A method for establishing a three-dimensional real scene model with autonomous updating function

Abstract

The present invention relates to the field of three-dimensional modeling technology, and specifically to a method for establishing a three-dimensional real-scene model with an autonomous updating function, comprising: collecting multi-viewpoint point cloud data of a scene to be modeled to obtain an initial three-dimensional point cloud set; estimating the curvature of the initial three-dimensional point cloud set according to a singular value decomposition algorithm to obtain a curvature value for each data point in the initial three-dimensional point cloud set; and estimating the normal of each data point in the initial three-dimensional point cloud set according to the curvature value of each data point in the initial three-dimensional point cloud set to obtain a normal for each data point in the initial three-dimensional point cloud set. By using a singular value decomposition algorithm for preliminary curvature estimation and normal estimation, the present invention can more accurately analyze details in the three-dimensional point cloud, avoid manual intervention, and improve the degree of automation in the modeling process.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional modeling, and in particular to a method for establishing a three-dimensional real scene model with an autonomous updating function. Background Art

[0002] Traditional methods usually rely on manual analysis and processing of three-dimensional point clouds, especially in curvature estimation and normal estimation, which lacks automated mechanisms and is prone to human errors or inconsistencies, affecting modeling accuracy. In traditional methods, the processing of non-planar areas usually requires manual selection, segmentation and labeling, and lacks intelligent automated processing tools, which leads to slow modeling and low efficiency when processing complex scenes or large-scale data. Traditional methods may have difficulties in processing complex non-planar areas because they require manual adjustment and optimization, which is not only time-consuming but may not perfectly adapt to complex geometric structures. Target recognition in traditional methods usually relies on manual judgment and lacks automated recognition mechanisms, which can easily lead to irrelevant areas being mismodeled or key areas being missed, affecting the accuracy of the final model. In traditional methods, redundant data often cannot be effectively deleted through automated means, resulting in the final generated three-dimensional model containing unnecessary parts, increasing the burden of data storage and reducing the accuracy of the model. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a method for establishing a three-dimensional real scene model with an autonomous updating function.

[0004] The technical solution adopted to solve the above technical problems is: a method for establishing a three-dimensional real scene model with an autonomous updating function, comprising:

[0005] Performing multi-viewpoint point cloud data collection on the scene to be modeled to obtain an initial three-dimensional point cloud set, and performing curvature estimation on the initial three-dimensional point cloud set according to a singular value decomposition algorithm to obtain a curvature value of each data point in the initial three-dimensional point cloud set;

[0006] estimating a normal for each data point in the initial three-dimensional point cloud set according to a curvature value of each data point in the initial three-dimensional point cloud set to obtain a normal for each data point in the initial three-dimensional point cloud set, and dividing the initial three-dimensional point cloud set into a planar region subset and a plurality of non-planar region subsets according to the curvature value and the normal of each data point in the initial three-dimensional point cloud set;

[0007] Performing point cloud mapping on the non-planar region subset using a principal component analysis method to obtain a two-dimensional projection image corresponding to the non-planar region subset, and performing contour extraction on the two-dimensional projection image to obtain a contour image corresponding to the two-dimensional projection image;

[0008] Performing feature encoding on the contour image to obtain image features corresponding to the contour image, matching the image features with a preset target feature library to obtain target features most similar to the image features, and obtaining the target type corresponding to the image features based on the most similar target features;

[0009] Determine whether the non-planar area subset is a target three-dimensional real scene model according to the target type; if the non-planar area subset is not a target three-dimensional real scene model, delete the non-planar area subset; after processing the multiple non-planar area subsets to obtain multiple target non-planar area subsets, perform three-dimensional modeling based on the planar area subset and the multiple target non-planar area subsets to obtain a three-dimensional real scene model.

[0010] Preferably, performing curvature estimation on the initial three-dimensional point cloud set according to a first singular value decomposition algorithm to obtain a curvature value of each data point in the initial three-dimensional point cloud set includes:

[0011] For each data point in the initial three-dimensional point cloud set, determine a neighborhood point set of the data point according to a K-nearest neighbor algorithm, and determine the centroid of the neighborhood point set of the data point;

[0012] Determining a covariance matrix of the data point according to the centroid of a neighborhood point set of the data point, and performing singular value decomposition on the covariance matrix to obtain three first singular values, wherein the three first singular values ​​represent the distribution of the neighborhood points in the directions of three principal axes;

[0013] The curvature of the initial three-dimensional point cloud set is estimated according to the smallest first singular value among the three first singular values ​​to obtain the curvature value of each data point in the initial three-dimensional point cloud set.

