Automatic three-dimensional building reconstruction method
Through drone shooting and SfM algorithm combined with 3D Gaussian Splatting technology, the efficient and high-precision automated reconstruction of building three-dimensional models is achieved, solving the problems of blurred model details and low reconstruction accuracy in large-scale scenarios. It is suitable for high insurance for large-scale architectural scenarios Real rendering.
Patent Information
- Application Number
- CN202411134555.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-19
- Publication Date
- 2024-12-03
AI Technical Summary
The existing three-dimensional model reconstruction methods of architectural 3D model have problems such as blurred model details, low reconstruction accuracy and low modeling efficiency in large-scale scenarios, especially in structural details areas, and rely on a large number of manual interventions and computing resources.
Using drones to take multi-view pictures, combined with Structure from Motion (SfM) algorithm and 3D Gaussian Splatting technology, the automatic reconstruction of high-precision building three-dimensional models is achieved through pseudo-view camera generation and Gaussian ellipsoid optimization, reducing manual intervention and improving calculations efficiency.
It realizes efficient and high-precision three-dimensional reconstruction of building, improves the reconstruction accuracy of structural details areas, reduces artifacts and surface floating objects, is suitable for high-fidelity rendering of large-scale architectural scenes, and reduces dependence on manual intervention and computing resources.
Smart Images

Figure CN119068114A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of civil engineering infrastructure health monitoring, and is a building automation three-dimensional reconstruction method. Background Art
[0002] In the process of urbanization, a large number of houses and infrastructure need to be built to meet the growing demand for urban functions. However, the service life of civil engineering buildings is generally decades or even hundreds of years. How to accurately evaluate the safety of buildings, guide building management and operation, and build a digital urban management system is a major challenge facing future urban development. Traditional building operation and maintenance methods make it difficult to detect and repair minor damage on the surface of buildings in a timely manner, which not only affects the service life of the building, but also increases the cost of maintenance. With the rapid development of big data, artificial intelligence and computer vision technology, digital twin technology has gradually emerged. It provides an effective solution to the above problems by building virtual copies of physical entities to achieve mapping and interaction between the physical and digital worlds.
[0003] The 3D model of a building is the foundation of a digital twin city, serving tasks such as visualization of building entities in the virtual world, performance simulation and evaluation during service. Based on high-precision 3D model reconstruction technology, it can not only restore the geometric shape and surface details of the building, but also form a digital twin system of the building by combining with IoT devices and various monitoring data. With this system, operation and maintenance personnel can monitor the condition of the building in real time, predict potential problems, and optimize maintenance strategies, thereby extending the service life of the building and reducing operating costs. For example, through high-precision 3D models, operation and maintenance personnel can detect fine cracks that are difficult to detect with the naked eye, and repair them in time to prevent small problems from turning into major accidents. In addition, 3D models can also be used to simulate and evaluate the performance of buildings under various environmental conditions, providing a scientific basis for disaster prevention and emergency response. Therefore, high-precision 3D reconstruction of buildings not only helps to improve the safety and reliability of buildings, but also realizes the intelligence and refinement of urban management and promotes the sustainable development of cities.
[0004] Typical traditional 3D reconstruction methods include Structure From Motion (SfM) and Multi-ViewStereo (MVS). They perform well when processing static scenes with rich textures and can generate high-precision 3D point cloud models. SfM does not require depth information in the initial stage and only relies on multi-view images to construct relative camera positions and sparse 3D point clouds through feature point detection and matching. MVS further generates dense point clouds through multi-view stereo matching based on sparse point clouds to achieve 3D reconstruction. These methods can provide high geometric accuracy and texture detail preservation when processing small-scale scenes. However, for large-scale scenes, traditional MVS methods need to process a large number of image pairs for stereo matching, and it is often difficult to achieve accurate depth estimation, resulting in incomplete or erroneous reconstruction. On the other hand, the MVS method requires multi-view and wide-coverage image data. The illumination changes and mismatches in the image data will seriously affect the reconstruction accuracy, thus hindering its widespread application in high-fidelity 3D scenes.
[0005] In the field of deep learning-based 3D reconstruction, Neural Radiance Field (NeRF) has developed rapidly and received widespread attention. It has performed well in promoting realistic reconstruction of scenes and synthesizing new views based on multi-view images. This method uses the powerful function approximation ability of neural networks to model the light transmission in the scene. It can reconstruct highly realistic scenes when given information about the camera perspective and achieve rendering of objects from any perspective. Compared with traditional SfM, NeRF can learn the 3D structure and radiation field of the scene from sparse image data without requiring a large coverage area and high-quality images. In addition, NeRF can reduce the error accumulation problem in traditional methods and improve reconstruction accuracy through end-to-end learning. Although NeRF has shown excellent results, it still has limitations, such as being difficult to expand to large scenes. In particular, for 3D reconstruction of buildings, which are large in scale and rich in geometric details, the training process of NeRF requires a lot of computing resources, which is difficult to meet the real-time needs of structural health monitoring and digital twins.
[0006] Currently, with the emergence of 3D Gaussian Splatting, significant progress has been made in the reconstruction of large scenes. This method uses a clear 3D Gaussian ellipsoid to represent the scene, and can render high-resolution images in real time by rasterizing the splashed 3D Gaussian image, with superior performance in visual quality and rendering speed. However, the data collected by drones often cannot fully cover the details of the building, so that the generated models usually have artifacts and surface floating objects, and lack fine texture details. In addition, conventional lidar equipment has a limited range in providing point cloud data. The above factors make it difficult to effectively optimize this method when it is applied to large-scene buildings, resulting in poor rendering effects. Therefore, an efficient and high-precision method for automated 3D reconstruction of buildings is proposed, which has important theoretical significance and practical value. Summary of the invention
[0007] In view of the above problems, the present invention proposes a method for automated 3D reconstruction of buildings. This method fully utilizes the advantages of 3D Gaussian Splatting in high-dimensional data modeling to achieve accurate reconstruction from multi-view images to building models, and realizes high-fidelity real-time rendering of large building scenes while ensuring computational efficiency.
