Land utilization monitoring model based on three-dimensional reconstruction
By combining 3D reconstruction technology with UAV aerial photography and semantic segmentation, dense point cloud data is generated, which solves the problem of insufficient accuracy and reliability of traditional monitoring methods in complex terrain and realizes high-precision land use change monitoring.
Patent Information
- Application Number
- CN202511043366.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-28
- Publication Date
- 2025-11-14
AI Technical Summary
Traditional land use monitoring methods cannot accurately reflect the three-dimensional characteristics of land use and the actual situation under complex terrain conditions, resulting in insufficient monitoring accuracy and reliability.
A land use monitoring model based on 3D reconstruction is adopted. Multi-angle and multi-scale image data are acquired through drone aerial photography. Semantic segmentation is performed using a fully convolutional neural network to generate dense point cloud data and construct a 3D model, thereby achieving deep fusion of 2D semantic information and 3D geometric information.
It enables high-precision monitoring of land use change, and can identify the types, spatial structures and shapes of land features, thus improving the accuracy and reliability of monitoring.
Smart Images
Figure CN120953529A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of three-dimensional scene reconstruction and real-time monitoring technology, and in particular to a land use monitoring model based on three-dimensional reconstruction. Background Technology
[0002] Against the backdrop of rapid technological advancements and increasing demands for intelligent applications, 3D scene reconstruction and real-time positioning technologies have become hot topics in research and industry. With breakthroughs in machine learning, artificial intelligence, and computer vision, more accurate and efficient scene understanding capabilities are crucial for various automated systems. Traditional 3D scene reconstruction methods often rely on complex hardware systems and limited sensor data processing capabilities, which restricts their application scope and processing efficiency.
[0003] In the field of land resource management, accurate and real-time monitoring of land use is crucial for rational land planning, ensuring sustainable resource utilization, and protecting the ecological environment. Taking the canal area as an example, land use along the canal exhibits a significant three-dimensional and complex characteristic, including ecological land such as river channels and embankment protection forests, as well as transportation facilities such as wharves and warehouses and riverside roads, and further involving three-dimensional spatial development scenarios such as bridges across the river and buildings along the river. With the acceleration of urbanization, land use along the canal is undergoing drastic changes, with frequent problems such as illegal filling of the river for land reclamation, disorderly development of waterfront spaces, and illegal construction of structures across the river, posing serious challenges to the canal's hydrological ecology, flood control safety, and cultural heritage protection.
[0004] Traditional land use monitoring methods (such as manual field surveys and 2D remote sensing image analysis) have gradually revealed significant limitations in canal area monitoring: 2D remote sensing images struggle to capture changes in the longitudinal profile of the canal, the structural characteristics of revetment projects, and the height of buildings along the banks, resulting in blind spots in the monitoring of three-dimensional development of the riverside space; manual patrols are limited by canal accessibility and monitoring frequency, making it impossible to promptly detect hidden land use changes (such as illegal underwater landfills and shoreline erosion under vegetation cover). Especially in areas with complex river network topography, traditional methods, lacking three-dimensional spatial structural details, cannot accurately reflect changes in canal embankment elevation, spatial conflicts between bridge clearance height and land development, leading to insufficient accuracy and reliability in monitoring canal land use changes. Summary of the Invention
[0005] The purpose of this application is to provide a land use monitoring model based on three-dimensional reconstruction, which solves the problem that traditional land use monitoring methods in the prior art cannot accurately reflect the three-dimensional characteristics of land use and the actual situation under complex terrain conditions, resulting in insufficient monitoring accuracy and reliability of land use change.
[0006] In a first aspect, embodiments of the present invention provide a land use monitoring model based on three-dimensional reconstruction, the model comprising a data acquisition module, a segmentation module, a reconstruction module, and a monitoring module: wherein...
[0007] The data acquisition module is used to take aerial photos of the target area using UAV measurement and photography technology to obtain image data of the target area from different angles and scales.
