Building position precision measuring system based on image recognition
By using an image recognition-based building location accuracy measurement system, which combines the Transformer-U-Net model and the Douglas-Peucker algorithm to optimize geographic coordinates, the system solves the problems of low efficiency and unstable accuracy in existing building location measurement technologies, and achieves efficient and accurate building location measurement.
Patent Information
- Application Number
- CN202511706814.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-20
- Publication Date
- 2026-02-17
AI Technical Summary
In existing technologies, traditional methods for measuring building location accuracy rely on total stations, which are inefficient and difficult to deploy in complex terrain. GPS measurements are susceptible to obstruction by tall buildings and electromagnetic interference, resulting in large accuracy fluctuations. LiDAR is expensive and data processing is complex. Image recognition-based methods often rely on a single image source, making it difficult to comprehensively capture building spatial information. Contour extraction is poorly adaptable to changes in lighting and complex backgrounds.
A building location accuracy measurement system based on image recognition is adopted, including image acquisition, preprocessing, contour extraction, feature matching and output modules. The Transformer-U-Net model and Douglas-Peucker algorithm are used to extract building contours, and the Levenberg-Marquardt algorithm is combined to optimize geographic coordinates. Multi-source image acquisition is used to make up for the limitations of single viewpoint and improve information integrity and accuracy.
It improves the accuracy and efficiency of building location measurement, adapts to complex environments, lighting changes and complex backgrounds, enhances the accuracy and smoothness of contour extraction, reduces equipment costs, and is suitable for building location measurement in various environments.
Smart Images
Figure CN121544704A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision technology, and more specifically to a building position accuracy measurement system based on image recognition. Background Technology
[0002] Accurate measurement of building locations is a key technology in fields such as urban planning and disaster monitoring. In traditional measurement methods, total stations rely on manual operation, which is inefficient and difficult to deploy in complex terrain. GPS measurements are easily affected by tall buildings and electromagnetic interference, resulting in large fluctuations in accuracy. LiDAR is expensive and data processing is complex. Existing image recognition-based methods mostly rely on a single image source, making it difficult to comprehensively capture building spatial information. Furthermore, contour extraction often uses traditional convolutional networks, which are poorly adaptable to changes in lighting and complex backgrounds, leading to insufficient contour accuracy. Summary of the Invention
[0003] The purpose of this invention is to solve the above problems by designing a building location accuracy measurement system based on image recognition.
[0004] The first aspect of this invention provides a building location accuracy measurement system based on image recognition, the system comprising: The image acquisition module is used to acquire remote sensing images and UAV aerial images containing the target building as images to be tested, and to obtain reference data for the target building area; The image preprocessing module is used to perform noise reduction, illumination equalization and distortion correction preprocessing on the image to be tested to obtain the preprocessed image to be tested. The contour extraction module is used to input the preprocessed image to be tested into the Transformer-U-Net model, extract the initial contour of the target building, and use the Douglas-Peucker algorithm to smooth the initial contour and output the target contour. The feature matching module is used to perform feature matching between the target contour and the reference data, determine the initial geographic coordinates corresponding to the target contour on the parameter data, and use the Levenberg-Marquardt algorithm to iteratively solve and optimize the initial geographic coordinates. The output module is used to calculate the reprojection error of the optimized geographic coordinates, and to calculate the geographic coordinates of the corner points of the target building through error propagation, and finally output the high-precision location information of the building.
[0005] Optionally, in a first implementation of the first aspect of the present invention, the image preprocessing module includes: The denoising submodule is used to suppress noise using an adaptive bilateral filtering algorithm to obtain a denoised image. The illumination equalization submodule is used to perform illumination equalization on the denoised image based on the Retinex algorithm to obtain an illumination-equalized image. The correction submodule is used to call the distortion parameters obtained by pre-calibration of the camera to correct the coordinates of each pixel in the image with equal illumination, so as to obtain the pre-processed image to be tested.
