Method and system for extracting building components based on three-dimensional model
By evaluating the complexity of the building based on DSM data within the exterior profile of the building and segmenting and regularizing it in combination with DOM and DSM data, the problem of automatic extraction of complex building roofs into multiple components is solved, and the precision of automatic building modeling is improved degree.
Patent Information
- Application Number
- CN202311146919.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-06
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2043-09-06
AI Technical Summary
The prior art has little research on the automatic extraction of complex building roofs into multiple components, and it is difficult to meet the needs of many application areas, such as the problem of distinguishing podiums from buildings of different heights in commercial complexes.
The building component extraction method based on three-dimensional model is adopted. By evaluating the complexity of the building based on DSM data within the outer contour range of the building, complex buildings are selected, and segmented with DOM and DSM data, a gradual regularization strategy is adopted. Simplify, regularize and update in turn to achieve decomposition and simplification of the overall model, thereby ensuring the correctness of vector topology of regularized building components.
It improves the accuracy and contour regularity of building components extraction, can automatically generate component vectors that express different structures and functions of complex buildings, and improves the precision of automatic building modeling.
Smart Images

Figure CN118037984B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of three-dimensional modeling, and in particular to a method and system for extracting building components based on a three-dimensional model. Background Art
[0002] Buildings are one of the most important types of urban features. The accurate extraction of building information is of great significance to urban planning management, population density estimation, and natural disaster assessment. With the continuous deepening of the application of LOD models in the process of digital city construction, the automatic construction of fine three-dimensional models of buildings has received more and more attention. Many scholars have conducted research on building outline extraction combining DOM and DSM data sources to give full play to the advantages of DOM in finely depicting the plane structure of buildings and the advantages of DSM data in containing building height information. At present, the research on automatic building extraction mainly focuses on improving the accuracy of building extraction and boundary regularity, while there is less research on automatically extracting complex building roofs into multiple components, which is difficult to meet the needs of many application fields. For example, commercial complexes often need to distinguish podiums from other buildings of different heights. Summary of the invention
[0003] In order to significantly improve the accuracy and contour regularity of building component extraction, the present application provides a building component extraction method and system based on a three-dimensional model.
[0004] The present application provides a method for extracting building components based on a three-dimensional model, which adopts the following technical solution: In a first aspect, a method for extracting building components based on a three-dimensional model is provided, comprising:
[0005] Evaluate the complexity of buildings based on DSM data within the outer contour of the building and select complex buildings; segment based on DOM and DSM data to obtain multiple component vectors of complex buildings;
[0006] A progressive regularization strategy is adopted for the vectors of segmented building components, which are simplified, regularized and updated in sequence to achieve continuous decomposition and simplification of the overall model, thereby ensuring the topological correctness of the vectors of the regularized building components.
[0007] Preferably, the evaluating the building complexity within the outer contour of the building based on DSM data comprises:
[0008] Statistics are performed on the DSM data within the outer contour of the building to obtain the building height data set;
[0009] Smooth the building height data set to remove noise data, and perform clustering calculation to obtain peak values and aggregation degrees;
[0010] If the building height dataset has multiple peaks, or has two peaks but the clustering difference is greater than the threshold, the building is marked as a complex building.
[0011] Preferably, the complex buildings are screened out and segmented based on DOM and DSM data to obtain multiple component vectors of the complex buildings, including:
[0012] Stretching DSM data to the dynamic range of DOM data, making full use of the height information provided by DSM data and the spectrum, shape, and texture information provided by DOM to improve segmentation accuracy;
[0013] Different weights are assigned to different bands according to their characteristics during the segmentation process. The calculation formula for the spectral information merging cost is as follows: h color =Ratio r *Color r +Ratio g *Color g +Ratio b *Color b +Ratio dsm *Color dsm .
[0014] Preferably, the segmented building component vectors are subjected to a progressive regularization strategy, including:
[0015] Comprehensively evaluate the shape area of the building component vector after segmentation and its relationship with the building outline vector to select the most important component vector for current regularization;
[0016] Regularize the most important component vectors at the moment;
[0017] Adjust the direction and position of the normalized vector to make it match the outer contour of the building as much as possible;
[0018] Delete the simplified patches from the set of building components to be simplified, and merge patches with an area smaller than the area threshold into adjacent patches. The merging rule is that the regularity of the patches is improved after merging, and a new set of patches to be simplified is generated for the next round of regularization.
[0019] Preferably, the above-mentioned comprehensive evaluation is performed based on the shape area of the building component vector after segmentation and its relationship with the building outer contour vector, and the most important component vector for the current regularization is screened out. The selection criteria are that the area of the pattern is as large as possible, the part on the building outer contour is as long as possible, and the pattern shape is as regular as possible. According to the above standards, a comprehensive evaluation is performed, and the evaluation is performed according to the following formula:
[0020] V total =V area +V border +V rect
[0021] Where V area 、V border 、V rect The measurement values of the current patch in terms of area, boundary, and shape
[0022]
[0023]
[0024] is the rectangularity of the building component pattern, where S0, S MER Respectively represent the area of the patch and the area of the minimum circumscribed rectangle of the patch;
[0025] According to the above formula, V total The largest building block is the one that needs to be simplified.
