Lightweight modeling method for building geometric model

Abstract

The invention discloses a lightweight modeling method for a building geometric model, and the method comprises the following steps: S1, employing a Point2Contour model, and carrying out the regression of a building contour point set and a building corner point set from a laser radar building point cloud set; s2, reconstructing a building contour line based on particle filtering, and constructing a lightweight geometric model based on the posterior probability of maximizing the global building contour line; and S3, comprehensively evaluating the lightweight geometric model from four dimensions of geometry, topology, semantics and lightweight. According to the method, the large-scene LoD3 building contour line model which is visual in structure, flexible in expression and light in weight can be constructed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of laser radar remote sensing real scene three-dimensional geometry reconstruction, and particularly relates to a building geometry model lightweight modeling method. BACKGROUND

[0002] According to the degree of dependence on laser radar point cloud, the modeling method can be divided into data-driven, model-driven and hybrid-driven; according to the difference of modeling methodology, it can be divided into boundary expression modeling, prior assumption modeling, dimension reduction modeling and nonlinear modeling, etc.; according to the intelligent degree of modeling algorithm, it can be divided into full-automatic modeling and man-machine interactive modeling; according to the model expression method, it can be divided into parameter modeling, surface modeling and volume modeling; according to the representation form of the modeling method, it can be divided into explicit modeling and implicit modeling. The above methods try to construct a three-dimensional building surface model or a three-dimensional building entity geometry model with an outer surface being a "two-dimensional manifold" embedded in a three-dimensional Euclidean space.

[0003] However, due to the differences in accuracy, density and dimension of laser radar point cloud, the complexity of building entity structure, the degree of human-computer interaction, the rationality of prior assumptions and the expression method of model data structure, the three-dimensional building surface model or entity geometry model constructed by current scholars still has many problems: 1) Over-reliance on prior constraints: the algorithm can only model specific style buildings, such as Manhattan world assumption, Atlanta world assumption, polyhedral geometry assumption, building facade segmented continuous assumption in vertical direction, etc., which greatly limits the adaptability and scalability of the algorithm in large scene (contains billions of point clouds, covers nearly 100 square kilometers of large-scale space) modeling. 2) Single model expression form, lack of flexibility: current research mainly constructs three-dimensional entity geometry model that meets the manifold constraint, but due to the strict topological restriction, it is easy to cause potential topological connection error, so that the flexibility of "manifold entity geometry" expression is limited in large-scale complex building modeling. 3) Low lightweight degree: in implicit representation, voxel field, symbolic distance function, neural implicit representation, etc. are usually used to store information. Although implicit representation is relatively compact in model expression, the calculation and storage overhead is large, which affects the lightweight degree of the algorithm; in explicit representation, due to excessive emphasis on geometric accuracy, the hierarchical details of the building model are too high, making the building entity geometry model bloated, containing a large number of redundant triangular facets, which seriously hinders the lightweight storage, progressive transmission and real-time rendering of large-scale three-dimensional building model. Lack of semantic expression: existing modeling methods pay too much attention to the geometric features of the model, often ignoring the structure, primitive, component semantics or other additional attributes of the model, which strictly limits the advanced calculation and analysis of the model in downstream applications. Therefore, improving the adaptability of automated modeling to different building styles, realizing multi-level detail expression of large-scale three-dimensional building model, optimizing the limitations of "manifold geometry" expression, enhancing the semantic expression ability of the model, and improving the adaptability of large-scale three-dimensional building model to downstream applications have become the core problems that need to be broken through in the field of large-scale three-dimensional building point cloud modeling in the current academic circle. SUMMARY

[0004] The purpose of the application is to provide a building geometry model lightweight modeling method, which can improve the real-time rendering efficiency, lightweight storage capacity of large scene building model, and enhance the computational and analytical ability of downstream applications.

[0005] Technical scheme: A building geometry model lightweight modeling method, comprising the following steps: S1, using Point2Contour model, regressing building contour point set and building corner point set from laser radar building point cloud set; S2, reconstructing the building contour line based on "particle filtering", and constructing a lightweight geometry model based on maximizing the posterior probability of the global building contour line; S3, comprehensive evaluation of lightweight geometric model from four dimensions of geometry, topology, semantics and lightweight.

[0006] Further, the Point2Contour model is used to realize the building contour point and building corner point double regression task. The implementation process of the building contour point regression task is as follows: taking the building point cloud as input, using the FPS algorithm to downsample the point cloud and extract the center point; then selecting the neighborhood points around each center point by the KNN algorithm, dividing the point cloud into several local blocks, and converting each block of point cloud to the relative coordinate system with the center point as the origin; Among them, random sampling is performed on the point cloud block, and part of the point cloud is marked as a "mask block", and the rest is marked as a "visible block"; for the "mask block", the coordinates of the mask part are embedded as a learnable mask label; for the "visible block", a lightweight PointNet composed of multiple layers of perceptron and max pooling layer is used for local feature embedding to generate a visible label; For each center point downsampled by the FPS algorithm, position embedding is performed to generate a position label, and the center point and the mask label or the visible label of the corresponding point cloud block are fused to form a low-dimensional embedding feature; then, the low-dimensional feature is input into the TF-1 model for encoding, and 6 self-attention heads are used in the TF-1 model, each outputting a 64-dimensional feature, so that all point cloud blocks are encoded into 384-dimensional feature vectors, and the encoding features are used to reconstruct the "mask block"; when predicting the contour point, all point cloud blocks are regarded as "visible blocks" for embedding and encoding, and further representing the features of each contour point, and the contour point is predicted by using the binary classification detection head Detector-1. The implementation process of the building corner point regression task is as follows: FPS sampling and blocking are performed on the building point cloud, and then random masking is performed on the point cloud block containing the contour point to mark it as a "mask block", and the remaining blocks are "visible blocks"; for the "mask block", embedding is performed to obtain a learnable mask label; for the "visible block", a lightweight PointNet is used again to embed a visible label; for the center point obtained by downsampling, embedding is performed to obtain a position label, and the center point and the corresponding mask label or visible label are fused; then, the low-dimensional feature is encoded by using the TF-2 model, and the encoding feature is used to reconstruct the "mask block"; when predicting the building corner point, all point cloud blocks are regarded as "visible blocks" for embedding and encoding, and further representing the features of each contour point, and the building corner point is predicted by using the binary classification detection head Detector-2, and the optimized "anchor corner point" is generated through clustering, and finally the corner point offset correction is combined to regress the corner point coordinates.

