A gaussian city surface reconstruction method with trusted partition and boundary coordination

Abstract

The application discloses a Gauss city surface reconstruction method based on trusted partition and boundary cooperation, relates to the field of three-dimensional city modeling, and solves the problem of insufficient three-dimensional city modeling precision, and comprises the following steps: S1, standard observation data sets are acquired; S2, a target city area is divided into regions, load calculation is performed in combination with the standard observation data sets, and a target sub-region is obtained; trusted views corresponding to the target sub-region are screened; S3, adjacent sub-regions are extracted for cooperative analysis, and shared observation relations between the adjacent sub-regions are recorded; S4, a global proxy Gauss field of the target city area is generated for cutting, and an initial local Gauss is obtained; pixel-level feasibility is analyzed; S5, Gauss information corresponding to an overlapping buffer zone is extracted, boundary conditions are analyzed, and Gauss optimization is performed; and S6, a three-dimensional model of the target city area is generated; and the application can effectively improve the accuracy of three-dimensional city modeling.

Description

Technical Field

[0001] This invention belongs to the field of 3D city modeling, and more specifically relates to 3D Gaussian splashing technology, specifically a Gaussian city surface reconstruction method based on reliable partitioning and boundary coordination. Background Technology

[0002] Existing 3D city modeling methods have the following drawbacks when performing city modeling: 1. Existing methods for 3D city modeling mainly divide the whole into multiple sub-regions by using a fixed grid. In practical applications, different urban areas have significant differences in building density, terrain undulation, number of available images, and structural complexity. The fixed grid can easily lead to excessive training load in some sub-regions and insufficient data in others. It may also cause the partition boundary to pass directly through high-rise buildings, complex roofs, or narrow streets, thereby increasing the risk of subsequent boundary cracks and geometric faults.

[0003] 2. The current method of selecting views mainly relies on the baseline angle between views, camera distance, and distance from the projection point to the image center. It does not fully consider dynamic semantics such as image blur, overexposure and underexposure, incident angle, occlusion, tree canopy and vehicle, glass reflection, and SfM reprojection residuals. This results in images with poor quality but satisfying geometric conditions being selected as reference views or source views. The error depth and normal further reduce the consistency constraints of urban modeling generation.

[0004] 3. Existing methods mainly train each sub-region independently, lacking a unified mechanism for sharing low-frequency colors, coordinate scales, and boundary parameters; adjacent sub-regions are prone to generating different Gaussian positions, depths, and colors for the same roof, road, or facade, resulting in boundary cracks, duplicate surfaces, color jumps, or slight scale differences after surface extraction; leading to low global synergy and poor accuracy in urban modeling.

[0005] To address this, a Gaussian method for urban surface reconstruction based on trusted partitioning and boundary coordination is proposed. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this invention is to provide a Gaussian urban surface reconstruction method based on reliable partitioning and boundary coordination, thereby improving modeling accuracy.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a Gaussian urban surface reconstruction method based on reliable partitioning and boundary coordination, the specific working process of each step of which is as follows: Step S1: Obtain spatial observation information and actual spatial distribution of the target urban area, and construct a local coordinate system for the target urban area; map the spatial observation information to the local coordinate system to obtain a standard observation dataset; Step S2: Based on the real spatial distribution and local coordinate system, the horizontal surface of the target urban area is collected to divide the area into regions, and the initial candidate regions are obtained. The load is calculated in combination with the standard observation dataset to optimize the initial candidate regions and obtain the target sub-regions. The standard observation dataset is then filtered to obtain the reliable view corresponding to the target sub-regions. Step S3: Traverse the target sub-region, extract adjacent sub-regions for collaborative analysis, obtain the overlapping buffer zone, and record the shared observation relationship between adjacent sub-regions; Step S4: Based on the standard observation dataset, generate a global proxy Gaussian field for the target urban area. Crop the global proxy Gaussian field and combine it with the standard observation dataset and shared observation relationships to generate an initial local Gaussian field. Analyze the pixel-level feasibility of the target sub-region using the initial local Gaussian field and control the geometric constraints of the target sub-region. Step S5: Initialize the local Gaussian, extract the Gaussian information corresponding to the overlapping buffer zone, analyze the boundary conditions of adjacent sub-regions, calculate the boundary compatibility loss, and optimize the Gaussian within the overlapping buffer zone; Step S6: Generate a 3D model of the target city area based on the core area Gaussian and the overlapping buffer zone Gaussian of the target sub-region.

[0008] Furthermore, the specific steps of step S1 are as follows: Step S11: Obtain the real spatial distribution of the target city, record the real spatial distribution of the target city through the world coordinate system, adjust the world coordinate system, set the origin of the coordinate system as the spatial center of the target city, and construct the local coordinate system of the target city. Step S12: Collect aerial oblique photographic images, SfM sparse points and point image observation trajectories of the target urban area to obtain spatial observation information of the target urban area. Based on the local coordinate system of the target urban area, spatially align the spatial observation information, spatially map the SfM sparse points to the local coordinate system, and record the aerial oblique photographic images and point image observation trajectories corresponding to the SfM sparse points to obtain a standard observation dataset.