[0014] Preferably, estimating the normal of each data point in the initial three-dimensional point cloud set according to the curvature value of each data point in the initial three-dimensional point cloud set to obtain the normal of each data point in the initial three-dimensional point cloud set includes:

[0015] Determine the Euclidean distance between each neighborhood point in the neighborhood point set and the data point, and determine the distance weight between each neighborhood point in the neighborhood point set and the data point according to the Euclidean distance;

[0016] Determining a weighted covariance matrix of the data points according to the distance weights, and performing singular value decomposition on the covariance matrix to obtain three second singular values;

[0017] The eigenvector corresponding to the smallest second singular value among the three second singular values ​​is used as the normal vector of the data point, wherein the normal vector is the normal line.

[0018] Preferably, dividing the initial three-dimensional point cloud set into a plane region subset and a non-plane region subset according to the curvature value and the normal of each data point in the initial three-dimensional point cloud set comprises:

[0019] Determining a region division threshold according to a curvature value of each data point in the initial three-dimensional point cloud set;

[0020] If the curvature value of each data point in the initial three-dimensional point cloud set is less than the region division threshold, the data point is a plane region subset; otherwise, the data point is a non-plane region subset;

[0021] An angle between a normal vector of the data point and a normal vector of the neighborhood point is determined; if the angle is less than a preset angle threshold, the neighborhood point and the neighborhood point belong to the same planar region subset or the same non-planar region subset.

[0022] Preferably, performing point cloud mapping on the non-planar region subset by a principal component analysis method to obtain a two-dimensional projection image corresponding to the non-planar region subset includes:

[0023] Determining a mean value of the data points in the non-planar region subset, subtracting the mean value from each data point to obtain a centered data point, and multiplying the centered data point with its transpose to obtain a centered covariance matrix of the data point;

[0024] Perform eigenvalue decomposition on the centralized covariance matrix to obtain three eigenvalues ​​and three corresponding eigenvectors, and select a plane formed by the eigenvectors corresponding to the largest eigenvalue and the second largest eigenvalue among the three eigenvalues ​​as a projection plane.

[0025] Preferably, performing point cloud mapping on the non-planar region subset by a principal component analysis method to obtain a two-dimensional projection image corresponding to the non-planar region subset further includes:

[0026] Determining a principal component direction matrix based on the eigenvectors corresponding to the largest eigenvalue and the second largest eigenvalue among the three eigenvalues, projecting each data point into a two-dimensional space based on the principal component direction matrix, calculating the projection coordinates, and determining a two-dimensional point set based on all the projected data points;

[0027] Normalizing the two-dimensional mapping data in the two-dimensional point set to obtain a normalized two-dimensional point set, smoothing the normalized two-dimensional point set according to kernel density estimation to obtain a data point density distribution corresponding to the normalized two-dimensional point set, and determining a two-dimensional distribution graph based on the data point density distribution, wherein the two-dimensional distribution graph is a two-dimensional projection image.

[0028] Preferably, feature encoding is performed on the contour image to obtain image features corresponding to the contour image, including:

[0029] Determining the centroid of the contour image, performing edge detection on the contour image, extracting a contour point set of a shape, and randomly downsampling the contour point set to a preset number to obtain a new contour point set;

[0030] Determine a Euclidean distance matrix and an angular distance matrix between the new contour point set and the centroid of the contour image, divide the 360° angle interval into a preset number of angle segments, and divide the distance interval [0, 1] into a preset number of distance segments;

[0031] According to the Euclidean distance matrix and the angular distance matrix, the number of data points in each of the angle segments and each of the distance segments is counted to obtain a two-dimensional histogram of angle-distance, wherein the value of each grid of the two-dimensional histogram of angle-distance represents the number of data point pairs in the angle-distance area, and the two-dimensional histogram of angle-distance is the image feature corresponding to the contour image.

[0032] Preferably, matching the image feature with a preset target feature library to obtain the target feature most similar to the image feature includes:

[0033] Determine the difference between the image feature and each target feature in a preset target feature library at each position;

[0034] Accumulating the differences between the image feature and the target feature at all positions to obtain a matching cost coefficient between the image feature and the target feature;

[0035] A target feature that is most similar to the image feature is determined according to the matching cost coefficient.