[0026] In a second aspect, a building component extraction system based on a three-dimensional model is also provided, comprising:
[0027] Segmentation module: used to evaluate the complexity of buildings within the outer contour of the building based on DSM data and screen out complex buildings; segment based on DOM and DSM data to obtain multiple component vectors of complex buildings;
[0028] Regularization module: It is used to adopt a progressive regularization strategy for the vectors of segmented building components, simplifying, regularizing, and updating in sequence, to achieve continuous decomposition and simplification of the overall model, thereby ensuring the topological correctness of the vectors of the regularized building components.
[0029] Preferably, the evaluating the building complexity within the outer contour of the building based on DSM data comprises:
[0030] Statistics module: used to perform statistics on DSM data within the outer contour of the building to obtain the building height data set;
[0031] Denoising module: used to smooth the building height data set to remove noise data, and perform clustering calculations to obtain peak values and aggregation degrees;
[0032] Labeling module: used to mark a building as a complex building if the building height dataset has multiple peaks, or has two peaks but the difference in aggregation is greater than a threshold.
[0033] Preferably, the complex buildings are screened out and segmented based on DOM and DSM data to obtain multiple component vectors of the complex buildings, including:
[0034] Stretching module: used to stretch DSM data to the dynamic range of DOM data, making full use of the height information provided by DSM data and the spectrum, shape, and texture information provided by DOM to improve segmentation accuracy;
[0035] Allocation module: used to assign different weights to different bands according to their characteristics during the segmentation process. The calculation formula for the spectral information merging cost is as follows:
[0036] h color =Ratio r *Color r +Ratio g *Color g +Ratio b *Color b +Ratio dsm *Color dsm .
[0037] Preferably, the segmented building component vectors are subjected to a progressive regularization strategy, including:
[0038] Screening module: used to make a comprehensive evaluation based on the shape area of the building component vector after segmentation and its relationship with the building outline vector, and screen out the most important component vector for the current regularization;
[0039] Vector module: used to regularize the most important component vectors at present;
[0040] Matching module: used to adjust the direction and position of the normalized vector to make it match the outer contour of the building as much as possible;
[0041] Merge module: used to delete the just simplified patches from the patch set of building components to be simplified, and merge patches with an area smaller than the area threshold into adjacent patches. The merging rule is that the regularity of the patches is improved after merging, and a new patch set to be simplified is generated for the next round of regularization.
[0042] Preferably, the above-mentioned comprehensive evaluation is performed based on the shape area of the building component vector after segmentation and its relationship with the building outer contour vector, and the most important component vector for the current regularization is screened out. The selection criteria are that the area of the pattern is as large as possible, the part on the building outer contour is as long as possible, and the pattern shape is as regular as possible. According to the above standards, a comprehensive evaluation is performed, and the evaluation is performed according to the following formula:
[0043] V total =V area +V border +V rect
[0044] Where V area 、V border 、V rectThe measurement values of the current patch in terms of area, boundary, and shape
[0045]
[0046]
[0047] is the rectangularity of the building component pattern, where S0, S MER Respectively represent the area of the patch and the area of the minimum circumscribed rectangle of the patch;
[0048] According to the above formula, V total The largest building block is the one that needs to be simplified.
[0049] In summary, the present application includes at least one of the following beneficial technical effects:
[0050] At the building outline extracted by the existing invention, DOM and DSM are integrated for segmentation and boundary processing, and component vectors expressing different structures and functions of complex buildings are automatically generated, which improves the precision of automatic building modeling. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 It is a step diagram of a first embodiment of a method for extracting building components based on a three-dimensional model;
[0052] Figure 2 is a step diagram of a second embodiment of a method for extracting building components based on a three-dimensional model;
[0053] Figure 3 The DOM data and DSM data of the building are used as a composite image, and the building vector data are combined to perform component segmentation on the building;
[0054] Figure 4 A step diagram of performing image segmentation within the building vector data range on the image to be segmented to obtain several components of the building;
[0055] Figure 5 A step diagram of extracting and regularizing the components of the building obtained by segmentation to obtain components of the building with different heights;
[0056] Figure 6 It is a composition diagram of the building component extraction system based on the three-dimensional model;
[0057] Figure 7 It is a composition diagram that evaluates the complexity of a building based on DSM data within the building's outer contour;
[0058] Figure 8It is to screen out complex buildings, segment them based on DOM and DSM data, and obtain the composition diagram of multiple component vectors of complex buildings;
[0059] Fig. 9 It is a composition diagram that uses a progressive regularization strategy for segmenting building component vectors;
[0060] Fig.10 It is a logic block diagram of a building component extraction method.