[0008] The purpose of the present invention is to solve the problems of fuzzy model details, low reconstruction accuracy, and low modeling efficiency caused by reliance on a large amount of manual intervention in existing building 3D model reconstruction methods due to point cloud quality and environmental interference. In order to achieve accurate and efficient building 3D reconstruction, especially the clarity of the model in areas such as structural details. The present invention provides a building automation 3D reconstruction method, and the present invention provides the following technical solutions:
[0009] A method for building automation 3D reconstruction comprises the following steps:
[0010] Step 1: Control the drone to fly around the building and take photos of the building facade from different perspectives to obtain a multi-perspective picture set;
[0011] Step 2: Based on the SfM algorithm, calculate the camera parameters of each image and obtain the initial point cloud;
[0012] Step 3: Based on the obtained camera parameters, calculate the spatial boundary of the building point cloud and generate a pseudo-perspective camera;
[0013] Step 4: According to the weights of the distance and direction of the camera, the Gaussian ellipsoid properties of the unknown view are defined and optimized by the Gaussian ellipsoids of the adjacent known viewpoints;
[0014] Step 5: Use the building space boundary calculated by camera parameters to extract the building point cloud, and use the density-based DBSCAN algorithm to filter the building point cloud to optimize the Gaussian model to obtain a high-precision building 3D model.
[0015] Preferably, the step 1 is specifically:
[0016] The target building was photographed from multiple perspectives using a drone. The drone was kept 3 meters away from the building surface and the photo overlap rate was controlled at 50%. Based on the drone's built-in fixed-point cruise algorithm and supplemented by manual camera posture adjustment, a multi-perspective picture set S{X i |i=1,...,N}, a total of N pictures.
[0017] Preferably, the step 2 is specifically:
[0018] According to the equivalent focal length f of the camera as the known initial parameters of the camera, the multi-view image set S{X i |i=1,...,N}N images to calculate camera parameters;
[0019] The SfM algorithm is as follows:
[0020] Preprocess the image and downsample it to 1 / 2 size;
[0021] Select the scale-invariant feature transformation algorithm to extract feature points. The extraction formula is:
[0022] L(x,y,σ)=G(x,y,σ)*I(x,y)
[0023]
[0024] D(x,y,σ)=L(x,y,σ)-L(x,y,kσ)
[0025] Where (x, y) is the pixel coordinate of the image, L(x, y, σ) is the Gaussian blurred image with σ as the parameter, and D(x, y, σ) is the Gaussian difference image. The local extremum detection algorithm compares each pixel in D(x, y, σ) with its 26 spatially adjacent pixels. If the pixel is larger or smaller than all adjacent pixels, it is marked as a SIFT feature point, thereby obtaining a feature point set.
[0026] The gradient size and direction of the 16×16 area near the feature point are formed into a 128-dimensional vector according to the histogram statistics method, and the vector is normalized to obtain the feature descriptor used to match the feature point The feature descriptor of the feature point is based on the following formula:
[0027]
[0028] θ(x,y)=arctan((L(x,y+1)-L(x,y-1)) / (L(x+1,y)-L(x-1,y)))
[0029] Where L is the Gaussian blurred image mentioned in (4b);
[0030] The feature points are matched by calculating the distance between the descriptors of all image pairs, thus obtaining a set of matched feature point pairs. in
[0031] The camera parameters are calculated based on the binocular stereo geometry algorithm. The camera parameters are represented by the matrix S λ K[R|t] represents, where is the camera internal reference; is the camera external parameter, where is the rotation matrix, is the translation vector; S λ =diag(1 / λ,1 / λ,1) is the radial distortion matrix of the lens, where is a nonlinear polynomial function of the distance d from a pixel to the center of the image, is the radial distortion parameter, where S λ , K is known, an estimate is obtained from the Exchange Image File Format EXIF, and a global optimization is performed in the final stage.
[0032] Preferably, the step 3 is specifically:
[0033] According to the obtained camera parameters S λ K[R|t] defines a sample space for generating pseudo cameras Among them, t min is the minimum camera extrinsic parameter, t max is the largest camera extrinsic parameter;
[0034] In order to obtain the sample space of the pseudo camera, all cameras are set to focus on the center of the scene, and spherical linear interpolation is used to generate the pseudo camera pose S P =(t+ε,q), where t∈S λ K[R|t], ε is a randomly added perturbation, q is a quaternion representing the average rotation of the two cameras;
[0035] According to the camera parameters S λ K[R|t] defines the spatial boundary of the building point cloud B = {(t min ,t max )}.
[0036] Preferably, the step 4 is specifically:
[0037] The weight of the pseudo-view Gaussian ellipsoid is defined according to the similarity of the distance and angle between the generated pseudo camera and the adjacent cameras:
[0038]
[0039] Among them, θ i is the angle between the pseudo camera and the i-th known camera direction, d i is the Euclidean distance between the pseudo camera and the i-th known camera, ε is a constant, and the weight is normalized to
[0040] Using the calculated weight ω′ i , perform weighted averaging on the Gaussian ellipsoid properties of adjacent perspectives to obtain the Gaussian ellipsoid properties of the pseudo perspective:
[0041]
[0042] Among them, x i is the position of the adjacent Gaussian ellipsoid, ∑ i is the covariance matrix representing the shape of the three-dimensional Gaussian ellipsoid, Represents the color of the three-dimensional Gaussian ellipsoid relative to the line of sight.
