial continuity of complex features.

[0048] First, using a drone equipped with LiDAR, point cloud data for each scene unit is acquired according to the point cloud data acquisition path corresponding to the scene unit, resulting in several scene unit point clouds.

[0049] For architectural scene units, the point cloud data acquisition methods based on UAVs include point cloud acquisition methods based on orthogonal flight paths and oblique flight paths, which can solve the problem of voids in the point cloud of walls and obtain complete point cloud data of the three-dimensional structure of the roof and facade.

[0050] For gentle slope scenarios, the point cloud data acquisition methods based on UAVs include contour line parallel flight path (along the slope direction) and cross-validation flight path point cloud acquisition methods, which uniformly cover the ground surface and avoid terrain distortion.

[0051] For steep slope scene units, the point cloud data acquisition methods based on UAVs include point cloud acquisition methods that use upward shooting routes from the valley floor and upward shooting routes from the ridge. These methods can offset the influence of slope, prevent radar obstruction, capture details of steep cliffs, and reduce the point cloud loss rate in steep slope areas.

[0052] For vegetation scene units, the point cloud data acquisition methods based on drones include point cloud acquisition methods that obtain canopy flight paths from high-altitude orthophotos and obtain ground flight paths from low-altitude penetration, which separate the vegetation canopy from the ground surface and improve the detail of the point cloud data of vegetation scene units.

[0053] For bare ground scene units, the point cloud data acquisition methods based on UAVs include grid-based region division and parallel snake-like navigation point cloud acquisition methods, which efficiently cover flat bare ground areas and improve the efficiency of UAV point cloud data acquisition.

[0054] Then, after preprocessing, the point clouds of several scene units are registered within each unit for point clouds at different angles within the same scene, resulting in several registered point clouds within scene units. The preprocessing includes point cloud denoising, filtering, and density unification of the point clouds in the scene units.

[0055] Taking the point clouds of different angles of the current scene unit as an example, the registration operation steps within the unit are as follows.

[0056] Step 1: After downsampling, the point clouds of different angles of the current scene unit are transformed into depth images and projected onto a two-dimensional plane to generate grayscale depth maps, resulting in multiple viewpoint depth images. Feature points are extracted from the multiple viewpoint depth images (which can be achieved through the SIFT feature method) to preserve the scale and edge features of the point cloud surface (such as building corners and vegetation outlines). After matching (which can be achieved through the FLANN matching algorithm), initial feature matching pairs (including correct matching pairs and incorrect matching pairs) are obtained.

[0057] Step 2: The initial feature matching pairs are purified (which can be achieved through the RANSAC algorithm) to remove mismatched pairs and obtain the filtered matching pairs; the filtered matching pairs are then subjected to spatial transformation matrix calculation (which can be achieved by solving the rotation matrix and translation vector using the least squares method) to obtain the intra-cell registration transformation parameters.

[0058] Step 3: Randomly select one angle point cloud from the point clouds of different angles in the current scene unit as the reference angle point cloud, and the rest as non-reference angle point clouds; the non-reference angle point clouds and their corresponding registration transformation parameters are subjected to coordinate transformation to eliminate point cloud misalignment caused by perspective differences (such as splicing gaps in different angle scans of building facades), resulting in aligned point clouds; the reference point cloud (selected from point clouds of different angles) and the aligned point cloud are fused to achieve uniform point cloud density, continuous structure, and no repetition or loss of ground feature details (such as slope texture of steep slopes and distribution of vegetation branches) in the scene unit, resulting in the registered point cloud in the current scene unit.

[0059] Next, based on the preset absolute positioning reference of the area to be surveyed, the registered point clouds in all scene units are aligned to a unified coordinate system, and then inter-unit registration processing is performed to obtain the initial point cloud model.

[0060] The specific steps to obtain the initial point cloud model are as follows: All point clouds within each scene unit are registered with a preset absolute positioning reference (including GNSS control point coordinates and coordinate system parameters). Coordinate transformation is performed to achieve coordinate system uniformity, ensuring absolute consistency in the spatial position of each unit's point cloud, resulting in a unified point cloud for each scene unit. The unified point clouds of adjacent scene units are then processed by extracting overlapping areas (based on spatial range intersection) to obtain overlapping point sets. These overlapping point sets of adjacent units are then registered using an iterative nearest-point method to eliminate splicing errors caused by the initial coordinate transformation, resulting in inter-unit transformation parameters. One scene unit's unified point cloud is randomly selected as the reference unit point cloud, and the rest are used as non-reference unit point clouds. The transformation parameters between the non-reference unit point clouds and their corresponding units are then processed by coordinate correction to ensure spatial continuity across unit features (such as continuous roads and through slopes), resulting in aligned unit point clouds. Finally, the reference unit point cloud and the aligned unit point cloud are fused to remove duplicate points and fill edge gaps, resulting in an initial point cloud model that covers the entire area to be mapped and has no obvious splicing marks between units.

