A regional adaptive DEM data generation method based on tilt model data

Abstract

The application provides a regional adaptive DEM data generation method based on tilt model data, ground points are identified based on a deep learning model and a CSF filtering algorithm respectively according to the tilt model data, full automation of ground point identification is realized, the identification results of the two are fused and complemented, and the identification precision is improved; in addition, in the process of identifying the ground points by using the CSF filtering algorithm, the region to be identified is divided into blocks, different CSF filtering parameters are set according to the terrain type of each block, and in the process of Kriging interpolation of the ground elevation points, different Kriging interpolation parameters are also set according to the terrain features of each block, the setting of adaptive parameters in different regions is realized, compared with the use of fixed parameters in different regions, the identification precision of the ground points and the precision of generated terrain grid data are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of surveying and mapping, more particularly to a regional adaptive DEM data generation method based on tilt model data. BACKGROUND

[0002] With the rapid development of low-altitude economy, digital twin city, real scene three-dimensional and other industries, the demand for large-scale, high-precision and high-timeliness DEM data production is steadily increasing. High-precision DEM data is the basic data and prerequisite for building high-precision and high-simulation digital twin cities.

[0003] Currently, the main production methods of high-precision DEM data for urban scenes include satellite remote sensing stereo image pair method, large-scale topographic map vectorization and interpolation method, ground field elevation point measurement, unmanned aerial photography measurement and Lidar technology. Among them, the satellite remote sensing stereo image pair method has low DEM precision due to the limitation of satellite image resolution, which cannot meet the requirements of urban three-dimensional scene construction. The precision of unmanned aerial photography measurement and Lidar technology in building-dense areas and water areas is low, usually needs to be combined with software indoor manual editing and correction, and cannot realize the full automation of DEM data production. The large-scale topographic map vectorization and interpolation method and the ground field elevation point measurement method are difficult to meet the requirements of high-precision and high-timeliness digital twin scene construction in terms of artificial cost, data production cycle and timeliness.

[0004] With the rapid development and progress of unmanned aerial photography technology, oblique photogrammetry and multi-view image three-dimensional reconstruction technology, oblique model data (OSGB) has gradually developed into the most common three-dimensional model data. Due to its low data acquisition cost, mature three-dimensional modeling technology, high model geometric precision, strong model texture authenticity, support for multi-detail level expression and lightweight, etc., it has been widely used in urban three-dimensional modeling, topographic surveying and mapping, environmental monitoring and emergency support, cultural heritage protection and other fields, and plays an important role.

[0005] However, there is still a lack of technology and method for fully automated and high-precision DEM data production based on oblique model data (OSGB). The development of a method for directly producing high-precision DEM data based on oblique model data has important significance in fully utilizing the value of existing real scene three-dimensional data, reducing the repeated production of basic surveying and mapping data, and meeting the high-timeliness requirements of digital twin scenes for terrain data. SUMMARY

[0006] The present application provides a regional adaptive DEM data generation method based on tilt model data to overcome the problems of low precision and inability to achieve full automation in generating DEM data in the prior art.

[0007] The application provides a regional adaptive DEM data generation method based on tilt model data, comprising:

[0008] Step S1, loading tilt model data of a specified range, generating three-dimensional color point cloud data by three-dimensional node sampling on the tilt model data, and generating true orthophoto data by z-axis vertical sampling on the tilt model data;

[0009] Step S2, denoising the three-dimensional color point cloud data based on a clustering algorithm;

[0010] Step S3, performing semantic segmentation on the denoised three-dimensional color point cloud data based on a deep learning model, and identifying first ground points;

[0011] Step S4, performing Kriging interpolation on ground elevation points to generate preliminary terrain grid data, and performing grid division on the preliminary terrain grid data according to a preset size to obtain multiple pieces of terrain grid data;

[0012] Step S5, extracting terrain features of each piece of the terrain grid data;

