Building facade structure line accurate extraction method based on laser point cloud feature depth

Abstract

The invention discloses a building facade structure line accurate extraction method based on laser point cloud feature depth, and is applied to the technical field of point cloud feature extraction. The method comprises the following steps: acquiring building point cloud data, preprocessing the acquired building point cloud data, and dividing the preprocessed building point cloud data into a training set and a test set; inputting the training set into a PCPNET network for training to obtain a PCPNET network training model, and extracting normal vector features of the test set through the PCPNET network training model; adopting a Gaussian weighting function to realize quadratic optimization of the normal vector features, and identifying edge feature points according to the difference of the normal vector features after quadratic optimization; and optimizing the edge feature points through an improved filtering algorithm, calculating normal vectors of the optimized edge feature points, and connecting the normal vectors into feature lines based on preset angle and distance thresholds. According to the method, the extraction precision and robustness of the normal vector features are remarkably improved, and a reliable data basis is provided for building facade structure analysis.

Description

Technical Field

[0001] The present invention relates to the technical field of point cloud feature extraction, and more specifically to a method for accurately extracting building facade structural lines based on laser point cloud feature depth. Background Art

[0002] At present, the commonly used methods in the field of feature extraction of point cloud models can be summarized into three categories: point-based feature extraction methods, surface-based feature extraction methods, and projection mapping-based feature extraction methods. 1) The point-based feature extraction method uses the basic properties of the point cloud, such as coordinates, color, reflection intensity and other information, and extracts the feature points or contour points of the point cloud as the building facade structure by calculating the normal vector, curvature, density, etc. For example, sharp edges are identified by normal vector estimation, and then closed by the minimum spanning tree to form feature segments, but this method is not robust enough to large-scale shape deformation. It is also possible to identify edge points with extreme curvature values ​​through curvature estimation, and connect these edge points with active contours that minimize energy to form closed contours, but the computational cost is high. Some scholars have also proposed a large-scale point cloud contour extraction algorithm based on supervoxel clustering analysis, which generates segmentation blocks through supervoxel clustering and extracts edge points using a-shape, and then uses the mutual nearest neighbor method to eliminate edge points between coplanar supervoxel blocks, but there are limitations in the extraction of fine edge points. 2) The surface-based feature extraction method analyzes the surface features of the building facade, and the intersection or relationship between the faces constitutes the final facade structure. For example, the feature line extraction is completed by relying on the dihedral angle between polygonal faces to judge the edge features, but the effectiveness for complex geometric shapes and topological structures is limited. Based on the tensor voting theory, the triangular mesh model feature edge extraction algorithm classifies vertices through the correspondence between the eigenvalue distribution of the tensor voting matrix and the vertex geometric features, and then uses the breakpoint connection method to distinguish the edge points and corner points on the smooth features. Finally, the boundary after region growth is extracted to obtain the mesh feature edge, but the effect is highly correlated with the parameters and has limitations in the face-to-face processing of buildings. 3) The feature extraction method based on projection mapping converts the complex calculation of three-dimensional space into two-dimensional space, and realizes the extraction of building facade structure through the projection mapping of two-dimensional images and three-dimensional point clouds. For example, some scholars have proposed a method for extracting straight line segments of large-scale buildings based on regional segmentation. Block point clouds are obtained through regional segmentation and projected to a two-dimensional plane. Line segments are extracted using the LSD line detection algorithm, and then back-projected to three-dimensional space, but the extracted line segments are messy. Other scholars use the Canny edge detection algorithm to realize edge feature extraction on two-dimensional images and three-dimensional point clouds, but the mapping mechanism has a great impact on accuracy. Some scholars have also proposed a method for precise registration of point clouds and panoramic images based on line features. The main body of the building is extracted through a two-dimensional straight line clustering algorithm, and the building facade structure is obtained by combining the two-dimensional orthophoto image and three-dimensional point cloud. However, errors are inevitably introduced in the conversion.

[0003] In summary, the current point cloud feature extraction methods have the following shortcomings: point-based methods are not robust enough to large-scale deformations, have high computational costs, and have limited extraction accuracy; surface-based methods have a large correlation between the effects and parameters, and are limited in applicability to complex geometric shapes; projection mapping-based methods lose information during the two-dimensional and three-dimensional conversion process, and the mapping mechanism has a great impact on accuracy. In summary, the current point cloud feature extraction methods have the following problems: insufficient robustness, high computational costs, limited extraction accuracy, information loss, and a great impact of the mapping mechanism on accuracy.

[0004] Therefore, it is an urgent problem for technical personnel in this field to propose a method for accurately extracting building facade structural lines based on laser point cloud feature depth to solve the difficulties existing in the prior art. Summary of the invention

[0005] In view of this, the present invention provides a method for accurately extracting building facade structural lines based on the feature depth of laser point cloud, which significantly improves the extraction accuracy and robustness of normal vector features and provides a reliable data basis for building facade structure analysis.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A method for accurately extracting building facade structure lines based on laser point cloud feature depth includes the following steps:

[0008] S1. Acquire building point cloud data, preprocess the acquired building point cloud data, and divide the preprocessed building point cloud data into a training set and a test set;

[0009] S2, inputting the training set into the PCPNET network for training to obtain a PCPNET network training model, and extracting the normal vector features of the test set through the PCPNET network training model;

[0010] S3, using Gaussian weight function to realize secondary optimization of normal vector features, and identifying edge feature points according to the difference of normal vector features after secondary optimization;

[0011] S4. Optimize edge feature points through an improved filtering algorithm, calculate normal vectors of the optimized edge feature points, and connect them into feature lines based on preset angle and distance thresholds.