[0061] Finally, after reconstructing the surface of the point cloud corresponding to each scene unit in the initial point cloud model, the three-dimensional point cloud model is obtained by point cloud boundary fusion.

[0062] The specific steps are as follows.

[0063] Step 1: Each unit in the initial point cloud model corresponds to a point cloud. After surface reconstruction, the discrete point cloud is transformed into a continuous topological surface, eliminating the sparsity and noise of the original point cloud and preserving the fine structure of the ground features (such as building surfaces and slope contours), thus obtaining the initial surface model of each unit.

[0064] For architectural scene units, the surface reconstruction operation is as follows: The point cloud of the architectural scene unit is processed by K-nearest neighbor outlier detection to remove noise points and measurement error points on the architectural surface, resulting in a denoised point cloud. Based on normal vector mutation and curvature threshold, key feature points (such as corner points, door and window edge points, roof edge points, etc.) are extracted from the denoised point cloud. Through interpolation calculation, the angular features of the building (such as corners and roof ridges) are accurately captured to reduce the normal vector error of the fitting plane and ensure that the geometric direction of the walls and roof is consistent with the actual situation, resulting in an initial fitting plane. The initial fitting plane and the denoised point cloud are optimized by least squares to minimize the distance error from the point to the plane, so that the fitting plane fits the details of the architectural surface (such as wall flatness and roof slope), resulting in an optimized fitting plane. The optimized fitting plane is then sampled by meshing to preserve the planar continuity and angular structural features of the building (such as no blurring of right-angle corners), resulting in the initial architectural surface model.

[0065] Step 2: The initial surface models of adjacent units are extracted based on boundary features derived from point cloud normal vectors and curvature abrupt changes to obtain a set of boundary points. Based on the distance weighting method (the closer the distance, the greater the point weight), the set of boundary points of adjacent units is weighted and smoothed to eliminate the splicing gaps between the surfaces of the units, realize boundary fusion, make the transition between adjacent units (such as the connection between buildings and gentle slopes) natural, avoid geometric abrupt changes, and obtain a transition surface.

[0066] The specific method for obtaining the boundary point set is as follows: the initial surface models of adjacent units are uniformly sampled to obtain several point clouds, forming a first point cloud set and a second point cloud set respectively; each point cloud in the first and second point cloud sets is processed by principal component analysis to obtain a normal vector used to describe the orientation of the point cloud surface, forming a first normal vector set and a second normal vector set respectively; the point cloud in the first or second normal vector set where the angle between the normal vectors of adjacent point clouds is greater than the angle threshold is taken as the normal vector mutation point; the point cloud in the first or second point cloud set where the curvature is greater than the curvature threshold is taken as the curvature mutation point set; the normal vector mutation points and curvature mutation points are intersected to obtain the initial boundary candidate points; the initial boundary candidate points are morphologically filtered to remove isolated noise points, thus obtaining the boundary point set of adjacent units.

[0067] Step 3: The initial surface model and the corresponding transition surface of each unit are spliced ​​together to obtain the global surface model; the global surface model is then resampled from the point cloud to obtain the three-dimensional point cloud model, thus achieving the geometric continuity and structural integrity of the entire land cover.

[0068] S3. Acquire RGB images of each scene unit from multiple angles to form their respective RGB image sets; for each triangular facet of the scene unit in the 3D point cloud model, obtain the visible image from the corresponding RGB image set; select the best viewpoint image from the visible images, and map the texture information of the best viewpoint image to the corresponding triangular facet to obtain the 3D model of the survey area.

[0069] By acquiring multi-angle RGB image sets of each scene unit, ensuring comprehensive coverage of texture information, extracting visible images, removing occlusion interference, selecting the best viewpoint image to reduce texture deformation, and finally mapping the texture to the triangular facets to obtain a 3D model, the model retains the real texture and color of fine features such as building edges and vegetation veins, making the model have both accurate geometric structure and high-fidelity visual features, providing high-precision and realistic 3D basic data for surveying visualization, engineering design, etc.