[0009] Furthermore, the specific steps of step S2 are as follows: Step S21: Collect the distribution boundary of the real spatial distribution of the target urban area in the local coordinate system to obtain the horizontal surface of the target urban area. Based on the horizontal surface of the target urban area, perform rasterization processing to divide the horizontal surface of the target urban area into multiple local cells. Based on the local cells, perform preliminary regional division of the target urban area to obtain the initial candidate region. Step S22: Extract the standard observation dataset corresponding to each initial candidate region, calculate the load status of the initial candidate region, iteratively segment the initial candidate region based on the load status to obtain segmented region units, fuse the segmented region units to obtain target sub-regions; filter the standard observation dataset of each target sub-region and record the reliable view associated with the sub-region.

[0010] Furthermore, the specific steps of step S22 are as follows: Step S221: Based on the standard observation dataset, aerial oblique photogrammetry images, SfM sparse points, and point image observation trajectories of the initial candidate region are collected. The number of SfM points xsl is extracted from the SfM sparse points. Based on the SfM sparse points and point image observation trajectories, the number of matching images of the initial candidate region is counted to obtain the number of images yxs. The elevation undulation gqf of the initial candidate region is recorded using the SfM sparse points. Based on the matching images of the initial candidate region, the structural edge density bmd of the initial candidate region is extracted. The load of the initial candidate region is calculated using the number of SfM points xsl, the number of images yxs, the elevation undulation gqf, and the structural edge density bmd to obtain the load value fzz. Step S222: Traverse the initial candidate regions to obtain the load values ​​of different initial candidate regions; obtain the CPU budget of the initial candidate regions, and compare the load value with the CPU budget. If the load value is lower than the CPU budget, it indicates that the CPU budget meets the load of the initial candidate region; if the load value is higher than the CPU budget, it indicates that the CPU budget cannot meet the load of the initial candidate region, and the initial candidate region needs to be segmented; traverse and merge the segmented initial candidate regions to obtain the target sub-region.

[0011] Furthermore, the specific steps of step S222 are as follows: Centered on the current initial candidate region, the load values ​​of the initial candidate region in the horizontal direction and the load values ​​in the vertical direction are statistically analyzed, and the horizontal load variance and the vertical load variance are collected. If the horizontal load variance is greater than the vertical load variance, the current initial candidate region is bisected along the horizontal direction to obtain segmented region units; if the horizontal load variance is less than the vertical load variance, the current initial candidate region is bisected along the vertical direction to obtain segmented region units. At the same time, the load status of the segmented region units is judged, and the segmented region units are iterated until all segmented region units meet the CPU budget, and the segmented region units are statistically analyzed. The segmented region units are traversed. Taking a segmented region unit as the center, if the sum of the load values ​​of the segmented region unit and its neighboring segmented region units is lower than the CPU budget, the segmented region unit is merged with its neighboring segmented region units. At the same time, the neighboring segmented region units of the merged segmented region unit are iterated. All segmented regions are merged into a set of multiple segmented region units that meet the CPU budget, and the target sub-region is obtained.

[0012] Furthermore, the specific steps of step S22 also include: Step S223: Collect the standard observation dataset corresponding to the spatial range of the target sub-region, extract the aerial oblique photography images based on the standard observation dataset, traverse the aerial oblique photography images, and collect the image indicators of each image. The SfM points corresponding to the target sub-region are collected, the aerial oblique photogrammetry images corresponding to the SfM points are statistically analyzed, the images are traversed, and the geometric parameters of the images are recorded. Step S224: Multiply the image index and the geometric index to obtain the overall view confidence level zkx; Sort the view comprehensive confidence scores zkx in descending order, extract the image views corresponding to the top cks largest view comprehensive confidence scores as the confidence views of the sub-region association, take the image view corresponding to the maximum value of the view comprehensive confidence score zkx in the confidence views as the reference view, and take the remaining image views as the source views.

[0013] Furthermore, the specific steps of step S3 are as follows: Step S31: Traverse the target sub-region, collect the elevation undulation gqf and structural edge density bmd of the target sub-region, and set the upper limit of the buffer band width b of the overlapping buffer band. max With the lower limit of the buffer band width b min The settings are configured, and the overlap buffer zone width hkd is calculated by combining the elevation undulation gqf of the target sub-region with the structural edge density bmd. Traverse the target sub-region and collect the overlapping buffer zone of the target region and its adjacent sub-regions; Step S32: Extract the SfM points within the overlapping buffer zone to obtain the boundary points within the overlapping buffer zone. Collect the distances from the boundary points to the target region and its adjacent sub-regions. If the distance from the boundary point to the target region is less than or equal to the distance from the boundary point to the adjacent sub-region, set the target region as the primary sub-region of the boundary point and the corresponding adjacent sub-region as the secondary sub-region of the boundary point. If the distance from the boundary point to the target region is greater than the distance from the boundary point to the adjacent sub-region, set the target region as the secondary sub-region of the boundary point and the corresponding adjacent sub-region as the primary sub-region of the boundary point. Traverse all overlapping buffer zones and record the boundary points within the overlapping buffer zones between adjacent sub-regions, as well as the primary and secondary sub-regions of their boundary points, to obtain the shared observation relationship between adjacent sub-regions.