[0012] The above method, optionally, S1-S2 specifically includes:

[0013] Firstly, the point cloud data of the building facade is obtained, the collected data is preprocessed and classified, the training set and the test set are determined, and they are input into the PCPNET network; the initial normal vector features of the training set are obtained, the neighborhood search function is used and a search radius of 0.03m is selected to estimate the normal vector of each point, and a variety of error analysis functions are used to calculate the normal vector prediction loss of the network to correct the prediction results, thereby realizing the construction of the sample library of the experimental data set.

[0014] The above method, optionally, S1-S2 further includes:

[0015] The local neighborhood of each point is identified and patched, the center point is transferred to the coordinate origin, and the scale is normalized; the input point cloud data is transformed into a standard posture through the spatial transformation network in the PCPNET network; the feature vector of the local area is processed through the functional network unit to obtain the normal vector estimation of the local point cloud; thus, the normal vector estimation result of the building facade is output.

[0016] The above method optionally uses multiple error analysis functions to calculate the normal vector prediction loss of the network. The multiple error analysis functions include: root mean square error, mean square minus cosine similarity loss and mean square Euclidean distance.

[0017] In the above method, optionally, the Gaussian weight function is used in S3 to implement secondary optimization of normal vector features, including:

[0018] (1) Check the matching between point cloud and normal vector features;

[0019] (2) Use Gaussian weight function to improve the network's estimation of normal vector features: For each point P in the point cloud i , calculate its K nearest neighbors {p j}, the normal vector n of each neighboring point j All are multiplied by the corresponding Gaussian weight; the Gaussian weight function w(d) assigns weights according to the Euclidean distance d between the point and its neighboring points;

[0020] (3) Using normal vector features to extract boundary points: First, calculate the target point P i The normal vector The normal vectors of other points in its local neighborhood The difference between them; then determine the boundary point: if for point P i There exists any adjacent point P ik The angle difference θ ik Exceeding the preset threshold value θ thresh , then P i are considered boundary points.

[0021] The above method may optionally include: determining the edge feature points obtained, denoising the edge feature points, and performing denoising and fitting using the following method:

[0022] A statistical outlier filter is used to process the noise in the point cloud; for each point P in the point cloud i , calculate the distance from a point to its k nearest neighbors Then calculate the average distance of all points The mean value μ and standard deviation σ are then used to define a threshold T to identify outliers. The threshold is calculated dynamically based on the mean and standard deviation. All points whose average distance exceeds this threshold are classified as outliers to filter out noise points.

[0023] The random sampling consistency algorithm is then used to fit the edge feature points. First, the smallest set of feature points is randomly selected to estimate the model parameters, and then the consistent set of internal points is evaluated. The optimal fitting model is obtained through continuous iteration.

[0024] The above method may further include: after optimizing the edge feature points, optimizing the feature lines after the merging process by the following method to obtain the building feature lines:

[0025] First, the point cloud is segmented into multiple blocks and feature line extraction is applied to each block;

[0026] Calculate the neighborhood features of the edge point cloud, calculate the angle between the two normal vectors by calculating the difference in the surface normal vector angle between two adjacent points in the point cloud; obtain the normal vectors of the two feature points P1 and P2 and Finally, calculate the dot product of the two normal vectors to determine whether they are connected;

[0027] Setting the angle threshold and distance connection threshold assists in generating feature lines. After calculating the normal vector angle between edge point clouds, the angle threshold judgment is added to determine whether two points belong to the same feature line. In three-dimensional space, the influence of spatial distance on line segment connection is considered, and the distance connection threshold judgment is added to accurately depict feature lines.

[0028] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a method for accurately extracting building facade structure lines based on laser point cloud feature depth, which has the following beneficial effects:

[0029] (1) This paper deeply explores the relationship between deep learning networks and building facade features, applies the PCPNET network to the extraction of building facade structures for the first time, and uses deep learning technology to accurately process building facade point cloud data. By optimizing the performance of the PCPNET network, the extraction accuracy and robustness of normal vector features are significantly improved, providing a reliable data basis for building facade structure analysis. This innovation has opened up a new research direction in the field of building facade structure extraction.

[0030] (2) This paper proposes an improved quadratic normal vector edge detection algorithm, which uses Gaussian weight function and weighted standard deviation allocation to smooth the normal vector data and accurately identify and extract the edge feature points of the building facade. This method significantly improves the extraction accuracy of edge feature points, and optimizes the extracted edge feature points through a filtering algorithm to remove noise points and ensure the stability and accuracy of the feature points.