[0007] Further, the total loss function L for training the Point2Contour model includes a self-supervised pre-training loss L self , a contour point loss L edge , and a corner point loss L corner , where L = L self + L edge + L corner . The self-supervised pre-training loss L self uses a “fillet distance” to design the self-supervised pre-training loss, L self The expression of the loss function is as follows: , where a represents a point from a predicted mask point set , and b represents a point from an original point cloud in a mask region; represents the Euclidean distance between a and b, represents the number of elements in the predicted mask point set, represents the number of elements in the original mask point set; The contour point loss L edge includes a binary cross-entropy of contour points, a distance error between a predicted contour point and a nearest contour line, and a consistency between a direction of the predicted contour point and a direction of the nearest contour line, L edge The expression of the loss function is as follows: , where CE(·) represents a cross-entropy of two variables, and MSE(·) represents a mean square error of two points; represents a laser radar building point, represents a nearest contour line point; represents a building point cloud set, and “|·|” represents an absolute value of a dot product of two vectors; represents a predicted point, represents a probability that a predicted point is a real contour point, represents a total number of building points; represents a consistency between a direction vector of a predicted point and a direction of a nearest contour line of the point, and α and β are weight coefficients, respectively; The corner point loss L cornerThe loss function includes a "candidate corner point" detection loss and an "anchor corner point" coordinate offset loss. After the point features are initially predicted by the detection head Detector-2, "candidate corner points" that are gathered near the real corner points are obtained, and the cross entropy of the "candidate corner points" and the real corner points is calculated at this time. Subsequently, the "candidate corner points" are clustered to obtain "anchor corner points", and the "anchor corner points" are corrected for coordinate offset to obtain a "precise corner point" set C. The mean square error between the "precise corner points" and the corresponding real corner points is calculated to optimize the corner point offset correction module; L corner The expression of the loss function is: , wherein, represents the probability of whether the predicted contour point is a corner point, represents the probability of the predicted contour point being a real corner point; represents a "precise corner point", represents the nearest real corner point to . , and respectively represent the number of elements in the predicted building contour point set B and the building corner point set C.

[0008] Further, the implementation steps of reconstructing the building contour line are as follows: S21, the Point2Contour model is used to regress to obtain a building contour point set and a building corner point set, and a KNN graph is constructed. All corner points and part of the contour points are selected to form an initial seed point set S; S22, for each seed point in the initial seed point set S, the KNN graph structure is expanded based on the KNN graph structure, and the building contour line is tracked based on "particle filtering". After m times of expansion, a series of contour path sets are obtained ; when the number of contour paths N = |S|, the contour path approximation is sampled to obtain the building precise contour line; S23, a non-maximum suppression algorithm is used to remove redundant or low-weight contour edges, and a global maximum contour edge posterior probability is used for optimization to generate a global building contour line; S24, pruning and simplification operations are adopted to construct a lightweight geometric model of the building contour line.

[0009] Further, the implementation steps of tracking the building contour line based on "particle filtering" include: S221, the particle filtering building contour line tracking problem is modeled; The contour path of each particle is composed of a state sequence , wherein represents the initial state of the particle in the contour path, represents the last state; for the seed point The state variable corresponding to the sampling time t of its contour path Defined as Seed point The coordinates are ; ,express Direction vector; according to Bayes' theorem, state variable The posterior probability estimate is: , in, Represents state variables Given a sequence of observed states The posterior probability, This represents the prior probability of a state transition. This represents the corresponding observation value after a given state. The observed likelihood probability; By using the weights of a series of particles to approximate the posterior probability of a certain state, the above formula is transformed into:

[0010] in, This represents the state variable sampled by the i-th particle at time t. Importance weights δ(.) is the Dirac function, and satisfies ; The problem of tracing building outlines is transformed into calculating the weight of particles at each state time step. The definition is as follows: , In the formula, the state of the i-th particle is sampled at time t. The weights are determined by the state transition probability of the particle from time t-1 to time t. And the likelihood observation of the particle at that moment probability; S222, predicts particle positions and updates particle weights; In the prediction step, given the state variables of the i-th particle at time t-1... Then the state variables of the contour line particles tracked at time t The definition is as follows: , Where k is the adjustment coefficient. State vector Corresponding particle point directional vector, For particle points in the KNN graph K nearest neighbors of K; In the update step, the depth feature vector set corresponding to the building contour point set B obtained by regression of the Point2Contour model , the depth feature vector set corresponding to the building corner point set C are all regarded as observation values of state vectors, and the observation values are subject to Gaussian distribution respectively; the Gaussian kernel density estimation based on the corner points and the contour points is defined as follows: , The observation likelihood probability is defined as follows: , Wherein, is a weight coefficient.

[0011] Further, the posterior probability of the particle is transmitted to the KNN graph, so that the particle is mapped to the posterior probability of the contour edge to which the particle belongs , and the global optimal contour structure line of the building is searched by optimization solution; The transmission of the particle point weight to the contour edge weight is defined as follows: , Wherein, d ij represents the Euclidean distance between particles, and lambda is a control parameter; The minimum spanning tree algorithm is adopted to search the optimal contour edge subset maximizing the posterior probability, and the definition is as follows: , Wherein, is the optimal contour edge subset maximizing the posterior probability generated by the MST, and the maximum energy function value corresponding to the optimal contour edge subset is ; Finally, the global optimal building contour line undirected graph is obtained, wherein represents the vertex in the undirected graph corresponding to the state variable of the particle.