[0006] Optionally, in a second implementation of the first aspect of the present invention, the noise reduction submodule includes: The image to be tested is processed pixel by pixel. A dynamic filtering kernel is constructed with the current pixel as the center. The weight of each pixel in the filtering kernel is dynamically adjusted according to the spatial distance and gray-level difference between pixels. The denoised gray-level value of the current pixel is calculated by weighted average to obtain the denoised image.
[0007] Optionally, in a third implementation of the first aspect of the present invention, the illumination equalization submodule includes: The denoised image is decomposed into illumination and reflection components. The illumination component is then subjected to logarithmic transformation and Gaussian filtering. The adjusted illumination and reflection components are then re-fused to obtain an image with balanced illumination.
[0008] Optionally, in the fourth implementation of the first aspect of the present invention, the denoised image is converted into a logarithmic form, and the product relationship of pixel values in the image is converted into an additive relationship; Gaussian filtering is applied to the logarithmic image to obtain the logarithmic estimate of the illumination component. The logarithmic estimate of the illumination component is then subtracted from the logarithmic original image to obtain the logarithmic form of the reflection component. The logarithmic estimate of the illumination component is subjected to dynamic range compression, and the compressed logarithmic estimate of the illumination component is converted back to the linear domain through exponential operation to obtain the adjusted illumination component. The logarithmic form of the reflection component is converted back to the linear domain through exponential operations to obtain the linear form of the reflection component. The adjusted illumination component and reflection component are multiplied in the linear domain and then fused to obtain an image with equal illumination.
[0009] Optionally, in a fifth implementation of the first aspect of the present invention, the contour extraction module includes: The capture submodule is used to input the preprocessed image to be tested into the Transformer-U-Net model. The encoder part captures the global semantic association of the image through a multi-layer Transformer attention mechanism, and extracts local detail features through convolutional layers to generate multi-scale feature maps. The connection submodule is used by the decoder to gradually restore the feature map resolution through upsampling operations and to make skip connections with the feature maps of the corresponding layers of the encoder to fuse global and local information. The generation submodule is used to generate a binary mask of the building region based on the multi-scale feature map using the sigmoid activation function, and to perform connectivity analysis on the binary mask to remove isolated edges of non-target buildings and obtain the initial outline of the target building. The smoothing submodule is used to smooth the initial contour using the Douglas-Peucker algorithm and output the target contour.
[0010] Optionally, in a sixth implementation of the first aspect of the present invention, the smoothing submodule includes: Select two endpoints on the initial contour as the start point and the end point, and calculate the perpendicular distance from all intermediate points on the initial contour to the line connecting the start point and the end point. Determine the point with the largest vertical distance. If the distance is greater than a preset threshold, retain the corresponding point and use it as the new endpoint to recursively process the sub-segments on both sides. If the distance is less than the threshold, the intermediate point is deleted, and the original line segment is replaced by the line connecting the two endpoints, finally obtaining a smooth target contour.
[0011] Optionally, in a seventh implementation of the first aspect of the present invention, the feature matching module includes: The extraction submodule is used to extract feature points from the target contour and calculate the local descriptor for each feature point, where the feature points include corner points, inflection points and equidistant sampling points; The matching submodule is used to extract the contour feature points and descriptors of the corresponding region from the reference data, calculate the similarity between the local descriptor of the target contour and the descriptor of the reference data through Euclidean distance, initially match the same feature points, and determine the initial geographic coordinates of the target contour on the reference data.
[0012] The iterative solution submodule is used to iteratively solve the initial geographic coordinates using the Levenberg-Marquardt algorithm to obtain the optimized geographic coordinates.
[0013] Optionally, in an eighth implementation of the first aspect of the present invention, the iterative solution submodule includes: Using the initial geographic coordinates as the starting point of the iteration, an error function is constructed, and the Gauss-Newton method is used to calculate the parameter update. If the error increases, a damping factor is introduced to switch to the gradient descent method to reduce the parameter update step size. The error function is recalculated after each iteration until the maximum number of iterations is reached, at which point the optimized geographic coordinates are output.