[0031] (3) Based on the optimized normal vector features and edge feature points, the normal vector angle and Euclidean distance between the feature points are calculated, and the qualified feature points are connected using the set angle and distance thresholds to form the structural line segments of the building facade. This process effectively filters out outliers through the improved RANSAC filtering algorithm, ensuring the continuity and accuracy of the structural line. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0033] Figure 1 A flow chart of a method for accurately extracting building facade structure lines based on laser point cloud feature depth provided by the present invention;

[0034] Figure 2 A flow chart of extracting the feature of the building facade point cloud normal vector of the PCPNET network of the present invention;

[0035] Figure 3 It is the normal vector data after processing of the secondary normal vector, the unprocessed normal vector data and the interpolation distribution diagram;

[0036] Figure 4 A schematic diagram of feature line connection for calculating the normal vector angle between optimized edge feature points;

[0037] Figure 5a This is a display of BirdFountain1 street point cloud data with RGB information;

[0038] Figure 5b This is a diagram showing the intensity value information of BirdFountain1 street point cloud data;

[0039] Figure 6a A schematic diagram of the training data set (Data1) collected by 3D laser scanning of Jiuzhang Building in Sanjiang Campus of Jiangxi University of Science and Technology;

[0040] Figure 6b A schematic diagram of the training data set (Data2) collected by 3D laser scanning of the Yuedao Building in Sanjiang Campus of Jiangxi University of Science and Technology;

[0041] Figure 6c A schematic diagram of the training data set (Data3) collected by 3D laser scanning of Dushi Building in Sanjiang Campus of Jiangxi University of Science and Technology;

[0042] Figure 7a A schematic diagram of the test data set (Data4) collected by 3D laser scanning of the soil survey building of the Hongqi Campus of Jiangxi University of Science and Technology;

[0043] Figure 7b A schematic diagram of the test data set (Data5) collected by 3D laser scanning of Tuwenshan Building in Sanjiang Campus of Jiangxi University of Science and Technology;

[0044] Figure 7c Schematic diagram of selecting the test data set (Data6) of open source point cloud;

[0045] Figure 7d Schematic diagram of selecting the test data set (Data7) of open source point cloud;

[0046] Figure 8a The result image of edge point extraction using unprocessed normal vectors for Data6 data;

[0047] Figure 8b The result graph of edge points extracted after optimizing Data6 data;

[0048] Figure 9a The result image of edge point extraction using unprocessed normal vectors for Data5 data;

[0049] Figure 9b The result map of edge points extracted after optimizing Data5 data;

[0050] Fig.10a This is the result of using this patented method to extract the point cloud elevation data (Data4) of the soil survey building of the Hongqi Campus of Jiangxi University of Science and Technology;

[0051] Fig.10bThis is the result of using this patented method to extract the point cloud facade data (Data5) of Tuwenshan Building in Sanjiang Campus of Jiangxi University of Science and Technology;

[0052] Fig.10c This is the result of extracting the point cloud elevation of the test data set (Data6) of open source point cloud using this patented method;

[0053] Fig.10d This is the result of extracting the point cloud elevation of the open source point cloud test data set (Data7) using this patented method;

[0054] Fig.11a This is a display diagram of the feature line results of the left and right sides of the main entrance of Data4 for quantitative evaluation of its structural line extraction effect;

[0055] Fig.11b This is a display diagram of the feature line results of the left and right sides of the main entrance of Data5 for quantitative evaluation of its structural line extraction effect;

[0056] Fig.11c This is a display diagram showing the feature line results of the left and right sides of the main entrance of Data6 for quantitative evaluation of its structural line extraction effect. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. 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.

[0058] Reference Figure 1 As shown, the present invention discloses a method for accurately extracting building facade structure lines based on laser point cloud feature depth, comprising the following steps:

[0059] S1. Acquire building point cloud data, preprocess the acquired building point cloud data, and divide the preprocessed building point cloud data into a training set and a test set;

[0060] S2, inputting the training set into the PCPNET network for training to obtain a PCPNET network training model, and extracting the normal vector features of the test set through the PCPNET network training model;

[0061] S3, using Gaussian weight function to realize secondary optimization of normal vector features, and identifying edge feature points according to the difference of normal vector features after secondary optimization;

[0062] S4. Optimize edge feature points through an improved filtering algorithm, calculate normal vectors of the optimized edge feature points, and connect them into feature lines based on preset angle and distance thresholds.

[0063] Specifically, the method of the present invention includes: step 1, obtaining building point cloud data, using the PCPNET network training test data set to obtain a training model.

[0064] 1. Training the network. The present invention pre-processes the acquired building facade point cloud data, including denoising, point cloud alignment and normalization. The network flow chart is as follows: Figure 2 As shown. The preprocessed data will be input into the PCPNET network as training samples. The training process of the PCPNET network is as follows: 1) The point cloud data is denoised and the statistical filtering algorithm is used to remove isolated points and noise points. Then the point cloud is aligned and the multi-frame point cloud data is aligned by the ICP algorithm to ensure data consistency. Finally, the point cloud data is normalized to standardize the point cloud coordinates to a uniform scale. 2) A series of deep learning models are used to process these normalized local neighborhood information. Through a spatial transformation network (STN1), the input point cloud is converted to a standard posture. This step ensures the robustness and generalization ability of the model. Then, through a series of functional network units FNN1 and STN2, the feature vector g of the local area is acted on to obtain a 64×64 transformation matrix, and then the function is obtained by using FNN2. The derivation formula is shown in formula (1).

[0065] h l (p j )=(FNN2°STN2)(g1(p j ),...,g 64 (p j )) (1)

[0066] 3) In order to deal with the disorder of point cloud data, PCPNET uses a symmetric function to eliminate the uncertainty caused by the disorder of point cloud data by summing. This method not only optimizes the stability of the algorithm, but also improves its ability to recognize local features. The summation formula is shown in formula (2).