[0036] The beneficial effects of the present invention are as follows: (1) The present invention uses the singular value decomposition algorithm to perform preliminary curvature estimation and normal estimation, which can more accurately analyze the details in the three-dimensional point cloud, avoid manual intervention, and improve the degree of automation of the modeling process. By performing point cloud mapping on a subset of non-planar areas through principal component analysis and extracting the contours of the two-dimensional projection image, it helps to extract key features from complex three-dimensional data more clearly and concisely, thereby improving the modeling efficiency of complex structures; (2) The present invention matches the feature coding of the contour image with the preset target feature library, and the system can automatically identify whether the target three-dimensional real scene model is included. This automated target recognition mechanism ensures the accuracy of the model during the modeling process. The invention is both scalable and targeted, avoiding the mismodeling of irrelevant areas, and when some non-planar area subsets do not belong to the target three-dimensional real scene model, these areas can be automatically deleted to reduce redundant data and ensure the modeling quality. This process improves the flexibility of model updating and optimization; (3) The invention can not only process simple planar areas but also effectively process complex non-planar areas, adapting to the three-dimensional modeling needs of different scenes, and has broad application potential, especially in the fields of architecture, urban planning, virtual reality, etc., and after comprehensive processing of the planar area subsets and the target non-planar area subsets, an accurate three-dimensional real scene model can be generated. This method can automatically generate high-quality three-dimensional models when processing large-scale and complex scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 The figure is a schematic flow chart of the overall method steps in one embodiment of the present invention. DETAILED DESCRIPTION

[0038] Example 1, as Figure 1 As shown, the present invention proposes a method for establishing a three-dimensional real scene model with an autonomous updating function, comprising:

[0039] S1. Collect multi-view point cloud data of the scene to be modeled to obtain an initial three-dimensional point cloud set, and perform curvature estimation on the initial three-dimensional point cloud set according to the singular value decomposition algorithm to obtain the curvature value of each data point in the initial three-dimensional point cloud set;

[0040] S2. estimating a normal of each data point in the initial three-dimensional point cloud set according to the curvature value of each data point in the initial three-dimensional point cloud set to obtain a normal of each data point in the initial three-dimensional point cloud set, and dividing the initial three-dimensional point cloud set into a planar region subset and multiple non-planar region subsets according to the curvature value and the normal of each data point in the initial three-dimensional point cloud set;

[0041] S3. performing point cloud mapping on the non-planar region subset using a principal component analysis method to obtain a two-dimensional projection image corresponding to the non-planar region subset, and performing contour extraction on the two-dimensional projection image to obtain a contour image corresponding to the two-dimensional projection image;

[0042] S4. Feature encoding is performed on the contour image to obtain image features corresponding to the contour image, the image features are matched with a preset target feature library to obtain target features most similar to the image features, and the target type corresponding to the image features is obtained based on the most similar target features;

[0043] S5. Determine whether the non-planar area subset is the target three-dimensional real scene model based on the target type. If the non-planar area subset is not the target three-dimensional real scene model, delete the non-planar area subset. After processing multiple non-planar area subsets to obtain multiple target non-planar area subsets, perform three-dimensional modeling based on the planar area subset and the multiple target non-planar area subsets to obtain a three-dimensional real scene model.

[0044] In the present invention, point cloud data refers to a three-dimensional coordinate data set obtained by laser scanning (LiDAR) or other sensors (such as cameras, depth cameras, etc.), and each data point generally represents a position in three-dimensional space; multi-perspective means obtaining point cloud data of multiple views from different angles and positions, so that the geometric shape of the scene can be fully captured; singular value decomposition refers to a matrix decomposition method used to analyze the characteristics of the matrix. Through the singular value decomposition algorithm, the curvature information of the point cloud can be obtained, which is used to estimate the geometric properties of the point cloud, such as the degree of curvature of the local surface; curvature is a quantity that describes the degree of curvature of a curve or surface. In three-dimensional point cloud processing, curvature estimation is used to determine whether the surface of each point is flat or curved, thereby providing information for the subsequent modeling process; the normal refers to the perpendicular direction of the tangent plane of a point on the surface. In 3D point cloud processing, the directional information of each point can be understood by estimating the normal of the point cloud data; the planar region subset refers to the relatively flat part of the surface in the point cloud; the non-planar region subset refers to the part of the surface with curved or complex shapes; principal component analysis is a dimensionality reduction method that reduces the dimension of the data by extracting the principal components of the data while retaining as much information as possible in the data. PCA is used here to map the point cloud data in the non-planar region and convert it into a two-dimensional projection image; a two-dimensional projection image refers to projecting the three-dimensional point cloud data onto a two-dimensional plane to form a planar image; extracting the outer boundary (contour) of the object in the two-dimensional image is crucial for identifying and analyzing the shape of the object; the target feature library refers to a preset database containing a large number of features of known objects. These feature libraries can help match the features of the current image with existing objects; the target three-dimensional real scene model refers to a three-dimensional model generated from point cloud data, usually used to represent objects or environments in the real world. By processing and analyzing the non-planar region subset, the goal is to identify which areas belong to specific three-dimensional objects and model them.