[0043] Preferably, the optimization of the pseudo-view Gaussian ellipsoid is accomplished by adding a regularization term to the loss function, and the regularization term consists of two parts: geometric regularization and color regularization:
[0044] In order to make the distribution of the pseudo-perspective Gaussian ellipsoid in three-dimensional space more reasonable, multi-perspective geometric regularization is used to constrain the position and radius of the pseudo-perspective Gaussian ellipsoid:
[0045]
[0046] Among them, ω′ i is the weight parameter, μ pseudo and r pseudo Respectively represent the position and radius of the pseudo-view Gaussian ellipsoid, μ i and r i Respectively represent the position and radius of the adjacent Gaussian ellipsoid;
[0047] Color regularization is used to constrain the color of the pseudo-view Gaussian ellipsoid to maximize its consistency with the adjacent view Gaussian ellipsoid:
[0048]
[0049] Among them, ω′ i is the weight parameter, c pseudo Indicates the color of the pseudo-view Gaussian ellipsoid, c i Indicates the color of the adjacent Gaussian ellipsoid;
[0050] The total loss function used combines image reconstruction, geometry and color losses, and divides the loss function into three terms, which are balanced by weights λ1, λ2, and λ3; the first term L image is an L1 loss, L D-SSIM is the error that measures the structural similarity between the rendered image and the real image. The second term L geometry is the geometric regularization term for the pseudo-perspective Gaussian ellipsoid, and the third term L color is the color regularization term for the pseudo-view Gaussian ellipsoid:
[0051] L=(1-λ1)L image +λ1L D-SSIM +λ2L geometry +λ3L color
[0052] 7. The method according to claim 6, characterized in that: the step 5 specifically comprises:
[0053] According to the spatial boundary B of the building point cloud, the initial point cloud P calculated from SfM is i} to extract points within the building space boundary:
[0054] P inside ={p i ∈P|b min ≤p i ≤ b max}
[0055] The extracted building point cloud is clustered using a density-based clustering algorithm to effectively remove floating objects on the building surface and optimize the final three-dimensional model.
[0056] A building automation three-dimensional reconstruction method system, the system comprising:
[0057] A data acquisition module, wherein the data acquisition module controls the drone to fly around the building and photograph the building facade from different perspectives to obtain a multi-perspective picture set;
[0058] A data calculation module, which calculates the camera parameters of each picture and obtains an initial point cloud based on the SfM algorithm;
[0059] A pseudo-view camera generation module, wherein the pseudo-view camera generation module calculates the spatial boundary of the building point cloud and generates a pseudo-view camera according to the obtained camera parameters;
[0060] An optimization module, wherein the optimization module defines and optimizes the properties of the Gaussian ellipsoid of the unknown perspective from the Gaussian ellipsoids of the adjacent known perspectives according to the weights of the distance and direction of the camera;
[0061] The reconstruction module uses the building space boundary obtained by camera parameter calculation to extract the building point cloud, and uses the density-based DBSCAN algorithm to filter the building point cloud to optimize the Gaussian model to obtain a high-precision building three-dimensional model.
[0062] A computer-readable storage medium stores a computer program, which is executed by a processor to implement a building automation three-dimensional reconstruction method.
[0063] A computer device comprises a memory and a processor, wherein the memory stores a computer program and the processor implements a building automation three-dimensional reconstruction method when executing the computer program.
[0064] The present invention has the following beneficial effects:
[0065] Compared with the prior art, the present invention has the following advantages:
[0066] The present invention proposes a method for reconstructing a three-dimensional building model, and realizes the whole process of drone cruise shooting, image preprocessing, camera pose acquisition, pseudo-perspective camera generation, Gaussian ellipsoid optimization and three-dimensional reconstruction. The whole process is convenient and fast, does not rely on manual parameter adjustment, has low requirements on the overlap rate of the images taken by the drone, is not affected by lighting, and improves the reconstruction efficiency of the building three-dimensional model and the accuracy of the geometric detail area.
[0067] The present invention has a high tolerance for errors in camera pose estimation, and effectively alleviates the problem of the inability to reconstruct the three-dimensional model due to large errors in camera parameter estimation caused by image quality and environmental factors. The present invention has good reconstruction effects on structural detail areas, and the proposed method has better reconstruction effects on structural details that are not fully covered by drone photography, achieving high-definition detail reconstruction effects. The Gaussian model output by the present invention has fewer floating objects on the surface, and while ensuring computational efficiency, high-fidelity rendering of large architectural scenes is achieved. The present invention relies less on manual intervention, and the model reconstruction process does not require manual parameter adjustment. BRIEF DESCRIPTION OF THE DRAWINGS
[0068] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0069] Figure 1 Rebuild the flow chart for the model;
[0070] Figure 2 This is a partial set of multi-view pictures taken;
[0071] Figure 3 Artifacts and floating objects generated by 3D Gaussian Splatting reconstruction results;
[0072] Figure 4 This is a schematic diagram of the Gaussian ellipsoid optimization process;
[0073] Figure 5 Optimize the schematic diagram for point cloud filtering;
[0074] Figure 6 This is the Gaussian ellipsoid output by the model;
[0075] Figure 7 A comparison diagram of the building 3D reconstruction result of the present invention and the original 3D Gaussian Splatting reconstruction result; Figure 7 The left is the reconstruction result of 3D GS; Figure 7 The middle is the reconstruction result of the present invention; Figure 7 The right is a real picture taken by a drone. DETAILED DESCRIPTION
[0076] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0077] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.
[0078] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0079] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0080] The present invention is described in detail below in conjunction with specific embodiments. Specific embodiment one:
[0082] according to Figure 1-Figure 7 As shown, the specific optimization technical solution adopted by the present invention to solve the above technical problems is: the present invention relates to a method for three-dimensional reconstruction of building automation.