[0067]

[0068] After processing by a three-layer fully connected network, PCPNET is able to output the normal vector of the center point of the local point cloud, providing an accurate estimate of the local 3D shape characteristics in the point cloud. The application of this deep learning method, especially in estimating the main curvature direction of the point cloud, demonstrates the powerful ability and high flexibility of PCPNET in processing complex 3D data.

[0069] 4) When designing the normal vector prediction loss function of PCPNET, the root mean square error (RMSE) is a standard measure of prediction error, which measures the square root of the average of the squared values ​​of the difference between the predicted value and the actual observed value. In the prediction of the normal vector, RMSE can be used to quantify the degree of deviation between the predicted normal and the actual normal. Where N is the number of samples, is the predicted normal vector, n i is the corresponding true normal vector, and the formula is shown in formula (3).

[0070]

[0071] 5) Optimization algorithm: Use the back propagation algorithm and gradient descent method to optimize the network parameters, and gradually reduce the loss function value through multiple iterations to improve the accuracy and robustness of normal vector estimation. 6) Model training: Input the preprocessed point cloud data into the PCPNET network and perform multiple iterations of training until the loss function converges to obtain the final training model. Through the above steps, the trained PCPNET model can estimate the normal vector features of the building facade point cloud with high precision, providing a reliable basis for subsequent feature line extraction.

[0072] Step 2: Use the trained model to extract the normal vector features of the test data set. After training the PCPNET network model, the model needs to be used to extract the normal vector features of the test data set. The specific steps are as follows.

[0073] 1) Identification of local neighborhood and patch construction: For each point p in the point cloud, a local patch centered on the point is selected. This local patch includes all points within a certain range around the center point to capture local geometric shape information. Define the local patch P i For i The set of all points within a fixed radius r, ||pp i || represents the Euclidean distance between points, and the formula is shown in formula (4).

[0074] P i ={p∈P|||pp i ||≤r} (4)

[0075] 2) Local patch preprocessing. The preprocessing step for each local patch includes translating the patch to the origin and normalizing the scale. The purpose is to eliminate the position and scale differences of the patch in space, so that the neural network can focus on learning the local characteristics of the geometric shape. This preprocessing process performs translation and scaling operations on each patch to ensure that the center of the patch is at the origin and all points in the patch are in a standardized neighborhood. For each point p in the patch i , p′i Represents the coordinates of the point after translation. Then determine the distance D from the farthest point in the patch to the origin max , and scale each point in the patch by this distance so that the maximum distance of all points in the patch does not exceed the predetermined standard radius R std , thereby determining the scaling in scale normalization, the formula is shown in formula (5).

[0076]

[0077] 3) Learn the feature representation of each local patch through a deep neural network. This process involves learning a set of complex nonlinear functions from the patch to capture the local structural information inside the patch. Define a function f in the network, which is responsible for converting the normalized patch P′ i Mapped to an n-dimensional feature vector, the formula is shown in formula (6).

[0078] f:P′ i →R n (6)

[0079] Among them, P′ i represents the normalized local patch, R n is the space where the feature vector is located, and n is the dimension of the feature vector. The normalization process usually involves translating the points within the patch relative to the center point P, and possibly scaling it to ensure that the model is invariant to the size and position of the patch.

[0080] 4) Normal vector regression. Using the learned local patch features to predict the normal vector of the center point is done by mapping the local feature vector to the normal vector space. This mapping process usually relies on fully connected layers (or dense layers). These layers can process the learned feature vectors and convert them into corresponding normal vectors, that is, mapping from high-dimensional feature space to normal vector space. Using a regression model g:R n →R 3 , maps the eigenvector to the normal vector n of the center point i The formula is shown in formula (7).

[0081] n i =g(f(P′ i )) (7)

[0082] Among them, g is a regression function, which consists of fully connected layers, and its purpose is to learn how to predict the normal vector from the input feature vector. The training process of this regression model involves minimizing the difference between the predicted normal vector and the actual normal vector, which is usually achieved by defining a loss function (such as mean squared error).

[0083] Through the above process, PCPNET can accurately extract local 3D shape attributes from point cloud data, use deep learning methods to handle complex geometric structures and noise, and demonstrate its high efficiency and accuracy in local shape analysis.

[0084] Step 3: Introduce Gaussian weight function to realize secondary optimization of normal vector features. Gaussian weight function is used to improve the estimation results. The normal vector estimation effect is as follows: Figure 3 As shown in Figure 8. The Gaussian weight function w(d) assigns weights based on the Euclidean distance d between a point and its neighboring points. Through this mechanism, the influence of neighboring points on the normal vector of the target point is weighted according to the distance between them, thereby achieving a distance-sensitive smoothing effect. The closer the point, the greater the influence on the normal vector, which helps to reduce the influence of noise points far away from the query point. The formula is shown in Equation (8).

[0085]

[0086] Here, σ is the standard deviation of the Gaussian function, which controls how fast the weight changes with distance. In this way, the contribution of each point's neighboring points to the point's normal vector is weighted according to the distance between them, and the closer the neighboring points are, the greater their influence. This weighting strategy can effectively suppress the influence of noise points far away from the query point.