[0012] Compared with the prior art, the present application has the following remarkable effects: 1. The present application proposes a lightweight contour line modeling theory aiming at the problems of strong prior dependence, strict topology and lack of lightweight and semantics in large scene building point cloud modeling; the modeling process is divided into self-supervised deep learning regression contour points and corner points, and Monte Carlo particle filtering tracking global topology line, the global consistency of the particle filtering structure line is maintained, the CityGML paradigm is used to integrate the line model geometry, topology and semantics, the downstream application ability is improved, and a new theoretical framework is provided for real scene three-dimensional modeling; 2, The self-supervised deep learning framework of Point2Contour model is proposed for regression of building contour points and corner points; through embedding a self-supervised mechanism, the sample production standardization problem caused by fuzzy structure line definition in supervised learning is effectively alleviated; based on particle filtering line modeling under Monte Carlo sampling, a contour line model is constructed under global structure line posterior probability maximization, and the topology connection reasonable tracking from contour points to lines is realized; 3, The large-scale lightweight building contour line CityGML model constructed by the application can meet the real scene three-dimensional standard specification and related application requirements; the lightweight line modeling scheme is suitable for efficient storage and rendering of large scene real scene three-dimensional, and can effectively support the computability and analysis of downstream applications; strictly following the international standard CityGML, and natively adapting to 3DCityDB and CityJSON database management system, further enhance the interoperability in storage management and downstream applications. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 The flowchart of the application; Figure 2 The overall architecture diagram of the self-supervised network Point2Contour is shown; Figure 3 The overall schematic diagram of the particle filtering building contour line lightweight modeling process is shown; Figure 4 The selected different building images are shown, wherein (a) is Building-1, (b) is Building-2, (c) is Building-3, (d) is Building-4, (e) is Building-5, (f) is Building-6, (g) is Building-7, and (h) is Building-8; Figure 5 The contour point cloud diagram extracted by the Point2Contour model is shown, wherein (a)~(h) respectively take the point cloud data of the 8 buildings in (a)~(h) as input data; Figure 4 Figure 6 The contour diagram obtained by the particle filtering building line modeling method of the application is shown, wherein (a)~(h) respectively take the 8 buildings in (a)~(h) as input data; Figure 4 Figure 7 The contour diagram obtained by the 3DLineDetection line modeling method is shown, wherein (a)~(h) respectively take the 8 buildings in (a)~(h) as input data; Figure 4 Figure 8 ​​​Figures showing the results of sensitivity experiments for Building-1 with different iteration numbers m, where (a) is m = 20, (b) is m = 50, (c) is m = 100, and (d) is m = 150; Figure 9 Figures showing the results of sensitivity experiments for Building-7 with different iteration numbers m, where (a) is m = 20, (b) is m = 50, (c) is m = 100, and (d) is m = 150; Figure 10 Figures showing the results of sensitivity experiments for Building-1 with different neighbor point thresholds k, where (a) is k = 10, (b) is k = 20, (c) is k = 30, and (d) is k = 40; Figure 11 Figures showing the results of sensitivity experiments for Building-7 with different neighbor point thresholds k, where (a) is k = 10, (b) is k = 20, (c) is k = 30, and (d) is k = 40. DETAILED DESCRIPTION

[0014] The present application will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0015] Using laser radar point cloud, a full-automatic three-dimensional building geometry model of a large scene, light weight, and high-fidelity can be constructed. Compared with a three-dimensional building entity geometry model strictly limited by manifold geometry constraints, a three-dimensional building contour line can accurately depict the complex structure of a large scene building entity, with the advantages of flexible expression, intuitive structure, easy operation and editing in CAD software, and light weight. In addition, the building contour line expression is not only an intuitive model expression, but also constitutes the core basis for the conversion of the entity building geometry model. A large number of studies have used building contour line models as intermediate features, which can be used as an independent light weight geometry expression form and further converted into a high-quality building surface model or entity geometry model. More importantly, compared with the original point cloud, the building contour line model can still provide a high-precision description of the building edge structure with a reduction of 3-4 orders of magnitude, and is particularly suitable for light weight expression, efficient storage, instant transmission, and real-time rendering of large scene building point clouds.

[0016] In view of the advantages of building contour line expression, the present application takes a three-dimensional building contour vector model as the starting point, takes a large scene ubiquitous laser radar building point cloud as the research object, adopts a deep feature representation method for large scene building contour points and corner points under self-supervision, and a particle filtering global tracking method for robustly tracking fine building contour lines in building contour points, deeply excavates the relevance between building contour lines and building fine structure components, gives rich component semantics to the three-dimensional building contour edge line, and improves the computability and analyzability in downstream applications.

[0017] Furthermore, in the task of reconstructing the building contour line of a large outdoor scene point cloud at the level of billions, compared with the task of predicting a structure that is accurate, topological and globally optimal, the current regression of the contour point is more reliable. In order to more stably realize building contour line reasoning, the present application introduces the "particle filtering" idea under the Bayesian framework into the field of building lightweight line modeling, guided by building corner points, robustly tracks the fine contour line of the building from the global scale of the contour point, then mines the semantic relevance of the contour line and the building structure components, gives the three-dimensional building contour line multi-level semantic information, and integrates the geometry, topology and multi-level semantics according to the GityGML mode, and finally stores and manages through the 3DCityDB or CityJSON database.

[0018] As Figure 1 The overall flowchart of the present application is shown, and the specific implementation steps are as follows: Step 1, reconstruct the building contour point and the building corner point by constructing a self-supervised network Point2Contour; In order to extract the building contour point and the building corner point from the building point cloud, the present application constructs a self-supervised network Point2Contour, and the specific steps include: Step 11, construct the overall architecture of the self-supervised network Point2Contour; In order to realize accurate analysis of the building contour point and the building corner point, the present application designs an end-to-end model Point2Contour model using a Transformer self-encoding backbone network based on the self-supervised learning idea, in order to regress the building contour point set B and the building corner point set C from the laser radar building point cloud set P, and the network architecture is as shown in Figure 2 .