[0083] A method for building automation 3D reconstruction comprises the following steps:
[0084] Step 1: Control the drone to fly around the building and take photos of the building facade from different perspectives to obtain a multi-perspective picture set;
[0085] Step 2: Based on the SfM algorithm, calculate the camera parameters of each image and obtain the initial point cloud;
[0086] Step 3: Based on the obtained camera parameters, calculate the spatial boundary of the building point cloud and generate a pseudo-perspective camera;
[0087] Step 4: According to the weights of the distance and direction of the camera, the Gaussian ellipsoid properties of the unknown view are defined and optimized by the Gaussian ellipsoids of the adjacent known viewpoints;
[0088] Step 5: Use the building space boundary calculated by camera parameters to extract the building point cloud, and use the density-based DBSCAN algorithm to filter the building point cloud to optimize the Gaussian model to obtain a high-precision building 3D model. Specific embodiment 2:
[0090] The difference between the second embodiment of the present invention and the first embodiment is that:
[0091] The step 1 is specifically as follows:
[0092] The target building was photographed from multiple perspectives using a drone. The drone was kept 3 meters away from the building surface and the photo overlap rate was controlled at 50%. Based on the drone's built-in fixed-point cruise algorithm and supplemented by manual camera posture adjustment, a multi-perspective picture set S{X i |i=1,...,N}, a total of N pictures. Specific embodiment three:
[0094] The difference between the third embodiment of the present invention and the second embodiment is that:
[0095] The step 2 is specifically as follows:
[0096] According to the equivalent focal length f of the camera as the known initial parameters of the camera, the multi-view image set S{X i |i=1,...,N}N images to calculate camera parameters;
[0097] The SfM algorithm is as follows:
[0098] Preprocess the image and downsample it to 1 / 2 size;
[0099] Select the scale-invariant feature transformation algorithm to extract feature points. The extraction formula is:
[0100] L(x,y,σ)=G(x,y,σ)*I(x,y)
[0101]
[0102] D(x,y,σ)=L(x,y,σ)-L(x,y,kσ)
[0103] Where (x, y) is the pixel coordinate of the image, L(x, y, σ) is the Gaussian blurred image with σ as the parameter, and D(x, y, σ) is the Gaussian difference image. The local extremum detection algorithm compares each pixel in D(x, y, σ) with its 26 spatially adjacent pixels. If the pixel is larger or smaller than all adjacent pixels, it is marked as a SIFT feature point, thereby obtaining a feature point set.
[0104] The gradient size and direction of the 16×16 area near the feature point are formed into a 128-dimensional vector according to the histogram statistics method, and the vector is normalized to obtain the feature descriptor used to match the feature point The feature descriptor of the feature point is based on the following formula:
[0105]
[0106] θ(x,y)=arctan((L(x,y+1)-L(x,y-1)) / (L(x+1,y)-L(x-1,y)))
[0107] Where L is the Gaussian blurred image mentioned in (4b);
[0108] The feature points are matched by calculating the distance between the descriptors of all image pairs, thus obtaining a set of matched feature point pairs. in
[0109] The camera parameters are calculated based on the binocular stereo geometry algorithm. The camera parameters are represented by the matrix S λ K[R|t] represents, where is the camera internal reference; is the camera external parameter, where is the rotation matrix, is the translation vector; S λ =diag(1 / λ,1 / λ,1) is the radial distortion matrix of the lens, where is a nonlinear polynomial function of the distance d from a pixel to the center of the image, is the radial distortion parameter, where S λ , K is known, an estimate is obtained from the Exchange Image File Format EXIF, and a global optimization is performed in the final stage. Specific embodiment four:
[0111] The difference between the fourth embodiment of the present invention and the third embodiment is that:
[0112] The step 3 is specifically as follows:
[0113] According to the obtained camera parameters S λ K[R|t] defines a sample space for generating pseudo cameras Among them, t min is the minimum camera extrinsic parameter, t max is the largest camera extrinsic parameter;
[0114] In order to obtain the sample space of the pseudo camera, all cameras are set to focus on the center of the scene, and spherical linear interpolation is used to generate the pseudo camera pose S P =(t+ε,q), where t∈S λ K[R|t], ε is a randomly added perturbation, q is a quaternion representing the average rotation of the two cameras;
[0115] According to the camera parameters S λ K[R|t] defines the spatial boundary of the building point cloud B = {(t min ,t max )}. Specific embodiment 5:
[0116] The difference between the fifth embodiment of the present invention and the fourth embodiment is that:
[0117] The step 4 is specifically as follows:
[0118] The weight of the pseudo-view Gaussian ellipsoid is defined according to the similarity of the distance and angle between the generated pseudo camera and the adjacent cameras:
[0119]
[0120] Among them, θ i is the angle between the pseudo camera and the i-th known camera direction, d i is the Euclidean distance between the pseudo camera and the i-th known camera, ε is a constant, and the weight is normalized to
[0121] Using the calculated weight ω′ i , perform weighted averaging on the Gaussian ellipsoid properties of adjacent perspectives to obtain the Gaussian ellipsoid properties of the pseudo perspective:
[0122]
[0123] Among them, x i is the position of the adjacent Gaussian ellipsoid, ∑ i is the covariance matrix representing the shape of the three-dimensional Gaussian ellipsoid, Represents the color of the three-dimensional Gaussian ellipsoid relative to the line of sight. Specific embodiment six:
[0125] The difference between the sixth embodiment of the present invention and the fifth embodiment is that:
[0126] The optimization of the pseudo-perspective Gaussian ellipsoid is completed by adding a regularization term to the loss function. The regularization term consists of two parts: geometric regularization and color regularization:
[0127] In order to make the distribution of the pseudo-perspective Gaussian ellipsoid in three-dimensional space more reasonable, multi-perspective geometric regularization is used to constrain the position and radius of the pseudo-perspective Gaussian ellipsoid:
[0128]
[0129] Among them, ω′ i is the weight parameter, μ pseudo and r pseudo Respectively represent the position and radius of the pseudo-view Gaussian ellipsoid, μ i and r i Respectively represent the position and radius of the adjacent Gaussian ellipsoid;
[0130] Color regularization is used to constrain the color of the pseudo-view Gaussian ellipsoid to maximize its consistency with the adjacent view Gaussian ellipsoid:
[0131]
[0132] Among them, ω′ i is the weight parameter, c pseudo Indicates the color of the pseudo-view Gaussian ellipsoid, c i Indicates the color of the adjacent Gaussian ellipsoid;
[0133] The total loss function used combines image reconstruction, geometry and color losses, and divides the loss function into three terms, which are balanced by weights λ1, λ2, and λ3; the first term L image is an L1 loss, L D-SSIM is the error that measures the structural similarity between the rendered image and the real image. The second term L geometryis the geometric regularization term for the pseudo-perspective Gaussian ellipsoid, and the third term L color is the color regularization term for the pseudo-view Gaussian ellipsoid:
[0134] L=(1-λ1)L image +λ1L D-SSIM +λ2L geometry +λ3L color . Specific embodiment seven:
[0136] The difference between the seventh embodiment of the present invention and the sixth embodiment is that:
[0137] The step 5 is specifically as follows:
[0138] According to the spatial boundary B of the building point cloud, the initial point cloud P calculated from SfM is i} to extract points within the building space boundary:
[0139] P inside ={p i ∈P|b min ≤p i ≤ b max}
[0140] The extracted building point cloud is clustered using a density-based clustering algorithm to effectively remove floating objects on the building surface and optimize the final three-dimensional model. Specific embodiment eight:
[0142] The difference between the eighth embodiment of the present invention and the seventh embodiment is that:
[0143] The present invention provides a building automation three-dimensional reconstruction method system, the system comprising:
[0144] A data acquisition module, wherein the data acquisition module controls the drone to fly around the building and photograph the building facade from different perspectives to obtain a multi-perspective picture set;
[0145] A data calculation module, which calculates the camera parameters of each picture and obtains an initial point cloud based on the SfM algorithm;
[0146] A pseudo-view camera generation module, wherein the pseudo-view camera generation module calculates the spatial boundary of the building point cloud and generates a pseudo-view camera according to the obtained camera parameters;
[0147] An optimization module, wherein the optimization module defines and optimizes the properties of the Gaussian ellipsoid of the unknown perspective from the Gaussian ellipsoids of the adjacent known perspectives according to the weights of the distance and direction of the camera;
[0148] The reconstruction module uses the building space boundary obtained by camera parameter calculation to extract the building point cloud, and uses the density-based DBSCAN algorithm to filter the building point cloud to optimize the Gaussian model to obtain a high-precision building three-dimensional model. Specific embodiment nine:
[0150] The difference between the ninth embodiment of the present invention and the eighth embodiment is that:
[0151] The present invention provides a computer-readable storage medium on which a computer program is stored. The program is executed by a processor to implement a method for three-dimensional reconstruction of building automation. Specific embodiment ten:
[0153] The difference between the tenth embodiment of the present invention and the ninth embodiment is that:
[0154] The present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements a building automation three-dimensional reconstruction method when executing the computer program. Specific embodiment eleven:
[0156] The difference between the eleventh embodiment of the present invention and the tenth embodiment is that:
[0157] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.
[0158] The present invention is a building automation 3D reconstruction method based on improved 3D Gaussian Splatting, and the specific implementation process includes:
[0159] (1) Control the drone to fly around the building and take photos of the building facade from different perspectives to obtain a multi-perspective image set;
[0160] (2) Calculate the camera parameters of each image based on the Structure from Motion (SfM) algorithm and obtain the initial point cloud;
[0161] (3) Calculate the spatial boundary of the building point cloud based on the obtained camera intrinsic parameters and generate a pseudo-view camera;
[0162] (4) defining and optimizing the Gaussian ellipsoid properties of the unknown viewpoint by the Gaussian ellipsoids of the adjacent known viewpoints according to the weights of the camera distance and direction;
[0163] (5) The building point cloud is extracted using the building spatial boundary calculated using the camera intrinsic parameters, and the density-based DBSCAN algorithm is used to filter the building point cloud to optimize the Gaussian model and obtain a high-precision three-dimensional building model.
[0164] The purpose of this embodiment is to propose an automated 3D reconstruction method for buildings based on improved 3D Gaussian Splatting, which can be applied to the 3D model reconstruction process of general buildings.
[0165] The core of the present invention is to reconstruct a high-precision, clear-texture 3D model from multi-view images of general buildings based on the 3D GS architecture and SfM algorithm. The method is mainly divided into two steps. The first step is to obtain a multi-view image dataset. This process requires manual control of the drone to fly around the target building to obtain a multi-view dataset with high image quality. The second step is model generation, that is, processing the obtained dataset and training the proposed model. Finally, the 3D model of the building is obtained from the model output, so as to facilitate applications in fields such as damage identification, disaster assessment, and heritage protection based on the model.
[0166] Step (1) is specifically as follows:
[0167] (2a) Choose a day with high visibility and use a consumer-grade drone equipped with a high-definition camera to photograph the target building from multiple perspectives.
[0168] (2b) The drone is kept about 3 meters away from the building surface, and the photo overlap rate is controlled at about 50%. Based on the built-in fixed-point cruise algorithm of the drone and supplemented by manual camera posture adjustment, the original multi-view image set S{X i |i=1,...,N}, a total of N pictures.
[0169] Step (2) is specifically as follows:
[0170] (3a) Obtain the equivalent focal length f from the camera manufacturer’s manual as the known initial camera intrinsic parameter.
[0171] (3b) Based on the SfM algorithm, the camera parameters are calculated to obtain a multi-view image set S{X i |i=1,...,N} to calculate the camera parameters.
[0172] The SfM algorithm is as follows:
[0173] (3c) First, the image is preprocessed and downsampled to 1 / 2 size.
[0174] (3d) Select the Scale-Invariant Feature Transform (SIFT) algorithm to extract feature points. The extraction formula is:
[0175] L(x,y,σ)=G(x,y,σ)*I(x,y)
[0176]
[0177] D(x,y,σ)=L(x,y,σ)-L(x,y,kσ)
[0178] Where (x, y) is the pixel coordinate of the image, L(x, y, σ) is the Gaussian blurred image with σ as the parameter, and D(x, y, σ) is the Gaussian difference image. The local extremum detection algorithm compares each pixel in D(x, y, σ) with its 26 spatially adjacent pixels. If the pixel is larger or smaller than all adjacent pixels, it is marked as a SIFT feature point, thereby obtaining a feature point set.
[0179] (3e) The gradient size and direction of the 16×16 area near the feature point are formed into a 128-dimensional vector according to the histogram statistics method, and the vector is normalized to obtain the feature descriptor used to match the feature point The feature descriptor of the feature point is based on the following formula:
[0180]
[0181] θ(x,y)=arctan((L(x,y+1)-L(x,y-1)) / (L(x+1,y)-L(x-1,y)))
[0182] Where L is the Gaussian blurred image mentioned in (4b).