[0087] During the optimization process, the algorithm maintains local details while reducing the influence of outliers, thereby achieving weighted smoothing of normal vectors. i , calculate its K nearest neighbors {p j}, the normal vector n of each neighboring point j They are all multiplied by the corresponding Gaussian weights, as shown in formula (9).

[0088]

[0089] where d ij It's point p i With neighboring point p j The distance between them, W is the sum of all weights, The smoothed normal vector The formula is shown in formula (10).

[0090]

[0091] According to the number of neighborhood points of the building facade point cloud and the standard deviation of the Gaussian function in the secondary optimization of the normal vector, the following is explained: 1) The number of neighborhood points k refers to the number of nearest neighbor points selected for each point when performing Gaussian weighted smoothing. This parameter has a significant impact on the accuracy and robustness of normal vector estimation. Smaller k values ​​may make the normal vector estimation susceptible to local noise, while larger k values ​​may lead to over-smoothing, thereby ignoring the subtle structure of the point cloud data. Therefore, the selection of the k value needs to be appropriately adjusted according to the density and feature details of the point cloud data to balance the relationship between noise suppression and detail retention. In the implementation process of this patent, the number of neighborhood points k is set to 10.

[0092] 2) The standard deviation σ of the Gaussian function determines the range and degree of influence of the neighborhood points on the calculation of the target point normal vector when performing Gaussian smoothing of the normal vector. The setting of the σ parameter reflects that when calculating the weighted average normal vector, points closer to the target point are given higher weights. The choice of the σ value is crucial to the performance of the algorithm, because when the σ value is large, points far away from the target point will also have a greater impact on the average normal vector, which may lead to over-smoothing; conversely, when the σ value is small, the neighborhood points near the target point are mainly considered, which can better preserve local details, but may increase the impact of noise. Therefore, the setting of σ needs to comprehensively consider the characteristics of the point cloud data and the degree of processing required. In the implementation process of this patent, the number of neighborhood points σ is set to 0.5.

[0093] Step 4: Use the normal vector feature difference after secondary optimization to identify edge feature points. Once the normal vector of the building facade point cloud is calculated using the PCPNET network, the boundary points can be identified by analyzing the spatial changes of the normal vector. Specifically, points at the edge of an object usually show obvious changes in the normal vector direction because the surrounding neighborhoods of these boundary points have significant differences in the internal and external surface characteristics. Through a detailed analysis of the spatial distribution of the normal vector and its changes, the boundary points in the point cloud data can be effectively identified and extracted, providing key information for subsequent data analysis and model reconstruction.

[0094] 1) Normal vector difference analysis. Calculate point P i The normal vector The normal vectors of other points in its local neighborhood The difference between the normal vectors can be measured using the angle θ between the normal vectors. ik To calculate, the formula is shown in formula (11).

[0095]

[0096] 2) Boundary point determination. If for point P i There exists any adjacent point P ik The angle difference θ ikExceeding the preset threshold value θ thresh , then P i is considered as a boundary point by reasonably choosing the neighborhood size K or radius r, and the threshold θ thresh , the algorithm can be flexibly applied to various point cloud data. In the implementation process of this patent, the number of neighborhood points K for boundary estimation is set to 100.

[0097] Step five, use the improved RANSAC algorithm to optimize the edge feature points. This patent adds a statistical outlier filter algorithm based on the RANSAC algorithm. 1) The statistical outlier filter (SOR) is a method used to remove noise from a data set in point cloud processing. It works by analyzing the neighborhood of each point and removing points that do not meet the conditions based on statistical criteria. The SOR filter assumes that the distribution of the average distance from each point to its neighboring points is a Gaussian distribution, and then calculates the mean and standard deviation of these distances. All points whose average distance exceeds the interval defined by the mean and standard deviation of the global distance are considered outliers and are removed from the data set. The statistical outlier filter (SOR) achieves denoising in point cloud processing through the following mathematical formula:

[0098] For each point P in the point cloud i , calculate its k nearest neighbor points The formula is shown in formula (12).

[0099]

[0100] Then calculate the average distance of all points The mean value μ and standard deviation σ of , see formula (13).

[0101]

[0102] Then a threshold T is defined to identify outliers. The threshold is calculated dynamically based on the mean and standard deviation. All points whose average distance exceeds this threshold are classified as outliers. T = μ + α·σ, where α is a user-defined standard deviation multiplier used to control the strictness of the filter. For all points P i ,if Then P i It is considered as an outlier and removed from the data set. In the implementation process of this patent, the threshold T of the filtering algorithm is set to 10.

[0103] 2) Random Sampling Consensus Algorithm (RANSAC) is used to identify model parameters in data with a large number of outliers. It is suitable for model parameter estimation in the fields of computer vision and machine learning. Given a data set and model parameterization, for each data point, its distance to the model is calculated. If the distance is less than a preset threshold, the point is considered an inlier, and the number of inliers is defined as the evaluation value of the model. In mathematical representation, suppose there is a data set D = {d1, d2, ..., d n} and a model M parameterized by θ. So that the data points consistent with M(θ) are maximized, for each data d i , calculate its distance dist(d i ,M(θ)). If the distance is less than the preset threshold T, then d i The number of inliers is defined as the evaluation value of the model M(θ). The estimated parameter θ of the random sampling consistency algorithm in the implementation process of this patent is set to 8000. The specific formula is shown in formula (14).