[0045] In a second embodiment, the present invention proposes a method for establishing a three-dimensional real scene model with an autonomous updating function. Compared to the first embodiment, this embodiment further includes: performing curvature estimation on the initial three-dimensional point cloud set according to a first singular value decomposition algorithm to obtain the curvature value of each data point in the initial three-dimensional point cloud set, including:

[0046] A1. For each data point in the initial 3D point cloud, determine the neighborhood point set of the data point using the K-nearest neighbor algorithm, and determine the centroid of the neighborhood point set of the data point;

[0047] A2. Determine the covariance matrix of the data point based on the centroid of the neighborhood point set of the data point, and perform singular value decomposition on the covariance matrix to obtain three first singular values, where the three first singular values ​​represent the distribution of the neighborhood points in the directions of the three principal axes;

[0048] A3. Estimating the curvature of the initial three-dimensional point cloud set according to the smallest first singular value among the three first singular values ​​to obtain the curvature value of each data point in the initial three-dimensional point cloud set.

[0049] In this embodiment, the K-nearest neighbor algorithm is a commonly used machine learning algorithm used for classification or regression. For data point processing, the KNN algorithm finds the K nearest neighbor points based on the distance relationship of a point. When used in point cloud processing, it can find the K nearest neighbor points of a data point, thereby constructing the neighborhood of the point. The centroid refers to the "center" position of a group of points, usually the average of the coordinates of these points. In three-dimensional space, the centroid is the arithmetic mean of the coordinates of all points in the area. For example, if a data point has multiple neighborhood points, the centroid is the average position of these points. The covariance matrix is ​​a matrix that describes how the various dimensions in a data set change in coordination. In a point cloud, the covariance matrix reflects the distribution of the neighborhood point set in different directions. It can show the variance (distribution along a certain axis) and covariance (relationship between different axes) of the point cloud data. In point cloud data analysis, the principal axes refer to the three main directions of the data set distribution, usually the three directions calculated by the covariance matrix through SVD. Each principal axis corresponds to a "main" direction of the point cloud and reflects the maximum or minimum change of the point cloud in that direction.

[0050] In an optional embodiment, estimating a normal of each data point in the initial three-dimensional point cloud set according to a curvature value of each data point in the initial three-dimensional point cloud set to obtain a normal of each data point in the initial three-dimensional point cloud set includes:

[0051] B1. Determine the Euclidean distance between each neighborhood point in the neighborhood point set and the data point, and determine the distance weight between each neighborhood point in the neighborhood point set and the data point based on the Euclidean distance;

[0052] B2. Determine a weighted covariance matrix of the data points according to the distance weights, and perform singular value decomposition on the covariance matrix to obtain three second singular values;

[0053] B3. Use the eigenvector corresponding to the smallest second singular value among the three second singular values ​​as the normal vector of the data point, where the normal vector is the normal line.

[0054] It should be noted that in the standard covariance matrix, the contribution of each point is equal, while in the weighted covariance matrix, the contribution of each point is weighted according to the distance (weight) between the points. A common practice is to multiply by a certain distance weight to reflect that points far away from the data points have less influence on the covariance; the normal vector refers to a vector perpendicular to a surface or curve. In three-dimensional space, the normal vector is a vector perpendicular to the local plane of the surface. By analyzing the covariance matrix and singular value decomposition of the point cloud, the normal vector of the surface where the data point is located can be obtained. The normal vector is usually associated with the eigenvector corresponding to the minimum singular value because the eigenvector represents the distribution of the data set in the direction of minimum change.

