Building height estimation method and apparatus, electronic device, and storage medium

Abstract

The application discloses a building height estimation method and device, electronic equipment and storage medium, and relates to the technical field of computer vision. The method comprises the following steps: acquiring monocular remote sensing images, building contour data and photon lidar data of a target area; performing overlapping block reasoning on the monocular remote sensing images and estimating uncertainty to obtain relative depth results and corresponding uncertainty results of each image block; performing view angle bias filtering on the relative depth results of each image block based on local ground candidate statistics to obtain relative height results of each image block; performing global scale unification on the relative height results of each image block according to overlapping area constraints between image blocks to obtain globally aligned relative height results; extracting at least one absolute height control point from the photon lidar data, and outputting a target building height based on the uncertainty results, the absolute height control point, the globally aligned relative height results and the building contour data.

Description

Technical Field

[0001] This application relates to the field of computer vision technology, and in particular to methods, apparatus, electronic devices and storage media for estimating building height. Background Technology

[0002] As urban development gradually shifts from planar expansion to three-dimensional expansion, building height has become a key fundamental parameter in urban spatial morphology analysis, population carrying capacity assessment, energy consumption calculation, thermal environment analysis, and 3D city modeling. Obtaining building height data with wide coverage, low update costs, and stable accuracy has become an important requirement for urban governance and the construction of digital twin cities.

[0003] Currently, traditional building height estimation methods generally suffer from problems such as high data acquisition costs, complex processing procedures, and insufficient cross-regional generalization ability. In addition, monocular depth estimation models typically output relative depth in the sense of perspective projection, which cannot directly correspond to the actual vertical height from the building roof to the ground, resulting in insufficient accuracy in estimating building height.

[0004] Therefore, improving the accuracy of building height estimation and its cross-regional generalization ability is an urgent problem to be solved. Summary of the Invention

[0005] The main objective of this application is to provide a method, apparatus, electronic device, and storage medium for estimating building height, aiming to solve the technical problem of how to improve the accuracy of building height estimation and its cross-regional generalization capability.

[0006] To achieve the above objectives, this application proposes a method for estimating building height, the method comprising: Acquire monocular remote sensing images, building outline data, and photonic lidar data of the target area; Overlapping block reasoning is performed on the monocular remote sensing image and uncertainty is estimated to obtain the relative depth result and the corresponding uncertainty result of each image block; Based on local ground candidate statistics, the relative depth results of each image block are subjected to viewpoint bias filtering to obtain the relative height results of each image block; Based on the constraints of overlapping areas between image blocks, the relative height results of each image block are globally scaled to obtain globally aligned relative height results; At least one absolute height control point is extracted from the photonic lidar data, and the target building height is output based on the uncertainty result, the absolute height control point, the globally aligned relative height result, and the building outline data.

[0007] In one embodiment, the step of performing overlapping block reasoning on the monocular remote sensing image and estimating uncertainty to obtain the relative depth result and corresponding uncertainty result of each image block includes: The monocular remote sensing image is subjected to overlapping and segmentation processing using a first sliding window to obtain multiple image blocks; For each image block, multiple test-time perturbation inputs are generated, and these multiple test-time perturbation inputs are respectively input into a pre-trained monocular depth estimation model to obtain multiple depth prediction results; The multiple depth prediction results are statistically aggregated pixel by pixel to obtain the relative depth results and corresponding uncertainty results for each image block.

[0008] In one embodiment, the step of performing viewpoint bias filtering on the relative depth results of each image patch based on local ground candidate statistics to obtain the relative height results of each image patch includes: For each target pixel in the relative depth result, a second sliding window is established centered on the target pixel; The depth values ​​within the second sliding window are sorted, and a preset proportion of pixels are selected as a local ground candidate set. The mean of the local ground candidate set is calculated to obtain the viewpoint offset value of the target pixel; The relative height value of the target pixel is obtained by subtracting the viewpoint offset value from the original relative depth value of the target pixel, and the relative height result of each image block is calculated based on the relative height value of the target pixel.

[0009] In one embodiment, the step of globally scaling the relative height results of each image block according to the overlap region constraint between image blocks to obtain a globally aligned relative height result includes: Obtain the scale parameters of each image block; The scale parameters of each image block are used as graph nodes, and image block pairs with overlapping areas are used as graph edges to construct a graph optimization model. Based on the relative high consistency within the overlapping regions, a relative constraint error function is constructed; By minimizing the relative constraint error function, the optimal scale parameter of each image block is solved, and the corresponding relative height result is scaled according to the optimal scale parameter to obtain the globally aligned relative height result.

[0010] In one embodiment, the step of extracting at least one absolute height control point from the photonic lidar data and outputting the target building height based on the uncertainty result, the absolute height control point, the globally aligned relative height result, and the building outline data includes: The photon lidar data is filtered based on a preset confidence threshold to obtain filtered photons, and the confidence level of the signal corresponding to the filtered photons is not lower than the preset confidence threshold. The filtered photons are then classified to separate ground photons from non-ground photons. Non-ground photons with a height difference exceeding a preset threshold for ground photons are classified as building photons; Calculate the elevation difference between the first average elevation and the second average elevation to obtain the absolute height control point corresponding to the building instance. The first average elevation is the average elevation of building photons belonging to the same building instance, and the second average elevation is the average elevation of ground photons adjacent to the building instance. Based on the uncertainty results, the absolute height control points, the globally aligned relative height results, and the building outline data, the target building height is output.

[0011] In one embodiment, before the step of outputting the target building height based on the uncertainty result, the absolute height control point, the globally aligned relative height result, and the building outline data, the method further includes: Based on the offset vector between the building roof area and the building footprint, the spatial position of the absolute height control point is corrected from the building footprint to the building roof area to obtain the corrected absolute height control point; Establish a nonlinear polynomial mapping model between the corrected absolute height control points and the globally aligned relative height results; The globally aligned relative height result is input into the nonlinear polynomial mapping model to obtain a preliminary absolute height result.