[0104] score(M(θ))=|{d i ∈D:dist(d i ,M(θ))<T}| (14)

[0105] The ultimate goal of RANSAC is to maximize the evaluation value score(M(θ)). In the application process, the choice of parameter threshold will affect the identification of inliers, and the more iterations, the greater the probability of finding the optimal model. The RANSAC algorithm, with its simplicity and robustness, shows significant advantages in dealing with data fitting problems containing a high proportion of outliers.

[0106] Step six, calculate the normal vector of the feature point and connect it into feature lines based on the angle and distance thresholds. After extracting and optimizing the edge feature points of the building facade point cloud, the next step is to calculate the normal vector of the edge feature points and connect these feature points into feature lines based on the angle and distance thresholds. First, calculate the normal vector of the feature point: input the preprocessed test data set into the trained PCPNET network model, and calculate the normal vector features of each point through the model. The normal vector represents the local geometric features of each point and can effectively reflect the structural information of the point cloud data. Then connect the feature points based on the angle and distance thresholds: After calculating the normal vector of each point, use the set angle and distance thresholds to connect the feature points into feature lines. The specific steps are as follows:

[0107] 1) Calculate the normal vector angle: For each pair of adjacent points in the point cloud data, calculate their normal vector angle. If the normal vector angle of two points is less than the preset angle threshold, it is considered that the normal vector directions of the two points are similar and can be used as candidate points for connection. Where θ is the angle between the two vectors, and is the modulus of the vector. In the implementation process of this patent, the angle threshold is set to 1, and the modulus length calculation formula using clicks and vectors is shown in (15).

[0108]

[0109] 2) Calculate the Euclidean distance: For each pair of adjacent points, calculate the Euclidean distance between them. If the Euclidean distance between two points is less than the preset distance threshold, the two points are considered to be close in space and can be used as candidate points for connection. ij represents the Euclidean distance between point i and point j, (x i ,y i ,z i ) and (x j ,y j ,z j ) represent the coordinates of point i and point j respectively. In the implementation process of this patent, the distance threshold is set to 0.35, and its specific formula is shown in (16).

[0110]

[0111] 3) Connect feature points: If the normal vector angle and Euclidean distance of two points meet the preset angle and distance thresholds, the two points are connected into a feature line, as shown in the schematic diagram Figure 4 As shown. Where L ij Indicates whether point i and point j are connected, θ thresh Expressed as the angle threshold, d thresh is the distance threshold, when L ij When it is 1, it means that point i and point j are connected, and L ij When it is 0, it means no connection. The formula is shown in (17).

[0112]

[0113] 4) Output feature lines: Finally, the connected feature lines are output for subsequent building facade structure analysis and 3D reconstruction. By connecting feature points based on angle and distance thresholds, the continuity and accuracy of feature lines can be guaranteed, significantly improving the quality of building facade feature line extraction.

[0114] In a specific embodiment, a simulation test is carried out, and Test 1 is used to verify the effectiveness of the entire method.

[0115] Test 1:

[0116] 1. Test platform and data

[0117] This experiment and this patent are all tested using ASUS Z690 computer (memory: 32GB, Intel(R) Core(TM) I7-12700KCPU@3.60GHz, NVIDIA GeForce GTX 3080) for case invention testing. Three sets of measured point cloud data were selected as training data, and the public point cloud data selected from the Semantic public data set ( Figure 5a This is a display of BirdFountain1 street point cloud data with RGB information. Figure 5b Data 1-3 is a display diagram of the intensity value information of the BirdFountain1 street point cloud data) and two sets of measured point cloud data to verify the effectiveness of the method of the present invention. Data 1-3 is the point cloud data of the facade of a building collected on a campus using a Rieg1VZ-1000 ground laser scanner, which is used as a training data set to verify the effectiveness of the PCPNET network training on the point cloud normal vector features. Figure 6a-6c Data 4-5 is a point cloud data of a building facade collected on a campus using a Rieg1VZ-1000 ground laser scanner to verify the feasibility of extracting building facade feature lines using the method of the present invention. Figure 7a-Figure 7b Data6-7 is a data set selected from Semantic, which is used to verify the feasibility of extracting building facade feature lines by the method of the present invention. Figure 7c-7d The detailed information of the four groups of experimental data is shown in Table 1.

[0118] Table 1

[0119]

[0120] The parameters involved in the method of the present invention are mainly distributed in three parts: feature point extraction, feature point optimization and feature line generation. The feature point extraction part mainly includes three parameters: the standard of the Gaussian function σ, the number of neighboring points k and the K value of the boundary estimation. The feature point optimization part mainly includes two parameters: the random sampling consistency parameter θ and the filtering threshold T. The feature line generation considers two parameters: the angle threshold α and the distance threshold β. Different parameters will produce different feature line results. The parameters used in this article are for building facades with many corners and a large amount of data. When facing relatively simple building facade data or low data accuracy, the relevant parameters can be reduced. Table 2 shows the parameter settings of the corresponding parts of the feature point extraction of the four groups of experimental data in this experiment.