[0013] Step S6, performing grid division on the denoised three-dimensional color point cloud data according to a preset size to obtain multiple pieces of three-dimensional color point cloud data, and performing merging on second ground points of each piece of the three-dimensional color point cloud data filtered based on a CSF filtering algorithm, wherein parameters of the corresponding CSF filtering algorithm are set according to terrain features corresponding to each piece of the three-dimensional color point cloud data;

[0014] Step S7, performing overlay analysis on the first ground points based on the second ground points, and correcting the first ground points;

[0015] Step S8, extracting vector data of building surfaces, forest surfaces and grassland surfaces from the true orthophoto data based on a ViTAE network model;

[0016] Step S9, performing misclassification correction on the ground points corrected in step S7 based on the vector data of the building surfaces, the forest surfaces and the grassland surfaces;

[0017] Step S10, setting Kriging interpolation parameters according to terrain features of a grid based on the ground points corrected in step S9, performing Kriging interpolation processing on ground elevation points of different grids to generate terrain grid data of each grid, merging terrain grid data of multiple grids, and outputting high-precision DEM data of a specified range.

[0018] The application provides a regional adaptive DEM data generation method based on tilt model data. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A regional adaptive DEM data generation method based on tilt model data is provided for the application. DETAILED DESCRIPTION

[0020] To make the objectives, technical solutions and advantages of the embodiments of the application clearer, the technical solutions in the embodiments of the application will be described below in detail with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are some but not all of the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the application. In addition, the technical features in each embodiment or in a single embodiment provided by the application can be combined with each other at will to form a feasible technical solution, and the combination is not restricted by the order of steps and / or structure mode, but should be based on the fact that a person of ordinary skill in the art can realize it. When the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist and is not within the protection scope of the application.

[0021] At present, the production of urban high-precision DEM data is mainly in the form of manual data processing after field data collection. This method has high internal and external measurement costs, low automation degree and long data production cycle, and is difficult to meet the requirements of rapid construction and updating of urban three-dimensional scenes. On the other hand, oblique photography model (OSGB) data production technology is mature, and the modeling cost is low, which has become the most widely used real scene three-dimensional data. However, there is still a lack of technology for automatically producing and deriving high-precision DEM data based on oblique photography model, which limits the application value of oblique photography model (OSGB) data in digital twin city, intelligent transportation and low-altitude economy.

[0022] The present application mainly solves the problem of automatically producing high-precision DEM data based on the above-mentioned oblique photography model (OSGB), and simultaneously solves the problem that in a large-scale urban scene, the terrain types of steep slopes, gentle slopes and flat ground coexist, and it is difficult to simultaneously meet the production requirements of all regional terrains by using a single terrain production algorithm parameter. A regional adaptive high-precision DEM generation method based on oblique model data is proposed. First, a variety of deep learning algorithms are used to perform three-dimensional scene element semantic segmentation and ground point extraction on the point cloud data derived from the oblique model and the true orthographic image data, to realize fully automated high-precision DEM generation based on oblique model data, and to reduce the cycle and cost of high-precision DEM data production. Second, for large-scale terrain generation, the different terrain regions of steep slopes, gentle slopes and flat ground are divided by using the slope calculation and statistical analysis based on the pre-generated DEM data, to realize the self-adaptation of the DEM generation algorithm parameters for different terrain regions, so as to improve the pertinence, adaptability and precision of large-scale terrain generation, and to meet the demand of various industries for fast production of large-scene high-precision DEM.