[0014] Furthermore, the specific steps of step S4 are as follows: Step S41: Obtain the global proxy Gaussian field of the target urban area, map the global proxy Gaussian field to the local coordinate system of the target urban area, collect the spatial distribution of the target sub-region in the local coordinate system, combine the mapping result of the global proxy Gaussian field in the local coordinate system, collect the global proxy Gaussian field of the corresponding spatial location of the target sub-region, combine the standard observation dataset corresponding to the target sub-region with the shared observation relationship between adjacent sub-regions, superimpose the SfM point of the target sub-region with the proxy Gaussian field, record the boundary points of the overlapping buffer zone through the shared observation relationship between adjacent sub-regions, align the boundaries between different target sub-regions, and generate an initial local Gaussian. Step S42: Obtain the reliable view associated with the sub-region, collect data on the reference view and the source view based on the reliable view, backproject the reference view to the source view, then backproject the depth of the source view to the reference view, collect the projection error, reduce the weight of the view based on the projection error, and optimize the generation of the initial local Gaussian.

[0015] Furthermore, the specific steps of step S5 are as follows: Step S51: Based on the initial local Gaussian of the target sub-region, collect the core region Gaussian and the overlapping buffer zone Gaussian of the target sub-region; mask the core region Gaussian and retain the overlapping buffer zone Gaussian; read the shared camera, shared ray, and boundary anchor point of the adjacent sub-region based on the overlapping buffer zone Gaussian; render the depth, normal, low-frequency color, and opacity of the adjacent sub-region under the shared camera, and calculate the boundary coordination loss value ssz; Step S52: Obtain the shared observation relationship between adjacent sub-regions, extract the primary and secondary sub-regions of the overlapping buffer zone, use the primary sub-region as the anchor point and the secondary sub-region as the update subject, adjust the rendering data of Gaussian in different sub-regions within the overlapping buffer zone, calculate the boundary coordination loss value in real time, and when the boundary coordination loss value is the minimum, record the corresponding rendering data to optimize Gaussian in the overlapping buffer zone.

[0016] Furthermore, the specific steps of step S6 are as follows: Step S61: Traverse the target sub-region, record the core area Gaussian of the corresponding target sub-region, traverse the overlapping buffer zone based on the shared observation relationship between adjacent sub-regions, record the corresponding Gaussian of the overlapping buffer zone region, and statistically analyze the core area Gaussian and the overlapping buffer zone Gaussian to generate a three-dimensional model of the target city. Step S62: Statistically analyze the Gaussian within the generated 3D model, record the standard observation dataset at the spatial location of the Gaussian, perform a backtracking process on the Gaussian generation based on the standard observation dataset, and record all processing steps; associate and store the processed data corresponding to the Gaussian with the standard observation dataset and the specific Gaussian.

[0017] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. The present invention divides and integrates the target urban area. It iteratively divides the sub-regions by comprehensively considering sparse points, available images, depth undulations, structural edges, and GPU budget. This reduces local memory overflow caused by fixed grids, ensures the stability of sub-region generation, and simultaneously traverses and integrates the divided sub-regions to eliminate idle sub-regions and improve the load utilization of sub-regions.

[0018] 2. This invention analyzes the image status of sub-regions (including sharpness, exposure, occlusion, incident angle, baseline angle, center distance, and reprojection residual) to filter images, ensuring the effectiveness of the sub-region view and reducing the contamination of depth and normal estimation by blurred, reflective, canopy, and edge distortion images; at the same time, it judges the projection error based on the sub-region view to ensure the accuracy and stability of Gaussian generation in the sub-region.

[0019] 3. This invention performs overlapping buffer zone analysis on adjacent sub-regions and coordinates the processing of sub-regions of the target city. Based on the positional relationship, it sets primary / secondary sub-regions for Gaussian sigma within the overlapping buffer zone to ensure the accuracy of Gaussian generation and global coordination. At the same time, it analyzes the rendering differences between adjacent sub-regions through overlapping buffer zones and adjusts and optimizes boundary losses (such as cracks, repeated surfaces, depth jumps, and color jumps) to ensure the accuracy of boundary generation. Attached Figure Description

[0020] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0021] Figure 1 This is a schematic diagram of the method of the present invention; Figure 2 This is a schematic diagram of the region division in this invention; Figure 3 This is a schematic diagram of the coordinated control of the present invention; Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0023] This application provides a Gaussian urban surface reconstruction method based on trusted partitioning and boundary coordination. The executing entity of this trusted partitioning and boundary coordination Gaussian urban surface reconstruction method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the trusted partitioning and boundary coordination Gaussian urban surface reconstruction method can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster.