[0121] Table 2

[0122]

[0123] 2. Feature result extraction and analysis

[0124] 1) Feature point extraction analysis: In order to evaluate the advantages of the secondary normal vector optimization used in the feature point extraction stage to extract edge feature points, experiments were conducted on the Data6 data using unprocessed normal vectors to extract edge points and optimized edge points to extract edge points. The comparison results are as follows: Figure 8a shows the unprocessed normal vector results, Figure 8b The feature point extraction result after the secondary normal vector is successfully captured, such as the location of water pipes in the original building point cloud, and the accurate mapping of the facade structure is achieved, and the effect of extracting complex corner structures is optimized. This set of comparative results fully verifies the work of this study in normal vector processing, not only proving the robustness and accuracy of the proposed method, but also showing significant advantages in improving the detail level.

[0125] 2) Feature point optimization analysis: In order to evaluate the advantages of the improved filtering algorithm used in the feature point optimization stage to screen edge feature points, experiments were conducted on Data5 data using unprocessed normal vectors to extract edge points and after optimization. The comparison results are shown in Figure 2. Figure 9a , Figure 9b The results show that when the neighborhood point parameter is 10, a total of 1923 noise points are effectively filtered out, reaching a filtering rate of 0.037, which greatly reduces the noise points and successfully avoids the misscreening of key line segment structures, thereby obtaining clearer window facade structure information.

[0126] 3) Feature line extraction analysis: The feature line results obtained by using this patent are as follows Figure 10a-Figure 10d As shown, it can be found that the extraction method used in this patent has the following advantages. 1. Structural richness and accuracy. Accurately identify the structural elements of the building facade, the extracted structure is rich and accurate, and the spatial distribution and normal information of the point cloud are comprehensively considered. 2. Continuity and integrity of the structural lines. The structural lines are continuous and complete, especially suitable for complex building facades, which improves the intuitiveness and practicality of model reconstruction. 3. Anti-interference ability. Effectively avoid non-structural line interference, reduce misidentification, and ensure the high robustness of the results through in-depth analysis.

[0127] In order to correctly evaluate the accuracy of building facade feature line extraction, this patent uses accuracy, precision and recall to calculate the evaluation indicators. Accuracy measures the proportion of correctly identified parts of the extracted feature lines to the total number of extracted features, reflecting the overall correctness of the model. The calculation formula for accuracy is shown in (18).

[0128]

[0129] Precision is an indicator to measure the performance of a classification or retrieval system, and is defined as the ratio of the number of correctly retrieved (or classified) positive examples to the total number of retrieved (or classified) positive examples. The formula for calculating precision is shown in (19).

[0130]

[0131] The recall rate measures the proportion of feature lines correctly identified by the model to the total number of feature lines that actually exist, reflecting the model's ability to capture the structural features of the facade. The calculation formula for the recall rate is shown in (20).

[0132]

[0133] In the above formula: TP represents the number of correctly extracted straight line segments; FP represents the number of incorrectly extracted straight line segments; FN represents the number of unextracted straight line segments. With the help of the above evaluation indicators, the accuracy and completeness of the extraction algorithm can be comprehensively evaluated, providing guidance for the subsequent accuracy of building facade structure extraction.

[0134] In order to quantitatively evaluate the effect of building facade structure extraction, the constructed building facade structure evaluation system is used to evaluate the extraction results of the three test data. The excerpts are the areas on the left and right sides of the main entrance of Data4, numbered 1 and 2 from left to right; the area selected from Data5 is numbered 3; the area selected from the public dataset Data6 is numbered 4, and the excerpts are as follows Fig.11a - Fig.11c This patent calculates the accuracy, recall and discontinuity of the selected measurement area respectively, and then calculates the average value as the overall evaluation result. The evaluation index of the building facade structure extraction result is recorded in Table 3.

[0135] Table 3

[0136]

[0137] The evaluation indexes of building facade structure extraction results were statistically analyzed, and the comprehensive average values ​​showed an accuracy of 0.953, a precision of 0.952, and a recall of about 0.956. This means that among the extracted structural segments, the correct identification of the extracted feature lines accounted for 95.3% of the total number of extracted segments, the correct structural segments accounted for 95.2%, and 95.6% of the structural lines of the entire facade were extracted. This patented method showed high accuracy and completeness in extracting building facade structures, providing a reliable basis for the extraction of building facade structures.

[0138] In summary, the present invention proposes a method based on deep learning for accurately extracting the structural features of building facades. The building facade point cloud data provides rich structural information, which is of great significance for the fine modeling of buildings, the restoration of ancient buildings and deformation monitoring. This patent proposes a building facade structure extraction method based on the PCPNET network by summarizing and analyzing the existing methods, achieving the goal of automatically identifying and accurately extracting key geometric features from a large amount of point cloud data, and significantly improving the processing speed and accuracy. First, this patent uses the PCPNET network to realize the point cloud normal vector feature calculation, and accurately estimates the normal vector of the input facade point cloud data by constructing training and test data sets and using the convolution computing power of the network. Experimental results show that the PCPNET network shows superiority in processing complex point cloud data, providing a solid data foundation for subsequent facade structure analysis. Secondly, an improved quadratic normal vector edge detection algorithm is proposed, which smoothes the normal vector data and distributes the weighted standard deviation, significantly improving the recognition accuracy of the building facade detail features and the accuracy of edge detection. This method not only reduces the demand for computing resources and time, but also optimizes the average value of the normal vector, providing a more effective and accurate means for processing complex building facade point cloud data. Finally, the facade contour feature line extraction method based on local normal similarity and distance threshold effectively filters out outliers and accurately connects feature points through an improved random sampling consensus (RANSAC) filtering algorithm, thereby extracting the precise contour line segments of the building facade. The experimental results show that the algorithm has significantly improved processing efficiency and contour extraction accuracy. In addition, this patent conducted experiments on different building facade point cloud data sets, and compared and analyzed the proposed method with existing methods through quantitative evaluation indicators such as accuracy, precision and recall. The results show that the method of this patent is significantly superior to existing methods in terms of accuracy, robustness and processing efficiency, further verifying the accuracy and superiority of this method.