[0014] Optionally, in a ninth implementation of the first aspect of the present invention, a method for implementing a building location accuracy measurement system based on image recognition is provided, the method comprising the following steps: Collect remote sensing images and drone aerial images containing the target building as the images to be measured, and obtain reference data for the target building area; The image to be tested is preprocessed by denoising, illumination equalization and distortion correction to obtain the preprocessed image to be tested. The preprocessed image to be tested is input into the Transformer-U-Net model to extract the initial contour of the target building, and the Douglas-Peucker algorithm is used to smooth the initial contour to output the target contour. The target contour is matched with the reference data to determine the initial geographic coordinates corresponding to the target contour on the parameter data, and the Levenberg-Marquardt algorithm is used to iteratively solve and optimize the initial geographic coordinates. The reprojection error of the optimized geographic coordinates is calculated, and the geographic coordinates of the corner points of the target building are obtained through error propagation. Finally, the high-precision location information of the building is output.
[0015] In the technical solution provided by this invention, remote sensing images and UAV aerial images containing the target building are acquired as the images to be tested, and reference data of the target building area is obtained; the images to be tested are preprocessed with denoising, illumination equalization, and distortion correction to obtain preprocessed images to be tested; the preprocessed images to be tested are input into the Transformer-U-Net model to extract the initial contour of the target building, and the Douglas-Peucker algorithm is used to smooth the initial contour to output the target contour; the target contour is matched with the reference data to determine the initial geographic coordinates corresponding to the target contour on the parameter data, and the Levenberg-Marquardt algorithm is used for iterative solution to optimize the initial geographic coordinates; the optimized coordinates are calculated. This invention addresses the reprojection error of building coordinates and calculates the geographic coordinates of the target building's corner points through error propagation, ultimately outputting high-precision building location information. It compensates for the limitations of single-view information by acquiring images from multiple sources, improving information integrity in complex scenes. The combination of the Transformer-U-Net model and the Douglas-Peucker algorithm enhances the accuracy and smoothness of contour extraction, adapting to changes in lighting and complex backgrounds. The Levenberg-Marquardt algorithm optimizes the initial coordinates, ensuring iterative convergence speed and accuracy. The combination of reprojection error and error propagation calculations further improves the reliability of corner coordinates. Requiring no complex equipment deployment, it improves measurement efficiency and accuracy, making it widely applicable to building location measurement in various environments and highly practical. Attached Figure Description
[0016] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.
[0017] Figure 1 A schematic diagram of the structure of a building position accuracy measurement system based on image recognition provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the image preprocessing module provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the contour extraction module provided in an embodiment of the present invention. Detailed Implementation
[0018] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.
[0019] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 A schematic diagram of the structure of the building position accuracy measurement system based on image recognition provided in this embodiment of the invention. The system includes: The image acquisition module is used to acquire remote sensing images and UAV aerial images containing the target building as images to be tested, and to obtain reference data for the target building area; The image preprocessing module is used to perform noise reduction, illumination equalization and distortion correction preprocessing on the image to be tested to obtain the preprocessed image to be tested. The contour extraction module is used to input the preprocessed image to be tested into the Transformer-U-Net model, extract the initial contour of the target building, and use the Douglas-Peucker algorithm to smooth the initial contour and output the target contour. The feature matching module is used to perform feature matching between the target contour and the reference data, determine the initial geographic coordinates corresponding to the target contour on the parameter data, and use the Levenberg-Marquardt algorithm to iteratively solve and optimize the initial geographic coordinates. The output module is used to calculate the reprojection error of the optimized geographic coordinates, and to calculate the geographic coordinates of the corner points of the target building through error propagation, and finally output the high-precision location information of the building.