[0070] First, acquire RGB images (including orthophotos and a large number of oblique images) from multiple angles for each scene unit. The RGB images cover the anisotropic features of the ground features, such as building facades and roofs, vegetation canopies and sides, providing comprehensive visual information for texture mapping to form their respective RGB image sets.

[0071] Then, for each triangular facet of each scene unit in the 3D point cloud model (the number and size of the triangular facets can be customized according to actual needs), the visible image is obtained from the corresponding RGB image set. A triangular facet is a geometric shape in the 3D point cloud model consisting of three vertices and three edges, with each edge connecting two vertices. As the simplest polygon, the triangular facet has stable geometric properties, which not only accurately approximates various complex terrain surfaces, such as building edges and slope surfaces, but also facilitates rapid calculation of subsequent texture coordinates, reducing mapping errors.

[0072] The method for obtaining the visible image of the current triangular facet is as follows: Based on the vertex coordinates (three) of the current triangular facet and camera parameters (including focal length, principal point coordinates, rotation matrix, and translation vector), multiple planar projection points (three) are obtained; the images of all planar projection points within the field of view of the corresponding RGB image set are retained to ensure that the current triangular facet is completely within the field of view of the RGB image, thus obtaining the first candidate RGB image set; the images of the facet facing the camera direction in the first candidate RGB image set are retained (the product of the triangular facet normal vector and the camera optical axis direction vector is greater than 0), thus obtaining the second candidate RGB image set; the images in the second candidate RGB image set that do not occlude the current triangular facet are taken as the visible images.

[0073] The image that does not obscure the current triangle is the image onto which the depth value of all pixels in the grayscale region of the image is greater than the depth value of the triangle, where the depth value of the triangle is the average value of the z-axis coordinates of the vertices of the triangle.

[0074] Then, the best viewing angle image is selected from the visible images. The specific method for selecting the best viewing angle image is as follows: obtain the angle value between the normal vector of the triangular facet and the optical axis direction vector of the visible image, and filter out the best viewing angle image based on the angle value and the preset angle threshold of the scene unit region corresponding to the triangular facet.

[0075] For the building top surface area (such as roof edges, skylights, etc.) in the architectural scene unit, the best viewing angle image is the visible image corresponding to the angle value being less than the first angle threshold (10°) and the angle value being the minimum value. Such a near-vertical viewing angle can avoid perspective distortion.

[0076] For building facade areas (such as doors and windows) within an architectural scene unit, the optimal viewing angle image is the visible image corresponding to the angle value between the second (85°) and third (95°) thresholds, with the angle value being the maximum value. This approximates a parallel viewing angle to preserve the integrity of vertical structures such as doors, windows, and decorative lines. Both the second and third angle thresholds are greater than the first angle threshold.

[0077] For steep slope scene units, the optimal viewing angle image is one with an angle value between the fourth angle threshold (30°) and the fifth angle threshold (60°). The closer the angle value is to the two extreme values, the more visible the image is. Such a tilted viewing angle covers the entire steep slope and avoids the shadow occlusion of the slope bottom caused by vertical scanning.

[0078] For a gently sloping scene unit, the optimal viewing angle image is the visible image corresponding to the angle value being less than the sixth angle threshold (20°) and the angle value being the minimum value. This allows for a wider vertical viewing angle range, balancing data acquisition efficiency and texture accuracy.

[0079] For the canopy in the vegetation scene unit, the best viewing angle image is the visible image corresponding to the angle value being less than the seventh angle threshold (15°) and the angle value being the minimum value. Such a near vertical viewing angle can penetrate the gaps in the tree canopy and capture the leaf distribution and canopy morphology.

[0080] For the surface area in the vegetation scene unit, the best viewing angle image is the visible image with the included angle value between the eighth included angle threshold (80°) and the ninth included angle threshold (100°), and the included angle value is the maximum value. This close parallel viewing angle reduces canopy occlusion and obtains the surface vegetation cover and soil texture.

[0081] For bare ground scene units, the optimal viewing angle image is the visible image corresponding to the angle value being less than the seventh angle threshold (15°) and the angle value being the minimum value. The near-vertical viewing angle can ensure the surface flatness and texture accuracy of micro-topography (such as potholes and ridges).

[0082] Finally, the texture information of the best-view image is mapped onto the corresponding triangular facets, and the average value of the mapping is taken for repeated areas to obtain the three-dimensional model of the surveyed area.