[0023] The present application proposes a method for automatically generating regional adaptive DEM data based on oblique model data (OSGB). In this method, the oblique model data is first preprocessed, the point cloud data is segmented by an AI deep learning model, and the building and vegetation data extracted from the true orthographic image (TDOM) are analyzed and corrected based on the point cloud classification results, to generate high-precision and high-reliability ground elevation point data. Then, the regional slope calculation and division are combined, different adaptive parameters are used for block DEM data generation, and finally the high-precision DEM data is output. The present application deeply integrates AI deep learning and traditional GIS spatial analysis algorithms, and the entire data production process is fully automated without human intervention, which significantly improves the efficiency of single reconstruction based on oblique model data, and enriches the methods and means for high-precision DEM data production based on multi-source surveying and mapping data. The following is a detailed introduction to the implementation steps and technical processes. The overall process is shown in Figure 1 The regional adaptive DEM data generation method based on oblique model data proposed by the present application includes the following steps:

[0024] Step S1, load the oblique model data of a specified range, perform three-dimensional node sampling on the oblique model data to generate three-dimensional color point cloud data, and perform z-axis vertical sampling on the oblique model data to generate true orthographic image data.

[0025] The tilt model data in the specified area range is loaded, the OpenSceneGraph library is used to read the tilt model data OSGB file, the vertex array in the geometry is traversed, the x, y and z coordinates of each vertex are extracted, the texture information is obtained from the geometry, the texture image is extracted and mapped to the vertex, then the extracted x, y, z vertex coordinates and color information are sampled at an interval of 0.1 m to generate color point cloud data, and the color point cloud data is stored in an xyz format. Similarly, the extracted x, y, z vertex coordinates and color information are vertically sampled along the Z axis to generate 0.1 m high-resolution true orthophoto image data (TDOM).

[0026] In step S2, the three-dimensional color point cloud data is denoised based on a clustering algorithm.

[0027] In an embodiment of the present application, the step S2 of denoising the three-dimensional color point cloud data based on a clustering algorithm comprises:

[0028] The DBSCAN clustering algorithm is used to identify core points, boundary points and noise points from the three-dimensional color point cloud data, wherein the core point refers to a point cloud with a number of point clouds in a neighborhood range greater than or equal to a minimum sample number, the boundary point refers to a point cloud located in the neighborhood range of the core point but with a number of point clouds in its own neighborhood range less than the minimum sample number, and the noise point refers to a point cloud that is neither a core point nor a boundary point.

[0029] The identified core points and boundary points are used as the denoised point cloud.

[0030] It can be understood that the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) clustering algorithm is used to remove suspended and isolated point noises from the three-dimensional color point cloud data generated in step S1, thereby obtaining the denoised three-dimensional color point cloud data. Since the original tilt model data is prone to produce abnormal data such as broken surfaces and suspended objects in weak texture areas, the removal of noise points is crucial for accurately obtaining ground points in the next step.

[0031] The DBSCAN clustering algorithm can discover clusters of arbitrary shapes and effectively identify noise points. The core idea is that points in a high-density area form a cluster, and low-density areas serve as the boundaries between clusters. By defining a neighborhood density threshold algorithm, the data is divided into the following three types of points:

[0032] (1) Core point: a point with at least a minimum sample number (MinPointsNum) of points within a neighborhood radius (ε). The neighborhood radius (ε) defines the neighborhood range, and the minimum sample number (MinPointsNum) defines the minimum number of points in the neighborhood, which is used to determine the core point. The formula for the ε neighborhood is as follows:

[0033] Nε(p) = {q e D | dist(p, q) < ε}

[0034] where dist(p, q) is a distance function (e.g., Euclidean distance).

[0035] And the core point is the condition that the number of points in the neighborhood radius is greater than or equal to the minimum sample number, that is:

[0036] | Nε ( p ) | ≥ MinPointsNum

[0037] (2) Boundary points: located in the neighborhood of the core point, but the number of points in its neighborhood is less than the minimum sample number (MinPointsNum).

[0038] (3) Noise points: points that are neither core points nor boundary points.

[0039] Through the above classification of core points, boundary points and noise points of the point cloud, the points composed of the core points and the boundary points are finally output as the three-dimensional color point cloud data after denoising.

[0040] Step S3, performing semantic segmentation on the three-dimensional color point cloud data after denoising based on a deep learning model to identify first ground points.