[0020] In this embodiment, a remote sensing satellite equipped with a high-resolution optical imaging system is selected to perform vertical remote sensing imaging of the area where the target building is located according to a preset orbit, acquiring remote sensing images covering the building and a range of 100-200 meters around it, ensuring that the images include the overall spatial distribution of the building and the surrounding terrain features; simultaneously, a multi-rotor UAV is activated, planning a circling flight path based on the height and shape of the building, with the flight altitude set at 2-3 times the height of the building, and the horizontal and vertical overlap rates not less than 75% and 80% respectively, acquiring aerial images from multiple angles such as the top and sides, and simultaneously recording the shooting time, camera intrinsic parameters and extrinsic parameters of each image; at the same time, a high-precision digital elevation model of the target area, at least 3 control point data with known geodetic coordinates, and a 1:500 scale topographic map of the area are retrieved from the mapping geographic information database as reference benchmark data for subsequent feature matching and coordinate calculation.
[0021] In this embodiment, the images to be tested specifically refer to remote sensing images and UAV aerial images containing the target building. These images are acquired through an image acquisition module and contain visual information about the building and its surrounding environment. The reference data refers to high-precision geographic information data obtained from a surveying and mapping geographic information database, including a high-precision digital elevation model of the target area, at least three control point data with known geodetic coordinates, and a 1:500 scale topographic map of the area. During feature matching, the image to be tested is first processed by a contour extraction module to obtain the target contour, i.e., the boundary contour of the building. Then, feature points, including corner points, inflection points, and equidistant sampling points, are extracted from this contour, and a local descriptor is calculated for each feature point. Simultaneously, contour feature points and descriptors for the corresponding area are extracted from the reference data. The feature matching module calculates the Euclidean distance between the local descriptor of the target contour and the descriptor of the reference data, matches similar feature points with high similarity, and thus determines the initial geographic coordinates of the target contour on the reference data. The contour features obtained from image processing are matched with features in the geographic reference data. This matching process enables the system to associate the building contour in the image with the geographic coordinate system.
[0022] In this embodiment, please refer to Figure 2 The image preprocessing module includes: The denoising submodule is used to suppress noise using an adaptive bilateral filtering algorithm to obtain a denoised image. The illumination equalization submodule is used to perform illumination equalization on the denoised image based on the Retinex algorithm to obtain an illumination-equalized image. The correction submodule is used to call the distortion parameters obtained by pre-calibration of the camera to correct the coordinates of each pixel in the image with equal illumination, so as to obtain the pre-processed image to be tested.
[0023] In this embodiment, the denoising submodule includes: processing the image to be tested pixel by pixel, constructing a dynamic filtering kernel with the current pixel as the center, dynamically adjusting the weight of each pixel in the filtering kernel according to the spatial distance and gray-level difference between pixels, and calculating the denoised gray-level value of the current pixel by weighted average to obtain the denoised image.
[0024] In this embodiment, the illumination equalization submodule includes: decomposing the denoised image into illumination components and reflection components, performing logarithmic transformation and Gaussian filtering on the illumination components, and re-fusing the adjusted illumination components and reflection components to obtain an illumination-equalized image.
[0025] In this embodiment, the denoised image is converted into logarithmic form, and the product relationship of pixel values in the image is converted into an additive relationship. Gaussian filtering is applied to the logarithmic image to obtain the logarithmic estimate of the illumination component. The logarithmic estimate of the illumination component is subtracted from the logarithmic original image to obtain the logarithmic form of the reflection component. Dynamic range compression is applied to the logarithmic estimate of the illumination component. The compressed logarithmic estimate of the illumination component is converted back to the linear domain through exponential operation to obtain the adjusted illumination component. The logarithmic form of the reflection component is converted back to the linear domain through exponential operation to obtain the linear form of the reflection component. The adjusted illumination component and the reflection component are multiplied in the linear domain and fused to obtain an image with balanced illumination.
[0026] In this embodiment, the distortion parameters obtained by the camera in advance through the Zhang Zhengyou calibration method are called to perform pixel-by-pixel coordinate correction on the image after illumination equalization processing: First, the original image coordinates of each pixel in the image are determined and converted into normalized coordinates with the camera principal point as the origin; then, based on the distortion model, the offset of the normalized coordinates caused by the radial bending of the lens is calculated using the radial distortion coefficient, and the tangential offset caused by the lens installation deviation is calculated using the tangential distortion coefficient, and the total distortion offset is obtained by summing them; the total distortion offset is subtracted from the original normalized coordinates to obtain the corrected normalized coordinates, and then converted back to image coordinates through the camera intrinsic parameters to obtain the corrected pixel position; for pixels whose corrected coordinates are not integers, bilinear interpolation is used to calculate their gray values to fill the information gaps after the pixel position adjustment; after all pixels have been corrected, the preprocessed image with geometric distortion eliminated is output to ensure that the geometric features such as the edges and corners of buildings in the image are consistent with the actual physical form.