[0008] The segmentation module is used to perform semantic segmentation on two-dimensional images of different regions in the image data using a segmentation algorithm to obtain multiple sub-regions, and to perform annotation processing on each sub-region so that the sub-region carries corresponding label information and obtains a two-dimensional segmentation result;
[0009] The reconstruction module is used to extract feature points from the image data according to the three-dimensional reconstruction technology of the image, match the feature points with corresponding points at different angles and scales in the image data to generate dense point cloud data, and construct a three-dimensional model of the target area based on the dense point cloud data.
[0010] The monitoring module is used to map the two-dimensional segmentation result onto the three-dimensional model so that the spatial geometric information of the sub-region and the target region are fused to monitor the land use information of the target region.
[0011] In one embodiment, after the data acquisition module uses UAV surveying and photography technology to take aerial photographs of the target area and obtain image data of the target area at different angles and scales, it further includes:
[0012] The visible light bands in the image data are extracted in a format that can be recognized by the machine model, and the visible light bands are paired with the original image data to confirm that the extracted visible light bands have the ability to characterize the image data.
[0013] In one embodiment, the segmentation module performs semantic segmentation on two-dimensional images of different regions in the image data using a segmentation algorithm to obtain multiple sub-regions, and performs annotation processing on each sub-region to carry corresponding label information and obtain a two-dimensional segmentation result. Specifically, this includes:
[0014] Using the image data and the labels associated with the image data as input, the image data and the labels associated with the image data are concatenated through a preset fully convolutional neural network to obtain labeled image data, and the labeled image data is output to complete sub-region segmentation and labeling.
[0015] In one embodiment, the segmentation module performs semantic segmentation on different regions of the two-dimensional image data using a segmentation algorithm, specifically including:
[0016] The loss function generated by the segmentation algorithm when performing semantic segmentation on the two-dimensional image is calculated. The loss function consists of weighted cross-entropy loss, and the calculation formula is as follows:
[0017]
[0018] Where L represents the cross-entropy loss, M represents the number of classes, and y c It is a one-hot vector where each element takes only two values: 0 and 1. The element is set to 1 if the class matches the sample's class, and 0 otherwise. c w represents the probability that a sample belongs to c. c The weighting coefficient is represented by the following formula:
[0019]
[0020] Where N represents the total number of sub-regions, and Nc represents the number of pixels in sub-region category c.
[0021] In one embodiment, after the reconstruction module extracts feature points from the image data using 3D image reconstruction technology, matches the feature points with corresponding points at different angles and scales in the image data to generate dense point cloud data and construct a 3D model of the target region, the method further includes:
[0022] The dense point cloud data is converted into a corresponding Gaussian kernel, and the parameters of the Gaussian kernel are adjusted through an optimization algorithm;
[0023] Based on the surface features of the Gaussian kernel, the point cloud corresponding to the Gaussian kernel is diffused to reconstruct the continuous surface in the dense point cloud, thereby obtaining a continuous and smooth three-dimensional surface.
[0024] In one embodiment, when the reconstruction module extracts feature points from the image data using 3D image reconstruction technology, and matches the feature points with corresponding points at different angles and scales in the image data, the process specifically includes:
[0025] The feature point cloud of the image data is extracted using ORBSLAM3 technology;
[0026] Based on preset additional information, the feature point cloud is matched with corresponding points at different angles and scales in the image data.
[0027] In one embodiment, after the reconstruction module uses ORBSLAM3 technology to extract the feature point cloud of the image data, it further includes:
[0028] Obtain illumination modeling information other than ORB feature point clouds;
[0029] The feature point cloud extracted by the ORBSLAM3 technology is stored, and the illumination modeling information of each key frame starting from the initial frame of the image data is calculated.
[0030] The surface fitting image is obtained by processing the optical rotation modeling information using Gaussian sputtering technology;
[0031] The three-dimensional scene is reconstructed based on the surface-fitted image.
[0032] In one embodiment, the step of processing the lighting modeling information using Gaussian sputtering technology to obtain a surface fitting image includes: projecting the lighting modeling information onto the three-dimensional model and calculating the Gaussian weight of each projection point; distributing each projection point in three-dimensional space based on the Gaussian weight and superimposing them to obtain a first fitting image; performing surface extraction on the first fitting image to obtain the isosurface of the first fitting image; converting the isosurface into a 3D texture layer using rendering technology; and generating the surface fitting image with a continuous surface based on the 3D texture layer.