[0019] The Point2Contour model is used to realize the building contour point and corner point double regression task, wherein the building contour point regression task is as shown in the first row in Figure 2 : taking the building point cloud as the input, first using the FPS (Farthest Point Sampling) algorithm to downsample the point cloud to extract the center point, then selecting the neighborhood points around each center point through the KNN algorithm, dividing the point cloud into several local blocks, and converting each block of point cloud to the relative coordinate system with the center point as the origin.

[0020] To introduce the self-supervised mechanism, random sampling is performed on the point cloud blocks, and part of the point cloud is marked as a "mask block", and the rest is marked as a "visible block". For the "mask block", the coordinates of the mask part are embedded into a learnable mask token (MaskedToken). For the "visible block", a lightweight PointNet composed of multiple layers of perception and maximum pooling layers is used for local feature embedding to generate a visible token (Visible Token).

[0021] For each center point sampled by the FPS algorithm, a positional token (PositionalToken) is generated by position embedding, and the center point and the mask token or visible token of the corresponding point cloud block are fused to form a low-dimensional embedding feature. Then, the low-dimensional feature is input into the Transformer-1 (abbreviated as TF-1) model for encoding. In the TF-1 model, 6 self-attention heads are used, and each self-attention head outputs a 64-dimensional feature, so that all point cloud blocks are encoded into a 384-dimensional feature vector. The encoded features are used to reconstruct the "mask block". When predicting the contour points, all point cloud blocks are treated as "visible blocks" for embedding and encoding, and further representing the features of each contour point, and the binary classification detection head Detector-1 is used to predict the contour points.

[0022] The building corner regression task is as shown in the second row in Figure 2 The building point cloud is again sampled by FPS and blocked, and then the point cloud blocks containing contour points are randomly masked and marked as "mask blocks", and the remaining blocks are "visible blocks". For the "mask block", it is embedded into a learnable mask token; for the "visible block", it is again embedded into a visible token using a lightweight PointNet. For the center points obtained by downsampling, they are embedded into a positional token, and the center points and the corresponding mask token or visible token are fused. Then, the low-dimensional feature is encoded using the Transformer-2 (abbreviated as TF-2) model, and the encoded feature is used to reconstruct the "mask block". When predicting the building corner points, all point cloud blocks are treated as "visible blocks" for embedding and encoding, and further representing the features of each contour point, and the binary classification detection head Detector-2 is used to predict the building corner points, and through clustering, the optimized "anchor corner" (Anchor Corner) is generated, and finally the corner point offset correction is combined to regress the corner point coordinates.

[0023] Step 12, train the self-supervised network Point2Contour; Since the Point2Contour model is based on self-supervised learning, to further adapt to the prediction task, the application fine-tunes the self-supervised pre-training result by using the building wireframe model in the Building3D public dataset on the basis of self-supervision. Therefore, the double regression tasks of contour points and corner points both need pre-training and fine-tuning, which are as follows: In the pre-training stage, the mask mechanism is used to process the point cloud block of the sampling point, the feature of the point cloud block is extracted through the Transformer, the mask area is reconstructed and predicted by using the coordinate regression head, and the reconstructed mask area point cloud is compared with the original point cloud coordinates of the mask block to improve the learning ability of the contour point and corner point self-supervised encoding feature.

[0024] In the fine-tuning stage, all point cloud blocks are regarded as "visible blocks", and the 384-dimensional features of all point cloud blocks are extracted by using the Transformer. In the building contour point regression task, the feature of each building point is represented by further encoding, and the contour point is predicted by Detector-1, and the predicted contour point is compared with the real contour point to calculate the loss function to optimize the model. In the building corner point regression task, the feature of each contour point is represented by further encoding, and the "candidate corner point" is predicted by Detector-2, and the refined "anchor corner point" is generated by local clustering, and finally the "precise corner point" is regressed by combining the corner point offset correction. The "candidate corner point" and "precise corner point" are compared with the real corner point respectively to learn, and the loss function is calculated.

[0025] The total loss function L of the Point2Contour model is composed of the self-supervised pre-training loss L self , the contour point loss L edge and the corner point loss L corner , that is, L = L self + L edge + L corner , and the specific definitions of the losses are as follows: The self-supervised pre-training loss L self : The application adopts "fillet distance" to design the self-supervised pre-training loss, L self , and the specific form of the loss function is shown in formula (1): (1) Wherein, a represents a point from the predicted mask point set , b represents a point from the original point cloud in the mask area, represents the Euclidean distance of a and b, , respectively represents the number of elements in the predicted mask point set and the original mask point set.

[0026] Self-supervised pre-training loss L self The self-supervised pre-training loss L is calculated by: for each predicted point, the squared distance to the nearest point in the original point cloud is calculated; at the same time, for each original point, the squared distance to the nearest point in the predicted point cloud is calculated; finally, the two parts of the distance are summed and normalized to evaluate the fit of the self-supervised predicted point cloud to the true point cloud.

[0027] Contour point loss L edge The contour point loss L is composed of three parts: the binary cross-entropy of the contour point, the distance error of the predicted contour point to the nearest contour line, and the consistency of the direction of the predicted contour point with the nearest contour line. edge The specific form of the loss function is shown in equation (2): (2) Where MSE(·) represents the mean squared error between two points, in order to make the predicted contour point closer to the true contour line; represents the laser radar building point, represents the nearest contour line point; represents the predicted point is a contour point, represents the probability that the contour point is a true contour point, is the set of building points, CE(·) is the cross-entropy of two variables, and the symbol " | · | " represents the absolute value of the dot product of two vectors, is the total number of building points. In addition, to strengthen the consistency of the direction, represents the direction vector of the predicted point and the consistency of the direction of the nearest contour line of the point, α and β are weight coefficients.