[0024] Reference Figure 1 The diagram shown is a flowchart illustrating a Gaussian urban surface reconstruction method based on trusted partitioning and boundary coordination, according to an embodiment of the present invention. In this embodiment, the Gaussian urban surface reconstruction method based on trusted partitioning and boundary coordination includes: Step S1: Obtain the spatial observation information and actual spatial distribution of the target urban area. Based on the spatial distribution of the target urban area and the world coordinate system, construct the local coordinate system of the target urban area. Map the spatial observation information of the target urban area to the local coordinate system to obtain the standard observation dataset. It should be noted that: the space observation information includes aerial oblique photogrammetry, SfM sparse points, and point image observation trajectories; Aerial oblique photogrammetry: Images acquired by a five-lens oblique camera mounted on a UAV / manned aerial survey aircraft, including vertical downward view images and four oblique view images (front / back / left / right), covering building facades, roofs, roads, and terrain, serving as the original visual data source for urban 3D reconstruction. SfM sparse points: A set of three-dimensional feature points obtained from multiple aerial oblique photographs; Point image observation trajectory: Completely record the matching relationship of which images each SfM sparse point was observed by.

[0025] Step S11: Obtain the real spatial distribution of the target city, record the real spatial distribution of the target city through the world coordinate system, adjust the world coordinate system, set the origin of the coordinate system as the spatial center of the target city, and construct the local coordinate system of the target city. Step S12: Collect aerial oblique photographic images, SfM sparse points and point image observation trajectories of the target urban area to obtain spatial observation information of the target urban area. Based on the local coordinate system of the target urban area, spatially align the spatial observation information, spatially map the SfM sparse points to the local coordinate system, and record the aerial oblique photographic images and point image observation trajectories corresponding to the SfM sparse points to obtain a standard observation dataset.

[0026] Step S2: Based on the actual spatial distribution and local coordinate system of the target urban area, collect the horizontal surface of the target urban area, perform preliminary regional division of the target urban area to obtain initial candidate areas, combine the standard observation dataset to perform load calculation on the initial candidate areas, optimize the initial candidate areas to obtain target sub-regions; filter the standard observation dataset from the target sub-regions to obtain the reliable view corresponding to the target sub-regions; Step S21: Collect the distribution boundary of the real spatial distribution of the target urban area in the local coordinate system to obtain the horizontal surface of the target urban area. Based on the horizontal surface of the target urban area, perform rasterization processing to divide the horizontal surface of the target urban area into multiple local cells. Based on the local cells, perform preliminary regional division of the target urban area to obtain the initial candidate region. Step S22: Extract the standard observation dataset corresponding to each initial candidate region, calculate the load status of the initial candidate region, iteratively segment the initial candidate region based on the load status to obtain segmented region units, fuse the segmented region units to obtain target sub-regions; filter the standard observation dataset of each target sub-region and record the reliable view associated with the sub-region.

[0027] Please see Figure 2 Step S221: Based on the standard observation dataset, aerial oblique photogrammetry images, SfM sparse points, and point image observation trajectories of the initial candidate region are collected. The number of SfM points xsl is extracted from the SfM sparse points. Based on the SfM sparse points and point image observation trajectories, the number of matching images of the initial candidate region is counted to obtain the number of images yxs. The elevation undulation gqf of the initial candidate region is recorded using the SfM sparse points. Based on the matching images of the initial candidate region, the structural edge density bmd of the initial candidate region is extracted. The load of the initial candidate region is calculated using the number of SfM points xsl, the number of images yxs, the elevation undulation gqf, and the structural edge density bmd to obtain the load value fzz. ; Where w1 to w4 are weights, representing the CPU resource consumption required to process different SfM point numbers xsl, image number yxs, elevation undulation gqf, and structure edge density bmd. It should be noted that: SfM sparse points are a set of three-dimensional feature points, where each point exists in a three-dimensional coordinate system, and the elevation fluctuation is the difference between the maximum and minimum values ​​of the three-dimensional coordinate height; structural edge density refers to the proportion of entity pixels in the edge part displayed by the image in the overall image.

[0028] Step S222: Traverse the initial candidate regions to obtain the load values ​​of different initial candidate regions; obtain the CPU budget of the initial candidate regions, and compare the load value with the CPU budget. If the load value is lower than the CPU budget, it indicates that the CPU budget meets the load of the initial candidate region; if the load value is higher than the CPU budget, it indicates that the CPU budget cannot meet the load of the initial candidate region, and the initial candidate region needs to be segmented; traverse and merge the segmented initial candidate regions to obtain the target sub-region.

[0029] The initial candidate regions are segmented and merged as follows: Centered on the current initial candidate region, the load values ​​of the initial candidate region in the horizontal direction and the load values ​​in the vertical direction are statistically analyzed, and the horizontal load variance and the vertical load variance are collected. If the horizontal load variance is greater than the vertical load variance, the current initial candidate region is bisected along the horizontal direction to obtain segmented region units; if the horizontal load variance is less than the vertical load variance, the current initial candidate region is bisected along the vertical direction to obtain segmented region units. At the same time, the load status of the segmented region units is judged, and the segmented region units are iterated until all segmented region units meet the CPU budget, and the segmented region units are statistically analyzed. The segmented region units are traversed. Taking a segmented region unit as the center, if the sum of the load values ​​of the segmented region unit and its neighboring segmented region units is lower than the CPU budget, the segmented region unit is merged with its neighboring segmented region units. At the same time, the neighboring segmented region units of the merged segmented region unit are iterated. All segmented regions are merged into a set of multiple segmented region units that meet the CPU budget, and the target sub-region is obtained.