[0061] Explanation of the accompanying drawings: 1. A building component extraction system based on a three-dimensional model; 11. Segmentation module; 12. Regularization module; 111. Statistics module; 112. De-noising module; 113. Labeling module; 114. Stretching module; 115. Allocation module; 121. Screening module; 122. Vector module; 123. Matching module; 124. Merging module. DETAILED DESCRIPTION
[0062] In order to make the purpose, technical solutions and advantages of this application more clear, the following Figure 1 -Attached Fig.10 It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0063] Terminology explanation:
[0064] DOM: Digital Orthophoto Map is an image data generated by using DEM to perform radiation correction, differential correction and mosaic on scanned digitized aerial photographs or remote sensing images (monochrome or color) pixel by pixel, and then cropping them according to the specified map range.
[0065] DSM: Digital Surface Model is a ground elevation model that includes the heights of surface buildings, bridges, and trees.
[0066] This application superimposes DOM and DSM to form a new logical image, making full use of the spectral information, texture information, structural information, and height information provided by the logical image to improve the accuracy of component extraction, and by constructing an angle one-dimensional encoding and topological iteration method, improves the regularity of component edges and realizes the recognition of building components.
[0067] First, as Figure 1 As shown, a method for extracting building components based on a three-dimensional model is provided, comprising:
[0068] S4: using the DOM data and the DSM data of the building as a synthetic image, and combining the building vector data to perform component segmentation on the building;
[0069] S5: Extract and regularize the building components obtained by segmentation to obtain building components with different heights.
[0070] Preferably, Figure 2 As shown, after the DOM data and DSM data of the building are processed, the building is segmented into components in combination with the building vector data, and before that, it also includes:
[0071] S1: stretch the DSM data of the building to the dynamic range of the DOM data of the building; input 8U, 16U or other types of DOM images and DSM data, calculate the distribution range of DSM data, and use linear stretching, piecewise linear stretching, nonlinear stretching and other image stretching methods to stretch the DSM data to the dynamic range of DOM data to ensure that reasonable weights are assigned to each band during segmentation. 8U and 16U are currently common remote sensing image data types. 8U data means that one pixel occupies 8 bits (i.e., one byte), so the value range of the pixel is 0-255; 16U data means that one pixel occupies 16 bits (i.e., two bytes), so the value range of the pixel is 0-65535; overall, 16U can store more ground object details than 8U, but it takes up more storage space. Common DOM images are 8U or 16U, while the DSM data type is Float, which can have decimal parts and a range of values up to -3.4E+38 to 3.4E+38. Of course, the meaningful DSM range is -11034 meters (Mariana Trench) to +8848 meters (Mount Everest). Taking the superposition and segmentation of 8U remote sensing images and DSM as an example, since the maximum value of 8U images is 255, the difference between adjacent pixels of 8U images will not be greater than 255, while the value range of DSM images is much greater than 255, and the difference between adjacent pixels is often greater. Therefore, when they are used for segmentation and classification, the difference in DSM data will "cover up" the difference in 8U images, reducing the role of 8U data. In order to avoid the above situation, the DSM data is stretched so that its dynamic range is basically consistent with the DOM dynamic range, and the weights of the two data are roughly the same during segmentation and classification.
[0072] S2: geographically matching the DOM data of the building with the stretched DSM data of the building; inputting the DOM image, reading the metadata information of the satellite data, adopting the projection difference correction algorithm, geographically matching the DOM data with the DSM data, and correcting the displacement of the building roof in the DOM and the DSM caused by the projection difference.
[0073] S3: The DOM data of the building after the geographic matching and the bands in the DSM data are superimposed together to form a synthetic image. The bands in the DOM image and the DSM image are superimposed together to form a synthetic image. The DOM is usually 3-band, so an image containing four bands, R, G, B, and elevation, will be formed, which will be used as an input image for subsequent segmentation of building components. For images with more than 3 bands, 3 bands can be selected to participate in the segmentation.
[0074] Preferably, Figure 3 As shown, the DOM data and DSM data of the building are used as a synthetic image, and the building vector data are combined to perform component segmentation on the building, including:
[0075] S41: Utilize the building outline vector in the building vector data and the synthetic image of the building to calculate and obtain the image to be segmented; only complex buildings need to be segmented, and complex buildings are often composed of multiple components of different shapes and heights. The application inputs the building outer outline vector, extracts the DSM data within the building outline, and counts the DSM data histogram. It uses but is not limited to a continuous-based fitting algorithm and a discrete-based detection algorithm to count the number of peaks and the degree of aggregation in the histogram. When the histogram has only one peak and the degree of aggregation is greater than the variance threshold, the building component is no longer segmented and the image spot is marked as 0, otherwise the building outline image spot is marked as 1.