[0028] Corner point loss L corner The loss function of the corner point detection is composed of two parts: "candidate corner point" detection loss and "anchor corner point" coordinate offset loss. After the point features are preliminarily predicted by the detection head Detector-2, "candidate corner points" are obtained which are gathered around the true corner points. At this time, the cross-entropy of the "candidate corner points" and the true corner points is calculated. Subsequently, the "candidate corner points" are clustered to extract "anchor corner points", and the "anchor corner points" are obtained. The "anchor corner point" coordinate offset correction module is optimized by calculating the mean squared error between the "accurate corner point" and the corresponding true corner point. corner The specific expression of the loss function is shown in equation (3): (3) wherein, denotes "precise corner point", denotes the nearest real corner point to ; , are the number of elements in the predicted building contour point set B and the building corner point set C, respectively, denotes the probability that a predicted contour point is a corner point, denotes the real probability that a predicted contour point is a corner point.

[0029] Step 2, reconstruct the building contour line based on "particle filtering"; The overall flowchart of reconstructing the building contour line is shown in Figure 3 , including the following steps: Step 21, use the Point2Contour model regression to obtain the building contour point and corner point set {B, C} (as shown in (a) of Figure 3 ), construct a KNN graph, and select all corner points and part of contour points from it to form an initial seed point set S (as shown in (b) of Figure 3 ).

[0030] Step 22, for each seed point in the initial seed point set S, based on the KNN graph structure, sequentially expand (sample) and track the building contour line based on "particle filtering", after m times of expansion, a series of contour path (i.e. particle line) sets (as shown in (c) of Figure 3 ) are obtained, when the number of contour paths N = |S| is sufficient, the sampled contour paths can approximate the building precise contour line.

[0031] Step 23, use the non-maximum suppression algorithm (NMS) to remove redundant or low-weight contour edges (as shown in (d) of Figure 3 ), and generate a global building contour structure line, i.e. a maximum weight generated tree, through global maximum contour edge posterior probability optimization (as shown in (e) of Figure 3 ).

[0032] Step 24, construct a lightweight geometric model of the building contour line through appropriate pruning and simplification operations (as shown in (f) of Figure 3 ).

[0033] Wherein, the implementation steps of building contour line tracking based on "particle filtering" include: Step 221, model the particle filtering building contour line tracking problem; The contour path of each particle is composed of a state sequence , wherein denotes the initial state of the particle in the contour path, This represents the last state. The definition of a state variable is associated with a seed point; for example, the seed point... The state variable corresponding to the sampling time t of its contour path Can be defined as Seed point The coordinates are ; ,express Direction vector. According to Bayes' theorem, state variables... The posterior probability estimate is: (4) in, Represents state variables Given a sequence of observed states The posterior probability, This represents the prior probability of a state transition. This represents the corresponding observation value after a given state. The observed likelihood probability; in this invention, the observed value High-dimensional feature vectors of contour points obtained from the Point2Contour model High-dimensional feature vectors of corner points .

[0034] Since formula (4) involves global integration over the state space, the prior probability of state transition is... Rather than the observed likelihood probability In this invention, it is defined as nonlinear, making it impossible to obtain an analytical solution. Therefore, this invention introduces the concept of "particle filtering," which uses the weights of a series of particles to approximate the posterior probability of a certain state. Thus, formula (4) can be transformed into: (5) in, This represents the state variable of the i-th particle sampled at time t. The importance weight is the relative confidence of the current particle as a contour line; N represents the number of particles, which determines how many particle lines (contour lines or contour paths) can be tracked. For normalized weights, i.e. δ(.) is the Dirac function, and satisfies .

[0035] Since the Dirac function only exists in The value is taken at a certain point, so formula (5) achieves the posterior estimation in a discrete approximation way through particle sampling based on the Monte Carlo idea. By weighting and summing the discrete particles, the posterior probability formula (4) is avoided. Continuous integration over the entire state space.

[0036] The problem of tracing building outlines naturally transforms into calculating the weight of the particle at each state time. The definition is as follows: (6) In formula (6), the state of the i-th particle is sampled at time t. The weights are determined by the state transition probability of the particle from time t-1 to time t. And the likelihood observation of the particle at time t Probability.

[0037] The state transition probability determines the amplification edge line of each particle at the next time step, thus enabling the prediction of the particle's position. The observation likelihood probability, on the other hand, determines the consistency between the predicted building outline point and the observed value at the next time step, thus enabling the update of the particle weights. Therefore, it is necessary to model the state transition probability and the observation probability separately, predicting the particle's position at the next time step in the particle amplification iteration process, and updating the particle's weights at the next time step in the update step.

[0038] Step 222: Predict the particle positions and update the particle weights; In the prediction step, given the state variables of the i-th particle at the previous time t-1... Then, the state variables of the contour line particles obtained by tracking at the next time step t The definition is as follows: (7) Where k is the adjustment coefficient. State vector Corresponding particle point directional vector, For particle points in the KNN graph The K-nearest neighbor.

[0039] In the update step, the set of depth feature vectors corresponding to the set of building contour points B obtained by the Point2Contour model regression. The set of depth feature vectors corresponding to the set of building corner points C All are considered as observations of the state vector, and the observations each follow a Gaussian distribution. This invention is aimed at the current particle. , respectively using and Construct two Gaussian kernel density estimation (KDE) functions. and The two Gaussian kernel density estimation functions are then fused to calculate the particle observation likelihood function. .

[0040] The Gaussian kernel density estimation based on the corner points and contour points is defined as follows: (8) The observation likelihood probability is finally defined as follows: (9) wherein, w is a weight coefficient, which can be set according to experience or obtained by training optimization. At this point, the particle state transition probability

[0041] of the next state t+1 can be calculated according to formula (7) and formula (9), and the particle weight observation likelihood probability , so that the particle weight is updated to the state of t+1 time: , so as to realize the iteration of the "prediction step" and the "update step", until the specified particle amplification length m is reached, so as to realize the parallel tracking of N particle contour lines. In the specific implementation of the application, the entire set of corner points is used, and part of the contour points are sampled to form the initial seed particle point set N =|S|, a multi-thread parallel algorithm is used to simultaneously amplify N particle lines, until the particle line amplification length m is reached. After the particle filter tracking is completed, N amplified particle lines can be obtained in the KNN graph, and

[0042] N weighted particle point sets composing these particle lines are obtained. .