[0030] Step S223: Collect the standard observation dataset corresponding to the spatial range of the target sub-region, extract the aerial oblique photography images based on the standard observation dataset, traverse the aerial oblique photography images, and collect the image indicators of each image, including sharpness yxz1, exposure quality yxz2, dynamic occlusion ratio yxz3 and static semantic ratio yxz4. The SfM points corresponding to the target sub-region are collected, the aerial oblique photogrammetry images corresponding to the SfM points are statistically analyzed, the images are traversed, and the geometric parameters of the images are recorded, including baseline angle jhz1, surface incident angle jhz2, normalized distance from the projection point to the image center jhz3, reprojection residual jhz4, and local coverage jhz5. It should be noted that: the baseline angle is the triangulated angle between the current camera and other co-viewing cameras; the surface incident angle is the angle between the camera's line of sight and the normal to the object's surface; the normalized distance from the projection point to the image center refers to the normalized distance between the projection of the SfM point and the center of the image; the reprojection residual refers to the error between the position of the projected pixel and the actual position of the point on the image; and the local coverage refers to the coverage percentage on the image (such as local occlusion by trees / structures).

[0031] Step S224: Multiply the image index and the geometric index to obtain the overall view confidence level zkx; ; Where qz1(i) is the corresponding weight of the image index, and qz2(j) is the corresponding weight of the geometric index; Sort the view comprehensive confidence scores zkx in descending order, extract the image views corresponding to the top cks largest view comprehensive confidence scores as the confidence views of the sub-region association, take the image view corresponding to the maximum value of the view comprehensive confidence scores zkx in the confidence views as the reference view, and take the remaining image views as the source views. It should be noted that cks is usually 4, which means that there is usually 1 reference view and 3 source views in the trusted view.

[0032] Step S3: Traverse the target sub-region, extract adjacent sub-regions for collaborative analysis, obtain the overlapping buffer zone, establish the primary and secondary sub-regions based on the boundary points of the adjacent sub-regions within the overlapping buffer zone, and record the shared observation relationship between adjacent sub-regions. Step S31: Traverse the target sub-region, collect the elevation undulation gqf and structural edge density bmd of the target sub-region, and set the upper limit of the buffer band width b of the overlapping buffer band. max With the lower limit of the buffer band width b min The settings are configured, and the overlap buffer zone width hkd is calculated by combining the elevation undulation gqf of the target sub-region with the structural edge density bmd. ; Where b0 is the initial width setting, and k d With k e Set weights for width collaboration; Traverse the target sub-region and collect the overlapping buffer zone of the target region and its adjacent sub-regions; Step S32: Extract the SfM points within the overlapping buffer zone to obtain the boundary points within the overlapping buffer zone. Collect the distances from the boundary points to the target region and its adjacent sub-regions. If the distance from the boundary point to the target region is less than or equal to the distance from the boundary point to the adjacent sub-region, set the target region as the primary sub-region of the boundary point and the corresponding adjacent sub-region as the secondary sub-region of the boundary point. If the distance from the boundary point to the target region is greater than the distance from the boundary point to the adjacent sub-region, set the target region as the secondary sub-region of the boundary point and the corresponding adjacent sub-region as the primary sub-region of the boundary point. Traverse all overlapping buffer zones and record the boundary points within the overlapping buffer zones between adjacent sub-regions, as well as the primary and secondary sub-regions of their boundary points, to obtain the shared observation relationship between adjacent sub-regions.

[0033] Step S4: Based on the standard observation dataset, generate a global proxy Gaussian field for the target urban area. Clip the global proxy Gaussian field through the target sub-region. Combine the standard observation dataset corresponding to the target sub-region with the shared observation relationship between adjacent sub-regions to generate an initial local Gaussian field. Analyze the pixel-level feasibility of the target sub-region using the initial local Gaussian field and control the geometric constraints of the target sub-region. Step S41: Obtain the global proxy Gaussian field of the target urban area, map the global proxy Gaussian field to the local coordinate system of the target urban area, collect the spatial distribution of the target sub-region in the local coordinate system, combine the mapping result of the global proxy Gaussian field in the local coordinate system, collect the global proxy Gaussian field of the corresponding spatial location of the target sub-region, combine the standard observation dataset corresponding to the target sub-region with the shared observation relationship between adjacent sub-regions, superimpose the SfM point of the target sub-region with the proxy Gaussian field, record the boundary points of the overlapping buffer zone through the shared observation relationship between adjacent sub-regions, align the boundaries between different target sub-regions, and generate an initial local Gaussian. Step S42: Obtain the reliable view associated with the sub-region, collect data on the reference view and the source view based on the reliable view, backproject the reference view to the source view, backproject the depth of the source view to the reference view, collect the projection error, reduce the weight of the view based on the projection error, and optimize the generation of the initial local Gaussian.