[0055] In an optional embodiment, the initial three-dimensional point cloud set is divided into a planar region subset and a non-planar region subset according to the curvature value and the normal of each data point in the initial three-dimensional point cloud set, including:

[0056] C1. Determine the region division threshold based on the curvature value of each data point in the initial 3D point cloud set;

[0057] C2. If the curvature value of each data point in the initial three-dimensional point cloud set is less than the region division threshold, the data point is a plane region subset; otherwise, the data point is a non-plane region subset;

[0058] C3. Determine the angle between the normal vector of the data point and the normal vector of the neighboring point. If the angle is less than a preset angle threshold, the neighboring point and the neighboring point belong to the same planar region subset or the same non-planar region subset.

[0059] It should be noted that the included angle refers to the angle between two vectors. Here, the included angle refers to the angle between the normal vector of a point and the normal vector of its neighboring points. This angle is used to determine whether the points belong to the same planar area or non-planar area. If the included angle is small, it means that the normal vectors of the two points are in similar directions and may belong to the same area.

[0060] In an optional embodiment, performing point cloud mapping on the non-planar region subset by a principal component analysis method to obtain a two-dimensional projection image corresponding to the non-planar region subset includes:

[0061] D1. Determine the mean of the data points in the non-planar region subset, subtract the mean from each data point to obtain the centered data point, and multiply the centered data point with its transposed value to obtain the centered covariance matrix of the data point.

[0062] D2. Perform eigenvalue decomposition on the centralized covariance matrix to obtain three eigenvalues ​​and their corresponding three eigenvectors. Select the plane formed by the eigenvectors corresponding to the largest eigenvalue and the second largest eigenvalue among the three eigenvalues ​​as the projection plane.

[0063] It should be noted that the projection plane is a two-dimensional plane used to represent the two main directions in the data. Here, it is a plane composed of the eigenvectors corresponding to the largest eigenvalue and the second largest eigenvalue. It is used to simplify the representation of data. Data can be reduced in dimension by projecting it onto this plane while retaining most of the important information.

[0064] In an optional embodiment, performing point cloud mapping on the non-planar region subset by a principal component analysis method to obtain a two-dimensional projection image corresponding to the non-planar region subset further includes:

[0065] D3. Determine the matrix of the principal component direction based on the eigenvectors corresponding to the largest and second largest eigenvalues ​​among the three eigenvalues. Project each data point into two-dimensional space based on the matrix of the principal component direction, calculate the projection coordinates, and determine the two-dimensional point set based on all the projected data points.

[0066] D4. Normalize the two-dimensional mapping data in the two-dimensional point set to obtain a normalized two-dimensional point set, smooth the normalized two-dimensional point set based on kernel density estimation to obtain a data point density distribution corresponding to the normalized two-dimensional point set, and determine a two-dimensional distribution map based on the data point density distribution, where the two-dimensional distribution map is the two-dimensional projection image.

[0067] It should be noted that kernel density estimation is a non-parametric method for estimating the distribution of data points. It generates a smooth probability density distribution by applying a kernel function (such as a Gaussian kernel) to the data points. This method is used to create a probability density curve of the data to help understand the distribution pattern of the data; the matrix of the principal component direction is a matrix composed of all eigenvectors (principal components), which represents the direction of data projection. Through this matrix, the data points can be projected into the new principal component space.

[0068] In an optional embodiment, feature encoding is performed on the contour image to obtain image features corresponding to the contour image, including:

[0069] E1. Determine the centroid of the contour image, perform edge detection on the contour image, extract the contour point set of the shape, and randomly downsample the contour point set to a preset number to obtain a new contour point set;

[0070] E2. Determine the Euclidean distance matrix and angular distance matrix between the new contour point set and the centroid of the contour image, divide the 360° angle interval into a preset number of angle segments, and divide the distance interval [0, 1] into a preset number of distance segments;

[0071] E3. Count the number of data points in each angle segment and each distance segment according to the Euclidean distance matrix and the angular distance matrix to obtain a two-dimensional histogram of angle-distance, wherein the value of each grid of the two-dimensional histogram of angle-distance represents the number of data point pairs in the angle-distance area, and the two-dimensional histogram of angle-distance is the image feature corresponding to the contour image.

