Building extraction method and system based on multi-dimensional point cloud structure entropy phase transition detection

Abstract

The application relates to the field of laser radar three-dimensional sensing technology and proposes a building extraction method and system based on multi-dimensional point cloud structure entropy phase change detection, which comprises the following steps: original point cloud data is denoised and non-ground point extraction is carried out, a multi-dimensional structure entropy field which fuses geometric structure, echo distribution and height consistency is constructed, and is used for quantifying the local structure order of the point cloud; building boundary identification is realized through structure entropy gradient field analysis combined with gradient flow tracking and persistent homology method, and a candidate building area is constructed; high-quality seed points are obtained by using multi-modal feature clustering, and a building point cloud aggregation area is formed through structure convergence driven directional growth; region fusion and contour extraction are carried out based on the minimum description length criterion, and a structured building vector expression meeting the quality evaluation requirement is generated. The application excludes the interference of high-confusion objects such as regularly arranged fences, single trees and vehicles, and realizes high-precision, high-robustness and high-explainability building extraction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of laser radar three-dimensional perception, and particularly relates to a building extraction method and system based on multi-dimensional point cloud structure entropy phase transition detection. BACKGROUND

[0002] The airborne laser radar building extraction technology refers to a technical system in which a laser radar sensor is carried on an aerial platform to obtain information such as the geometric shape, spatial distribution and structural characteristics of buildings in a region through three-dimensional point cloud data. Compared with the traditional building extraction method based on two-dimensional images or manual surveying, the airborne laser radar building extraction technology has the advantages of higher three-dimensional precision, stronger structural recognition ability and wider operation coverage. The airborne laser radar can quickly obtain large-scale, high-precision three-dimensional point cloud data, realize efficient and accurate automatic extraction of buildings and three-dimensional reconstruction, and provide timely and accurate data support for city geographic information systems and three-dimensional modeling.

[0003] In the prior art, the airborne laser radar building extraction mainly adopts traditional methods based on geometric features, statistical features and deep learning. Although the basic detection and extraction of buildings are realized, in actual application, there is a lack of general mathematical tools that can effectively quantify the internal order of the local structure of the point cloud, and the existing feature extraction methods cannot accurately depict the structural complexity and order difference of the building point cloud. Therefore, it is difficult to accurately distinguish the structural differences between buildings and other ground objects in high-demand scenarios such as complex urban environments, especially under the conditions of building and vegetation adhesion and irregular building shapes, resulting in a high miss detection rate for complex structural buildings and a high misjudgment rate for mixed ground objects. SUMMARY

[0004] Therefore, the present application provides a building extraction method and system based on multi-dimensional point cloud structure entropy phase transition detection, which solves the problem that the prior art cannot accurately distinguish the structural differences between buildings and other ground objects in high-demand scenarios such as complex urban environments, especially under the conditions of building and vegetation adhesion and irregular building shapes, resulting in a high miss detection rate for complex structural buildings and a high misjudgment rate for mixed ground objects.

[0005] The technical scheme of the present application is implemented as follows: on the one hand, the present application provides a building extraction method based on multi-dimensional point cloud structure entropy phase transition detection, which includes the following steps:

[0006] The original point cloud data is denoised and a progressive triangular net encryption filtering algorithm is used to obtain non-ground points, and a structure entropy field is constructed by weighted fusion of geometric structure entropy, echo distribution entropy and elevation consistency entropy;

[0007] The structural entropy field quantifies the structural order of the local neighborhood of the point cloud by fusing three dimensions of geometric anisotropy, echo distribution complexity, and elevation fluctuation consistency.

[0008] Gradient field calculation is performed on the non-ground points based on the structural entropy field, boundary structures are identified based on gradient flow tracking and persistent homology, and a candidate building region is constructed based on a closed-loop boundary ring.

[0009] A multi-modal feature field is constructed for the candidate building region, feature clustering is performed based on a Gaussian mixture model and a density peak value, and a high-quality seed point set is obtained.