[0034] Please see Figure 3 Step S5: Initialize the local Gaussian, mask the core Gaussian region of the target sub-region, extract the Gaussian information corresponding to the overlapping buffer zone, analyze the boundary conditions of adjacent sub-regions, calculate the boundary coordination loss, and optimize the Gaussian within the overlapping buffer zone based on the boundary coordination loss. Step S51: Based on the initial local Gaussian of the target sub-region, collect the core region Gaussian and the overlapping buffer zone Gaussian of the target sub-region; mask the core region Gaussian and retain the overlapping buffer zone Gaussian; read the shared camera, shared ray, and boundary anchor point of the adjacent sub-region based on the overlapping buffer zone Gaussian; render the depth, normal, low-frequency color, and opacity of the adjacent sub-region under the shared camera, and calculate the boundary coordination loss value ssz; ; Where: sd1 and sd2 are the depth values ​​of Gaussian in adjacent sub-regions, fxj1 and fxj2 are the normal angles, jq1 and jq2 are the weights of Gaussian in the geometric direction, and their sum is 1; dpy1 and dpy2 are the RGB color values ​​of Gaussian in adjacent sub-regions, tmd1 and tmd2 are the opacity, and yq1 and yq2 are the weights of Gaussian in the color direction, and their sum is 1; Step S52: Obtain the shared observation relationship between adjacent sub-regions, extract the primary and secondary sub-regions of the overlapping buffer zone, use the primary sub-region as the anchor point and the secondary sub-region as the update subject, adjust the rendering data of Gaussian in different sub-regions within the overlapping buffer zone, calculate the boundary coordination loss value in real time, and when the boundary coordination loss value is the minimum, record the corresponding rendering data to optimize Gaussian in the overlapping buffer zone.

[0035] Step S6: Based on the core area Gaussian and overlapping buffer zone Gaussian of the target sub-region, merge the sub-regions of the target urban area to generate a 3D model of the target urban area and output the 3D surface mesh of urban buildings, roads and ground surface. Step S61: Traverse the target sub-region, record the core area Gaussian of the corresponding target sub-region, traverse the overlapping buffer zone based on the shared observation relationship between adjacent sub-regions, record the corresponding Gaussian of the overlapping buffer zone region, and statistically analyze the core area Gaussian and the overlapping buffer zone Gaussian to generate a three-dimensional model of the target city. Step S62: Statistically analyze the Gaussians within the generated 3D model, record the standard observation dataset at the spatial location of the Gaussians, perform backtracking processing on the Gaussian generation based on the standard observation dataset, and record all processing steps; associate and store the processed data corresponding to the Gaussians with the standard observation dataset and the specific Gaussians. It should be noted that the Gaussian generation is backed up based on the standard observation dataset. For example, if the view confidence of the standard observation dataset is low, the Gaussian generation is supplemented with views.

[0036] Compared to the problems described in the background art, the present invention divides and integrates the target urban area region. It iteratively divides sub-regions by comprehensively considering sparse points, available imagery, depth undulations, structural edges, and GPU budget, reducing local memory overflow caused by fixed grids and ensuring the stability of sub-region generation. Simultaneously, it traverses and integrates the divided sub-regions, eliminating idle sub-regions and improving the load utilization of sub-regions. Furthermore, the present invention analyzes the image status of sub-regions (including sharpness, exposure, occlusion, incident angle, baseline angle, center distance, and reprojection residuals) to filter images, ensuring the view validity of sub-regions and reducing the contamination of depth and normal estimation by blurred, reflective, canopy, and edge distortion images. The invention employs several techniques: first, it analyzes the overlapping buffer zones of adjacent sub-regions to collaboratively process sub-regions within the target city area; second, it assigns primary / secondary sub-regions to Gaussians within the overlapping buffer zones based on their positional relationships, ensuring the accuracy and global synergy of Gaussian generation; and third, it analyzes the rendering differences between adjacent sub-regions through overlapping buffer zones to regulate and optimize boundary losses (such as cracks, repeated surfaces, depth jumps, and color jumps), ensuring the accuracy of boundary generation. Therefore, the Gaussian urban surface reconstruction method based on reliable partitioning and boundary collaboration provided by this invention can improve the accuracy of 3D city modeling.