[0083] This embodiment also provides a 3D model construction system based on UAV mapping, used to implement the above-mentioned 3D model construction method based on UAV mapping, including:

[0084] The scene unit generation module is used to perform pixel-level segmentation of the image of the area to be surveyed to obtain the land feature categories; and to perform scene unit segmentation of the land feature categories based on the land feature categories and the terrain information of the area to be surveyed to obtain several scene units.

[0085] The 3D point cloud model generation module is used to acquire point cloud data for each scene unit according to the point cloud data acquisition method corresponding to the scene unit, resulting in several scene unit point clouds. After preprocessing, the point clouds of several scene units are registered within the same unit for point clouds at different angles of the same scene, resulting in several registered point clouds within the scene units. Based on the preset absolute positioning reference of the area to be mapped, all registered point clouds within the scene units are aligned to a unified coordinate system, and then inter-unit registration is performed to obtain an initial point cloud model. After surface reconstruction of the point cloud corresponding to each scene unit in the initial point cloud model, the point cloud boundary is fused to obtain a 3D point cloud model.

[0086] The 3D model generation module for the survey area is used to acquire RGB images of each scene unit from multiple angles to form their respective RGB image sets. For each triangular facet of the scene unit in the 3D point cloud model, a visible image is obtained from the corresponding RGB image set. From the visible images, the best viewpoint image is selected, and the texture information of the best viewpoint image is mapped to the corresponding triangular facet to obtain the 3D model of the survey area.

[0087] This embodiment also provides a 3D model building device based on UAV mapping, including a processor and a memory, wherein the processor executes the computer program stored in the memory to implement the above-described 3D model building method based on UAV mapping.

[0088] This embodiment also provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described method for constructing a 3D model based on UAV mapping.

[0089] This embodiment provides a method for constructing a 3D model based on UAV mapping. First, the images of the building area acquired by the UAV are segmented at the pixel level to obtain the land feature categories. Combined with terrain information, scene units are segmented to obtain scene units. Then, the point cloud of each scene unit is preprocessed and internally matched to ensure the fusion of details of the point cloud from multiple angles within the same scene. Combined with absolute positioning reference, inter-unit registration is performed to achieve accurate alignment of the point cloud across the entire domain. Then, surface reconstruction and boundary fusion are performed to obtain a 3D point cloud model that retains the detailed features of the land features. Finally, by acquiring multi-angle RGB image sets of each scene unit, the texture information is fully covered. Visible image extraction is performed to remove occlusion interference, and the best viewpoint image is selected to reduce texture deformation. Finally, the texture of the best viewpoint image is mapped onto triangular patches. The resulting 3D model of the surveyed area has both accurate geometric structure and high-fidelity visual features, providing high-precision and realistic 3D basic data for surveying visualization, engineering design, etc.

Claims

1. A method for constructing a 3D model based on UAV mapping, characterized in that, This includes the following operations: S1. Based on pixel brightness values ​​and brightness histograms, obtain the image quality assessment results of the image of the area to be mapped; if the image quality assessment result is that the image is overexposed, perform brightness suppression and detail restoration processing on the image of the area to be mapped to obtain an enhanced image of the area to be mapped; if the image quality assessment result is that the image is underexposed, perform shadow brightening and noise suppression processing on the image of the area to be mapped to obtain an enhanced image of the area to be mapped; if the image quality assessment result is that the image is backlit, perform dual-region adaptive enhancement processing on the image of the area to be mapped to obtain an enhanced image of the area to be mapped. Pixel-level segmentation is performed on the enhanced image of the area to be surveyed to obtain the land cover categories; By combining the land feature categories and the topographic information of the area to be surveyed, the land feature categories are segmented into scene units to obtain several scene units; S2. Based on the point cloud data acquisition method corresponding to the scene unit, acquire the point cloud data of each scene unit to obtain several scene unit point clouds; After preprocessing, the point clouds of several scene units are registered within the same scene from different angles to obtain several registered point clouds within the same scene unit. Based on the preset absolute positioning reference of the area to be surveyed, after aligning the registered point clouds in all scene units to a unified coordinate system, perform inter-unit registration processing to obtain the initial point cloud model. After reconstructing the surface of the point cloud corresponding to each scene unit in the initial point cloud model, the three-dimensional point cloud model is obtained by merging the point cloud boundaries. S3. Obtain RGB images of each scene unit from multiple angles to form their respective RGB image sets; for each triangular facet of the scene unit in the 3D point cloud model, obtain the visible image from the corresponding RGB image set; select the best viewpoint image from the visible images, and map the texture information of the best viewpoint image to the corresponding triangular facet to obtain the 3D model of the survey area. The method for obtaining the visible image of the current triangular facet is as follows: based on the vertex coordinates and camera parameters of the current triangular facet, multiple planar projection points are obtained; the images of all planar projection points in the corresponding RGB image set within the image field of view are retained to obtain the first candidate RGB image set; The images of the facets facing the camera are retained from the first candidate RGB image set to obtain the second candidate RGB image set; the images in the second candidate RGB image set that do not occlude the current triangular facets are taken as the visible images; The method for selecting the best viewpoint image is as follows: obtain the angle between the normal vector of the triangular facet and the optical axis direction vector of each visible image, and select the best viewpoint image based on the angle value and the preset angle threshold of the scene unit region corresponding to the triangular facet.