[0076] S42: Perform image segmentation within the building vector data range on the image to be segmented to obtain several components of the building. Select the spots marked as 1 in the building outline vector to form the building outline vector to be segmented, calculate the pixel position of the corresponding synthetic image based on the building outline vector to be segmented, and extract all synthetic image pixels surrounded by the vector spots to obtain the image to be segmented. Perform image segmentation within the building vector range on the image to be segmented so as to segment the complex building into multiple building components.
[0077] Preferably, Figure 4 As shown, the image segmentation is performed within the building vector data range on the image to be segmented to obtain several components of the building, including:
[0078] S421: Multi-scale segmentation is adopted, and the patches that make up the building are gradually merged starting from the pixels in the building vector data based on a bottom-up strategy; during the execution process, the patch merging cost calculation formula is as follows:
[0079] f=w*h color +(1-w)*h shape
[0080] The image to be segmented consists of four bands: R, G, B, and elevation, which contain the radiation information and elevation information of the ground objects respectively. They have different importance in segmentation, so when calculating h color Different weights need to be set for them, and the calculation formula is as follows:
[0081] h color =Ratio r *Color r +Ratio g *Color g +Ratio b *Color b +Ratio dsm *Color dsm
[0082] The Ratio r 、Ratio g 、Ratio b 、Ratio dsm They are the weights of the r, g, b, and dsm bands respectively.
[0083] S422: Input the image to be segmented, and select typical building samples according to the characteristics of building components; including low-rise residential buildings, high-rise residential buildings, industrial buildings, large commercial buildings, landmark buildings, etc.
[0084] S423: within the segmentation scale range [m,n], segmentation is performed at intervals of a as a segmentation scale, and (nm) / a segmentation result sequences are obtained for calculating the optimal segmentation scale;
[0085] S424: Calculate the mean variance, maximum area, and homogeneity-heterogeneity eigenvalue of the object at each segmentation scale respectively.
[0086] Statistical eigenvalue results;
[0087] S425: Calculate the scale value corresponding to the maximum area value difference of three adjacent scales within 10%, and then calculate the mean variance and homogeneity and heterogeneity index values on the corresponding scale. If they are all in the first 25% of the value range, the scale is considered to be the optimal scale; otherwise, repeat the above steps until a scale value that meets the requirements is found.
[0088] Preferably, Figure 5 As shown, the components of the building obtained by segmentation are extracted and regularized to obtain components of buildings with different heights, including:
[0089] S51: Screening the building component spots to be regularized; Screening the building component spots to be simplified, the screening indicators include but are not limited to the spot area, the overlapping length with the building outer contour line, and the spot regularity. According to the importance of the above three indicators, weights are assigned to design the screening cost function, and the component with the largest cost function value among all the components to be simplified is the component to be simplified in this round.
[0090] S52: Regularizing the contours of the building components to be regularized; and generating contour vectors of the building components to be simplified by using a 4-connected contour tracking algorithm.
[0091] S53: Update the component spots of the building to be regularized; adopt a new building boundary optimization processing logic, convert the component vector data from 2D to 1D through distance angle encoding, and then perform polygon fitting and optimization, which can greatly reduce the complexity of the processing process. At the same time, prior knowledge can be incorporated into the optimization process to retain the original feature information of the building. Calculate the main direction of the building vector, and the direction perpendicular to the main direction is recorded as the secondary direction; adjust the direction of each line segment of the regularized component vector. If the difference between the direction of the line segment and the main direction of the current building is less than the direction threshold, adjust the direction of the line segment to the main direction of the building; if the difference between the direction of the line segment and the secondary direction of the current building is less than the direction threshold, adjust the direction of the line segment to the secondary direction of the building. Adjust the position of each line segment of the regularized component vector. If the line segment is parallel to the direction of a line segment of the building outer contour vector and the distance is less than the distance threshold, adjust the position of the line segment to the outer contour of the building.
[0092] S54: Assign heights to the components of the building; calculate the mean, median or mode value of the DSM within the building component patch range. The mode value means the value with the highest frequency.
[0093] S55: Repeat the above steps, each time selecting the most important image patch for regularization, iteratively executing to achieve continuous decomposition and simplification of building components, and obtaining building components of different heights.
[0094] This application adopts a progressive topology construction strategy, each time selecting the most important patch for regularization, iterative execution to achieve continuous decomposition and simplification of the overall model, thereby limiting error-prone connections to local parts of the model. The building component patch adds the attribute field "DSM high" and assigns the mean value to the field.
[0095] Second, as Figure 6 As shown, a building component extraction system 1 based on a three-dimensional model is also provided, comprising: a segmentation module 11: used to evaluate the complexity of the building based on the DSM data within the outer contour of the building, and screen out the complex buildings; segment based on the DOM and DSM data to obtain multiple component vectors of the complex building;
[0096] Regularization module 12: It is used to adopt a progressive regularization strategy for the segmented building component vectors, simplify, regularize and update in sequence, so as to realize the continuous decomposition and simplification of the overall model, thereby ensuring the vector topological correctness of the regularized building components.