[0037] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0038] Finally, it should be noted that deleting any one of the above embodiments does not affect the technical solutions of other embodiments. The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A Gaussian urban surface reconstruction method based on reliable partitioning and boundary coordination, characterized in that, include: Step S1: Obtain spatial observation information and actual spatial distribution of the target urban area, and construct a local coordinate system for the target urban area; By mapping spatial observation information to a local coordinate system, a standard observation dataset is obtained; Step S2: Based on the real spatial distribution and local coordinate system, the horizontal surface of the target urban area is collected to divide the area into regions, and the initial candidate regions are obtained. The load is calculated in combination with the standard observation dataset to optimize the initial candidate regions and obtain the target sub-regions. The standard observation dataset is then filtered to obtain the reliable view corresponding to the target sub-regions. Step S3: Traverse the target sub-region, extract adjacent sub-regions for collaborative analysis, obtain the overlapping buffer zone, and record the shared observation relationship between adjacent sub-regions; Step S4: Obtain the global proxy Gaussian field of the target urban area, map the global proxy Gaussian field to the local coordinate system of the target urban area, collect the spatial distribution of the target sub-region in the local coordinate system, combine the mapping result of the global proxy Gaussian field in the local coordinate system, collect the global proxy Gaussian field of the corresponding spatial location of the target sub-region, combine the standard observation dataset corresponding to the target sub-region with the shared observation relationship between adjacent sub-regions, superimpose the SfM points of the target sub-region with the proxy Gaussian field, record the boundary points of the overlapping buffer zone through the shared observation relationship between adjacent sub-regions, align the boundaries between different target sub-regions, and generate an initial local Gaussian. Obtain a reliable view associated with the sub-region, collect data on the reference view and the source view based on the reliable view, backproject the reference view to the source view, then backproject the depth of the source view back to the reference view, collect the projection error, reduce the weight of the view based on the projection error, and optimize the generation of the initial local Gaussian. Step S5: Based on the initial local Gaussian of the target sub-region, collect the core Gaussian and overlapping buffer band Gaussian of the target sub-region; mask the core Gaussian and retain the overlapping buffer band Gaussian; read the shared camera, shared ray, and boundary anchor points of adjacent sub-regions based on the overlapping buffer band Gaussian; render the depth, normal, low-frequency color, and opacity of adjacent sub-regions under the shared camera, and calculate the boundary coordination loss value ssz. ； Where: sd1 and sd2 are the depth values ​​of Gaussian in adjacent sub-regions, fxj1 and fxj2 are the normal angles, jq1 and jq2 are the weights of Gaussian in the geometric direction, and their sum is 1; dpy1 and dpy2 are the RGB color values ​​of Gaussian in adjacent sub-regions, tmd1 and tmd2 are the opacity, and yq1 and yq2 are the weights of Gaussian in the color direction, and their sum is 1; The shared observation relationship between adjacent sub-regions is obtained, the primary and secondary sub-regions of the overlapping buffer zone are extracted, the primary sub-region is used as the anchor point and the secondary sub-region is used as the update subject, the rendering data of Gaussian in different sub-regions in the overlapping buffer zone is adjusted, the boundary coordination loss value is calculated in real time, and when the boundary coordination loss value is the minimum, the corresponding rendering data is recorded to optimize Gaussian in the overlapping buffer zone. Step S6: Generate a 3D model of the target city area based on the core area Gaussian and the overlapping buffer zone Gaussian of the target sub-region.

2. The Gaussian urban surface reconstruction method based on trusted partitioning and boundary coordination according to claim 1, characterized in that, The specific steps of step S1 are as follows: Step S11: Obtain the real spatial distribution of the target city, record the real spatial distribution of the target city through the world coordinate system, adjust the world coordinate system, set the origin of the coordinate system as the spatial center of the target city, and construct the local coordinate system of the target city. Step S12: Collect aerial oblique photographic images, SfM sparse points and point image observation trajectories of the target urban area to obtain spatial observation information of the target urban area. Based on the local coordinate system of the target urban area, spatially align the spatial observation information, spatially map the SfM sparse points to the local coordinate system, and record the aerial oblique photographic images and point image observation trajectories corresponding to the SfM sparse points to obtain a standard observation dataset.

3. The Gaussian urban surface reconstruction method based on trusted partitioning and boundary coordination according to claim 1, characterized in that, The specific steps of step S2 are as follows: Step S21: Collect the distribution boundary of the real spatial distribution of the target urban area in the local coordinate system to obtain the horizontal surface of the target urban area. Based on the horizontal surface of the target urban area, perform rasterization processing to divide the horizontal surface of the target urban area into multiple local cells, perform preliminary regional division of the target urban area, and obtain the initial candidate region. Step S22: Extract the standard observation dataset corresponding to each initial candidate region, calculate the load status of the initial candidate region, iteratively segment the initial candidate region to obtain segmented region units, fuse the segmented region units to obtain target sub-regions; filter the standard observation dataset of each target sub-region and record the reliable view associated with the sub-region.

4. The Gaussian urban surface reconstruction method based on trusted partitioning and boundary coordination according to claim 3, characterized in that, The specific steps of step S22 are as follows: Step S221: Based on the standard observation dataset, aerial oblique photogrammetry images, SfM sparse points, and point image observation trajectories of the initial candidate region are collected. The number of SfM points xsl is extracted from the SfM sparse points. Based on the SfM sparse points and point image observation trajectories, the number of matching images of the initial candidate region is counted to obtain the number of images yxs. The elevation undulation gqf of the initial candidate region is recorded using the SfM sparse points. Based on the matching images of the initial candidate region, the structural edge density bmd of the initial candidate region is extracted. The load of the initial candidate region is calculated using the number of SfM points xsl, the number of images yxs, the elevation undulation gqf, and the structural edge density bmd to obtain the load value fzz. Step S222: Traverse the initial candidate regions to obtain the load values ​​of different initial candidate regions; obtain the CPU budget of the initial candidate regions, and compare the load value with the CPU budget. If the load value is lower than the CPU budget, it indicates that the CPU budget meets the load of the initial candidate region; if the load value is higher than the CPU budget, it indicates that the CPU budget cannot meet the load of the initial candidate region, and the initial candidate region needs to be segmented; traverse and merge the segmented initial candidate regions to obtain the target sub-region.