[0010] Seed activation is performed based on the high-quality seed point set, absorption decision is made on candidate points based on a structural convergence dynamic evaluation method, and directional growth is performed based on the structural entropy field, and a building point cloud aggregation region is obtained.

[0011] Optimal fusion is performed on the building point cloud aggregation region based on a minimum description length metric method, the fused three-dimensional point cloud building object is converted into a two-dimensional vector contour, quality evaluation is performed on the two-dimensional vector contour, and a structured building expression is obtained.

[0012] Based on the above technical solutions, preferably, the gradient field calculation is performed on the non-ground points based on the structural entropy field, the boundary structures are identified based on gradient flow tracking and persistent homology, and the candidate building region is constructed based on a closed-loop boundary ring, which includes:

[0013] Gradient field calculation and significance analysis of the three-dimensional structural entropy field are performed on the non-ground points, a scalar entropy value field of the point neighborhood is fitted using a moving least squares method, local linear fitting coefficients are obtained by solving a weighted least squares problem, entropy gradient vectors and gradient amplitudes at each point are calculated, and a significant gradient point set that meets the gradient amplitude condition is screened out.

[0014] The significant gradient point set is used as a seed point, bidirectional streamline tracking is performed along the gradient direction, boundary segments are obtained by numerically integrating the streamline equation using a fourth-order Runge-Kutta method, significant boundary structures in the boundary segments are identified based on a topological feature importance evaluation scheme of persistent homology, and the candidate building region is constructed based on a closed-loop boundary ring.

[0015] Based on the above technical solutions, preferably, the gradient field calculation and significance analysis of the three-dimensional structural entropy field are performed on the non-ground points, the scalar entropy value field of the point neighborhood is fitted using a moving least squares method, the local linear fitting coefficients are obtained by solving a weighted least squares problem, the entropy gradient vectors and gradient amplitudes at each point are calculated, and the significant gradient point set that meets the gradient amplitude condition is screened out, which includes:

[0016] The point product of the gradient direction and the local normal vector determines the positive and negative gradients, the positive gradient of the positive and negative gradient represents the corresponding building outer boundary from the building to the environment, and the negative gradient of the positive and negative gradient represents the corresponding hole or courtyard boundary from the environment to the inside of the building, the statistical quantity of the global gradient amplitude distribution is calculated, and the significant gradient point set satisfying the gradient amplitude greater than the mean value plus the standard deviation multiple is screened out.

[0017] On the basis of the above technical scheme, preferably, the significant gradient point set is taken as a seed point, bidirectional streamline tracking is carried out along the gradient direction, a fourth-order Runge-Kutta method numerical integral streamline equation is used to obtain a boundary segment, a topological feature importance evaluation scheme based on persistent homology is used to identify significant boundary structures in the boundary segment, and the candidate building area is constructed based on a closed loop boundary ring, comprising:

[0018] The positive gradient points are integrated along the positive gradient direction to simulate an entropy increase path from the inside of the building to the outside, the negative gradient points are integrated along the negative gradient direction to simulate an entropy decrease path from the outside to the inside of the building, an ordered point set is generated for each streamline, and spline fitting is performed to obtain a smooth parameterized curve representation, and a α complex is constructed, the birth time and death time of the generators of each dimension homology group are calculated, topological features satisfying a persistence threshold are screened out, and corresponding boundary segments are extracted according to the streamline point set corresponding to the topological features.

[0019] On the basis of the above technical scheme, preferably, the building point cloud aggregation area is optimally fused based on a minimum description length metric method, the fused three-dimensional point cloud building object is converted into a two-dimensional vector contour, and the quality of the two-dimensional vector contour is evaluated to obtain a structured building expression, comprising:

[0020] The building point cloud aggregation area is optimally fused based on an inter-regional minimum description length metric method, the total description length is defined as the weighted sum of the geometric description length and the structure entropy description length, the merging benefit and merging confidence of a candidate merging pair are calculated, the merging of regions satisfying the conditions is performed, and a final building object set is obtained;

[0021] The final building object set is subjected to contour polygon extraction based on an adaptive alpha-shape contour extraction, the inner angle size is calculated by detecting the contour angle, the inner angle close to 90 degrees is adjusted to 90 degrees, the two-dimensional vector contour is obtained, the geometric fidelity and area consistency are calculated to perform the quality evaluation, and the structured building expression is generated.