[0033] Secondly, the present invention provides a land use monitoring system based on three-dimensional reconstruction, comprising:
[0034] A processor is configured with the land use monitoring model described in any one of the first aspects; a memory for storing processor-executable instructions; wherein the processor is configured to execute the instructions to control the land use monitoring model to complete land use monitoring.
[0035] Thirdly, the present invention provides a computer-readable storage medium storing instructions for loading by a processor and controlling the land use monitoring model described in any one of the first aspects to perform land use monitoring.
[0036] The land use monitoring model based on 3D reconstruction provided in this application utilizes UAV aerial photography to acquire multi-angle, multi-scale image data, enabling high-precision imaging of surface targets with higher spatial resolution and the ability to identify more detailed land feature features. Image processing via semantic segmentation algorithms not only distinguishes different land feature categories but also achieves precise labeling of sub-regions, resulting in more detailed and comprehensive land use classification. Dense point clouds and 3D models are generated using 3D reconstruction technology, and 2D semantic information is mapped to 3D space, achieving deep fusion of 2D semantic and 3D geometric information. By combining the 3D model with the semantic segmentation model, not only can the categories of land features be identified, but their spatial structure, shape, and height differences can also be reflected, facilitating more accurate monitoring of land use change and identification of illegal land use. This application integrates UAV data acquisition, image processing, 3D reconstruction, and intelligent analysis algorithms, accurately reflecting the three-dimensional characteristics of land use and the actual situation under complex terrain conditions, thus improving the accuracy and reliability of land use change monitoring. Attached Figure Description
[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0038] Figure 1 The module intent of the land use monitoring model based on three-dimensional reconstruction provided in the embodiments of the present invention. Detailed Implementation
[0039] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.
[0040] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, components, features, and elements with the same names in different embodiments of this application may have the same meaning or different meanings, the specific meaning of which must be determined by its interpretation in that specific embodiment or further in conjunction with the context of that specific embodiment.
[0041] See Figure 1 This is a schematic diagram illustrating the module composition of a land use monitoring model based on 3D reconstruction, provided in an embodiment of this application. The land use monitoring model based on 3D reconstruction includes a data acquisition module, a segmentation module, a reconstruction module, and a monitoring module, wherein:
[0042] The data acquisition module is used to take aerial photos of the target area using UAV measurement and photography technology to obtain image data of the target area from different angles and scales.
[0043] The segmentation module is used to perform semantic segmentation on two-dimensional images of different regions in the image data using a segmentation algorithm to obtain multiple sub-regions, and to perform annotation processing on each sub-region so that the sub-region carries corresponding label information and obtains a two-dimensional segmentation result;
[0044] The reconstruction module is used to extract feature points from the image data according to the three-dimensional reconstruction technology of the image, match the feature points with corresponding points at different angles and scales in the image data to generate dense point cloud data, and construct a three-dimensional model of the target area based on the dense point cloud data.
[0045] The monitoring module is used to map the two-dimensional segmentation result onto the three-dimensional model so that the spatial geometric information of the sub-region and the target region are fused to monitor the land use information of the target region.
[0046] Here, drones are used to conduct multi-angle, multi-scale aerial photography of the target area, acquiring high-resolution image data, which is then transmitted to the segmentation module. The segmentation module performs semantic segmentation on two-dimensional features in the images (such as river channels, hard revetments, ecological revetments, docks and warehouses, riverside roads, bridges, waterfront buildings, and protective forests on embankments), and incorporates prior knowledge such as river curvature and embankment slope to optimize the identification of irregular river boundaries and hidden illegal constructions (such as reclaimed areas under vegetation cover). Sub-region labeling (such as plot names, numbers, and usage markings) is also performed. After segmenting and labeling the image data, the segmentation module transmits the data to the monitoring module. The monitoring module integrates the two-dimensional segmentation results with the spatial geometric information corresponding to the target area to achieve spatial analysis and statistical analysis of land use changes, ultimately realizing land use monitoring.