[0097] Preferably, Figure 7 As shown, the building complexity is evaluated based on DSM data within the outer contour of the building, including:
[0098] The statistical module 111 is used to perform statistics on the DSM data within the outer contour of the building to obtain a building height data set; the denoising module 112 is used to smooth the building height data set to remove noise data, and perform clustering calculation to obtain peak values and aggregation degrees;
[0099] The labeling module 113 is used to mark the building as a complex building if the building height dataset has multiple peaks, or has two peaks but the difference in concentration is greater than a threshold.
[0100] Preferably, Figure 8 As shown, the complex buildings are screened out and segmented based on DOM and DSM data to obtain multiple component vectors of the complex buildings, including:
[0101] Stretching module 114: used to stretch the DSM data to the dynamic range of the DOM data, making full use of the height information provided by the DSM data and the spectrum, shape, and texture information provided by the DOM to improve the segmentation accuracy;
[0102] The allocation module 115 is used to allocate different weights to different bands of data according to their characteristics during the segmentation process. The calculation formula for the spectral information merging cost is as follows:
[0103] h color =Ratio r *Color r +Ratio g *Color g +Ratio b *Color b +Ratio dsm *Color dsm .
[0104] Preferably, Fig. 9 As shown, the progressive regularization strategy for segmented building component vectors includes: a screening module 121: for performing a comprehensive evaluation based on the shape area of the segmented building component vectors and their relationship with the building outer contour vectors, and screening out the most important component vectors for current regularization;
[0105] Vector module 122: used to regularize the most important component vector at present;
[0106] Matching module 123: used to adjust the direction and position of the normalized vector so that it matches the outer contour of the building as much as possible;
[0107] Merging module 124: used to delete the just simplified patches from the patch set of building component patches to be simplified, and merge patches with an area smaller than the area threshold into adjacent patches. The merging rule is that the regularity of the patches is improved after merging, and a new patch set to be simplified is generated for the next round of regularization.
[0108] Preferably, the above-mentioned comprehensive evaluation is performed based on the shape area of the building component vector after segmentation and its relationship with the building outer contour vector, and the most important component vector for the current regularization is screened out. The selection criteria are that the area of the pattern is as large as possible, the part on the building outer contour is as long as possible, and the pattern shape is as regular as possible. According to the above standards, a comprehensive evaluation is performed, and the evaluation is performed according to the following formula:
[0109] V total =V area +V border +V rect
[0110] Where V area 、V border 、V rect The measurement values of the current patch in terms of area, boundary, and shape
[0111]
[0112]
[0113] is the rectangularity of the building component pattern, where S0, S MER They represent the area of the patch and the area of the minimum circumscribed rectangle of the patch respectively.
[0114] According to the above formula, V total The largest building block is the one that needs to be simplified.
[0115] like Fig.10As shown, the present application designs an automatic extraction model of building components based on DSM. The present application uses building vectors as input sources (building vectors can be automatically generated, manually drawn, or historical results), performs corresponding processing on DOM and DSM within the building vector range, and then uses spectral information and height information to divide the building into multiple component spots, and finally processes these component spots to obtain regular building component vectors that maintain topological correctness. Now, taking the Beijing-3 Agile Imaging Satellite data as an example, it is described as follows to facilitate relevant personnel to better understand the present application. The input data of the present application includes three categories: images, DSM, and building outer contour vectors. The coverage area is within the Sixth Ring Road of Beijing, among which the image data is the Beijing-3 satellite L4-level DOM image (3 bands 8U), the DSM data is the standard DSM product (1 band 32F) generated based on the Beijing-3 satellite stereo imaging processing, and the building vector data is the building vector standard result provided by other departments of the company, which is obtained by extracting the building outer contour using a deep learning method on the L4-level DOM and undergoing vector regularization processing.
[0116] Step 1: Split data preparation
[0117] 1. DSM data stretching: Input 8U DOM image and float type DSM data, calculate DSM histogram and remove 1% above and below, and then use linear stretching method to adjust its value range to 0-255. When using Beijing No. 3 L1 image for experiment, since the DN value range of L1 image data is 0-4096, the DN value range of DSM data after stretching is also 0-4096.
[0118] 2. Image registration: Input DOM image, read the metadata information of satellite data, use the projection difference correction algorithm to geographically match DOM data with DSM data, and correct the displacement of building roofs in DOM and DSM due to projection difference.
[0119] 3. Band overlay: Overlay the bands in the DOM and DSM images to form a composite image. Since the Beijing No. 3 L4 DOM image is a 3-band image, a composite image containing four bands, R, G, B, and elevation, is formed, which is used as the input image for the subsequent segmentation of building components.