[0041] It can be understood that the deep learning model based on the improved RandLA-Net and PointNet++ is combined to perform point cloud semantic segmentation on the three-dimensional color point cloud data after denoising in step S2, to obtain point cloud data after segmentation and classification of buildings, ground, woodland, grassland, water surface, etc. RandLA-Net has significant improvement in sampling strategy, feature extraction method, receptive field and computing efficiency compared with PointNet++ model, especially through the cascaded local feature aggregation (LFA) module to gradually expand the receptive field, which can more effectively learn complex geometric structures and is more suitable for semantic segmentation of large-scale point cloud. But PointNet++ uses multi-scale method MSG and multi-level method MRG to solve the problem of uneven sampling of sparse point cloud samples, which can capture features of different scales and levels, and is more adaptable to point clouds of different densities. Therefore, the present application uses RandLA-Net as a point cloud semantic segmentation framework and combines PointNet++ point cloud sampling algorithm to improve the accuracy, robustness and efficiency of large-scale color point cloud classification. Because the grassland, water surface point cloud and the actual terrain are relatively close, and the grassland and water surface points are merged into the ground point classification, more terrain feature points are reserved for subsequent high-precision DEM generation. This step performs semantic segmentation on the three-dimensional color point cloud data based on a deep learning model to identify ground points from the three-dimensional color point cloud data.

[0042] In step S4, Kriging interpolation is performed on the ground elevation points to generate preliminary terrain raster data, and the preliminary terrain raster data is divided into grids according to a preset size to obtain terrain raster data.

[0043] According to the ground points identified in step S4, the ground elevation points are derived, and then Kriging interpolation is performed on the ground elevation points to pre-generate preliminary terrain raster data. Compared with inverse distance weighting interpolation (Inverse Distance Weighting, IDW), nearest neighbor interpolation (Nearest Neighbor Interpolation) and other algorithms, Kriging interpolation performs better in processing complex terrain and can better capture the subtle changes and structures of the terrain. In order to meet the requirements of high-precision DEM data production, Kriging interpolation is performed for terrain raster interpolation generation, and then the preliminary terrain raster data generated by Kriging interpolation is divided into 200*200m grids and cut into multiple pieces of terrain raster data.

[0044] In step S5, the terrain features of each piece of terrain raster data are extracted.

[0045] It can be understood that due to different terrains in different regions, such as steep slope terrain, gentle slope terrain and flat terrain. Among them, the slope value of each piece of terrain raster data is calculated, and the terrain features of each piece of terrain raster data are obtained according to the slope value.

[0046] Specifically, the slope of each piece of terrain raster data is calculated separately, and the slope value of each grid pixel is calculated. All the slope values of each piece of terrain raster data are arranged from small to large, and the slope values of the larger 10% and the smaller 10% are removed, and the average slope of the remaining slope values is calculated to obtain the comprehensive slope value of the terrain in the terrain raster. The comprehensive slope greater than 10 degrees is determined as steep slope terrain, 2 degrees to 10 degrees is gentle slope terrain, and less than 2 degrees is flat terrain. The slope map generated based on the terrain raster and the slope analysis result.

[0047] In step S6, the three-dimensional color point cloud data after denoising is divided into grids according to a preset size to obtain multiple pieces of three-dimensional color point cloud data, and the second ground points of each piece of three-dimensional color point cloud data are merged based on the CSF filtering algorithm, wherein the parameters of the corresponding CSF filtering algorithm are set according to the corresponding terrain features of each piece of three-dimensional color point cloud data.

[0048] It can be understood that the three-dimensional color point cloud data after denoising in step S2 is divided according to the grid of the same size as step S4. Specifically, after the three-dimensional color point cloud in step S2 is divided according to the 200*200 meter grid in the previous step, a CSF (Cloth Simulation Filter) filtering algorithm is used to filter to obtain preliminary ground points. The CSF filtering algorithm is a point cloud ground filtering algorithm based on cloth simulation. The core idea is to simulate the physical process of cloth covering the inverted terrain, and adjust the cloth node position to approximate the real ground. The main implementation steps of the CSF filtering algorithm are as follows:

[0049] (1) Original point cloud preprocessing: input the original point cloud, take the inverse of the Z coordinate to invert the terrain, and grid the point cloud to reduce the computational complexity.