[0022] On the basis of the above technical scheme, preferably, the inter-regional minimum description length measurement method is used to optimally fuse the building point cloud aggregation region, the total description length is defined as the weighted sum of the geometric description length and the structure entropy description length, the merging benefit and the merging confidence evaluation of the candidate merging pair are calculated, the region merging meeting the condition is performed, and the final building object set is obtained, including:

[0023] The geometric description length includes boundary complexity and internal structure complexity, the structure entropy description length is calculated based on the region point entropy value distribution covariance and the KL divergence of the region distribution and the global building point distribution, the region pairs with effective adjacent distance less than a preset value are retained based on the spatial adjacency relationship, and the region merging is performed on the regions with merging benefit less than a preset value and merging confidence greater than a preset value.

[0024] On the basis of the above technical scheme, preferably, the original point cloud data is denoised and the progressive triangulation encryption filtering algorithm is used to obtain the non-ground points, the structure entropy field is constructed by weighted fusion of geometric structure entropy, echo distribution entropy and elevation consistency entropy, including:

[0025] The statistical outlier removal method is used on the original point cloud data to eliminate sensor noise and abnormal points, and the voxel grid method is used for regularized resampling of the point cloud to obtain resampled point cloud;

[0026] The resampled point cloud is separated by using the progressive triangulation encryption filtering algorithm to obtain the non-ground points;

[0027] According to the characteristic scale of the building component, a multi-scale adaptive neighborhood is constructed, the final representative entropy value of each point is generated by weighted geometric mean fusion of the geometric structure entropy, echo distribution entropy and elevation consistency entropy, and the reliability weight of each scale entropy value is dynamically determined based on the local terrain complexity, and the structure entropy field is constructed.

[0028] On the basis of the above technical scheme, preferably, the multi-modal feature field is constructed for the candidate building region, feature clustering is performed based on Gaussian mixture model and density peak value to obtain a high-quality seed point set, including:

[0029] The local point density, normalized relative elevation and final structure entropy are calculated for each point in the candidate building region, the robust statistical quantity calculation and winsorization processing are performed on each feature dimension, the mapping from the three-dimensional coordinate space to the feature space is established, and the feature point set distribution is obtained.

[0030] The optimal component number is determined by using Bayesian information criterion to fit a Gaussian mixture model, the local density and minimum distance of each point in the feature space are calculated, the posterior probability of the Gaussian mixture model is fused with the density peak value information, similar cluster clusters are merged based on Mahalanobis distance, and the high-quality seed point set is obtained.

[0031] On the basis of the above technical scheme, preferably, the seed activation is performed based on the high-quality seed point set, the absorption decision of the candidate point is performed based on a structure convergence dynamic evaluation method, and the directional growth is performed based on the structure entropy field, so as to obtain a building point cloud aggregation region, including:

[0032] A weighted undirected graph is constructed based on the high-quality seed point set, and confidence propagation iteration is performed, so as to obtain an activation sequence, an initial growth region is created starting from a seed point with the highest confidence, a region structure entropy reference value and a region normal vector dominant direction are calculated, an active boundary set, a candidate point buffer area, a growth history record and a dynamic convergence threshold are initialized;

[0033] An unmarked candidate point is searched from each point in the active boundary set, local structure compatibility, region entropy change prediction and topological continuity index are calculated for each unmarked candidate point, an absorption decision state vector is constructed, absorption decision is performed in combination with a pre-training strategy network and a rule threshold, and batch absorption is performed to update the growth region.

[0034] In another aspect, the application also provides a building extraction system based on multi-dimensional point cloud structure entropy phase transition detection, the system comprising:

[0035] A data preprocessing and structure entropy field construction module is configured to denoise original point cloud data, acquire non-ground points by using a progressive triangulation encryption filtering algorithm, and construct a structure entropy field by weighted fusion of geometric structure entropy, echo distribution entropy and elevation consistency entropy.