[0120] Step 2: Component segmentation
[0121] 1. Segmentation vector screening: Only complex buildings need to be segmented. Complex buildings are often composed of multiple components of different shapes and heights. This application inputs the building outline vector, extracts the DSM data within the building outline, counts the DSM data histogram, and then fits the histogram data after filtering, and then calculates the number of peaks and aggregation in the histogram. When the histogram has only one peak and the aggregation is greater than the variance threshold, the building component is no longer segmented and the spot is marked as 0, otherwise the building outline spot is marked as 1.
[0122] 2. Obtaining component segmented images; selecting the spots marked as 1 in the building outline vector to form the building outline vector to be segmented, calculating the pixel position of the corresponding synthetic image based on the building outline vector to be segmented, and extracting all synthetic image pixels surrounded by the vector spots to obtain the image to be segmented.
[0123] 3. Image segmentation: Perform image segmentation within the building vector range on the image to be segmented so as to segment complex buildings into multiple building components.
[0124] In step 1, this application adopts multi-scale segmentation, and gradually merges the patches that make up the features from the pixels based on a bottom-up strategy. The calculation formula for the patch merging cost during the segmentation process is as follows:
[0125] f=w*h color +(1-w)*h shape
[0126] The image to be segmented consists of four bands: R, G, B, and elevation, which represent the radiation characteristics and elevation characteristics of the ground objects respectively. They have different importance in segmentation, so when calculating h color Different weights need to be set for them respectively, and the calculation formula is as follows:
[0127] h color =Ratio r *Color r +Ratio g *Color g +Ratio b *Color b +Ratio dsm *Color dsm
[0128] The Ratio r 、Ratio g 、Ratio b 、Ratio dsm They are the weights of the r, g, b and dsm bands respectively. For the Beijing-3 L4 DOM and DSM data, the weights of the above four bands are 0.15, 0.15, 0.15 and 0.55 respectively.
[0129] In step 2, input the image to be segmented, and select typical building samples based on the characteristics of building components, including low-rise residential buildings, high-rise residential buildings, industrial buildings, large commercial buildings, landmark buildings, etc.
[0130] In step three, for the synthetic image to be segmented consisting of Beijing No. 3 L4 DOM and DSM, the maximum reasonable scale of multi-scale segmentation is more than 200 after experiments. Therefore, we use 10 as the starting scale and 230 as the ending scale, and perform multiple segmentations with a scale interval of 10 to obtain 23 segmentation result sequences for calculating the optimal segmentation scale.
[0131] In step four, the mean variance, maximum area, and homogeneity-heterogeneity eigenvalues of the objects at each segmentation scale are calculated, and the eigenvalue results are statistically analyzed.
[0132] In step five, calculate the scale value corresponding to the maximum area value difference of three adjacent scales within 10%, and then calculate the mean variance and homogeneity and heterogeneity index values on the corresponding scale. If they are all in the first 25% of the value range, the scale is considered to be the optimal scale. Otherwise, repeat step five until a scale value that meets the requirements is found.
[0133] Step 3: Component extraction and regularization
[0134] This application adopts a progressive topology construction strategy, sequentially selecting important building components for regularization, realizing continuous decomposition and simplification of the overall model, thereby limiting error-prone connections to local parts of the model. It mainly includes the following four steps: Step 1: Simplify component selection
[0135] Select the building component spots to be simplified. The selection criteria are that the spot area is as large as possible, the part on the building outline is as long as possible, and the spot shape is as regular as possible. According to the above criteria, a comprehensive evaluation is performed according to the following formula:
[0136] V total =V area +V border +V rect
[0137] Where V area 、V border 、V rect The measurement values of the current patch in terms of area, boundary, and shape
[0138]
[0139]
[0140] is the rectangularity of the building component pattern, where S0, SMER They represent the area of the patch and the area of the minimum circumscribed rectangle of the patch respectively.
[0141] According to the above formula, V total The largest building block is the one that needs to be simplified.
[0142] Step 2: Simplify the outline of the parts
[0143] 1. Vectorize the simplified component pattern
[0144] The 4-connected contour tracking algorithm is used to generate the contour vectors of the building components to be simplified.
[0145] 2. Regularize component vectors
[0146] A new building boundary optimization processing logic is adopted. The component vector data is converted from 2D to 1D through distance angle encoding, and then polygon fitting and optimization are performed. This can greatly reduce the complexity of the processing process. At the same time, prior knowledge can be incorporated into the optimization process to retain the original feature information of the building. The following is a regularization example.
[0147] Step 3: Component vector adjustment based on building vector constraints
[0148] 1. Calculate the main direction of the building vector, and the direction perpendicular to the main direction is recorded as the secondary direction;
[0149] 2. Adjust the direction of each line segment of the normalized component vector. If the difference between the line segment direction and the current main direction of the building is less than the direction threshold, adjust the line segment direction to the main direction of the building; if the difference between the line segment direction and the current secondary direction of the building is less than the direction threshold, adjust the line segment direction to the secondary direction of the building;
[0150] The direction of each line segment of the normalized component vector is adjusted. If the line segment is parallel to a line segment of the building outer contour vector and the distance is less than the distance threshold, the position of the line segment is adjusted to the building outer contour.