[0183] The feature points are then matched by calculating the distance between the descriptors of all image pairs, thus obtaining a set of matched feature point pairs. in
[0184] (3f) The camera parameters are calculated based on the binocular stereo geometry algorithm. The camera parameters are given by the matrix S λ K[R|t] represents, where is the camera internal reference; is the camera external parameter, where is the rotation matrix, is the translation vector; S λ =diag(1 / λ,1 / λ,1) is the radial distortion matrix of the lens, where is a nonlinear polynomial function of the distance d from a pixel to the center of the image, is the radial distortion parameter. λ , K is often known, and an estimate can be obtained from the Exchange Image File Format (EXIF) and globally optimized in the final stage.
[0185] Step (3) is specifically as follows:
[0186] (5a) Based on the camera parameters S obtained from step (3f) λK[R|t] defines a sample space for generating pseudo cameras Among them, t min is the minimum camera extrinsic parameter, t max is the largest camera extrinsic parameter.
[0187] (5b) In order to obtain the sample space of the pseudo camera, assume that all cameras are focused on the center of the scene and use spherical linear interpolation to generate the pseudo camera pose S P =(t+ε,q), where t∈S λ K[R|t], ε is a randomly added perturbation, and q is a quaternion representing the average rotation of the two cameras.
[0188] (5c) Based on the camera parameters S obtained in step (3f) λ K[R|t] defines the spatial boundary of the building point cloud B = {(t min ,t max )}.
[0189] Step (4) is specifically as follows:
[0190] (6a) The weight of the pseudo-view Gaussian ellipsoid is defined according to the similarity of the distance and angle between the generated pseudo camera and the adjacent cameras:
[0191]
[0192] In the formula, θ i is the angle between the pseudo camera and the i-th known camera direction, d i is the Euclidean distance between the pseudo camera and the i-th known camera, and ε is a small constant. And the weight is normalized to (6b) Using the calculated weight ω i ', perform weighted averaging on the Gaussian ellipsoid properties of adjacent perspectives to obtain the Gaussian ellipsoid properties of the pseudo perspective:
[0193]
[0194] In the formula, x i is the position of the adjacent Gaussian ellipsoid, ∑ i is the covariance matrix representing the shape of the three-dimensional Gaussian ellipsoid, Represents the color of the three-dimensional Gaussian ellipsoid relative to the line of sight.
[0195] Step (5) is specifically as follows:
[0196] (7a) According to the building point cloud space boundary B defined in step (3f), the initial point cloud P calculated from SfM is obtained as follows: i} to extract the points that belong to the spatial boundary of the building.
[0197] Pinside ={p i ∈P|b min ≤p i ≤ b max}
[0198] (7b) Finally, the density-based clustering algorithm (DBSCAN) is used to cluster the extracted building point cloud to effectively remove floating objects on the building surface and optimize the final three-dimensional model.
[0199] In step (4):
[0200] (8a) In order to make the distribution of the pseudo-perspective Gaussian ellipsoid in three-dimensional space more reasonable, multi-perspective geometric regularization is used to constrain the position and radius of the pseudo-perspective Gaussian ellipsoid.
[0201]
[0202] In the formula, ω′ i is the weight parameter, μ pseudo and r pseudo Respectively represent the position and radius of the pseudo-view Gaussian ellipsoid, μ i and r i Respectively represent the position and radius of the adjacent Gaussian ellipsoid.
[0203] (8b) Color regularization is used to constrain the color of the pseudo-view Gaussian ellipsoid so that it is as consistent as possible with the adjacent view Gaussian ellipsoids.
[0204]
[0205] In the formula, ω′ i is the weight parameter, c pseudo Indicates the color of the pseudo-view Gaussian ellipsoid, c i Indicates the color of the adjacent Gaussian ellipsoid.
[0206] (8c) The total loss function used by the model combines image reconstruction, geometry and color losses, and divides the loss function into three terms, which are balanced by weights λ1, λ2, and λ3. image is an L1 loss, L D-SSIM is the error that measures the structural similarity between the rendered image and the real image. The second term L geometry is the geometric regularization term for the pseudo-perspective Gaussian ellipsoid, and the third term L color is the color regularization term for the pseudo-view Gaussian ellipsoid.
[0207] L=(1-λ1)L image +λ1L D-SSIM +λ2L geometry +λ3L color
[0208] The invention is completed by a computer program written in Python and C++, and the target of the implementation is a scientific research building of a certain school. The following specific examples illustrate the effect of the invention.
[0209] When conducting experiments, pilots should try to choose weather with high visibility to operate drones. After arriving at the location, observe the environment around the building and pre-plan the flight route through the built-in fixed-point cruise function of the drone. Experiments are conducted based on the expected image overlap rate to set the drone flight speed and timed shooting interval. The proposed method has lower requirements for overlap rate and image quality than traditional methods. Therefore, it is not sensitive to errors in flight speed and shooting interval caused by environmental and human factors during actual flight. The parameters of the drone and onboard camera used in the experiment are shown in Table 1. Some sample set images are shown in the figure below. Figure 2 shown.
[0210] Table 1 UAV and camera parameters used in the experiment
[0211]
[0212] After shooting, the original image resolution (5280×3956) is high, which makes the process of calculating camera parameters slow. Therefore, the image is first downsampled to half of the original resolution. This operation is only performed during the camera parameter calculation process and will not affect the subsequent training and optimization process. After preprocessing, the camera parameters are calculated based on the SfM algorithm. The hardware environment used by the method of the present invention is shown in Table 2.
[0213] Table 2 Hardware environment parameters
[0214]
[0215] Then, a set of target camera poses is defined based on the known camera poses, and the minimum and maximum coordinates of the camera poses are taken as the bounding box of the building point cloud to determine the possible camera position space. The pseudo-perspective camera poses are generated by spherical linear interpolation and position interpolation of the known camera poses. Then, the initial Gaussian ellipsoid and the Gaussian ellipsoid corresponding to the pseudo-perspective are interpolated as follows: Figure 4 The series of optimizations shown results in the final 3D model.