[0047] For example, suppose the target area is a typical section of a canal, which includes five core landforms: waterway, hard revetment, ecological revetment, two-story waterfront shops, and pedestrian bridge across the river. The segmentation module segments and labels the 2D image data, resulting in the following sub-regions: river channel (main channel, 500-tonnage navigation, current water depth 3.2m), hard revetment sub-region (revetment crest elevation 7.5m, blue line boundary coordinates of the prohibited construction zone: X=120345.6, Y=45678.9), ecological revetment sub-region (vegetation coverage 85%, revetment type: herbaceous + shrub mixed, 2022 restoration section), waterfront commercial sub-region (2 stories, first floor cantilever length 1.8m, planned height limit 8m, located within 30m of the revetment management area), and cross-river pedestrian bridge sub-region (clearance height 5m, bridge bottom elevation 8.5m, prohibited navigation zone: 50m upstream and downstream of the bridge). The labeled sub-regions are then mapped onto the 3D model of the target area to check if the 2D segmentation results match the 3D spatial information. If a match is successful, the utilization status of each sub-region is determined based on its 3D state.
[0048] Overall, the land use monitoring model based on 3D reconstruction provided in this application, by combining data acquisition, semantic segmentation, 3D modeling, and fusion analysis, enables each module to perform its function, ultimately constructing a land use monitoring system with high intelligence, high resolution, and spatial semantic attributes, thus realizing the upgrade from traditional remote sensing analysis to an intelligent geographic information system.
[0049] In one embodiment, after the data acquisition module uses UAV surveying and photography technology to take aerial photographs of the target area and obtain image data of the target area at different angles and scales, it further includes:
[0050] The visible light bands in the image data are extracted in a format that can be recognized by the machine model, and the visible light bands are paired with the original image data to confirm that the extracted visible light bands have the ability to characterize the image data.
[0051] Understandably, machine learning models typically have strict requirements regarding the format, dimensionality, and normalization of input data when processing images. However, raw aerial images may be in formats such as JPEG or RAW, which suffer from compression, embedded information, or uneven lighting, and may include useless bands, color interference, and sensor errors. Therefore, converting images to a "machine-model-recognizable format," such as unifying them to three-channel (RGB) tensors, normalizing pixel values, and removing compression artifacts, ensures the model's recognition accuracy and efficiency. Extracting visible light bands from image data in a machine-model-recognizable format ensures efficient and accurate model operation, enhancing the robustness and practicality of the entire land monitoring system. Furthermore, to confirm that the extracted visible light bands indeed originate from the target area's image data and can characterize the target area, the visible light bands are matched with the original image data. If a match is successful, the visible light bands are confirmed as usable; otherwise, they should be extracted again.
[0052] In one embodiment, when the segmentation module performs semantic segmentation on two-dimensional images of different regions in the image data using a segmentation algorithm to obtain multiple sub-regions, and performs annotation processing on each sub-region to carry corresponding label information and obtain a two-dimensional segmentation result, the specific steps include: taking the image data and the labels associated with the image data as input, processing the image data and the labels associated with the image data in series through a preset fully convolutional neural network to obtain labeled image data, and outputting the labeled image data to complete the sub-region segmentation and annotation.
[0053] The fully convolutional neural network in this embodiment can be a classic or improved FCN network (such as FCN-8s, U-Net, DeepLab, etc.). FCN networks are suitable for pixel-level semantic segmentation tasks, and since they have no fully connected layers, they support image inputs of arbitrary sizes and can classify every pixel in an image, thus finely segmenting different sub-regions. By adding a layer (which could be a concatenation, transformation, or processing module) to align or concatenate image data with labels in some dimension, feature-label pairing processing is achieved. This allows training data to enter subsequent network layers in a network-recognizable form, thereby improving the stability and accuracy of segmentation and labeling.
[0054] Preferably, the layer structure of the fully convolutional neural network in this embodiment may include a data layer, a concatenation layer, and an output layer, wherein: the data layer is used to read image data and identify the label corresponding to the image data; the concatenation layer is used to concatenate the image data with the corresponding label to obtain labeled image data; and the output layer is used to output the labeled image data obtained by the concatenation process.