[0012] In one embodiment, the step of outputting the target building height based on the uncertainty result, the absolute height control point, the globally aligned relative height result, and the building outline data includes: Relative constraints are constructed based on the relative high consistency within overlapping regions; An absolute constraint is constructed based on the difference between the absolute height control point and the preliminary absolute height result; The relative constraints and the absolute constraints are combined into a joint constraint objective function; The preliminary absolute height result is optimized by minimizing the joint constraint objective function to obtain the optimized absolute height result; Based on the uncertainty results of each image block, the pixel values ​​of the overlapping areas in the optimized absolute height results are weighted and fused to generate a target absolute building height map. The target absolute building height map is spatially overlaid with the building outline data, the height statistics corresponding to each building instance are extracted, and the target building height is output based on the height statistics.

[0013] Furthermore, to achieve the above objectives, this application also proposes a building height estimation device, which includes: The data acquisition module is used to acquire monocular remote sensing images, building outline data, and photon lidar data of the target area; The inference estimation module is used to perform overlapping block inference on the monocular remote sensing image and estimate the uncertainty, so as to obtain the relative depth result and the corresponding uncertainty result of each image block; The bias filtering module is used to perform viewpoint bias filtering on the relative depth results of each image block based on local ground candidate statistics to obtain the relative height results of each image block. The global unification module is used to perform global scale unification on the relative height results of each image block according to the overlapping area constraints between image blocks, so as to obtain a globally aligned relative height result; The target output module is used to extract at least one absolute height control point from the photonic lidar data, and output the target building height based on the uncertainty result, the absolute height control point, the globally aligned relative height result, and the building outline data.

[0014] In addition, to achieve the above objectives, this application also proposes an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the building height estimation method as described above.

[0015] In addition, to achieve the above objectives, this application also proposes a non-transitory storage medium on which a computer program is stored, which, when executed by a processor, implements the steps of the building height estimation method described above.

[0016] One or more technical solutions proposed in this application have at least the following technical effects: (1) Acquire monocular remote sensing images, building outline data, and photonic lidar data of the target area; perform overlapping block reasoning on the monocular remote sensing images and estimate the uncertainty to obtain the relative depth results and corresponding uncertainty results of each image block; based on local ground candidate statistics, perform viewpoint bias filtering on the relative depth results of each image block to obtain the relative height results of each image block; according to the overlap area constraints between image blocks, perform global scale unification on the relative height results of each image block to obtain globally aligned relative height results; extract at least one absolute height control point from the photonic lidar data, and output the target building height based on the uncertainty results, absolute height control points, globally aligned relative height results, and building outline data. By adopting the overlapping block reasoning strategy, the model can process remote sensing images of any large area, and the output relative depth results express the distance relationship of the scene, while the accompanying uncertainty results quantify the prediction confidence. This method utilizes building outline data to locate the building and ground region. By applying a viewpoint offset filter based on local ground candidate statistics to the relative depth of each block, it effectively eliminates projection offset caused by off-axis satellite observations, converting the relative depth results into relative height results that better reflect the vertical undulations of the building itself. Furthermore, by using pre-defined overlapping area constraints between blocks to achieve global scale unification, it solves the problem of scale and benchmark mismatch between prediction results of different image blocks, thus obtaining a seamless and internally scale-consistent globally aligned relative height result. Finally, by fusing sparse but absolutely accurate absolute height control points extracted from photonic lidar data, and using the aforementioned uncertainties as a fusion weight reference, the globally consistent relative height field is calibrated to an absolute elevation benchmark, outputting the target building height. This application effectively overcomes the challenges of block differences, projection offset, and benchmark inconsistency, requiring no additional training or large-scale annotation. It can achieve absolute building height estimation in different regions and architectural styles, and has the advantages of low cost, strong generalization, and wide applicability.

[0017] (2) The monocular remote sensing image is processed by overlapping and segmenting using the first sliding window to obtain multiple image blocks. For each image block, multiple test-time perturbation inputs are generated and input into the pre-trained monocular depth estimation model to obtain multiple depth prediction results. The multiple depth prediction results are statistically aggregated pixel by pixel to obtain the relative depth results and corresponding uncertainty results of each image block. Through multiple forward inferences and statistical aggregation based on test-time perturbations, the robustness of relative depth estimation can be improved by using the prediction results under multiple perturbation conditions. On the other hand, the local prediction instability can be explicitly characterized by the standard deviation, thus providing a basis for the subsequent weighted stitching of overlapping areas based on uncertainty.

[0018] (3) Obtain the scale parameters of each image block; use the scale parameters of each image block as graph nodes and the image block pairs with overlapping areas as graph edges to construct a graph optimization model; construct a relative constraint error function based on the relative height consistency within the overlapping area; solve for the optimal scale parameters of each image block by minimizing the relative constraint error function, and scale the corresponding relative height results according to the optimal scale parameters to obtain the globally aligned relative height results. This method can maintain the local continuity and consistency of overlapping areas between image blocks on the one hand, and anchor the overall height results to the real physical scale on the other hand using high-precision sparse absolute control points, thereby obtaining more accurate building absolute height estimation results, which is suitable for building height estimation tasks in complex urban environments and large-scale scenes. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the first embodiment of the building height estimation method of this application; Figure 2 This is a schematic diagram of the modular structure of the building height estimation device of this application; Figure 3 This is a schematic diagram of the device structure for the hardware operating environment of the electronic device involved in this application.