[0216] Some parameter settings during the experiment are shown in Table 3.
[0217] Table 3 Model related parameter settings
[0218]
[0219] The weight coefficient is determined by calculating the pseudo camera pose and the distance between adjacent cameras, and the correlation of adjacent view information is used to regularize the parameters of the unknown view Gaussian ellipsoid. During the training process, the loss function quickly drops below 0.3 after 10 iterations. As the iterations proceed, the loss function stabilizes after 30 iterations, that is, the model converges. The loss function of the model is 0.128 at 6000 iterations. After 30,000 iterations, the final loss function stabilizes at around 0.012. The total training time is about 20 minutes. After training, the model is tested as follows: Figure 5 The filtering optimization shown in the figure is finally output through the optimized Gaussian ellipsoid model as shown in the figure. Figure 6 shown.
[0220] In order to illustrate that the reconstruction effect of the present invention is better than that of general methods, a comparative experiment was conducted in this example, using the same picture and preprocessing method as the control model input. Through experimental comparison, the PSNR (peak signal-to-noise ratio) value of the reconstruction result of the present invention is 27.22, which is 1.14 higher than that of the control model, indicating that the present invention can more effectively reduce the reconstruction error. The SSIM (structural similarity index) value is 0.039 higher than that of the control model, and the LPIPS (perceptual similarity loss) value is 0.268, which is lower than 0.301 of the control model, indicating that the present invention has a better effect in maintaining image structure and details, and is closer to the real image in visual perception. The reconstruction results of the control model are shown in the figure. Figure 3 As shown in Figure 4, the reconstructed results have obvious artifacts and floating objects. The specific evaluation indicators are shown in Table 4.
[0221] Table 4 Comparison of model quality evaluation indicators
[0222]
[0223] Figure 7 The figure shows a local rendering diagram of the three-dimensional model finally generated by the method. The left figure is the reconstruction result of 3D Gaussian Splatting, the middle figure is the reconstruction result of the model proposed in the present invention, and the right figure is a real picture taken by a drone. It can be seen that the model reconstructed by this method still has a high resolution in the geometric detail area, and the reconstruction result is still clear and accurate when the lighting environment is poor when shooting by a drone.
[0224] In the description of this specification, the description of reference terms such as "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples without contradiction. In addition, the terms "first" and "second" are used only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, the features defined as "first" and "second" can explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "N" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined. Any process or method description in the flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code including one or more executable instructions for implementing the steps of a custom logic function or process, and the scope of the preferred embodiment of the present invention includes other implementations, in which the functions may not be performed in the order shown or discussed, including in a substantially simultaneous manner or in a reverse order according to the functions involved, which should be understood by a person skilled in the art of the art to which the embodiments of the present invention belong. The logic and / or steps represented in the flowchart or otherwise described herein, for example, may be considered as a sequenced list of executable instructions for implementing the logic function, and may be specifically implemented in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in combination with these instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" may be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in combination with these instruction execution systems, devices or apparatuses. More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection having one or N wirings (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disk read-only memory (CDROM).In addition, the computer readable medium can even be paper or other suitable medium on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, then editing, interpreting or processing in other suitable ways as necessary, and then storing it in a computer memory. It should be understood that the various parts of the present invention can be implemented with hardware, software, firmware, or a combination thereof. In the above-mentioned embodiment, N steps or methods can be implemented with software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented with hardware, as in another embodiment, any one of the following technologies known in the art or their combination can be used to implement: a discrete logic circuit with a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit with a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.
[0225] The above is only a preferred implementation of a method for 3D reconstruction of building automation. The protection scope of a method for 3D reconstruction of building automation is not limited to the above embodiments. All technical solutions under this idea belong to the protection scope of the present invention. It should be pointed out that for those skilled in the art, several improvements and changes without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.
Claims
1. A method for 3D reconstruction of building automation, characterized by: The following steps are involved: Step 1: Control the drone to fly around the building and take photos of the building facade from different perspectives to obtain a multi-perspective picture set; Step 2: Based on the SfM algorithm, calculate the camera parameters of each image and obtain the initial point cloud; Step 3: Based on the obtained camera parameters, calculate the spatial boundary of the building point cloud and generate a pseudo-perspective camera; Step 4: According to the weights of the distance and direction of the camera, the Gaussian ellipsoid properties of the unknown view are defined and optimized by the Gaussian ellipsoids of the adjacent known viewpoints; Step 5: Use the building space boundary calculated by camera parameters to extract the building point cloud, and use the density-based DBSCAN algorithm to filter the building point cloud to optimize the Gaussian model to obtain a high-precision building 3D model.
2. The method according to claim 1, characterized in that: The step 1 is specifically as follows: The target building was photographed from multiple perspectives using a drone. The drone was kept 3 meters away from the building surface and the photo overlap rate was controlled at 50%. Based on the drone's built-in fixed-point cruise algorithm and supplemented by manual camera posture adjustment, a multi-perspective picture set S{X i |i=1,...,N}, a total of N pictures.