[0139] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0140] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for accurately extracting building facade structure lines based on laser point cloud feature depth, characterized in that: The following steps are involved: S1. Acquire building point cloud data, preprocess the acquired building point cloud data, and divide the preprocessed building point cloud data into a training set and a test set; S2, inputting the training set into the PCPNET network for training to obtain a PCPNET network training model, and extracting the normal vector features of the test set through the PCPNET network training model; S3, using Gaussian weight function to realize secondary optimization of normal vector features, and identifying edge feature points according to the difference of normal vector features after secondary optimization; S4. Optimize edge feature points through an improved filtering algorithm, calculate normal vectors of the optimized edge feature points, and connect them into feature lines based on preset angle and distance thresholds.

2. The method for accurately extracting building facade structure lines based on laser point cloud feature depth according to claim 1, characterized in that: S1-S2 specifically includes: Firstly, the point cloud data of the building facade is obtained, the collected data is preprocessed and classified, the training set and the test set are determined, and they are input into the PCPNET network; the initial normal vector features of the training set are obtained, the neighborhood search function is used and a search radius of 0.03m is selected to estimate the normal vector of each point, and a variety of error analysis functions are used to calculate the normal vector prediction loss of the network to correct the prediction results, thereby realizing the construction of the sample library of the experimental data set.

3. The method for accurately extracting building facade structure lines based on laser point cloud feature depth according to claim 2, characterized in that: S1-S2 also includes: The local neighborhood of each point is identified and patched, the center point is transferred to the coordinate origin, and the scale is normalized; the input point cloud data is transformed into a standard posture through the spatial transformation network in the PCPNET network; the feature vector of the local area is processed through the functional network unit to obtain the normal vector estimation of the local point cloud; thus, the normal vector estimation result of the current building facade is output.

4. The method for accurately extracting building facade structure lines based on laser point cloud feature depth according to claim 2, characterized in that: A variety of error analysis functions are used to calculate the normal vector prediction loss of the network, including root mean square error, mean square minus cosine similarity loss, and mean square Euclidean distance.

5. The method for accurately extracting building facade structure lines based on laser point cloud feature depth according to claim 1, characterized in that: The secondary optimization of normal vector features using Gaussian weight function in S3 includes: (1) Check the matching between point cloud and normal vector features; (2) Use Gaussian weight function to improve the network's estimation of normal vector features: For each point P in the point cloud i , calculate the average normal vector between its K nearest neighbors {pj}, and the normal vector n of each neighborhood point j All are multiplied by the corresponding Gaussian weight; the Gaussian weight function w(d) assigns weights according to the Euclidean distance d between the point and its neighboring points; (3) Using normal vector features to extract boundary points: First, calculate the target point P i The normal vector The normal vectors of other points in its local neighborhood The difference between them; then determine the boundary point: if for point P i There exists any adjacent point P ik The angle difference θ ik Exceeding the preset threshold value θ thresh , then P i are considered boundary points.

6. The method for accurately extracting building facade structure lines based on laser point cloud feature depth according to claim 1, characterized in that: S4 includes: determining the edge feature points obtained, denoising the edge feature points, and performing denoising and fitting using the following method: A statistical outlier filter is used to process the noise in the point cloud; for each point P in the point cloud i , calculate the distance from a point to its k nearest neighbors Then calculate the average distance of all points The mean value μ and standard deviation σ are then used to define a threshold T to identify outliers. The threshold is calculated dynamically based on the mean and standard deviation. All points whose average distance exceeds this threshold are classified as outliers to filter out noise points. The random sampling consistency algorithm is then used to fit the edge feature points. First, the smallest set of feature points is randomly selected to estimate the model parameters, and then the consistent set of internal points is evaluated. The optimal fitting model is obtained through continuous iteration.

7. The method for accurately extracting building facade structure lines based on laser point cloud feature depth according to claim 6, characterized in that: Also includes: After optimizing the edge feature points, the following method is used to optimize the feature lines after merging to obtain the building feature lines: First, the point cloud is segmented into multiple blocks and feature line extraction is applied to each block; Calculate the neighborhood features of the edge point cloud, calculate the angle between the two normal vectors by calculating the difference in the surface normal vector angle between two adjacent points in the point cloud; obtain the normal vectors of the two feature points P1 and P2 and Finally, calculate the dot product of the two normal vectors to determine whether they are connected; Setting the angle threshold and distance connection threshold assists in generating feature lines. After calculating the normal vector angle between edge point clouds, the angle threshold judgment is added to determine whether two points belong to the same feature line. In three-dimensional space, the influence of spatial distance on line segment connection is considered, and the distance connection threshold judgment is added to accurately depict feature lines.

Images