[0055] In one embodiment, the segmentation module performs semantic segmentation on different regions of the two-dimensional image data using a segmentation algorithm, specifically including:
[0056] The loss function generated by the segmentation algorithm when performing semantic segmentation on the two-dimensional image is calculated. The loss function consists of weighted cross-entropy loss, and the calculation formula is as follows:
[0057]
[0058] Where L represents the cross-entropy loss, M represents the number of classes, and y c It is a one-hot vector where each element takes only two values: 0 and 1. The element is set to 1 if the class matches the sample's class, and 0 otherwise. c w represents the probability that a sample belongs to c. c The weighting coefficient is represented by the following formula:
[0059]
[0060] Where N represents the total number of sub-regions, and Nc represents the number of pixels in sub-region category c.
[0061] In one embodiment, after the reconstruction module extracts feature points from the image data using image 3D reconstruction technology, matches the feature points with corresponding points at different angles and scales in the image data to generate dense point cloud data and construct a 3D model of the target region, the method further includes: converting the dense point cloud data into a corresponding Gaussian kernel, and adjusting the parameters of the Gaussian kernel using an optimization algorithm; and diffusing the point cloud corresponding to the Gaussian kernel according to the surface features of the Gaussian kernel to reconstruct the continuous surface in the dense point cloud, thereby obtaining a continuous and smooth 3D surface.
[0062] The point cloud surface reconstruction method using Gaussian kernel density modeling, optimization, and diffusion can transform dense point clouds into continuous and smooth 3D surfaces, exhibiting good smoothness, noise resistance, and controllability.
[0063] In one embodiment, when the reconstruction module extracts feature points from the image data using three-dimensional reconstruction technology and matches the feature points with corresponding points at different angles and scales in the image data, the specific steps include: using ORBSLAM3 technology to extract feature point clouds from the image data; and matching the feature point clouds with corresponding points at different angles and scales in the image data based on preset additional information.
[0064] Understandably, ORBSLAM3 is a feature-point-based visual SLAM (Simultaneous Localization and Mapping) system, the third generation in the ORBSLAM series. ORB feature point clouds are a fast, efficient, rotation-invariant feature descriptor. ORBSLAM3 extracts these feature point clouds from images and performs matching and tracking across multiple frames.
[0065] In one embodiment, after the reconstruction module uses ORBSLAM3 technology to extract the feature point cloud of the image data, it further includes: acquiring illumination modeling information other than the ORB feature point cloud; storing the feature point cloud extracted by the ORBSLAM3 technology, and calculating the illumination modeling information for each key frame starting from the initial frame of the image data; processing the illumination modeling information using Gaussian sputtering technology to obtain a surface fitting image; and completing the three-dimensional scene reconstruction based on the surface fitting image.
[0066] Here, the efficient feature extraction capabilities of ORBSLAM3 technology, along with illumination modeling information in addition to ORB feature point clouds, can significantly improve feature matching performance under different viewpoints, scales, and illumination conditions, thereby achieving more stable, robust, and accurate 3D mapping and pose estimation.
[0067] In this embodiment, the lighting modeling information may include pixel brightness, normal vectors, and reflection intensity. Pixel brightness includes image contrast, depth, and detail. By adjusting pixel brightness, various image processing operations such as brightening, darkening, and contrast enhancement can be performed to improve the visual quality of the image and highlight important information. Normal vectors are mainly used to determine the interaction between light and the object's surface. When calculating lighting, normal vectors are used to determine the angle of incidence of light relative to the object's surface, thereby calculating the light intensity received by each point on the surface based on the lighting model. Reflection intensity is used in lighting modeling to simulate the reflection characteristics of different material surfaces, thereby achieving a realistic rendering effect.
[0068] In one embodiment, the step of processing the lighting modeling information using Gaussian sputtering technology to obtain a surface fitting image includes: projecting the lighting modeling information onto the three-dimensional model and calculating the Gaussian weight of each projection point; distributing each projection point in three-dimensional space based on the Gaussian weight and superimposing them to obtain a first fitting image; performing surface extraction on the first fitting image to obtain the isosurface of the first fitting image; converting the isosurface into a 3D texture layer using rendering technology; and generating the surface fitting image with a continuous surface based on the 3D texture layer.