[0022] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0023] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0024] This application provides a method for estimating building height. Specifically, refer to... Figure 1 , Figure 1This is a flowchart illustrating the first embodiment of the building height estimation method of this application. In this embodiment, the building height estimation method includes steps S10 to S50:

[0025] Step S10: Acquire monocular remote sensing images, building outline data, and photon lidar data of the target area.

[0026] It should be noted that the target area can be understood as the geographic spatial range within which the building height is to be estimated. Monocular remote sensing imagery refers to high-resolution two-dimensional images acquired from a single viewpoint by optical remote sensing satellites or airborne platforms. Building outline data is polygonal vector data used to describe the planar shape and location of the building's base within the target area; it can be obtained through image interpretation or existing geographic information databases. Building outline data can be understood as geographic spatial data that accurately describes the area of ​​contact between the building's base and the ground in vector polygon form. Photon lidar data refers to data acquired by photon-counting lidar systems aboard satellites such as ICESat-2, providing high-precision three-dimensional location information of the Earth's surface.

[0027] Step S20: Perform overlapping block reasoning on the monocular remote sensing image and estimate the uncertainty to obtain the relative depth results and corresponding uncertainty results for each image block.

[0028] It should be noted that overlapping block inference refers to the process of dividing a large single remote sensing image into multiple small image blocks with overlapping areas, and then inputting each block into a trained depth estimation model for processing. The relative depth result includes the relative depth value for each pixel, reflecting the relative distance between different points in the scene and the camera; for example, a larger value indicates a greater distance. The uncertainty result can be understood as a quantitative indicator representing the reliability or confidence level of the depth value provided by the depth estimation model when outputting the relative depth value for each pixel.

[0029] Step S30: Based on local ground candidate statistics, viewpoint bias filtering is applied to the relative depth results of each image block to obtain the relative height results of each image block.

[0030] It should be noted that local ground candidate statistics can be understood as selecting pixels that may represent the ground within a certain neighborhood of the building outline within a single image block, based on building outline data, and performing statistical analysis on the relative depth values ​​of these pixels, such as sorting and quantile calculation. Due to off-axis satellite imaging, the position of the building top on the image is projected to be offset from its bottom footprint. Viewpoint offset filtering utilizes local ground statistics to systematically adjust the relative depth values ​​of the building area, aiming to convert "relative depth" into "relative height" that better represents the building's vertical direction, thereby reducing systematic biases caused by projection geometry.

[0031] Step S40: Based on the overlap area constraints between image blocks, perform global scale unification on the relative height results of each image block to obtain globally aligned relative height results.

[0032] It should be noted that the overlapping region constraint refers to the spatial overlap between adjacent image blocks. By utilizing the overlapping region constraint, optimization methods can be used to calculate adjustment parameters for the height value of each image block, such as scaling factors and offsets, ensuring that the height prediction of all image blocks in the overlapping region is consistent, thereby eliminating the scale and baseline inconsistency problem caused by block processing.

[0033] Step S50: Extract at least one absolute height control point from the photonic lidar data, and output the target building height based on the uncertainty result, the absolute height control point, the globally aligned relative height result, and the building outline data.

[0034] It should be noted that the absolute height control point refers to the sparse absolute height control information calculated from the elevation difference between the building photon and the corresponding ground photon and associated with the building instance. The target building height is the final output value of the vertical height from the building roof to the ground.

[0035] In this embodiment, by employing an overlapping block inference strategy, the model can process remote sensing images of arbitrarily large areas. The output relative depth result expresses the distance relationship of the scene, while the accompanying uncertainty result quantifies the prediction reliability. Building outline data is used to locate the building and ground areas. By applying a viewpoint bias filter based on local ground candidate statistics to the relative depth of each block, the projection offset caused by off-axis satellite observation is effectively eliminated, converting the relative depth result into a relative height result that better reflects the vertical undulations of the building itself. Furthermore, global scale unification is achieved using pre-set overlapping area constraints between blocks, solving the scale and reference mismatch problem between prediction results of different image blocks, thus obtaining a seamless and internally scale-consistent globally aligned relative height result. Finally, by fusing sparse but absolutely accurate absolute height control points extracted from photonic lidar data, and using the aforementioned uncertainty as a fusion weight reference, the globally consistent relative height field is calibrated to an absolute elevation reference, outputting the target building height. This embodiment effectively overcomes the challenges of block differences, projection offsets, and inconsistent benchmarks, achieving accurate estimation of the absolute height of buildings over a large area with high consistency and availability. It possesses superior cross-regional generalization capabilities and engineering promotion potential.

[0036] In one implementation, step S20 includes: Step S201: The monocular remote sensing image is overlaid and segmented using the first sliding window to obtain multiple image blocks.

[0037] It should be noted that, since large-format, high-resolution remote sensing images cannot usually be directly input into a monocular depth estimation model in their entirety, this embodiment can use a first sliding window to perform overlapping block processing on the input image, obtaining a set of image blocks, represented as: .in, Let represent the i-th image patch, and n represent the total number of image patches. For example, the image patch size can be set to 1024×1024 pixels, and the overlap ratio between adjacent image patches can be set to 50%. The purpose of setting the overlap area is to: alleviate the memory and computational overhead caused by direct inference of large-format remote sensing images; and provide a spatial overlap basis for subsequent construction of relative height scale uniform constraints based on the overlap area and for weighted stitching based on uncertainty.

[0038] Step S202: For each image block, generate multiple test-time perturbation inputs and input the multiple test-time perturbation inputs into the pre-trained monocular depth estimation model to obtain multiple depth prediction results.