[0027] In this embodiment, please refer to Figure 3 The contour extraction module includes: The capture submodule is used to input the preprocessed image to be tested into the Transformer-U-Net model. The encoder part captures the global semantic association of the image through a multi-layer Transformer attention mechanism, and extracts local detail features through convolutional layers to generate multi-scale feature maps. The connection submodule is used by the decoder to gradually restore the feature map resolution through upsampling operations and to make skip connections with the feature maps of the corresponding layers of the encoder to fuse global and local information. The generation submodule is used to generate a binary mask of the building region based on the multi-scale feature map using the sigmoid activation function, and to perform connectivity analysis on the binary mask to remove isolated edges of non-target buildings and obtain the initial outline of the target building. The smoothing submodule is used to smooth the initial contour using the Douglas-Peucker algorithm and output the target contour.
[0028] In this embodiment, the preprocessed image to be tested is input into the encoder part of the Transformer-U-Net model, which segments the image into feature blocks of fixed size. The association weights of each feature block with all other feature blocks are calculated through the multi-head self-attention mechanism of the multi-layer Transformer, and the semantic information of the global scope is aggregated in a weighted manner, such as the distinction between buildings and background and the overall shape features of buildings. At the same time, the encoder embeds multiple sets of convolutional layers, which perform sliding convolution through 3×3 convolutional kernels from the bottom layer to the top layer to extract local detail features of the image, such as building edge texture and corner contours. The feature map size is gradually reduced through max pooling operation, and finally 4-5 multi-scale feature maps with different resolutions are generated. The low-resolution feature map focuses on global semantics, while the high-resolution feature map retains local details. The decoder starts with the lowest resolution feature map output by the encoder and gradually enlarges the resolution of the feature map to twice that of the previous layer through deconvolution or bilinear interpolation combined with convolution, restoring it to the same size as the corresponding layer feature map of the encoder. After the resolution of each layer is restored, skip connections are achieved by feature concatenation or element-wise addition, fusing the feature map of the current layer of the decoder with the feature map of the same resolution layer of the encoder. This allows the decoder to obtain both the global semantic association information transmitted by the encoder and the local detailed features of the layer. Through multi-level fusion, a feature representation containing complete global-local information is gradually constructed. The fused feature map is input to the output layer. The sigmoid activation function maps the feature value of each pixel to the 0-1 range. A threshold of 0.5 is set to convert the feature map into a binary mask. Regions with a value of 1 in the mask correspond to the predicted building area, and regions with a value of 0 correspond to the background. Then, connectivity analysis is performed on the binary mask. The 8-neighborhood labeling method is used to identify independent connected regions in the mask, and the pixel area of each connected region is calculated. An area threshold is set according to the approximate size of the target building. Isolated connected regions with an area smaller than the threshold are removed, such as small light spots in the background and misidentified small object edges. The largest connected region with an area that meets the expectation is retained. The edge contour of this region is the initial contour of the target building.
[0029] In this embodiment, the smoothing submodule includes: selecting two endpoints on the initial contour as the starting point and the ending point, calculating the vertical distance from all intermediate points on the initial contour to the line connecting the starting point and the ending point; determining the point with the largest distance from the vertical distances; if the distance is greater than a preset threshold, retaining the corresponding point and using it as the new endpoint, recursively processing the sub-line segments on both sides; if the distance is less than the threshold, deleting the intermediate point, replacing the original line segment with the line connecting the two endpoints, and finally obtaining the smooth target contour.