[0050] (2) Initialize cloth grid: generate a cloth grid according to the point cloud range, and define cloth parameters, including particle mass, gravitational acceleration, stiffness coefficient, etc.

[0051] (3) Iterative simulation of cloth motion: perform particle force calculation and collision detection, check whether each particle collides with the inverted terrain point, stop the particle motion after collision, and record the current position.

[0052] (4) Extract ground points: invert the final position of the cloth to the original coordinate system, generate a ground surface model, and separate the ground points in the original point cloud by distance threshold (such as the vertical distance from the point to the cloth surface).

[0053] In the process of separating the ground points in the three-dimensional color point cloud data based on the CSF filtering algorithm, different CSF filtering parameters are set for each grid according to the terrain features of each grid extracted in step S5. The CSF filtering algorithm provides three types of terrain parameters: steep slope, gentle slope, and flat ground, to adapt to the extraction of ground points in different slope scenarios. The setting of terrain parameters is important for accurate extraction of ground points. The present application uses the classification results (steep slope, gentle slope, and flat ground) of the terrain blocks in step S5 to extract ground points with different CSF terrain parameters, and finally obtains reliable ground point data with adaptive CSF filtering.

[0054] Step S7, based on the second ground points, the first ground points are analyzed and corrected.

[0055] It can be understood that the ground points identified based on the CSF filtering algorithm are generally more accurate, and the ground points segmented based on the deep learning model are more comprehensive. Therefore, the CSF adaptive filtering ground points of step S6 are used to perform overlay analysis on the point cloud semantic segmentation results obtained based on the deep learning model of step S3, correct the ground points missed or misclassified based on the deep learning model, and finally output the adaptive CSF filtering corrected point cloud classification result. The present application uses the CSF filtering ground points with adaptive terrain parameters and the ground point results of step S3 AI point cloud semantic segmentation to perform fusion correction. Based on the improved RandLA-Net deep learning algorithm for different element semantic segmentation and classification results in complex scenes, the high-reliability ground points extracted by the CSF filtering algorithm according to the steep slope, gentle slope and flat ground parameters are combined, and the accuracy and robustness of the ground point acquisition are further improved.

[0056] Step S8, extracting vector data of building surface, forest surface and grass surface from the true orthophoto image data based on the ViTAE network model.

[0057] It can be understood that the true orthophoto image data (TDOM) obtained in step S1 is used to extract vector data related to buildings, trees, grass and ground in the image using the ViTAE (Vision Transformer with Advanced Convolutional Embedding) network model, to further correct the point cloud classification results generated and corrected in step S7, and to improve the ground point classification accuracy. ViTAE is a hybrid model combining the advantages of visual Transformer (ViT) and convolutional neural network (CNN), which significantly improves the extraction accuracy of buildings and vegetation in remote sensing images through multi-scale feature extraction and efficient attention mechanism. Compared with traditional network models such as U-Net and DeepLabV3+, the extraction accuracy of buildings and vegetation in 0.1m high-precision TDOM generated by the oblique photography model is more than 85%.

[0058] Step S9, misclassification correction of the ground points corrected in step S7 based on the vector data of the building surface, the forest surface and the grass surface.

[0059] It can be understood that the ground points corrected in step S7 are further corrected based on the vector data of the building surface, forest land surface and grassland surface extracted in step S8. Among them, according to the building surface extracted from the true orthographic image data, the area belonging to the building is removed from the ground points corrected in step S7 to obtain a hollow area; a buffer range is generated for the building vector surface extracted from the true orthographic image data, and the median of the elevation values of each point in the buffer range is calculated as the bottom elevation of the building, and the building bottom point cloud is generated by interpolation; the hollow area is filled based on the building bottom point cloud, and the correction of the building area is completed.