[0039] For example, the pre-trained monocular depth estimation model can be a Distill Any Depth (DAD) model, more preferably a DAD large model, to output pixel-level relative depth results for each image patch. This can be done for each image patch. A test-time perturbation set is constructed, and multiple forward inferences are performed based on the same pre-trained monocular depth estimation model to obtain more stable relative depth estimation results and their corresponding uncertainty measures. Specifically, let... This represents the perturbation operator at the t-th test. Then from the original image block Generate the t-th perturbation input. The corresponding formula is: .in, This represents the input transformation operation performed on the image patch during the testing phase. The purpose of the input transformation operation is to simulate the response fluctuations of remote sensing imagery in local textures, building boundaries, occluded areas, and regions with incomplete context, thereby improving the inference robustness of the model in off-axis remote sensing building scenes.

[0040] It should be noted that at least one perturbation from a preset set of perturbation operators can be applied to the image block to generate a corresponding perturbation input. When the perturbation includes geometric transformation, after obtaining the depth prediction result, the depth prediction result is also restored to the original coordinate system of the image block through the corresponding inverse transformation. During testing, the perturbation can include one or more of the following: brightness change, blurring, flipping operation, and rotation operation.

[0041] For example, suppose This represents a pre-trained monocular depth estimation model with fixed parameters, where, This represents the model parameters. During the test-inference phase, the model parameters remain unchanged, and no training updates are performed. For each perturbation input... Perform forward inference once each to obtain the corresponding depth prediction results. The corresponding formula is:

[0042] Disturbance during testing When geometric transformations are involved, to ensure that the results of multiple inferences are under the same spatial reference, the corresponding prediction results need to be inversely transformed. Restore to the original image block Using a coordinate system, we obtain depth prediction results under a unified coordinate expression. , represented as: Step S203: Perform pixel-by-pixel statistical aggregation on multiple depth prediction results to obtain the relative depth results and corresponding uncertainty results for each image block.

[0043] For example, when obtaining the same image block The corresponding T depth prediction results Then, pixel-by-pixel statistics were performed on multiple prediction results, and the mean value was used as the relative depth result for the i-th image patch. , represented as: The standard deviation is used as the uncertainty result for the i-th image patch. , represented as: This yields the relative depth result set D and the uncertainty result set U, respectively, as follows: In this embodiment, overlapping block processing using a first sliding window transforms large-scale remote sensing images into a scale that the model can handle, providing a spatial basis for subsequent inter-block alignment and fusion. By generating a test-time perturbation input set containing various perturbations for each image block and performing forward inference separately, the model response fluctuations caused by factors such as local texture, occlusion, and incomplete boundaries in real images are simulated, thereby enhancing the model's inference robustness in complex off-axis remote sensing scenarios. When the perturbation includes geometric transformations, performing the corresponding inverse transformation on the prediction results ensures that all prediction results are in a unified spatial coordinate system, creating conditions for subsequent statistical aggregation. Finally, by performing pixel-by-pixel statistical aggregation on multiple prediction values ​​at the same location and using the mean as the relative depth result, the overall stability of depth estimation is improved. For large-format off-axis remote sensing images, the contextual integrity of the same building differs between the central and edge regions of the image block, and building boundaries, occluded areas, and areas with weak textures are more prone to inference fluctuations. Through the above-mentioned multiple forward inferences and statistical aggregations based on test-time perturbations, on the one hand, the robustness of relative depth estimation can be improved by utilizing the prediction results under multiple perturbation conditions; on the other hand, the degree of local prediction instability can be explicitly characterized by the standard deviation, thereby providing a basis for subsequent weighted splicing of overlapping regions based on uncertainty.

[0044] Since the monocular depth estimation model outputs relative depth in the sense of perspective projection, which represents the distance relationship between the pixel-corresponding target and the observation viewpoint; while building height estimation focuses on the vertical height difference between the roof and the ground. Therefore, it is necessary to convert the relative depth result into a relative height result with the ground as a reference. In one implementation, step S30 includes:

[0045] Step S301: For each target pixel in the relative depth result, establish a second sliding window centered on the target pixel.

[0046] For example, for any pixel position in the relative depth map d Create a sliding window centered on that pixel. The corresponding formula is: Where s represents the sliding window size parameter, and the relative depth map d consists of multiple relative depth results. composition.

[0047] Step S302: Sort the depth values ​​within the second sliding window and select a preset proportion of pixels as a local ground candidate set.

[0048] For example, the depth values ​​within the window are flattened and sorted in ascending order, and pixels with a preset proportion are selected as a local ground candidate set. The corresponding formula is: in, This indicates a preset pixel ratio, where p represents the preset ratio parameter. This indicates a floor operation. The Flatten operation prepares the user for the subsequent Sort function. The sorting operation needs to be performed on a one-dimensional sequence. Only by "flattening" all depth values ​​representing different positions within the window into the same list can the depth values ​​of the entire local region be uniformly sorted in ascending order, thus facilitating the selection of the smallest value. Each value serves as a local ground candidate set. .

[0049] Step S303: Calculate the mean of the local ground candidate set to obtain the view offset value of the target pixel.

[0050] For example, the mean of the local ground candidate set is calculated as the view offset value for the current location. The corresponding formula is: It should be noted that the target pixel is the current position. This represents an element in the local ground candidate set.

[0051] Step S304: Subtract the view offset value from the original relative depth value of the target pixel to obtain the relative height value of the target pixel, and calculate the relative height result of each image block based on the relative height value of the target pixel.

[0052] For example, the relative height is obtained by subtracting the viewpoint offset value from the original relative depth value. : By performing the above processing pixel-by-pixel on the entire relative depth map, a relative height map can be obtained; after performing the above processing on all image blocks, a set of relative height results can be obtained, represented as: in, This represents the relative height result corresponding to the i-th image block.