[0036] A gradient field calculation and boundary identification module is configured to perform gradient field calculation on the non-ground points based on the structure entropy field, identify boundary structures based on gradient flow tracking and persistent homology, and construct a candidate building region based on a closed-loop boundary ring.

[0037] A multi-modal feature clustering module is configured to construct a multi-modal feature field for the candidate building region, perform feature clustering based on a Gaussian mixture model and a density peak value, and obtain a high-quality seed point set.

[0038] A seed activation and region growth module is configured to perform seed activation based on the high-quality seed point set, perform absorption decision of a candidate point based on a structure convergence dynamic evaluation method, and perform directional growth based on the structure entropy field, so as to obtain a building point cloud aggregation region.

[0039] The region fusion and contour extraction module is used for optimal fusion of the building point cloud aggregation region based on a minimum description length metric method, converting the fused three-dimensional point cloud building object into a two-dimensional vector contour, and performing quality evaluation on the two-dimensional vector contour to obtain a structured building expression.

[0040] The building extraction method and system based on multi-dimensional point cloud structure entropy phase transition detection of the present application have the following beneficial effects relative to the prior art:

[0041] (1) By constructing a multi-dimensional point cloud structure entropy phase transition detection mechanism, a paradigm shift from traditional geometric feature segmentation to structure order phase transition detection is realized, which can accurately distinguish the structural differences between buildings and vegetation and other features in complex urban environments, effectively exclude the interference of regularly arranged fences, single trees, vehicles and other high confusion objects, and realize high-precision, high-robustness and high-explainability building extraction.

[0042] (2) By fusing gradient field analysis and topological homology theory, entropy field gradient calculation and significance screening are performed using the moving least squares method, the building outer boundary and inner courtyard boundary are distinguished by combining positive and negative gradient direction judgment, and the boundary segments are obtained by bidirectional streamline tracking using the fourth-order Runge-Kutta method, realizing high-precision automatic detection of building boundaries and complete closed loop region construction.

[0043] (3) By fusing the minimum description length metric method of geometric description length and structure entropy description length, optimal fusion of building aggregation regions is realized using merging benefits and confidence evaluation, the contour parameters are dynamically adjusted using the adaptive alpha-shape contour extraction method, and the inner angles close to 90 degrees are regularized according to angle detection, which improves the integrity and geometric normativity of building vectorization expression, and at the same time, the quantitative quality evaluation index meets the high-precision application requirements of geographic information systems and three-dimensional modeling. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0045] Figure 1 The flowchart of the building extraction method based on multi-dimensional point cloud structure entropy phase transition detection of the present application. DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work belong to the protection scope of the present application.

[0047] Please refer to Figure 1 The present application provides a building extraction method based on multi-dimensional point cloud structure entropy phase detection, comprising the following steps:

[0048] The original point cloud data is denoised and the non-ground points are obtained by using the progressive triangulation encryption filtering algorithm, and the structure entropy field is constructed by the weighted fusion of geometric structure entropy, echo distribution entropy and elevation consistency entropy;

[0049] The structure entropy field is fused with geometric anisotropy, echo distribution complexity and elevation fluctuation consistency in three dimensions to quantify the structure order of the local neighborhood of the point cloud;

[0050] The gradient field of the non-ground points is calculated based on the structure entropy field, the boundary structure is identified based on gradient flow tracking and persistent homology, and the candidate building area is constructed based on the closed loop boundary ring;

[0051] A multi-modal feature field is constructed for the candidate building area, feature clustering is performed based on Gaussian mixture model and density peak value, and a high-quality seed point set is obtained;

[0052] Seed activation is performed based on the high-quality seed point set, absorption decision is made on the candidate points based on the structure convergence dynamic evaluation method, and directional growth is performed based on the structure entropy field, and the building point cloud aggregation area is obtained;

[0053] The optimal fusion of the building point cloud aggregation area is performed based on the minimum description length metric method, the fused three-dimensional point cloud building object is converted into a two-dimensional vector contour, the quality of the two-dimensional vector contour is evaluated, and the structured building expression is obtained.