[0151] Step 4: Update the building component pattern
[0152] 1. Delete the image tiles that have just been simplified from the image tiles of the building component set to be simplified.
[0153] 2. Merge the new set of patches to be simplified, and merge patches with an area smaller than the area threshold into adjacent patches. The merging rule is that the regularity of the patches is improved after merging.
[0154] Step 4: Add height to building components
[0155] 1. Calculate the mean of the DSM within the vector range of the building component.
[0156] 2. Add an attribute field "DSM height" to the building component vector and assign the mean value to the field, which can be used for building component height calculation.
[0157] This application designs a method for automatically extracting building components that combines DOM and DSM. This method comprehensively utilizes building spectral information, texture information, structural information, and elevation information to segment within the building boundary vector range, and divides complex buildings into multiple components representing different structures and functions, and then regularizes the building components to obtain building component vectors that maintain correct topological relationships.
[0158] In the 3-data superposition of step 1 data segmentation preparation, a method of superimposing DOM and DSM data into logical images for segmentation is proposed, and in the 3-image segmentation of step 2 component segmentation, different segmentation weights are set for DOM and DSM data according to their characteristics and importance.
[0159] In step three, component extraction and regularization, a progressive vector regularization strategy is designed to ensure the topological correctness of the result vector; the correctness of the component result vector is ensured by automatically extracting the direction and position information of the building boundary and driving each line segment to move closer to the building boundary during regularization.
[0160] In addition, other segmentation algorithms may be used for the segmentation of building components in the present application, including but not limited to threshold-based segmentation, edge-based segmentation, region-based segmentation, deep learning-based segmentation methods, etc.
[0161] Other weights can be assigned to the input DSM and DOM according to their different characteristics of the building component information they contain.
[0162] Different topological construction strategies and regularization methods can be used to regularize building component vectors.
[0163] In summary, the present application includes at least one of the following beneficial technical effects:
[0164] For the building outlines that have been extracted through applications, DOM and DSM are integrated for segmentation and boundary processing, and component vectors that express the different structures and functions of complex buildings are automatically generated, which improves the precision of automatic building modeling.
[0165] The above are all preferred embodiments of the present application, and are not intended to limit the protection scope of the present application. Any feature disclosed in this specification (including the abstract and drawings), unless otherwise stated, can be replaced by other equivalent or alternative features with similar purposes. That is, unless otherwise stated, each feature is only an example of a series of equivalent or similar features.
Claims
1. A building component extraction method based on a three-dimensional model, characterized in that: include: Evaluate the building complexity based on DSM data within the building outline and screen out complex buildings; Segmentation is performed based on the DOM and DSM data of the building to obtain multiple component vectors of complex buildings; A progressive regularization strategy is adopted for the vectors of segmented building components, which are simplified, regularized, and updated in sequence to achieve continuous decomposition and simplification of the overall model, thereby ensuring the topological correctness of the vectors of the regularized building components. The segmentation based on the DOM and DSM data of the building to obtain multiple component vectors of the complex building includes: S1: stretching the DSM data of the building to the dynamic range of the DOM data of the building; S2: geographically matching the DOM data of the building with the stretched DSM data of the building; S3: superimposing the DOM data of the building after the geographic matching and the bands in the DSM data to form a synthetic image; S41: using the building outline vector in the building vector data and the synthetic image of the building to calculate and obtain the image to be segmented; S42: Perform image segmentation within the building vector data range on the image to be segmented to obtain a number of component vectors of the building.
2. The building component extraction method based on a three-dimensional model according to claim 1, characterized in that: The building complexity is assessed based on DSM data within the building outline, including: Statistics are performed on the DSM data within the outer contour of the building to obtain the building height data set; Smooth the building height data set to remove noise data, and perform clustering calculation to obtain peak values and aggregation degrees; If the building height dataset has multiple peaks, or has two peaks but the clustering difference is greater than the threshold, the building is marked as a complex building.
3. The building component extraction method based on a three-dimensional model according to claim 1, characterized in that: The complex buildings are screened out and segmented based on the DOM and DSM data of the buildings to obtain multiple component vectors of the complex buildings, including: Stretching DSM data to the dynamic range of DOM data, making full use of the height information provided by DSM data and the spectrum, shape, and texture information provided by DOM to improve segmentation accuracy; Different weights are assigned to different bands according to their characteristics during the segmentation process. The calculation formula for the spectral information merging cost is as follows: h color =Ratio r *Color r +Ratio g *Color g +Ratio b *Color b +Ratio dsm *Color dsm .