[0043] The weight of the particle point set is the posterior probability corresponding to the particle point set. In order to further construct the global optimal building contour structure line, it is necessary to further transfer the posterior probability of the particle to the KNN graph, so that the posterior probability of the particle to the contour edge , and search for the global optimal building contour structure line by optimization. The application adopts a "Gaussian weighting" strategy, and when calculating the posterior probability of the contour edge , the posterior probabilities of the particle points at both ends of the edge and are comprehensively considered. The transfer of the particle point weight to the contour edge weight is defined as follows: (10) wherein, d ij represents the Euclidean distance between particles; and λ is a control parameter for balancing the influence of edge length on weight, and a short edge is more likely to belong to the building contour edge line.

[0044] ​After determining the weight of the edge of the particle line, the building contour line is finally generated by maximizing the posterior probability of the edge. (11) By solving formula (11), the globally optimal building contour line undirected graph can be obtained , wherein represents the vertex in the undirected graph corresponding to the particle state variable, is the posterior probability maximization optimal contour edge subset generated by the MST, and the maximum energy function value corresponding to this time is .

[0045] On this basis, further pruning and simplifying are carried out to improve the lightweight expression ability of the building contour line.

[0046] Step 3, quality evaluation of building three-dimensional contour line modeling; The present application comprehensively evaluates the model quality from the four dimensions of geometry, topology, semantics and lightweight of the large-scale building contour line model: C1, geometric accuracy The present application adopts comprehensive evaluation indexes including MAE, RMSE, NRMSE, Hausdorff distance and three-dimensional volume intersection ratio mIoU to verify the geometric accuracy of the large-scale scene building line model. The evaluation true value benchmark can refer to the original building point cloud. In addition, the experimental data set listed in the "feasibility analysis" of the present application contains part of the building line frame model or Mesh model, which can be directly used or used as a true value reference model after indirect conversion processing. At the same time, the present application manually labels part of the building structure line model by means of 3DLABELME tool, so as to objectively evaluate the geometric accuracy of the building line model.

[0047] C2, topological accuracy The present application comprehensively evaluates the topological accuracy of the line model by traversing the building contour line MST graph and the semantic unit graph in CityGML and other data structures: including the redundancy of contour points, the alignment of vertices, the topological correctness of contour point connection, the coplanarity of linear plane primitive contour points, and whether there are trivial or pseudo contour line units and other topological problems.

[0048] C3, semantic accuracy The present application extracts the embedded hierarchical semantic content by analyzing the hierarchical paradigm of the monomer model CityGML, and evaluates the hierarchical granularity of the semantics and the correctness of the semantic content at each level in combination with artificial inspection.

[0049] C4, lightweight degree The application directly evaluates the lightweight degree of the wireframe model by constructing two indexes of "lightweight rate" and "storage rate". The "lightweight rate" and "storage rate" are respectively defined as the ratio of the number of structure lines and corner points in the building structure line model to the number of face patches in the Mesh model (or the number of original point clouds of the building). The comparison results are as follows: (I) Self-supervised network Point2Contour building contour point modeling results The original data used in the application is the open source data set Semantic3D. The building point cloud is extracted from different scenes through the building semantic category, and the point cloud data of a single building is obtained through cropping, as shown in Table 1: Table 1 Semantic3D single building point cloud experimental data table

[0050] D1, qualitative results The application takes the point cloud data of 8 buildings in Table 1 as input data, trains the self-supervised network Point2Contour to extract the contour point cloud of the 8 buildings, and uses it as the input of the subsequent particle filtering line modeling. The qualitative results of the self-supervised network Point2Contour extracting the contour points are shown in (a)-(h) of Figure 5

[0051] D2, quantitative results According to the three-dimensional visualization of the original point cloud, the CloudCompare software is used to extract a contour point cloud from each point cloud building as the true contour point cloud as the true value, and then calculate the consistency of the building contour points extracted by the self-supervised network Point2Contour and the true value, including the maximum deviation, minimum deviation, average deviation, Hausdorff distance deviation. The quantitative accuracy analysis is shown in Table 2: Table 2 Point2Contour extraction building contour line quantitative accuracy table

[0052] (II) Qualitative and quantitative analysis of particle filtering building geometric model lightweight line modeling ​This invention evaluates the particle filter lightweight architectural geometry modeling algorithm in terms of runtime, reconstruction effect, and number of reconstructed line segments, and compares it with the classic 3DlineDetection algorithm to demonstrate the advantages of the particle filter lightweight architectural geometry representation of this invention. The 3DlineDetection algorithm first performs planar segmentation on a large-scale unorganized point cloud through region growing and merging. Then, it projects the points in each plane onto the plane to form a sparse 2D image and extracts contours. It then uses the least squares method to fit and obtain 2D line segments, and projects these line segments back onto the original 3D plane to generate preliminary 3D line segments. Finally, it optimizes the results by removing outlier line segments and merging spatially collinear adjacent line segments. Figure 4 The results obtained using the particle filter building line modeling method of this invention in (a) to (h) are shown in the figure. Figure 6 Qualitative results using 3DlineDetection line modeling, as shown in (a) to (h) of the table, are as follows: Figure 7 (a) to (h) in the text.

[0053] Depend on Figure 7 It can be seen that 3DLineDetection retains more details in the reconstruction, such as the roof tiles of Building-4 and Building-6. 3DLineDetection uses a large number of line segments to simulate the unevenness of the tiles. Although the reconstructed building outlines are richer in detail, the quantization is low, and some noise lines are easily introduced, requiring removal or merging during post-processing. Figure 6 It can be seen that the particle filter line reconstruction algorithm of this invention is relatively stable in terms of overall contour extraction and has a higher degree of line quantization. Even with a high degree of quantization, it can still retain the details of the building result lines. For example, the roof ridge, the main edges of the tower and the wall transitions of Building-3 are well preserved.