[0054] Specifically, the multi-dimensional point cloud structure entropy phase detection mechanism is constructed, the paradigm change from traditional geometric feature segmentation to structure order phase detection is realized, the problem of lack of general structure order quantization tool in the existing building extraction method is solved, the structural differences between buildings and vegetation and other ground objects can be accurately distinguished in complex urban environment, the interference of high confusion objects such as regularly arranged fences, single trees and vehicles is effectively excluded, the building extraction with high precision, high robustness and high interpretability is realized without pre-setting geometric model, manual experience rule or a large amount of labeled data, and the output result is an independent building polygon with attributes.

[0055] The original point cloud data is denoised and the non-ground points are obtained by using a progressive triangulation densification filtering algorithm, and a structure entropy field is constructed by weighted fusion of geometric structure entropy, echo distribution entropy and elevation consistency entropy, including:

[0056] The original point cloud data is denoised and the non-ground points are obtained by using a progressive triangulation densification filtering algorithm, and a structure entropy field is constructed by weighted fusion of geometric structure entropy, echo distribution entropy and elevation consistency entropy, including:

[0057] The original point cloud data is denoised and the non-ground points are obtained by using a progressive triangulation densification filtering algorithm, and a structure entropy field is constructed by weighted fusion of geometric structure entropy, echo distribution entropy and elevation consistency entropy, including:

[0058] According to the characteristic scale of the building component, a multi-scale adaptive neighborhood structure entropy is constructed, the final representative entropy value of each point is generated by weighted geometric mean fusion of the geometric structure entropy, echo distribution entropy and elevation consistency entropy, and the structure entropy field is constructed based on the local terrain complexity and dynamically determining the reliability weight of each scale entropy value.

[0059] In a specific embodiment, the statistical outlier removal method is used to eliminate sensor noise and abnormal points for the obtained original point cloud data set, and in order to avoid the influence of uneven point cloud density on local calculation, the voxel grid method is used for regularized resampling of the point cloud, thereby providing a uniform data basis for subsequent steps. In this embodiment, the voxel grid size is set to 0.1m;

[0060] The ground points and non-ground points are separated by using an improved progressive triangulation densification filtering algorithm (ATIN) to obtain a non-ground point set ;

[0061] According to the characteristic scale of the building component, a multi-scale adaptive neighborhood structure entropy is constructed, the final representative entropy value of each point is generated by weighted geometric mean fusion of the geometric structure entropy, echo distribution entropy and elevation consistency entropy, and the structure entropy field is constructed based on the local terrain complexity and dynamically determining the reliability weight of each scale entropy value.

[0062] Step 1, based on the typical size of urban buildings, three analysis scales are preset:

[0063]

[0064] Among them, the first analysis scale is suitable for detail scale (such as window frame, railing, etc.), the second analysis scale is suitable for component scale (such as wall surface, standard roof surface piece, etc.), and the third analysis scale is suitable for structure scale (such as complete roof, building main body, etc.). For each scale , the corresponding KD-tree space index is established;

[0065] Step 2, for each point in the non-ground point set , search the ball center under each analysis scale , radius of all points in the field point set . If the number of points in the field point set is less than (typically set ), the entropy value of the analysis scale is marked as invalid.

[0066] Step 3, for each valid field , the geometric structure entropy , echo distribution entropy and elevation consistency entropy are calculated as follows:

[0067] For the geometric structure entropy : the covariance matrix of the points in the field is calculated to obtain the first, second and third eigenvalues , and is defined, where , respectively represent the strength of the local structure showing linear, planar and volumetric distribution. Then:

[0068]

[0069] where is a very small positive number to prevent calculation errors of logarithm. At the same time, a confidence weight is introduced, is the number of points in the field, is an empirical threshold, which is selected as in this embodiment. The physical meaning of the geometric structure entropy is that it corresponds to when it is an ideal plane (such as a roof); and it corresponds to when it is a random distribution (such as the interior of a tree canopy).