[0053] It should be noted that the viewpoint offset value is directly obtained from a preset low-proportion of pixels sorted within a local window of the relative depth map, rather than being calculated based on the world coordinates of the building's ground or an explicit ground elevation model. Therefore, it is possible to convert the "distance representation affected by viewpoint" into a "height difference representation from the roof to the ground" without needing to restore the ground elevation in the world coordinate system. This method effectively eliminates the continuous influence of off-axis observation and gradually varying viewpoint offset on the relative depth results.

[0054] In this embodiment, a second sliding window is established for each pixel in the depth map, providing a context for local statistical analysis. By sorting all depth values ​​within the window and selecting the lowest value (based on a preset ratio) as a local ground candidate set, the depth level representing the local ground is directly estimated from the relative depth map based on the spatial prior that the ground is typically the lowest locally. Next, the mean of this candidate set is calculated as a viewpoint bias value, thereby quantifying the systematic depth deviation caused by satellite off-axis imaging and perspective projection in this local area. Finally, the original relative depth value of each pixel is subtracted from its corresponding viewpoint bias value, effectively eliminating the systematic bias and obtaining a relative height value that more realistically reflects the vertical undulations of the object itself. This lays a crucial foundation for subsequent fusion of absolute control points and restoration of the true physical height.

[0055] Although viewpoint bias filtering has eliminated local reference drift in the relative height results, making different image blocks more consistent in height reference, scale differences may still exist between image blocks. Therefore, in one embodiment, step S40 includes:

[0056] Step S401: Obtain the scale parameters of each image block.

[0057] For example, a relative height image block can be randomly selected. As a global reference image block, for the remaining image blocks Introducing scale parameters The relative height result after global alignment Represented as: Step S402: The scale parameters of each image block are used as graph nodes, and image block pairs with overlapping areas are used as graph edges to construct a graph optimization model.

[0058] For example, the scale parameters of each image block are used as graph nodes V, and the pairs of image blocks with overlapping regions are used as graph edges E, to construct a graph G = (V, E).

[0059] Step S403: Based on the relative high consistency within the overlapping regions, construct a relative constraint error function.

[0060] Specifically, a weighting coefficient can be assigned to each pixel within the overlapping area. When the relative height value of any related image block at the pixel location exceeds a preset height threshold, a first weighting coefficient is assigned to the pixel; otherwise, a second weighting coefficient is assigned.

[0061] For example, for any two image blocks i and j with overlapping regions, let their overlapping pixel set be denoted as . A relative constraint error function is constructed based on the overlapping region. , represented as:

[0062] Where w(p) represents the weighting coefficient corresponding to pixel p. To highlight the role of buildings in scale uniformity, for heights exceeding a preset threshold... Pixels with higher weights are assigned the weight function as follows:

[0063] in, This represents the enhanced weight of the building area, i.e., the first weighting coefficient, while the second weighting coefficient can be 1.

[0064] Step S404: By minimizing the relative constraint error function, the optimal scale parameter of each image block is solved, and the corresponding relative height result is scaled according to the optimal scale parameter to obtain the globally aligned relative height result.

[0065] For example, by minimizing the relative constraint error function The optimal scale parameters for each image patch are obtained. This generates a globally aligned set of relative height results. This approach allows for the elimination of scale drift caused by independent reasoning by utilizing consistency constraints in overlapping regions, and enhances the contribution of building regions to the global scale unification process.

[0066] In this implementation, by introducing scale parameters for each independently inferred image block, the problem of scale inconsistency between block results is transformed into a problem of solving these parameters. By constructing a graph optimization model, image blocks are abstracted as graph nodes, and their spatial overlap relationships are abstracted as graph edges, thus transforming the physical spatial consistency constraint into a computable mathematical optimization problem. When constructing the core relative constraint error function, a weighted coefficient allocation mechanism based on a preset height threshold is introduced, assigning a higher first weighting coefficient to areas that may belong to buildings, ensuring that the optimization focuses more on the scale uniformity of building heights rather than secondary areas such as the ground. Finally, the optimal scale parameters are solved by minimizing this error function, and the block results are scaled accordingly to obtain globally aligned relative height results. Through a rigorous global optimization framework, utilizing only the overlap information of the image blocks themselves, adaptive scale correction and seamless stitching without manual intervention are achieved, fundamentally solving the scale drift and seam problems caused by block inference, and generating a high-quality, internally consistent, large-scale relative height field for subsequent fusion of absolute control points.

[0067] In one implementation, step S50 includes: Step S501: Filter the photon lidar data based on a preset confidence threshold to obtain the filtered photons.

[0068] The selected photons correspond to signal confidence levels no lower than a preset confidence threshold. For example, photon data can be segmented along the ICESat-2 trajectory direction into segments of a preset length, preferably 1 kilometer. Subsequently, photons are filtered based on the built-in signal confidence tags in the ICESat-2 ATL03 data, retaining photons with confidence levels no lower than 2 as high-confidence photons.

[0069] Step S502: The filtered photons are classified to separate ground photons from non-ground photons.

[0070] For example, anomaly points can be eliminated using a sliding window histogram filter. A 1-meter sliding window can be used for elevation histogram statistics. Then, iterative RANSAC linear fitting is used to obtain candidate building photons and candidate ground photons, and a 10-meter sliding window is used to perform RANSAC linear fitting. Subsequently, DBSCAN clustering, region growing, and cloth simulation filtering operations are used to separate ground photons from non-ground photons.

[0071] Step S503: Non-ground photons with a height difference threshold higher than that of ground photons are identified as building photons.

[0072] For example, for photons above a preset height difference threshold on the ground... Non-ground photons are classified as building photons, with a preset height difference threshold. A length of 3 meters is acceptable.

[0073] Step S504: Calculate the elevation difference between the first average elevation and the second average elevation to obtain the absolute height control point corresponding to the building instance.