[0072] It should be noted that downsampling refers to selecting a portion of the original data to reduce the amount of data. By randomly downsampling the contour point set to a preset number, a smaller point set can be obtained. This is usually used to reduce the amount of computation or simplify the problem. Edge detection is an image processing technique used to identify edges in an image. Edge detection algorithms (such as the Sobel operator and the Canny operator) can find areas in the image with large grayscale variations, thereby determining the contours of objects. The angle-distance two-dimensional histogram is a two-dimensional array in which each grid represents the number of data points within a specific angle and distance range. This histogram is obtained by counting the number of contour points in each angle and distance segment. It describes the distribution characteristics of contour points in angle and distance space. Data point pairs are points within a specific angle and distance segment. Each grid contains all contour points within that interval, thus representing the distribution density of contour points within that angle and distance interval.

[0073] In an optional embodiment, the image features are matched with a preset target feature library to obtain the target features that are most similar to the image features, including:

[0074] Determine the difference between the image features and each target feature in the preset target feature library at each position;

[0075] The differences between the image features and the target features at all positions are accumulated to obtain the matching cost coefficient between the image features and the target features. The calculation formula of the matching cost coefficient is as follows:

[0076] ;

[0077] in, Represents the matching cost coefficient between image features and target features, and represents a preset number of angle segments and a preset number of distance segments, Indicates the angle-distance two-dimensional histogram corresponding to the image feature Rank The values ​​in the column, Indicates the angle-distance two-dimensional histogram corresponding to the target feature Rank The values ​​within the column;

[0078] The target feature that is most similar to the image feature is determined based on the matching cost coefficient.

[0079] It should be noted that the difference at each position refers to the difference between the image features and the target features at each position in the image; the matching cost coefficient is a numerical value that measures the difference between the image features and the target features. Generally, the smaller the cost coefficient, the higher the degree of matching. It is usually used to evaluate the similarity between the image and the target. The lower the cost coefficient, the more similar the image features and the target features are.

[0080] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for establishing a three-dimensional real scene model with an autonomous updating function, characterized in that: include: Performing multi-viewpoint point cloud data collection on the scene to be modeled to obtain an initial three-dimensional point cloud set, and performing curvature estimation on the initial three-dimensional point cloud set according to a singular value decomposition algorithm to obtain a curvature value of each data point in the initial three-dimensional point cloud set; estimating a normal for each data point in the initial three-dimensional point cloud set according to a curvature value of each data point in the initial three-dimensional point cloud set to obtain a normal for each data point in the initial three-dimensional point cloud set, and dividing the initial three-dimensional point cloud set into a planar region subset and a plurality of non-planar region subsets according to the curvature value and the normal of each data point in the initial three-dimensional point cloud set; Performing point cloud mapping on the non-planar region subset using a principal component analysis method to obtain a two-dimensional projection image corresponding to the non-planar region subset, and performing contour extraction on the two-dimensional projection image to obtain a contour image corresponding to the two-dimensional projection image; Performing feature encoding on the contour image to obtain image features corresponding to the contour image, matching the image features with a preset target feature library to obtain target features most similar to the image features, and obtaining the target type corresponding to the image features based on the most similar target features; determining whether the non-planar area subset is a target three-dimensional real scene model according to the target type; if the non-planar area subset is not a target three-dimensional real scene model, deleting the non-planar area subset; after processing the multiple non-planar area subsets to obtain multiple target non-planar area subsets, performing three-dimensional modeling based on the planar area subset and the multiple target non-planar area subsets to obtain a three-dimensional real scene model; The point cloud mapping is performed on the non-planar region subset by a principal component analysis method to obtain a two-dimensional projection image corresponding to the non-planar region subset, including: Determining a mean value of the data points in the non-planar region subset, subtracting the mean value from each data point to obtain a centered data point, and multiplying the centered data point with its transpose to obtain a centered covariance matrix of the data point; Perform eigenvalue decomposition on the centralized covariance matrix to obtain three eigenvalues ​​and three corresponding eigenvectors, and select a plane formed by the eigenvectors corresponding to the largest eigenvalue and the second largest eigenvalue among the three eigenvalues ​​as a projection plane.