[0069] Specifically, the lighting modeling information of keyframes in the image data is projected onto the point cloud or voxel mesh in the 3D model. The spatial distance and directional consistency of each projected point are determined based on its Gaussian weight, and the contribution of each projected point to its neighborhood is also determined. Here, a suitable Gaussian function needs to be selected; for example, in this embodiment, an isotropic or anisotropic Gaussian function can be chosen. The Gaussian function smooths the lighting information in the neighborhood to handle noise and discontinuities. Next, each projected point is "splashed" into a Gaussian distribution. These distributions can be superimposed on each other in 3D space based on color or intensity values to form a continuous surface, i.e., the first fitted image. Then, a surface fitted image is obtained based on the first fitted image.
[0070] The above describes the reconstruction of the land's three-dimensional structure. Surface fitting using Gaussian sputtering allows for the rendering of smooth, continuous, and realistic scene images. This yields high-quality surface-fitted images, providing effective support for subsequent land use monitoring.
[0071] Optionally, after the reconstruction module completes the 3D scene reconstruction, it further includes: visualizing the reconstructed 3D scene using a visualization tool.
[0072] Here, visualization tools can include MeshLaba, CloudCompare, ParaView, Blender, etc. By visualizing the 3D scene, one can determine whether the 3D model is complete and whether the texture is accurate. The model can also be observed from different angles and scaled and rotated to analyze the building structure and environmental layout of the target area.
[0073] Specifically, each point in the point cloud is considered as the center of a Gaussian distribution to describe the uncertainty or density distribution of the local space. The point cloud is transformed into a set of multiple Gaussian kernel functions, each with its own parameters, such as mean, the position of the corresponding point, covariance, and the diffusion range of the corresponding point. Then, algorithms such as gradient descent, EM algorithm, and variational inference can be used to adjust the parameters of the Gaussian kernels to better fit the point cloud data. In this way, it is possible to capture the implicit structure in the point cloud, such as the smoothness or continuity of the surface.
[0074] The land use monitoring model based on 3D reconstruction provided in this application utilizes UAV aerial photography to acquire multi-angle, multi-scale image data, enabling high-precision imaging of surface targets with higher spatial resolution and the ability to identify more detailed land feature features. Image processing via semantic segmentation algorithms not only distinguishes different land feature categories but also achieves precise labeling of sub-regions, resulting in more detailed and comprehensive land use classification. Dense point clouds and 3D models are generated using 3D reconstruction technology, and 2D semantic information is mapped to 3D space, achieving deep fusion of 2D semantic and 3D geometric information. By combining the 3D model with semantic segmentation, not only can the categories of land features be identified, but their spatial structure, shape, and elevation differences can also be reflected, facilitating more accurate monitoring of land use change and identification of illegal land use. Integrating UAV data acquisition, image processing, 3D reconstruction, and intelligent analysis algorithms, this approach solves the problem that traditional land use monitoring methods in existing technologies cannot accurately reflect the three-dimensional characteristics of land use and the actual situation under complex terrain conditions, leading to insufficient accuracy and reliability in monitoring land use change.
[0075] Based on the same concept as the foregoing embodiments, this embodiment of the invention provides a land use monitoring system based on three-dimensional reconstruction, comprising: a processor configured with a land use monitoring model as described in any of the first aspects; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute the instructions to control the land use monitoring model to complete land use monitoring.
[0076] Based on the same inventive concept as the foregoing embodiments, the present invention also provides a computer-readable storage medium storing instructions for loading and controlling the above-mentioned land use monitoring model to complete land use monitoring by a processor.
[0077] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0078] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.
Claims
1. A land use monitoring model based on three-dimensional reconstruction, characterized in that, Includes a data acquisition module, a segmentation module, a reconstruction module, and a monitoring module: The data acquisition module is used to take aerial photos of the target area using UAV measurement and photography technology to obtain image data of the target area from different angles and scales. The segmentation module is used to perform semantic segmentation on two-dimensional images of different regions in the image data using a segmentation algorithm to obtain multiple sub-regions, and to perform annotation processing on each sub-region so that the sub-region carries corresponding label information and obtains a two-dimensional segmentation result; The reconstruction module is used to extract feature points from the image data according to the three-dimensional reconstruction technology of the image, match the feature points with corresponding points at different angles and scales in the image data to generate dense point cloud data, and construct a three-dimensional model of the target area based on the dense point cloud data. The monitoring module is used to map the two-dimensional segmentation result onto the three-dimensional model so that the spatial geometric information of the sub-region and the target region are fused to monitor the land use information of the target region.