3. The method according to claim 2, characterized in that: The step 2 is specifically as follows: According to the equivalent focal length f of the camera as the known initial parameters of the camera, the multi-view image set S{X i |i=1,...,N}N images to calculate camera parameters; The SfM algorithm is as follows: Preprocess the image and downsample it to 1 / 2 size; Select the scale-invariant feature transformation algorithm to extract feature points. The extraction formula is: L(x,y,σ)=G(x,y,σ)*I(x,y) D(x,y,σ)=L(x,y,σ)-L(x,y,kσ) Where (x, y) is the pixel coordinate of the image, L(x, y, σ) is the Gaussian blurred image with σ as the parameter, and D(x, y, σ) is the Gaussian difference image. The local extremum detection algorithm compares each pixel in D(x, y, σ) with its 26 spatially adjacent pixels. If the pixel is larger or smaller than all adjacent pixels, it is marked as a SIFT feature point, thereby obtaining a feature point set. The gradient size and direction of the 16×16 area near the feature point are formed into a 128-dimensional vector according to the histogram statistics method, and the vector is normalized to obtain the feature descriptor used to match the feature point The feature descriptor of the feature point is based on the following formula: θ(x,y)=arctan((L(x,y+1)-L(x,y-1)) / (L(x+1,y)-L(x-1,y))) Where L is the Gaussian blurred image mentioned in (4b); The feature points are matched by calculating the distance between the descriptors of all image pairs, thus obtaining a set of matched feature point pairs. in The camera parameters are calculated based on the binocular stereo geometry algorithm. The camera parameters are represented by the matrix S λ K[R|t] represents, where is the camera internal reference; is the camera external parameter, where is the rotation matrix, is the translation vector; S λ =diag(1 / λ,1 / λ,1) is the radial distortion matrix of the lens, where is a nonlinear polynomial function of the distance d from a pixel to the center of the image, is the radial distortion parameter, where S λ , K is known, an estimate is obtained from the Exchange Image File Format EXIF, and a global optimization is performed in the final stage.
4. The method according to claim 3, characterized in that: The step 3 is specifically as follows: According to the obtained camera parameters S λ K[R|t] defines a sample space for generating pseudo cameras Among them, t min is the minimum camera extrinsic parameter, t max is the largest camera extrinsic parameter; In order to obtain the sample space of the pseudo camera, all cameras are set to focus on the center of the scene, and spherical linear interpolation is used to generate the pseudo camera pose S P =(t+ε,q), where t∈S λ K[R|t], ε is a randomly added perturbation, q is a quaternion representing the average rotation of the two cameras; According to the camera parameters S λ K[R|t] defines the spatial boundary of the building point cloud B = {(t min ,t max )}.
5. The method according to claim 4, characterized in that: The step 4 is specifically as follows: The weight of the pseudo-view Gaussian ellipsoid is defined according to the similarity of the distance and angle between the generated pseudo camera and the adjacent cameras: Among them, θ i is the angle between the pseudo camera and the i-th known camera direction, d i is the Euclidean distance between the pseudo camera and the i-th known camera, ε is a constant, and the weight is normalized to Using the calculated weight ω′ i , perform weighted averaging on the Gaussian ellipsoid properties of adjacent perspectives to obtain the Gaussian ellipsoid properties of the pseudo perspective: Among them, x i is the position of the adjacent Gaussian ellipsoid, ∑ i is the covariance matrix representing the shape of the three-dimensional Gaussian ellipsoid, Represents the color of the three-dimensional Gaussian ellipsoid relative to the line of sight.
6. The method according to claim 5, characterized in that: The optimization of the pseudo-perspective Gaussian ellipsoid is completed by adding a regularization term to the loss function. The regularization term consists of two parts: geometric regularization and color regularization. composition: In order to make the distribution of the pseudo-perspective Gaussian ellipsoid in three-dimensional space more reasonable, multi-perspective geometric regularization is used to constrain the position and radius of the pseudo-perspective Gaussian ellipsoid: Among them, ω′ i is the weight parameter, μ pseudo and r pseudo They represent the position and radius of the pseudo-view Gaussian ellipsoid, μi and ri represent the position and radius of the adjacent Gaussian ellipsoid; Color regularization is used to constrain the color of the pseudo-view Gaussian ellipsoid to maximize its consistency with the adjacent view Gaussian ellipsoid: Among them, ω′ i is the weight parameter, c pseudo Indicates the color of the pseudo-view Gaussian ellipsoid, c i Indicates the color of the adjacent Gaussian ellipsoid; The total loss function used combines image reconstruction, geometry and color losses, and divides the loss function into three terms, which are balanced by weights λ1, λ2, and λ3; the first term L image is an L1 loss, L D-SSIM is the error that measures the structural similarity between the rendered image and the real image. The second term L geometry is the geometric regularization term for the pseudo-perspective Gaussian ellipsoid, and the third term L color is the color regularization term for the pseudo-view Gaussian ellipsoid: L=(1-λ1)L image +λ1L D-SSIM +λ2L geometry +λ3L color 。 7. The method according to claim 6, characterized in that: The step 5 is specifically as follows: According to the spatial boundary B of the building point cloud, the initial point cloud P calculated from SfM is i } to extract points within the building space boundary: P inside ={p i ∈P|b min ≤p i ≤b max } The extracted building point cloud is clustered using a density-based clustering algorithm to effectively remove floating objects on the building surface and optimize the final three-dimensional model.
8. A building automation 3D reconstruction method system, characterized by: The system comprises: A data acquisition module, wherein the data acquisition module controls the drone to fly around the building and photograph the building facade from different perspectives to obtain a multi-perspective picture set; A data calculation module, which calculates the camera parameters of each picture and obtains an initial point cloud based on the SfM algorithm; A pseudo-view camera generation module, wherein the pseudo-view camera generation module calculates the spatial boundary of the building point cloud and generates a pseudo-view camera according to the obtained camera parameters; An optimization module, wherein the optimization module defines and optimizes the properties of the Gaussian ellipsoid of the unknown perspective from the Gaussian ellipsoids of the adjacent known perspectives according to the weights of the distance and direction of the camera; The reconstruction module uses the building space boundary obtained by camera parameter calculation to extract the building point cloud, and uses the density-based DBSCAN algorithm to filter the building point cloud to optimize the Gaussian model to obtain a high-precision building three-dimensional model.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: The program is executed by a processor to implement the method according to claims 1-7.
10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the method of claims 1-7 is implemented.
Citation Information
Cited By
Binocular vision and Gaussian sputtering fused three-dimensional depth reconstruction method, device, equipment and medium
CN119693437A
Three-dimensional automatic modeling method for high-precision transformer substation
CN119919578A
Building three-dimensional reconstruction method based on 3D Gaussian sputtering guidance
CN119991961A
Fire scene three-dimensional reconstruction method and device and electronic equipment
CN120431245A