[0054] The quantitative comparison between the building outline reconstruction based on particle filtering and the 3DLineDetection reconstruction of this invention is shown in Table 3.

[0055] Table 3. Quantitative Comparison of Two Methods: Results

[0056] It can be seen from Table 3 that the number of line segments reconstructed based on the method of the application is significantly more than that of 3DLineDetection, because the line segments reconstructed by the application are all short line segments, and a long line segment is fitted by a large number of repeated short line segments, while long line segments can be generated in 3DLineDetection, so the number of line segments reconstructed by 3DLineDetection is significantly less than that of the line segments reconstructed by the application. In terms of reconstruction time, the time for 3DLineDetection to reconstruct the contour line is significantly shorter than the time for the method of the application, because the strategy adopted by the application is to generate a path for each seed point according to the method of particle filtering, and the number of seed point clouds will affect the reconstruction accuracy and reconstruction time of the method of the application, and in the process of particle filtering iteration, each iteration particle is in the neighborhood of the point, so the number of steps moved each iteration is small, which is also the reason why the method of the application generates a large number of short lines.

[0057] (Three) Key parameter sensitivity analysis The key reference of the particle filtering building line modeling of the application includes the number of amplification of each particle line, i.e. the number of iterations (m) and the number of query neighbors (K) required to realize the amplification of each particle line from the KNN graph, in order to explore the influence of the two parameters on the final reconstruction accuracy and reconstruction time, and to provide a scientific basis for parameter selection.

[0058] First, the sensitivity experiment about the number of iterations m is carried out, and the point cloud data of Building-1 and Building-7 are used in the experiment process, and the number of neighbor points is unchanged in the experiment process, and the experimental results are shown in Table 4.

[0059] Table 4: Ablation experiment table of the number of iterations m

[0060] As can be seen from Table 4, as the number of iterations increases, the effect of point cloud reconstruction becomes better. When the number of iterations m is 20, the building eave contour line is composed of a large number of noisy short lines. When the number of iterations is 150, it can be seen that the Building-1 eave contour line has no noisy short lines, but is replaced by longer line segments, and these longer line segments are all parallel, and the reconstruction effect is better than that when the number of iterations is 20. In terms of time cost, as the number of iterations increases, more and more time is spent, and the effect of the number of iterations 100 and the number of iterations 150 is not much different, so it can be considered that when the number of iterations m exceeds a certain value, the parameter has little effect on the reconstruction. Considering the time cost and the reconstruction effect, it is appropriate to select the number of iterations 100 as the parameter.

[0061] Next, the parameter sensitivity analysis on the neighbor point threshold k is carried out, and the Building-1 and Building-7 point cloud data are used in the experiment process, and the iteration number is 100 in the experiment process, and the experimental results are shown in Table 5.

[0062] Table 5 Neighbor point threshold k sensitivity analysis table

[0063] From the ablation experiment of the neighbor point threshold 5, it can be concluded that as the neighbor point threshold increases, the reconstruction effect is better and better. When the neighbor point threshold is 10, it can be seen that the reconstruction effect is very poor, and the contour line of the building produces a broken phenomenon. When the neighbor point threshold is 20, the broken phenomenon is alleviated, and when the neighbor point threshold is 30 and 40, the reconstruction of the building contour can achieve a better effect. As the neighbor point threshold increases, the time cost of reconstruction will also increase. Considering the reconstruction result and the time cost, it is more reasonable to select the neighbor point threshold of 30 to reconstruct the building.

Claims

1. A lightweight modeling method for architectural geometry, characterized in that, The steps include the following: S1 uses the Point2Contour model to regress the set of building outline points and the set of building corner points from the LiDAR building point cloud set; S2 reconstructs building outlines based on "particle filtering" and constructs a lightweight geometric model based on the posterior probability of maximizing the global building outline. S3 comprehensively evaluates lightweight geometric models from four dimensions: geometry, topology, semantics, and lightweighting.

2. The lightweight modeling method for architectural geometry according to claim 1, characterized in that, The Point2Contour model is used to perform a dual regression task on building outline points and building corner points; The implementation process of the building outline point regression task is as follows: taking the building point cloud as input, the FPS algorithm is used to downsample the point cloud and extract the center point; then the KNN algorithm is used to select neighborhood points around each center point, divide the point cloud into several local blocks, and transform each block of the point cloud to a relative coordinate system with the center point as the origin. In this process, point cloud blocks are randomly sampled, and some point clouds are marked as "mask blocks" and the rest are marked as "visible blocks". For "mask blocks", the coordinates of the mask part are embedded into learnable mask labels. For "visible blocks", a lightweight PointNet consisting of a multilayer perceptron and a max pooling layer is used to embed local features and generate visible labels. For each center point downsampled by the FPS algorithm, position embedding is performed to generate a position label. The center point is then fused with the mask label or visible label of the corresponding point cloud block to form a low-dimensional embedding feature. This low-dimensional feature is then input into the TF-1 model for encoding. The TF-1 model employs six self-attention heads, each of which is intended to output 64-dimensional features, thus encoding all point cloud blocks into 384-dimensional feature vectors. The encoded features are then used to reconstruct the "mask block." When predicting contour points, all point cloud blocks are treated as "visible blocks" for embedding and encoding, and the features of each contour point are further characterized. The Detector-1 binary classification head is then used to predict the contour points. The implementation process of the building corner regression task is as follows: The building point cloud is sampled and divided into blocks using FPS. Point cloud blocks containing contour points are then randomly masked and labeled as "mask blocks," while all remaining blocks are designated as "visible blocks." For "mask blocks," learnable mask labels are embedded. For "visible blocks," lightweight PointNet is used again to embed them as visible labels. The downsampled center point is embedded as a location label, and the center point is fused with the corresponding mask label or visible label. Then, the TF-2 model is used to encode low-dimensional features, and the encoded features are used to reconstruct the "mask blocks." When predicting building corners, all point cloud blocks are treated as "visible blocks" for embedding and encoding, and the features of each contour point are further characterized. The Detector-2 binary classification head is used to predict building corners, and optimized "anchor corners" are generated through clustering. Finally, corner offset correction is combined to regress the corner coordinates.