5. The Gaussian urban surface reconstruction method based on trusted partitioning and boundary coordination according to claim 4, characterized in that, The specific steps of step S222 are as follows: Centered on the current initial candidate region, the load values ​​of the initial candidate region in the horizontal direction and the load values ​​in the vertical direction are statistically analyzed, and the horizontal load variance and the vertical load variance are collected. If the horizontal load variance is greater than the vertical load variance, the current initial candidate region is bisected along the horizontal direction to obtain segmented region units; if the horizontal load variance is less than the vertical load variance, the current initial candidate region is bisected along the vertical direction to obtain segmented region units. At the same time, the load status of the segmented region units is judged, and the segmented region units are iterated until all segmented region units meet the CPU budget, and the segmented region units are statistically analyzed. The segmented region units are traversed. Taking a segmented region unit as the center, if the sum of the load values ​​of the segmented region unit and its neighboring segmented region units is lower than the CPU budget, the segmented region unit is merged with its neighboring segmented region units. At the same time, the neighboring segmented region units of the merged segmented region unit are iterated, and all segmented regions are merged into a set of multiple segmented region units that meet the CPU budget, thus obtaining the target sub-region.

6. The Gaussian urban surface reconstruction method based on trusted partitioning and boundary coordination according to claim 3, characterized in that, The specific steps of step S22 also include: Step S223: Collect the standard observation dataset corresponding to the spatial range of the target sub-region, extract the aerial oblique photography images based on the standard observation dataset, traverse the aerial oblique photography images, and collect the image indicators of each image. The SfM points corresponding to the target sub-region are collected, the aerial oblique photogrammetry images corresponding to the SfM points are statistically analyzed, the images are traversed, and the geometric parameters of the images are recorded. Step S224: Multiply the image index and the geometric index to obtain the overall view confidence level zkx; Sort the view comprehensive confidence scores zkx in descending order, extract the image views corresponding to the top cks largest view comprehensive confidence scores as the confidence views of the sub-region association, take the image view corresponding to the maximum value of the view comprehensive confidence score zkx in the confidence views as the reference view, and take the remaining image views as the source views.

7. The Gaussian urban surface reconstruction method based on trusted partitioning and boundary coordination according to claim 1, characterized in that, The specific steps of step S3 are as follows: Step S31: Traverse the target sub-region, collect the elevation undulation gqf and structural edge density bmd of the target sub-region, and set the upper limit of the buffer band width b of the overlapping buffer band. max With the lower limit of the buffer band width b min The settings are configured, and the overlap buffer zone width hkd is calculated by combining the elevation undulation gqf of the target sub-region with the structural edge density bmd. Traverse the target sub-region and collect the overlapping buffer zone of the target region and its adjacent sub-regions; Step S32: Extract the SfM points within the overlapping buffer zone to obtain the boundary points within the overlapping buffer zone. Collect the distances from the boundary points to the target region and its adjacent sub-regions. If the distance from the boundary point to the target region is less than or equal to the distance from the boundary point to the adjacent sub-region, set the target region as the primary sub-region of the boundary point and the corresponding adjacent sub-region as the secondary sub-region of the boundary point. If the distance from the boundary point to the target region is greater than the distance from the boundary point to the adjacent sub-region, set the target region as the secondary sub-region of the boundary point and the corresponding adjacent sub-region as the primary sub-region of the boundary point. Traverse all overlapping buffer zones and record the boundary points within the overlapping buffer zones between adjacent sub-regions, as well as the primary and secondary sub-regions of their boundary points, to obtain the shared observation relationship between adjacent sub-regions.

8. The Gaussian urban surface reconstruction method based on trusted partitioning and boundary coordination according to claim 1, characterized in that, The specific steps of step S6 are as follows: Step S61: Traverse the target sub-region, record the core area Gaussian of the corresponding target sub-region, traverse the overlapping buffer zone based on the shared observation relationship between adjacent sub-regions, record the corresponding Gaussian of the overlapping buffer zone region, and statistically analyze the core area Gaussian and the overlapping buffer zone Gaussian to generate a three-dimensional model of the target city. Step S62: Statistically analyze the Gaussian within the generated 3D model, record the standard observation dataset at the spatial location of the Gaussian, perform a backtracking process on the Gaussian generation based on the standard observation dataset, and record all processing steps; associate and store the processed data corresponding to the Gaussian with the standard observation dataset and the specific Gaussian.

Images