[0060] Specifically, a 0.5m buffer zone is generated for the building vector surface extracted from the TDOM image in step S8, and the second quartile point (Q2, i.e. the median) of the elevation points in the buffer range is calculated as the bottom elevation of the building, and the building bottom point cloud is generated by interpolation at the 0.1m grid center point to fill the ground point hollow area at the building bottom. The reason why the second quartile point is used in the calculation and filling of the building bottom area elevation points is that in the case of reasonably statistical building surrounding area elevation distribution characteristics, the reasonable building foundation height is calculated, which maintains the relative flatness of the building bottom while trying to ensure the accuracy of the terrain generation.

[0061] Similarly, the forest land and grassland vector data extracted from the TDOM image are subjected to overlay analysis and correction. Among them, the method for correcting the point cloud classification result based on the forest land vector data is the same as the method for correcting based on the building surface vector data, and the grassland is also considered as the ground and does not need to be corrected. The point cloud classification result is further corrected based on the forest land and grassland vector data, which solves the problems of missing classification of ground points in dense vegetation areas, low shrubs and green belts being incorrectly classified as ground, and finally obtains the accurate point cloud classification result after comprehensive correction.

[0062] In step S10, based on the ground points corrected in step S9, the Kriging interpolation parameters are set according to the terrain characteristics of the grid, the Kriging interpolation processing is performed on the ground elevation points of different grids, the terrain raster data of each grid is generated, the terrain raster data of multiple grids is merged, and the high-precision DEM data of a specified range is output.

[0063] It can be understood that based on the ground points corrected in step S9, the ground elevation points are extracted, and the ground elevation points are divided according to a 200*200m grid to obtain multiple grids.

[0064] In the first embodiment of the present application, according to the terrain features of each grid region, Kriging interpolation parameters are set, including neighborhood search radius and nugget value parameters; based on the Kriging interpolation parameters, Kriging interpolation processing is performed on the ground elevation points of each grid to generate terrain raster data corresponding to each grid region.

[0065] Based on the terrain features of each grid, according to the determination results of steep slopes, gentle slopes and flat ground in the block region, adaptive parameter generation of the terrain is performed by using small-to-large neighborhood search radius and high-to-low nugget value parameters during Kriging interpolation. That is, terrain data of the block with steep slopes uses smaller neighborhood search radius and higher nugget value to ensure more terrain undulations and detailed expression. The block with flat ground and gentle slopes uses larger neighborhood search radius and smaller nugget value to obtain smoother and smoother terrain results. The adaptive terrain interpolation generation method of the present application can effectively avoid the problem of poor adaptability of terrain precision in different regions caused by using fixed and unified parameters in large-scale scene terrain generation, and can significantly improve the precision and reliability of the terrain generation results of scenes with large terrain feature changes.

[0066] The terrain raster data generated for each grid is merged, and the terrain block edge region is smoothed, and finally high-precision DEM data (better than 0.2m) for the entire oblique photography model range is output. The DEM data generated finally well preserves the terrain details and undulation features in steep regions such as rivers and riverbeds, generates relatively flat and smooth terrain in road and farmland regions, and generates flat building foundations at the bottom of buildings, fully utilizing the original oblique model to generate regionally adaptive high-precision DEM data.

[0067] The regionally adaptive DEM data generation method based on oblique model data provided by the embodiment of the present application has the following beneficial effects:

[0068] (1) The characteristics and advantages of oblique model data are fully utilized, and high-precision ground elevation points are extracted based on data such as color point cloud, TDOM true orthophoto derived from the oblique model.

[0069] (2) DEM data production is fully automated based on oblique model data, and the entire process does not require human intervention.