2. The method for constructing a 3D model based on UAV mapping according to claim 1, characterized in that, The specific operation of intra-cell registration in S2 is as follows: After downsampling, the point clouds of the current scene unit at different angles are converted into depth images to obtain multiple view depth images. Feature points are extracted from the multiple view depth images and matched to obtain initial feature matching pairs. The initial feature matching pairs are purified to obtain the filtered matching pairs; the filtered matching pairs are then subjected to spatial transformation matrix calculation to obtain the intra-cell registration transformation parameters. Randomly select one angle point cloud as the reference angle point cloud, and the rest as non-reference angle point clouds; The non-reference angle point cloud and the corresponding registration transformation parameters are transformed by coordinate transformation to obtain the aligned point cloud; the reference point cloud and the aligned point cloud are fused to obtain the registered point cloud within the current scene unit.

3. The method for constructing a 3D model based on UAV mapping according to claim 1, characterized in that, The operation to obtain the initial point cloud model in S2 is as follows: All point clouds within each scene unit are registered with the preset absolute positioning reference. After coordinate transformation, a unified point cloud for each scene unit is obtained. The point clouds of adjacent scene units are unified, and overlapping regions are extracted to obtain overlapping point sets; the overlapping point sets of adjacent units are registered based on the iterative nearest point method to obtain the inter-unit transformation parameters. Randomly select one scene unit point cloud as the reference unit point cloud, and the rest as non-reference unit point clouds; The transformation parameters between the non-reference element point cloud and the corresponding element are processed by coordinate correction to obtain the aligned element point cloud; The reference unit point cloud and the aligned unit point cloud are fused to obtain the initial point cloud model.

4. In the UAV-based 3D model construction method according to claim 1, dark area brightening and noise suppression can be achieved through Gamma correction and Zero-DCE network, respectively.

5. A 3D model construction system based on UAV mapping, characterized in that, The method for constructing a 3D model based on UAV mapping as described in claim 1 is characterized by comprising: The scene unit generation module is used to perform pixel-level segmentation of the image of the area to be surveyed to obtain the land feature categories; and to perform scene unit segmentation of the land feature categories based on the land feature categories and the terrain information of the area to be surveyed to obtain several scene units. The 3D point cloud model generation module is used to acquire point cloud data for each scene unit according to the point cloud data acquisition method corresponding to the scene unit, resulting in several scene unit point clouds. After preprocessing, the point clouds of several scene units are registered within the same unit for point clouds at different angles of the same scene, resulting in several registered point clouds within the scene units. Based on the preset absolute positioning reference of the area to be mapped, all registered point clouds within the scene units are aligned to a unified coordinate system, and then inter-unit registration is performed to obtain an initial point cloud model. After surface reconstruction of the point cloud corresponding to each scene unit in the initial point cloud model, the point cloud boundary is fused to obtain a 3D point cloud model. The 3D model generation module for the survey area is used to acquire RGB images of each scene unit from multiple angles to form their respective RGB image sets. For each triangular facet of the scene unit in the 3D point cloud model, a visible image is obtained from the corresponding RGB image set. From the visible images, the best viewpoint image is selected, and the texture information of the best viewpoint image is mapped to the corresponding triangular facet to obtain the 3D model of the survey area.

6. A 3D model construction device based on UAV mapping, characterized in that, It includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the three-dimensional model construction method based on UAV mapping as described in any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, Used to store computer programs, wherein the computer programs, when executed by a processor, implement the method for constructing a 3D model based on UAV mapping as described in any one of claims 1-4.