[0030] In this embodiment, two endpoints are selected on the initial contour as the start and end points: the initial contour consists of a series of sequentially arranged pixels. First, the start and end points of the contour are determined, and these two points are connected to form a straight line segment. Then, all intermediate pixels on the initial contour between the start and end points are traversed, and the vertical distance from each intermediate point to the straight line segment is calculated one by one. By determining the spatial positional relationship of the straight line segment, the shortest perpendicular length from each intermediate point to the straight line segment is found, which is taken as the vertical distance value of that point. The point with the largest vertical distance is determined. If the distance is greater than a preset threshold, the corresponding point is retained and used as the new endpoint. The sub-segments on both sides are recursively processed: the vertical distances of all intermediate points are compared, and the intermediate point with the largest distance value is selected. The maximum distance is compared with the preset threshold. If the maximum distance is greater than the threshold, it indicates that the point has a significant impact on the contour shape and the point needs to be retained. As the new endpoints, the original line segment is then divided into two sub-segments: one with the starting point and the maximum distance point as endpoints, and the other with the maximum distance point and the ending point as endpoints. The point selection operation is repeated for these two sub-segments, that is, the vertical distance from each midpoint to the corresponding line segment is recalculated. If the distance is less than the threshold, the midpoint is deleted, and the line connecting the two endpoints is used to replace the original line segment, finally obtaining a smooth target contour. When the maximum vertical distance of the selected sub-segments is less than the preset threshold, it means that all midpoints between the current two endpoints have little impact on the overall shape of the contour and can be regarded as redundant points. At this time, these midpoints are deleted, and the line connecting the current two endpoints is used to replace the original continuous contour line segment. The above process is repeated until the maximum vertical distance of all sub-segments is less than the threshold, and no new endpoints are generated. Finally, the combination of line segments formed by connecting all the retained endpoints in sequence is the smoothed target contour.
[0031] In this embodiment, the feature matching module includes: The extraction submodule is used to extract feature points from the target contour and calculate the local descriptor for each feature point, where the feature points include corner points, inflection points and equidistant sampling points; The matching submodule is used to extract the contour feature points and descriptors of the corresponding region from the reference data, calculate the similarity between the local descriptor of the target contour and the descriptor of the reference data through Euclidean distance, initially match the same feature points, and determine the initial geographic coordinates of the target contour on the reference data.
[0032] The iterative solution submodule is used to iteratively solve the initial geographic coordinates using the Levenberg-Marquardt algorithm to obtain the optimized geographic coordinates.
[0033] In this embodiment, the iterative solution submodule includes: using the initial geographic coordinates as the starting point of the iteration, constructing an error function, using the Gauss-Newton method to calculate the parameter update amount, and if the error increases, introducing a damping factor to switch to the gradient descent method to reduce the parameter update step size; recalculating the error function after each iteration until the maximum number of iterations is reached, and outputting the optimized geographic coordinates.
[0034] In this embodiment, the initial geographic coordinates obtained through feature matching are set as the starting parameter vector for iterative optimization. This vector contains the longitude, latitude, and other coordinate information of the key feature points of the target building. Then, an error function is constructed, based on the deviation between the geographic coordinates of the feature points in the reference data and the theoretical coordinates obtained through image projection. The deviations of all feature points are summarized to form an overall error value. Subsequently, the Gauss-Newton method is used to construct the Jacobian matrix by calculating the partial derivatives of the error function with respect to each coordinate parameter. Then, the Jacobian matrix is used to solve a system of linear equations to obtain the parameter update amount, i.e., the adjustment value of the initial geographic coordinates. This update amount is superimposed on the current coordinate parameters, and the error value corresponding to the new parameters is calculated. If the new error is greater than the error before the update, it indicates that the current step size is too large. At this time, a damping factor is introduced. By adding a damping term to the relevant calculation of the Jacobian matrix, the Gauss-Newton method is converted into an approximate form of the gradient descent method, reducing the parameter update step size to avoid error divergence. After each parameter update, the current geographic coordinate parameters are substituted back into the error function to calculate and record the new error value. The current iteration number is compared with the preset maximum iteration number. If the maximum number has not been reached, the process returns to the previous step to continue iterating. The damping factor and update method are adjusted based on the new error value. If the maximum number of iterations is reached, the iteration process stops regardless of whether the current error continues to change. The geographic coordinate parameters obtained from the last iteration are output as the optimized result. This result is the high-precision geographic coordinates of the target building after multiple rounds of iteration and adjustment.