[0070] (3) The DEM terrain generation precision is high, the output terrain data has high consistency with the original oblique model data, and the three-dimensional spatial relationship consistency between the subsequent single modeling results based on the oblique model and the DEM terrain can also be ensured.

[0071] (4) The relationship between the terrain and the building is considered, the terrain area under the building is supported to be interpolated and flattened according to the height of the surrounding terrain, and when the building model and the terrain are superimposed in the three-dimensional scene, the problem that the building is suspended above the ground or penetrates into the ground is avoided.

[0072] (5) When the DEM terrain data is generated, the high-precision TDOM true projection image is synchronously output, and all basic data for the construction of the three-dimensional scene terrain bottom plate are provided.

[0073] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0074] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0075] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system), and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing devices to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks Figure 1 The means for performing the functions specified in one or more blocks or flows.

[0076] These computer program instructions can also be stored in a computer-readable memory capable of guiding the computer or other programmable data processing devices to work in a specific manner, so that the instructions stored in the computer-readable memory produce a product including instruction means, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks Figure 1 The means for performing the functions specified in one or more blocks or flows.

[0077] These computer program instructions can also be loaded into a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 one or more flowcharts and / or blocks

[0078] Although preferred embodiments of the application have been described herein, it will be apparent to those skilled in the art that various modifications can be made within the scope of the application without departing from the spirit of the application. Accordingly, it is intended that all such possible modifications be included within the scope of the application as defined in the following claims in which the use of the singular is deemed to include the plural, unless specifically stated otherwise.

[0079] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described herein.

Claims

1. A method for generating region adaptive DEM data based on tilt model data, characterized by, The method comprises the following steps: Step S1, loading the specified range of tilt model data, performing three-dimensional node sampling on the tilt model data to generate three-dimensional color point cloud data, and performing z-axis vertical sampling on the tilt model data to generate true orthographic image data; Step S2, denoising the three-dimensional color point cloud data based on a clustering algorithm; Step S3, performing semantic segmentation on the denoised three-dimensional color point cloud data based on a deep learning model to identify first ground points; Step S4, performing Kriging interpolation on the ground elevation points to generate preliminary terrain grid data, and performing grid division on the preliminary terrain grid data according to a preset size to obtain multiple pieces of terrain grid data; Step S5, extracting terrain features of each piece of terrain grid data; Step S6, performing grid division on the denoised three-dimensional color point cloud data according to a preset size to obtain multiple pieces of three-dimensional color point cloud data, and performing filtering based on a CSF filtering algorithm to obtain second ground points of each piece of three-dimensional color point cloud data, wherein the parameters of the corresponding CSF filtering algorithm are set according to the corresponding terrain features of each piece of three-dimensional color point cloud data; Step S7, performing overlay analysis on the first ground points based on the second ground points to correct the first ground points; Step S8, extracting vector data of building surfaces, forest surfaces and grassland surfaces from the true orthographic image data based on a ViTAE network model; Step S9, based on the vector data of the building surfaces, the forest surfaces and the grassland surfaces, correcting misclassified ground points after the correction in step S7; Step S10, based on the corrected ground points in step S9, setting Kriging interpolation parameters according to the terrain features of the grid, performing Kriging interpolation processing on the ground elevation points of different grids to generate terrain grid data of each grid, merging the terrain grid data of multiple grids, and outputting high-precision DEM data of the specified range.

2. The method of claim 1, wherein, The step S1 of loading the specified range of tilt model data, performing three-dimensional node sampling on the tilt model data to generate three-dimensional color point cloud data, and performing z-axis vertical sampling on the tilt model data to generate true orthographic image data comprises: loading the specified range of tilt model data, reading the tilt model data file using the OpenSceneGraph library, traversing the vertex array in the geometry, and extracting the x, y and z coordinates of each vertex; obtaining texture information from the geometry and mapping the texture information to the vertices; sampling the extracted x, y and z vertex coordinates and color information according to a sampling interval to generate three-dimensional color point cloud data and storing the three-dimensional color point cloud data in xyz format; performing Z-axis vertical sampling on the extracted x, y and z vertex coordinates and color information to generate true orthographic image data TDOM.