[0070] For the echo distribution entropy : the distribution of the number of echoes of the points in the field is counted. Let the proportion of the points of the th echo be , then . The physical meaning of the echo distribution entropy is that the surface of a rigid building mainly produces single or first echoes, and the distribution is concentrated . The vegetation canopy is prone to multiple echoes, and the distribution is uniform, with a high value.

[0071] The elevation consistency entropy : the elevation value range of the points in the field is equally divided into B intervals, and the probability of the elevation value falling into each interval is calculated. Then . The physical meaning of the elevation consistency entropy is that the flat or regularly inclined roof has a concentrated The vegetation canopy is uneven and distributed in a scattered manner. The value is high.

[0072] The three sub-entropies are fused by weighted geometric mean:

[0073]

[0074] Record points Multi-scale entropy values at three analysis scales: first analysis scale entropy value , second analysis scale entropy value , and third analysis scale entropy value .

[0075] Based on local terrain features, the multi-scale entropy values are fused to generate the final representative entropy value of each point, and the main steps include:

[0076] Step 4, for each point , calculate the terrain complexity index in the second analysis scale ( ) field:

[0077]

[0078] wherein, are the first, second, and third eigenvalues of the covariance matrix of the field point coordinates ( ), is the field point elevation variance. Normalized to interval, as terrain complexity coefficient .

[0079] Step 5, dynamically determine the reliability weight of each scale entropy value according to the terrain complexity coefficient :

[0080]

[0081] wherein, is the decay coefficient, which is 0.3 in this embodiment. is the first reliability weight, is the second reliability weight, is the third reliability weight. At this time, when the point is in a simple platform area, , the large-scale information is more reliable, large; in a complex and fine area, , the small-scale information is more reliable, large.

[0082] Step 6, generate the final structural entropy value of the point :

[0083]

[0084] in, This represents the final structural entropy value. As the first credibility weight, As the second credibility weight, As the third credibility weight; , , These are the entropy values ​​of the first analytical scale. Second analytical scale entropy value The third analytical scale is the entropy value.

[0085] The steps to construct a pyramid representation of the structural entropy field include:

[0086] Step 7, set the original resolution as... Generated by stepwise downsampling First low-resolution layer, second low-resolution layer, third low-resolution layer. For each layer... Perform voxelization (voxel size is For each downsampled point, the entropy distribution statistical characteristics (mean, median, minimum) of its original neighborhood (corresponding to multiple points in the original point cloud) are transferred to that point.

[0087] Step 8, for the original resolution level For each point, check its position via spatial coordinate mapping. Entropy statistics for the corresponding region. If a point exhibits significantly low entropy across multiple resolution levels (i.e., below the global threshold for each sub-level), then the confidence level of that point as a building point is increased by multiplying its final entropy value by a decay factor. Attenuation factor In this embodiment, 0.9 is used; if the point only exhibits low entropy at the finest level but high entropy at the coarse level, it may be an object that is ordered at a small scale but disordered at a large scale, such as a single vehicle, and the final entropy value of the point can be appropriately increased.

[0088] The process of calculating the gradient field for the non-ground points based on the structural entropy field, identifying boundary structures based on gradient flow tracing and persistent cohomology, and constructing candidate building regions based on closed-loop boundary rings includes:

[0089] The gradient field of the three-dimensional structural entropy field of the non-ground points is calculated and its significance is analyzed. The scalar entropy field of the neighborhood of the point is fitted by the moving least squares method. The local linear fitting coefficient is obtained by solving the weighted least squares problem. The entropy gradient vector and gradient magnitude at each point are calculated, and the set of significant gradient points that meet the gradient magnitude condition is selected.

[0090] Using the set of significant gradient points as seed points, bidirectional streamline tracing is performed along the gradient direction. The fourth-order Runge-Kutta method is used to numerically integrate the streamline equation to obtain the boundary segments. The significant boundary structures in the boundary segments are identified based on the persistent homology topological feature importance evaluation scheme, and the candidate building regions are constructed based on the closed-loop boundary ring.