4. The building component extraction method based on a three-dimensional model according to claim 1, characterized in that: The stepwise regularization strategy for segmenting the building component vectors includes: Comprehensively evaluate the shape area of the building component vector after segmentation and its relationship with the building outline vector to select the most important component vector for current regularization; Regularize the most important component vectors at the moment; Adjust the direction and position of the normalized vector to make it match the outer contour of the building as much as possible; Delete the simplified patches from the set of building components to be simplified, and merge patches with an area smaller than the area threshold into adjacent patches. The merging rule is that the regularity of the patches is improved after merging, and a new set of patches to be simplified is generated for the next round of regularization.
5. The method for extracting building components based on a three-dimensional model according to claim 4, characterized in that: The above-mentioned comprehensive evaluation is performed based on the shape area of the building component vector after segmentation and its relationship with the building outer contour vector, and the most important component vector for the current regularization is screened out. The selection criteria are that the area of the pattern is as large as possible, the part on the building outer contour is as long as possible, and the pattern shape is as regular as possible. According to the above standards, a comprehensive evaluation is performed, and the evaluation is performed according to the following formula: V total =V area +V border +V rect Where V area 、V border 、V rect The measurement values of the current patch in terms of area, boundary, and shape is the rectangularity of the building component pattern, where S0, S MER Respectively represent the area of the patch and the area of the minimum circumscribed rectangle of the patch; According to the above formula, V total The largest building block is the one that needs to be simplified.
6. A building component extraction system based on a three-dimensional model, characterized in that: The method for extracting building components based on a three-dimensional model according to any one of claims 1 to 5 comprises: Segmentation module: used to evaluate the complexity of buildings within the outer contour of the building based on DSM data and screen out complex buildings; segment the building based on DOM and DSM data to obtain multiple component vectors of the complex building; Regularization module: It is used to adopt a progressive regularization strategy for the vectors of segmented building components, simplifying, regularizing, and updating in sequence, to achieve continuous decomposition and simplification of the overall model, thereby ensuring the topological correctness of the vectors of the regularized building components.
7. The building component extraction system based on three-dimensional model according to claim 6, characterized in that: The building complexity is assessed based on DSM data within the building outline, including: Statistics module: used to perform statistics on DSM data within the outer contour of the building to obtain the building height data set; Denoising module: used to smooth the building height data set to remove noise data, and perform clustering calculations to obtain peak values and aggregation degrees; Labeling module: used to mark a building as a complex building if the building height dataset has multiple peaks, or has two peaks but the difference in aggregation is greater than a threshold.
8. The building component extraction system based on three-dimensional model according to claim 6, characterized in that: The complex buildings are screened out and segmented based on the DOM and DSM data of the buildings to obtain multiple component vectors of the complex buildings, including: Stretching module: used to stretch DSM data to the dynamic range of DOM data, making full use of the height information provided by DSM data and the spectrum, shape, and texture information provided by DOM to improve segmentation accuracy; Allocation module: used to assign different weights to different bands according to their characteristics during the segmentation process. The calculation formula for the spectral information merging cost is as follows: h color =Ratio r *Color r +Ratio g *Color g +Ratio b *Color b +Ratio dsm *Color dsm 。 9. The building component extraction system based on three-dimensional model according to claim 6, characterized in that: The stepwise regularization strategy for segmenting the building component vectors includes: Screening module: used to make a comprehensive evaluation based on the shape area of the building component vector after segmentation and its relationship with the building outline vector, and screen out the most important component vector for the current regularization; Vector module: used to regularize the most important component vectors at present; Matching module: used to adjust the direction and position of the normalized vector to make it match the outer contour of the building as much as possible; Merge module: used to delete the just simplified patches from the patch set of building components to be simplified, and merge patches with an area smaller than the area threshold into adjacent patches. The merging rule is that the regularity of the patches is improved after merging, and a new patch set to be simplified is generated for the next round of regularization.
10. The building component extraction system based on three-dimensional model according to claim 9, characterized in that: The above-mentioned comprehensive evaluation is performed based on the shape area of the building component vector after segmentation and its relationship with the building outer contour vector, and the most important component vector for the current regularization is screened out. The selection criteria are that the area of the pattern is as large as possible, the part on the building outer contour is as long as possible, and the pattern shape is as regular as possible. According to the above standards, a comprehensive evaluation is performed, and the evaluation is performed according to the following formula: V total =V area +V border +V rect Where V area 、V border 、V rect The measurement values of the current patch in terms of area, boundary, and shape is the rectangularity of the building component pattern, where S0, S MER Respectively represent the area of the patch and the area of the minimum circumscribed rectangle of the patch; According to the above formula, V total The largest building block is the one that needs to be simplified.
Citation Information
Patent Citations
Automatic building information extraction method and system based on high-scene satellite image
CN111178175A
Method, device and equipment for extracting three-dimensional building contour based on aerospace data
CN115620168A
Rural building extraction and vectorization method based on inclined three-dimensional model
CN115994987A
Building extraction method and device for homologous high-resolution image and DSM data
CN116258958A
KR20200082611A