[0074] The first average elevation is the average elevation of building photons belonging to the same building instance, and the second average elevation is the average elevation of ground photons adjacent to the building instance.

[0075] For example, for each building instance, the average elevation of the building photons inside it and the average elevation of the adjacent ground photons outside it are calculated, and the difference between the two is used as the absolute height control value corresponding to that building instance. The corresponding formula is: in, This represents the average elevation of building photons within the i-th building instance. It represents the average elevation of the ground photons in the vicinity of the i-th building instance.

[0076] This yields the set CH of sparse absolute height control points for buildings, denoted as: in, This represents the absolute height control value corresponding to the i-th building instance. This indicates the spatial location of the control point, and m represents the total number of absolute height control points.

[0077] It should be noted that sparse control constraints can be used subsequently, so it is not required to extract absolute control points for all buildings. It is only necessary to obtain sparse control points that can be anchored at the global scale within the study area.

[0078] Given that most publicly available high-resolution remote sensing imagery is off-axis, there is often a spatial offset between the visible roof positions and the actual footprints of buildings in the imagery. ICESat-2 photon positions, however, possess high spatial accuracy and do not exhibit this type of roof projection offset. Therefore, spatial alignment needs to be performed before using absolute height control points for relative-to-absolute height mapping.

[0079] Based on this, before step S505, the method further includes: correcting the spatial position of the absolute height control point from the building footprint to the building roof area based on the offset vector between the building roof area and the building footprint, to obtain the corrected absolute height control point; establishing a nonlinear polynomial mapping model between the corrected absolute height control point and the globally aligned relative height result; and inputting the globally aligned relative height result into the nonlinear polynomial mapping model to obtain the preliminary absolute height result.

[0080] For example, for a set of buildings B, the building footprints F, roof areas R, and offset vector V from the roof to the footprints are extracted from the remote sensing image. Then, ICESat-2 building photons are matched with the building footprints F, the average elevation of the building photons within each building footprint is calculated, and the difference is made with the average elevation of the ground photons outside the building to obtain the absolute height control value corresponding to each building instance. Then, based on the offset vector V, the absolute height control points are corrected from the building footprint positions to the visible roof positions to achieve spatial alignment with the relative height results.

[0081] After spatial alignment, the globally aligned relative height results are paired with the absolute height control points to construct a multinomial mapping model θ(h) from relative height to absolute height. Considering that the monocular depth estimation base model is trained under disparity space supervision, there is a non-linear relationship between its output relative depth or relative height and the true absolute height; therefore, a multinomial function is used for fitting. The preliminary absolute height result ah can be expressed as:

[0082] The nonlinear polynomial mapping model θ(h) can be expressed as: in, The coefficients are the polynomial coefficients to be fitted. Polynomial fitting can be performed by a processor calling a numerical computation program on a server, workstation, desktop computer, or other electronic device.

[0083] Step S505: Based on the uncertainty results, absolute height control points, globally aligned relative height results, and building outline data, output the target building height.

[0084] It should be noted that the preliminary absolute height result may be affected by factors such as initial mapping error, local area scale propagation error, and sparse distribution of control points. To further improve the accuracy of absolute height estimation, joint optimization can be performed under the combined effect of relative and absolute constraints. Therefore, step S505 includes: constructing relative constraints based on the relative height consistency within the overlapping area; constructing absolute constraints based on the difference between the absolute height control points and the preliminary absolute height result; combining the relative and absolute constraints into a joint constraint objective function; optimizing the preliminary absolute height result by minimizing the joint constraint objective function to obtain the optimized absolute height result; performing weighted fusion of pixel values ​​in the overlapping area of ​​the optimized absolute height result based on the uncertainty results of each image block to generate a target absolute building height map; spatially overlaying the target absolute building height map with the building outline data, extracting the height statistics corresponding to each building instance, and outputting the target building height based on the height statistics.

[0085] For example, applying the nonlinear polynomial mapping model θ(h) to the globally aligned relative height result yields the initial absolute height image patch set AH, expressed as: in, Let represent the initial absolute height result of the i-th image patch. The height consistency constraint built based on the overlapping region is retained as a relative constraint; simultaneously, the difference between the absolute height control points and the initial absolute height result is constructed as an absolute constraint. Let This represents all points that spatially cover the i-th absolute height control point. The initial set of absolute height image patch indices, Indicates absolute altitude control point The j-th initial absolute height image block The corresponding pixel position within the range then absolutely controls the error. It can be represented as:

[0086] in, This represents the scale parameter corresponding to the j-th image block.

[0087] Furthermore, absolute constraints With relative constraints Combine them to construct a joint constraint objective function. , represented as: in, The weighting coefficients are used to balance the contributions of relative and absolute constraints. 5 is acceptable.

[0088] By minimizing the joint constraint objective function The optimized absolute height result set FH is obtained, which is represented as: in, This represents the optimized absolute height result of the i-th image block.

[0089] This approach maintains local continuity and consistency of overlapping areas between image blocks while using high-precision sparse absolute control points to anchor the overall height result to the real physical scale, thus obtaining a more accurate estimate of the absolute height of buildings.

[0090] Since the absolute height prediction results for the same spatial location may be given by multiple overlapping image blocks, the prediction quality of different image blocks varies. Weighted stitching of overlapping regions can be performed based on the uncertainty of each image block to prioritize the preservation of prediction results in areas of low uncertainty and suppress error propagation from edge truncation and areas with insufficient context.

[0091] Specifically, at pixel location (x, y), there are K overlapping image blocks providing optimized absolute height predictions. The corresponding uncertainty is The final absolute height value after fusion It can be represented as: in, This is used to convert the result after inverse normalization of uncertainty into fusion weights.