2. The model according to claim 1, characterized in that, After acquiring image data of the target area from different angles and scales using UAV surveying and photography technology, the data acquisition module further includes: The visible light bands in the image data are extracted in a format that can be recognized by the machine model, and the visible light bands are paired with the original image data to confirm that the extracted visible light bands have the ability to characterize the image data.
3. The model according to claim 1, characterized in that, The segmentation module performs semantic segmentation on two-dimensional images of different regions in the image data using a segmentation algorithm to obtain multiple sub-regions, and performs annotation processing on each sub-region to carry corresponding label information and obtain a two-dimensional segmentation result. Specifically, this includes: Using the image data and the labels associated with the image data as input, the image data and the labels associated with the image data are concatenated through a preset fully convolutional neural network to obtain labeled image data, and the labeled image data is output to complete sub-region segmentation and labeling.
4. The model according to claim 1, characterized in that, The segmentation module performs semantic segmentation on different regions of the two-dimensional image data using a segmentation algorithm, specifically including: The loss function generated by the segmentation algorithm when performing semantic segmentation on the two-dimensional image is calculated. The loss function consists of weighted cross-entropy loss, and the calculation formula is as follows: Where L represents the cross-entropy loss, M represents the number of classes, and y c It is a one-hot vector where each element takes only two values: 0 and 1. The element is set to 1 if the class matches the sample's class, and 0 otherwise. c w represents the probability that a sample belongs to c. c The weighting coefficient is represented by the following formula: Where N represents the total number of sub-regions, and Nc represents the number of pixels in sub-region category c.
5. The model according to claim 1, characterized in that, After extracting feature points from the image data using the image 3D reconstruction technology via the reconstruction module, matching the feature points with corresponding points at different angles and scales in the image data to generate dense point cloud data, and constructing a 3D model of the target region based on the dense point cloud data, the method further includes: The dense point cloud data is converted into a corresponding Gaussian kernel, and the parameters of the Gaussian kernel are adjusted through an optimization algorithm; Based on the surface features of the Gaussian kernel, the point cloud corresponding to the Gaussian kernel is diffused to reconstruct the continuous surface in the dense point cloud, thereby obtaining a continuous and smooth three-dimensional surface.
6. The model according to claim 1, characterized in that, When the reconstruction module extracts feature points from the image data using 3D image reconstruction technology, and matches the feature points with corresponding points at different angles and scales in the image data, the specific steps include: The feature point cloud of the image data is extracted using ORBSLAM3 technology; Based on preset additional information, the feature point cloud is matched with corresponding points at different angles and scales in the image data.
7. The model according to claim 6, characterized in that, After the reconstruction module uses ORBSLAM3 technology to extract the feature point cloud of the image data, it also includes: Obtain illumination modeling information other than ORB feature point clouds; The feature point cloud extracted by the ORBSLAM3 technology is stored, and the illumination modeling information of each key frame starting from the initial frame of the image data is calculated. The surface fitting image is obtained by processing the illumination modeling information using Gaussian sputtering technology; The three-dimensional scene is reconstructed based on the surface-fitted image.
8. The model according to claim 7, characterized in that, The process of processing the illumination modeling information using Gaussian sputtering technology to obtain a surface fitting image includes: The lighting modeling information is projected onto the 3D model, and the Gaussian weight of each projection point is calculated. The projection points are distributed in three-dimensional space based on the Gaussian weights and superimposed to each other to obtain a first fitted image; Surface extraction is performed on the first fitted image to obtain the isosurface of the first fitted image; The isosurface is converted into a 3D texture layer using rendering techniques; The surface fitting image with a continuous surface is generated based on the 3D texture layer.
9. A land use monitoring system based on three-dimensional reconstruction, characterized in that, include: The processor is configured with the land use monitoring model as described in any one of claims 1-8; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to control the land use monitoring model to complete land use monitoring.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions for loading by a processor and controlling the land use monitoring model of any one of claims 1-8 to perform land use monitoring.