3. The lightweight modeling method for architectural geometry according to claim 1, characterized in that, The total loss function L for training the Point2Contour model includes the self-supervised pre-training loss L. self Contour point loss L edge and corner loss L corner , where L = L self + L edge + L corner ; The self-supervised pre-training loss L self The self-supervised pre-training loss is designed using "beveled distance". self The expression for the loss function is: ， Where 'a' represents the value from the predicted mask point set. A point b represents the original point cloud from the masked region. One of the points; This represents the Euclidean distance between a and b. This indicates the number of elements in the predicted mask point set. Indicates the number of elements in the original mask point set; The contour point loss L edge This includes the binary cross-entropy of contour points, the distance error between the predicted contour point and the nearest contour line, and the consistency between the direction of the predicted contour point and the direction of the nearest contour line, L edge The expression for the loss function is: ， Where CE(·) represents the calculation of the cross entropy between two variables; MSE(·) represents the calculation of the mean square error between two points; Indicates the building points of the lidar. Indicates the nearest point on the contour line; For the set of building point clouds, "∣·∣" represents the absolute value of the dot product of two vectors; Indicates prediction The probability that a point is a contour point. This represents the probability that the predicted point is a true contour point. This represents the total number of building points. Indicates prediction The direction vector of the point and Point nearest contour line Consistency of direction; α and β are weighting coefficients, respectively; The corner loss L corner This includes "candidate corner point" detection loss and "anchor corner point" coordinate offset loss. After the point features are initially predicted by the Detector-2 head, "candidate corner points" clustered near the real corner points are obtained. At this point, the cross-entropy between the "candidate corner points" and the real corner points is calculated. Subsequently, the "candidate corner points" are clustered to obtain "anchor corner points," and the coordinate offset of the "anchor corner points" is corrected to obtain the set C of "precise corner points." The corner point offset correction module is optimized by calculating the mean square error between the "precise corner points" and the corresponding real corner points. corner The expression for the loss function is: ， in, This represents the probability of predicting whether a contour point is a corner point. This represents the probability that a predicted contour point is a true corner point. Indicates "precise corner point", Indicates and The nearest real corner point; , These represent the number of elements in the set of building outline points B and the set of building corner points C, respectively.

4. The lightweight modeling method for architectural geometry according to claim 1, characterized in that, The steps to reconstruct the building outline are as follows: S21. Use the Point2Contour model to regress and obtain the set of building outline points and the set of building corner points to construct a KNN graph; and select all corner points and some outline points from it to form the initial seed point set S; S22: For each seed point in the initial seed point set S, it is sequentially amplified based on the KNN graph structure, and the building outline is tracked based on "particle filtering"; after m amplifications, a set of outline paths is obtained. When the number of contour paths N = |S|, the sampled contour paths approximate the precise contour lines of the building. S23 uses a non-maximum suppression algorithm to remove redundant or low-weight contour edges and optimizes them by globally maximizing the posterior probability of contour edges to generate global building contour lines. S24 employs pruning and simplification techniques to construct a lightweight geometric model of the building outline.

5. The lightweight modeling method for architectural geometry according to claim 4, characterized in that, The implementation steps for building outline tracing based on "particle filtering" include: S221, Modeling the particle filter building outline tracking problem; The contour path of each particle is determined by a state sequence. Composition, in which This indicates the initial state of the particles in the contour path. This represents the last state; for the seed point The state variable corresponding to the sampling time t of its contour path Defined as Seed point The coordinates are ; ,express Direction vector; according to Bayes' theorem, state variable The posterior probability estimate is: ， in, Represents state variables Given a sequence of observed states The posterior probability, This represents the prior probability of a state transition; This represents the corresponding observation value after a given state. The observed likelihood probability; By using the weights of a series of particles to approximate the posterior probability of a certain state, the above formula is transformed into: ， in, This represents the state variable of the i-th particle sampled at time t. Importance weight, δ(.) is the Dirac function, and satisfies ; The problem of tracing building outlines is transformed into calculating the weight of particles at each state time step. The definition is as follows: ， In the formula, the state of the i-th particle is sampled at time t. The weights are determined by the state transition probability of the particle from time t-1 to time t. And the likelihood observation of the particle at that moment probability; S222, predicts particle positions and updates particle weights; In the prediction step, given the state variables of the i-th particle at time t-1... Then the state variables of the contour line particles tracked at time t The definition is as follows: ， Where k is the adjustment coefficient. State vector Corresponding particle point directional vector, For particle points in the KNN diagram The K-nearest neighbor; In the update step, the set of depth feature vectors corresponding to the set of building contour points B obtained by the Point2Contour model regression are... The set of depth feature vectors corresponding to the set of building corner points C All are considered as observations of the state vector, and the observations follow Gaussian distributions; therefore, the Gaussian kernel density estimation based on corner points and contour points is defined as follows: ， Observational likelihood probability The definition is as follows: ， in, These are the weighting coefficients.

6. The lightweight modeling method for architectural geometry according to claim 5, characterized in that, The posterior probability of a particle is passed to the KNN graph, so that the posterior probability of a particle being mapped to the edge of its corresponding contour is... And by optimizing the solution, the optimal global outline structure line of the building is searched; The transfer of particle point weights to contour edge weights is defined as follows: ， Where, d ij λ represents the Euclidean distance between particles, and λ is a control parameter. The minimum spanning tree algorithm is used to search for the optimal subset of contour edges that maximizes the posterior probability, defined as follows: ， in, The optimal subset of contour edges generated for posterior probability maximization for MST has the corresponding maximum energy function value. ; Finally, the globally optimal undirected graph of building outlines is obtained. ,in This represents the vertex in the undirected graph corresponding to the particle's state variable.

Images