[0092] The above method yields the final absolute building height map corresponding to the entire image. Subsequently, the final target absolute building height map is spatially overlaid with the building outline data, and the height statistics for each building instance are extracted. These height statistics can be the average, maximum, median, or other statistical measures of pixel heights within the building area. Based on the target absolute building height map and building outline data, individual building height results, building height classification results, or 3D building models based on building outlines and height extrusion can be further output.

[0093] The target absolute building height map can be output in raster data format, the individual building height results can be output in building instance attribute table format, the building height classification results can be output in thematic classification map format, and the 3D building model can be output in LoD1 building model format. The output process can be completed by the processor of a server, workstation, desktop computer, laptop, or other electronic device.

[0094] In this embodiment, the original photon data is filtered by setting a pre-set confidence threshold, effectively removing noise and improving the data quality used for control point extraction. Subsequently, a series of classification operations accurately separate ground photons from non-ground photons, and building photons are robustly identified from them using a preset elevation difference threshold. By calculating the elevation difference between the first average elevation of building photons within the same building instance and the second average elevation of adjacent ground photons, sparse but high-precision absolute height control points are obtained. To overcome the problem of rooftops and footprints not coinciding in off-axis images, the control point positions are corrected from the footprints to the visible rooftop area in the image based on the offset vector, achieving precise spatial alignment between sparse absolute control and dense relative prediction. On this basis, a nonlinear polynomial mapping model is established to initially transform the globally aligned relative height results to an absolute scale, obtaining preliminary absolute height results. Furthermore, in the joint optimization, relative constraints based on overlap consistency and absolute constraints based on control point conformity are introduced simultaneously. A joint constraint objective function is constructed and solved by minimization, ensuring that the optimized absolute height results are strictly anchored to the absolute elevation benchmark while maintaining internal smoothness and continuity. Finally, the uncertainties of each image patch are used to weightedly fuse multiple optimized predictions for overlapping areas, prioritizing predictions with high confidence levels. This effectively suppresses errors in edge or low-quality areas, generating a seamless target absolute building height map. Height statistics are then extracted by overlaying the map with contour data, outputting the final building height product. This implementation systematically solves core challenges such as the difficulty in utilizing control points in off-axis imagery, relative-to-absolute scale conversion, and stitching defects caused by uneven patch quality. It achieves automated, high-precision estimation of building heights over a large area with high consistency and absolute benchmark.

[0095] Based on the same inventive concept, such as Figure 2 As shown, this application also provides a building height estimation device, which includes: Data acquisition module 10 is used to acquire monocular remote sensing images, building outline data and photon lidar data of the target area; The inference estimation module 20 is used to perform overlapping block inference on the monocular remote sensing image and estimate the uncertainty to obtain the relative depth result and the corresponding uncertainty result of each image block; The bias filtering module 30 is used to perform viewpoint bias filtering on the relative depth results of each image block based on local ground candidate statistics to obtain the relative height results of each image block. The global unification module 40 is used to perform global scale unification on the relative height results of each image block according to the overlapping area constraints between image blocks, so as to obtain a globally aligned relative height result. The target output module 50 is used to extract at least one absolute height control point from the photonic lidar data, and output the target building height based on the uncertainty result, the absolute height control point, the globally aligned relative height result, and the building outline data.

[0096] It should be noted that the building height estimation device provided in this application embodiment and the building height estimation method provided in this application embodiment are based on the same application concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned building height estimation method, and the repeated parts will not be described again.

[0097] In some embodiments, this application provides an electronic device 2000, which includes: at least one processor 2001; and a memory 2003 communicatively connected to the at least one processor 2001; wherein the memory 2003 stores instructions executable by the at least one processor 2001, the instructions being executed by the at least one processor 2001 to enable the at least one processor 2001 to perform the building height estimation method in the above embodiments.

[0098] The following is for reference. Figure 3 This diagram illustrates a structural schematic of an electronic device 2000 suitable for implementing embodiments of this application. The electronic device 2000 in this application embodiment may include a processor 2001 and a memory 2003. The processor 2001 and the memory 2003 are connected, for example, via a bus 2002. Optionally, the electronic device 2000 may also include a transceiver 2004. It should be noted that in practical applications, the transceiver 2004 is not limited to one, and the structure of this electronic device 2000 does not constitute a limitation on the embodiments of this application.

[0099] Processor 2001 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 2001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0100] Bus 2002 may include a pathway for transmitting information between the aforementioned components. Bus 2002 may be a PCI bus or an EISA bus, etc. Bus 2002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0101] The memory 2003 may be ROM or other type of static storage device capable of storing static information and instructions, RAM or other type of dynamic storage device capable of storing information and instructions, or EEPROM, CD-ROM or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0102] The memory 2003 stores the application code that executes the scheme of this application, and its execution is controlled by the processor 2001. The processor 2001 executes the application code stored in the memory 2003 to implement... Figure 2 The operation of the building height estimation device provided in the illustrated embodiment.

[0103] This application also provides a non-transitory storage medium storing a computer program that, when executed by a processor, implements the aforementioned building height estimation method. This storage medium may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not be assembled into that device / apparatus / system. The aforementioned non-transitory storage medium carries one or more programs, which, when executed, implement the method as described in the embodiments or implementations of this application.

[0104] According to embodiments of this application, a non-transitory storage medium can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. A computer-readable signal medium can also be any storage medium other than a computer-readable storage medium that can transmit, propagate, or transfer a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the storage medium can be transmitted using any suitable medium, including but not limited to: wireless, wired, optical fiber, radio frequency signals, etc., or any suitable combination thereof.