2. The method for establishing a three-dimensional real scene model with an autonomous updating function according to claim 1, characterized in that: Performing curvature estimation on the initial three-dimensional point cloud set according to a singular value decomposition algorithm to obtain a curvature value of each data point in the initial three-dimensional point cloud set includes: For each data point in the initial three-dimensional point cloud set, determine a neighborhood point set of the data point according to a K-nearest neighbor algorithm, and determine the centroid of the neighborhood point set of the data point; Determining a covariance matrix of the data point according to the centroid of a neighborhood point set of the data point, and performing singular value decomposition on the covariance matrix to obtain three first singular values, wherein the three first singular values ​​represent the distribution of the neighborhood points in the directions of three principal axes; The curvature of the initial three-dimensional point cloud set is estimated according to the smallest first singular value among the three first singular values ​​to obtain the curvature value of each data point in the initial three-dimensional point cloud set.

3. The method for establishing a three-dimensional real scene model with an autonomous updating function according to claim 2, characterized in that: Estimating a normal of each data point in the initial three-dimensional point cloud set according to a curvature value of each data point in the initial three-dimensional point cloud set to obtain a normal of each data point in the initial three-dimensional point cloud set, including: Determine the Euclidean distance between each neighborhood point in the neighborhood point set and the data point, and determine the distance weight between each neighborhood point in the neighborhood point set and the data point according to the Euclidean distance; Determining a weighted covariance matrix of the data points according to the distance weights, and performing singular value decomposition on the covariance matrix to obtain three second singular values; The eigenvector corresponding to the smallest second singular value among the three second singular values ​​is used as the normal vector of the data point, wherein the normal vector is the normal line.

4. The method for establishing a three-dimensional real scene model with an autonomous updating function according to claim 3, characterized in that: Dividing the initial three-dimensional point cloud set into a plane region subset and a non-plane region subset according to the curvature value and the normal of each data point in the initial three-dimensional point cloud set includes: Determining a region division threshold according to a curvature value of each data point in the initial three-dimensional point cloud set; If the curvature value of each data point in the initial three-dimensional point cloud set is less than the region division threshold, the data point is a plane region subset; otherwise, the data point is a non-plane region subset; An angle between a normal vector of the data point and a normal vector of the neighborhood point is determined; if the angle is less than a preset angle threshold, the neighborhood point and the neighborhood point belong to the same planar region subset or the same non-planar region subset.

5. The method for establishing a three-dimensional real scene model with an autonomous updating function according to claim 4, characterized in that: Performing point cloud mapping on the non-planar region subset by a principal component analysis method to obtain a two-dimensional projection image corresponding to the non-planar region subset further includes: Determining a principal component direction matrix based on the eigenvectors corresponding to the largest eigenvalue and the second largest eigenvalue among the three eigenvalues, projecting each data point into a two-dimensional space based on the principal component direction matrix, calculating the projection coordinates, and determining a two-dimensional point set based on all the projected data points; Normalizing the two-dimensional mapping data in the two-dimensional point set to obtain a normalized two-dimensional point set, smoothing the normalized two-dimensional point set according to kernel density estimation to obtain a data point density distribution corresponding to the normalized two-dimensional point set, and determining a two-dimensional distribution graph based on the data point density distribution, wherein the two-dimensional distribution graph is a two-dimensional projection image.

6. The method for establishing a three-dimensional real scene model with an autonomous updating function according to claim 5, characterized in that: Performing feature encoding on the contour image to obtain image features corresponding to the contour image includes: Determining the centroid of the contour image, performing edge detection on the contour image, extracting a contour point set of a shape, and randomly downsampling the contour point set to a preset number to obtain a new contour point set; Determine a Euclidean distance matrix and an angular distance matrix between the new contour point set and the centroid of the contour image, divide the 360° angle interval into a preset number of angle segments, and divide the distance interval [0, 1] into a preset number of distance segments; According to the Euclidean distance matrix and the angular distance matrix, the number of data points in each of the angle segments and each of the distance segments is counted to obtain a two-dimensional histogram of angle-distance, wherein the value of each grid of the two-dimensional histogram of angle-distance represents the number of data point pairs in the angle-distance area, and the two-dimensional histogram of angle-distance is the image feature corresponding to the contour image.

7. The method for establishing a three-dimensional real scene model with an autonomous updating function according to claim 6, characterized in that: Matching the image features with a preset target feature library to obtain the target features most similar to the image features, including: Determine the difference between the image feature and each target feature in a preset target feature library at each position; Accumulating the differences between the image feature and the target feature at all positions to obtain a matching cost coefficient between the image feature and the target feature; A target feature that is most similar to the image feature is determined according to the matching cost coefficient.

Images