[0035] In this embodiment, when calculating the reprojection error of the optimized geographic coordinates, the camera's intrinsic and extrinsic parameters are first called to project the optimized geographic coordinates of the target building feature points onto the pixel coordinate system of the corresponding image to be measured through a spatial coordinate transformation model, thus obtaining the theoretical pixel coordinates. Then, the actual pixel coordinates of these feature points in the preprocessed image are extracted, and the Euclidean distance between the theoretical and actual pixel coordinates is calculated. The average of all distances is taken as the reprojection error, reflecting the reliability of the coordinate optimization. When calculating the geographic coordinates of corner points through error propagation, error sources such as reprojection error, camera calibration error, and feature matching error are comprehensively considered. Based on the error propagation law, the mapping relationship between each error source and the corner point coordinates is established, and the influence of each error source on the corner point coordinates is quantified to obtain the error range of the geographic coordinates of each corner point of the target building. Combined with the optimized coordinate results, the high-precision geographic coordinates of each corner point are determined, including longitude, latitude, and elevation information. The final output of the high-precision location information of the building includes the specific geographic coordinates of all corner points and their corresponding error ranges. At the same time, the geometric center of the corner point coordinates is calculated as the overall location coordinates of the building, along with the reprojection error value and measurement accuracy level, forming a complete location measurement report.
[0036] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. An image recognition-based building position accuracy measurement system, characterized by, The system comprises: An image acquisition module, configured to acquire remote sensing images and unmanned aerial vehicle aerial images containing a target building as test images, and obtain reference data of a target building region; An image preprocessing module, configured to perform denoising, illumination equalization and distortion correction preprocessing on the test images to obtain preprocessed test images; A contour extraction module, configured to input the preprocessed test images into a Transformer-U-Net model, extract an initial contour of the target building, and perform smoothing processing on the initial contour by using a Douglas-Peucker algorithm to output a target contour; A feature matching module, configured to perform feature matching on the target contour and the reference data, determine initial geographic coordinates corresponding to the target contour on parameter data, and perform iterative solving on the initial geographic coordinates by using a Levenberg-Marquardt algorithm to optimize the initial geographic coordinates; An output module, configured to calculate a re-projection error of the optimized geographic coordinates, and obtain geographic coordinates of corner points of the target building by error propagation calculation, and finally output high-precision position information of the building.
2. The image recognition based building location accuracy measurement system of claim 1, wherein, The image preprocessing module comprises: A denoising submodule, configured to perform noise suppression on the test images by using an adaptive bilateral filtering algorithm to obtain denoised images; An illumination equalization submodule, configured to perform illumination equalization on the denoised images by using a Retinex algorithm to obtain illumination-equalized images; A correction submodule, configured to correct the coordinates of each pixel in the illumination-equalized images by using distortion parameters obtained by pre-calibration of a camera to obtain the preprocessed test images.
3. The image recognition based building position accuracy measurement system of claim 2, wherein, The denoising submodule comprises: The test images are processed pixel by pixel, a dynamic filter kernel is constructed with the current pixel as the center, the weights of the pixels in the filter kernel are dynamically adjusted according to the spatial distance and gray difference between the pixels, the denoised gray value of the current pixel is calculated by weighted average, and the denoised images are obtained.
4. The image recognition based building location accuracy measurement system of claim 2, wherein, The illumination equalization submodule comprises: The denoised images are decomposed into illumination components and reflection components, the illumination components are logarithmically transformed and Gaussian filtered, the adjusted illumination components and reflection components are re-fused to obtain the illumination-equalized images.