[0105] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0106] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application. Therefore, the scope of this application should not be limited to the above embodiments. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for estimating building height, characterized in that, The method includes: Acquire monocular remote sensing images, building outline data, and photonic lidar data of the target area; Overlapping block reasoning is performed on the monocular remote sensing image and uncertainty is estimated to obtain the relative depth result and the corresponding uncertainty result of each image block; Based on local ground candidate statistics, the relative depth results of each image block are subjected to viewpoint bias filtering to obtain the relative height results of each image block; Based on the constraints of overlapping areas between image blocks, the relative height results of each image block are globally scaled to obtain globally aligned relative height results; At least one absolute height control point is extracted from the photonic lidar data, and the target building height is output based on the uncertainty result, the absolute height control point, the globally aligned relative height result, and the building outline data.

2. The method as described in claim 1, characterized in that, The step of performing overlapping block reasoning on the monocular remote sensing image and estimating uncertainty to obtain the relative depth result and corresponding uncertainty result of each image block includes: The monocular remote sensing image is subjected to overlapping and segmentation processing using a first sliding window to obtain multiple image blocks; For each image block, multiple test-time perturbation inputs are generated, and these multiple test-time perturbation inputs are respectively input into a pre-trained monocular depth estimation model to obtain multiple depth prediction results; The multiple depth prediction results are statistically aggregated pixel by pixel to obtain the relative depth results and corresponding uncertainty results for each image block.

3. The method as described in claim 1, characterized in that, The step of performing viewpoint bias filtering on the relative depth results of each image patch based on local ground candidate statistics to obtain the relative height results of each image patch includes: For each target pixel in the relative depth result, a second sliding window is established centered on the target pixel; The depth values ​​within the second sliding window are sorted, and a preset proportion of pixels are selected as a local ground candidate set. The mean of the local ground candidate set is calculated to obtain the viewpoint offset value of the target pixel; The relative height value of the target pixel is obtained by subtracting the viewpoint offset value from the original relative depth value of the target pixel, and the relative height result of each image block is calculated based on the relative height value of the target pixel.

4. The method as described in claim 1, characterized in that, The step of performing global scale unification on the relative height results of each image block based on the overlap region constraint between image blocks to obtain globally aligned relative height results includes: Obtain the scale parameters of each image block; The scale parameters of each image block are used as graph nodes, and image block pairs with overlapping areas are used as graph edges to construct a graph optimization model. Based on the relative high consistency within the overlapping regions, a relative constraint error function is constructed; By minimizing the relative constraint error function, the optimal scale parameter of each image block is solved, and the corresponding relative height result is scaled according to the optimal scale parameter to obtain the globally aligned relative height result.

5. The method as described in claim 1, characterized in that, The step of extracting at least one absolute height control point from the photonic lidar data, and outputting the target building height based on the uncertainty result, the absolute height control point, the globally aligned relative height result, and the building outline data, includes: The photon lidar data is filtered based on a preset confidence threshold to obtain filtered photons, and the confidence level of the signal corresponding to the filtered photons is not lower than the preset confidence threshold. The filtered photons are then classified to separate ground photons from non-ground photons. Non-ground photons with a height difference exceeding a preset threshold for ground photons are classified as building photons; Calculate the elevation difference between the first average elevation and the second average elevation to obtain the absolute height control point corresponding to the building instance. The first average elevation is the average elevation of building photons belonging to the same building instance, and the second average elevation is the average elevation of ground photons adjacent to the building instance. Based on the uncertainty results, the absolute height control points, the globally aligned relative height results, and the building outline data, the target building height is output.

6. The method as described in claim 5, characterized in that, Before the step of outputting the target building height based on the uncertainty result, the absolute height control point, the globally aligned relative height result, and the building outline data, the method further includes: Based on the offset vector between the building roof area and the building footprint, the spatial position of the absolute height control point is corrected from the building footprint to the building roof area to obtain the corrected absolute height control point; Establish a nonlinear polynomial mapping model between the corrected absolute height control points and the globally aligned relative height results; The globally aligned relative height result is input into the nonlinear polynomial mapping model to obtain a preliminary absolute height result.

7. The method as described in claim 6, characterized in that, The step of outputting the target building height based on the uncertainty result, the absolute height control point, the globally aligned relative height result, and the building outline data includes: Relative constraints are constructed based on the relative high consistency within overlapping regions; An absolute constraint is constructed based on the difference between the absolute height control point and the preliminary absolute height result; The relative constraints and the absolute constraints are combined into a joint constraint objective function; The preliminary absolute height result is optimized by minimizing the joint constraint objective function to obtain the optimized absolute height result; Based on the uncertainty results of each image block, the pixel values ​​of the overlapping areas in the optimized absolute height results are weighted and fused to generate a target absolute building height map. The target absolute building height map is spatially overlaid with the building outline data, the height statistics corresponding to each building instance are extracted, and the target building height is output based on the height statistics.

8. A building height estimation device, characterized in that, The building height estimation device includes: The data acquisition module is used to acquire monocular remote sensing images, building outline data, and photon lidar data of the target area; The inference estimation module is used to perform overlapping block inference on the monocular remote sensing image and estimate the uncertainty, so as to obtain the relative depth result and the corresponding uncertainty result of each image block; The bias filtering module is used to perform viewpoint bias filtering on the relative depth results of each image block based on local ground candidate statistics to obtain the relative height results of each image block. The global unification module is used to perform global scale unification on the relative height results of each image block according to the overlapping area constraints between image blocks, so as to obtain a globally aligned relative height result; The target output module is used to extract at least one absolute height control point from the photonic lidar data, and output the target building height based on the uncertainty result, the absolute height control point, the globally aligned relative height result, and the building outline data.

9. An electronic device, characterized in that, The electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the building height estimation method as described in any one of claims 1 to 7.

10. A non-transitory storage medium, characterized in that, The non-transitory storage medium stores a computer program, which, when executed by a processor, implements the steps of the building height estimation method as described in any one of claims 1 to 7.

Images