3. The method of claim 1, wherein, The step S2 of denoising the three-dimensional color point cloud data based on a clustering algorithm comprises: identifying core points, boundary points and noise points from the three-dimensional color point cloud data based on a DBSCAN clustering algorithm, wherein the core points refer to point clouds with a number of point clouds in a neighborhood range greater than or equal to a minimum sample number, the boundary points refer to point clouds located in the neighborhood range of the core points but with a number of point clouds in their own neighborhood range less than the minimum sample number, and the noise points refer to point clouds that are neither core points nor boundary points; taking the identified core points and boundary points as the point cloud after denoising.

4. The method of claim 1, wherein, In the step S3, the three-dimensional color point cloud data after denoising is subjected to semantic segmentation based on a deep learning model to identify first ground points, including: performing semantic segmentation on the three-dimensional color point cloud data after denoising based on a deep learning model combined with an improved RandLA-Net and PointNet++, to obtain point cloud data classified and segmented for buildings, ground, woodland, grassland and water surface; merging the grassland point cloud and the water surface point cloud into the ground point cloud as the identified first ground points.

5. The method of claim 1, wherein, In the step S5, the terrain features of each block of the terrain grid data are extracted, including: calculating the slope value of each grid cell in each block of the terrain grid data, eliminating abnormal slope values and averaging to obtain the slope value of each block of the terrain grid data; obtaining the terrain features of each block of the terrain grid data according to the slope value of each block of the terrain grid data, wherein the terrain features include steep terrain, gentle terrain and flat terrain.

6. The method of claim 1, wherein, In the step S6, the second ground points of each block of the three-dimensional color point cloud data are filtered based on a CSF filtering algorithm and merged, including: setting the CSF filtering algorithm parameters according to the terrain features corresponding to each block of the three-dimensional color point cloud data; performing CSF filtering on each block of the three-dimensional color point cloud data according to the set CSF filtering algorithm parameters, identifying second ground points from each block of the three-dimensional color point cloud data, and merging all the second ground points as the second ground points in the specified range.

7. The method of claim 1, wherein, In the step S9, the ground points corrected in the step S7 are subjected to misclassification correction based on the vector data of the building surface, the woodland surface and the grassland surface, including: eliminating the areas belonging to buildings and / or woodlands from the ground points corrected in the step S7 according to the building surface and / or the woodland extracted from the true aerial image data, to obtain a hollow area; generating a buffer range based on the building vector surface and / or the woodland vector surface extracted from the true aerial image data, calculating the median of the elevation values of each point in the buffer range as the bottom elevation of the building and / or the woodland, and interpolating to generate the building bottom point cloud and / or the woodland bottom point cloud; filling the hollow area based on the building bottom point cloud and / or the woodland bottom point cloud to obtain the corrected ground points.

8. The method of claim 7, wherein, In the step S10, the ground points corrected in the step S9 are subjected to Kriging interpolation processing of the ground elevation points of different grids according to the Kriging interpolation parameters set according to the terrain features of the grids, to generate the terrain grid data of each grid, including: For the ground points corrected in step S9, grid division is performed according to a preset size to obtain a plurality of grids; According to the terrain features of each grid area, Kriging interpolation parameters are set, the Kriging interpolation parameters including a neighborhood search radius and a nugget value parameter; Based on the Kriging interpolation parameters, Kriging interpolation processing is performed on the ground elevation points of each grid to generate terrain raster data corresponding to each grid area.

9. The method of claim 1, wherein, In step S10, the terrain raster data of the plurality of grids is merged to output high-precision DEM data of a specified range, including: The terrain raster data of the plurality of grids is merged, and the adjacent grid edge areas are smoothed to obtain high-precision DEM data of a specified range.

Images