5. The image recognition based building position accuracy measurement system of claim 4, wherein, The denoised images are converted into logarithmic form, and the product relationship of the pixel values in the images is converted into an addition relationship; The logarithmic form of the images is Gaussian filtered to obtain the logarithmic estimate value of the illumination components, and the logarithmic form of the reflection components is obtained by subtracting the logarithmic estimate value of the illumination components from the original images in logarithmic form; The logarithmic estimate value of the illumination components is dynamically range compressed, and the compressed logarithmic estimate value of the illumination components is converted back to the linear domain by exponential operation to obtain the adjusted illumination components; The logarithmic form of the reflection components is converted back to the linear form by exponential operation; The adjusted illumination components and the reflection components are multiplied in the linear domain to obtain the illumination-equalized images.
6. The image recognition based building position accuracy measurement system of claim 1, wherein, The contour extraction module comprises: The capturing submodule is configured to input the preprocessed image to be tested into a Transformer-U-Net model, and the encoder part captures global semantic correlation of the image through a multi-layer Transformer attention mechanism, and extracts local detail features through a convolution layer to generate a multi-scale feature map; The connecting submodule is configured to restore the feature map resolution through upsampling operation in the decoder part, and perform a skip connection with the feature map of the corresponding level of the encoder to fuse global and local information; The generating submodule is configured to generate a binary mask of the building region based on the multi-scale feature map through a sigmoid activation function, and perform connectivity analysis on the binary mask to remove isolated edges of non-target buildings to obtain an initial contour of the target building. The smoothing submodule includes:
7. The image recognition based building position accuracy measurement system of claim 6, wherein, selecting two end points on the initial contour as a starting point and an ending point, and calculating the perpendicular distance of all intermediate points on the initial contour to the line connecting the starting point and the ending point; determining the point with the maximum distance from the perpendicular distance, and if the distance is greater than a preset threshold, retaining the corresponding point as a new end point and recursively processing the two side line segments; if the distance is less than the threshold, deleting the intermediate point and replacing the original line segment with the line connecting the two end points to obtain a smooth target contour. The feature matching module includes:
8. The image recognition based building location accuracy measurement system of claim 1, wherein, The extraction submodule is configured to extract feature points from the target contour and calculate local descriptors of each feature point, wherein the feature points include corner points, inflection points and equidistant sampling points. The matching submodule is configured to extract contour feature points and descriptors of the corresponding region from the reference data, calculate the similarity of the local descriptors of the target contour and the descriptors of the reference data through the Euclidean distance, preliminarily match the feature points with the same name, and determine the initial geographic coordinates of the target contour on the reference data. The iterative solving submodule is configured to use the Levenberg-Marquardt algorithm to iteratively solve the initial geographic coordinates to obtain optimized geographic coordinates. The iterative solving submodule includes:
9. The image recognition based building location accuracy measurement system of claim 1, wherein, taking the initial geographic coordinates as the iteration starting point, constructing an error function, calculating the parameter update amount using the Gauss-Newton method, introducing a damping factor to convert to a gradient descent method if the error increases, and reducing the parameter update step; re-calculate the error function after each iteration until the iteration reaches the maximum number, and output the optimized geographic coordinates. The method includes the following steps:
10. A method of implementing an image recognition based building position accuracy measurement system as claimed in claim 1, characterized by, collecting remote sensing images and unmanned aerial vehicle aerial images containing target buildings as images to be tested, and obtaining reference data of the target building region; performing denoising, light equalization and distortion correction preprocessing on the image to be tested to obtain a preprocessed image to be tested; inputting the preprocessed image to be tested into a Transformer-U-Net model, extracting an initial contour of the target building, and smoothing the initial contour using the Douglas-Peucker algorithm to output a target contour; The target contour is matched with reference data, initial geographic coordinates corresponding to the target contour are determined on parameter data, and the initial geographic coordinates are iteratively solved using a Levenberg-Marquardt algorithm to optimize the initial geographic coordinates; The re-projection error of the optimized geographic coordinates is calculated, and the geographic coordinates of the target building corner points are calculated through error propagation, and finally the high-precision position information of the building is output.