[0091] The gradient field calculation and significance analysis of the three-dimensional structural entropy field for the non-ground points are performed. The moving least squares method is used to fit the scalar entropy field of the point's neighborhood. Local linear fitting coefficients are obtained by solving a weighted least squares problem. The entropy gradient vector and gradient magnitude at each point are calculated, and a set of significant gradient points satisfying the gradient magnitude condition is selected, including:

[0092] The positive and negative gradients are determined by calculating the dot product of the gradient direction and the local normal vector. The positive gradient represents the direction from the building to the outer boundary of the corresponding building in the environment, and the negative gradient represents the direction from the environment to the boundary of the corresponding hole or courtyard inside the building. The statistics of the global gradient magnitude distribution are calculated, and the significant gradient point set that satisfies the requirement that the gradient magnitude is greater than a multiple of the mean plus the standard deviation is selected.

[0093] The process involves using the set of significant gradient points as seed points, performing bidirectional streamline tracing along the gradient direction, numerically integrating the streamline equations using the fourth-order Runge-Kutta method to obtain boundary segments, identifying significant boundary structures within these segments based on a persistent homology-based topological feature importance assessment scheme, and constructing the candidate building regions based on closed-loop boundary rings.

[0094] For points with positive gradients, the entropy-increasing path from the building's interior to the exterior is simulated by integrating along the positive gradient direction; for points with negative gradients, the entropy-decreasing path from the exterior to the building's interior is simulated by integrating along the negative gradient direction. An ordered set of points is generated for each streamline, and spline fitting is performed to obtain a smooth, parametric curve representation, thus constructing a candidate point set. α The complex is used to calculate the birth and death times of homology group generators in each dimension, select topological features that meet the persistence threshold, and extract the corresponding boundary segments based on the streamline point set corresponding to the topological features.

[0095] In one specific embodiment, the gradient field calculation and significance analysis of the three-dimensional structural entropy field are performed on the non-ground points. The specific steps include:

[0096] Step 9, based on non-ground points , The structural entropy value is Build a KD-tree spatial index for each point. Search its radius k nearest neighbors , Typically, a distance of 1.0 to 1.5 times the average point spacing is used. The moving least squares method is employed to fit the points. Scalar entropy field of the domain :

[0097] ;

[0098] in, basis functions Gaussian weighting function , By solving this weighted least squares problem, the local linear fitting coefficients are obtained. ,point The entropy gradient vector at point is:

[0099] ;

[0100] Gradient magnitude is .

[0101] Step 10: Calculate the gradient direction and local normal vector using PCA. dot product:

[0102] ;

[0103] Positive gradient ( ): This indicates the direction of increasing entropy outward along the normal, that is, from the building (low entropy) to the environment (high entropy), corresponding to the outer boundary of the building;

[0104] negative gradient ( ): This indicates the direction inward along the normal direction, where the entropy value decreases, i.e., from the environment towards the interior of the building, corresponding to the boundary of a hole or inner courtyard.

[0105] Step 11, calculate the statistics of the global gradient magnitude distribution: mean. and standard deviation Select the set of significant gradient points that meet the following conditions. : ,in, The sensitivity parameter is set to 1.0 in this embodiment. The location of each significant gradient point is also recorded. gradient vector Gradient magnitude Positive and negative .

[0106] The candidate boundary curve segments are extracted by using gradient flow tracing to extract significant gradient points. The steps include:

[0107] Step 12, set the significant gradient points As seed points, bi-directional streamline tracing is performed along the gradient direction (considering the positive and negative nature):

[0108] For positive gradient points, forward tracing is performed: integrating along the direction, simulating the entropy-increasing path from the interior to the exterior of the building.

[0109] For negative gradient points, backward tracing is performed: integrating along the direction, simulating the entropy-decreasing path from the exterior to the interior of the building.

[0110] The streamline equation is numerically integrated using the fourth-order Runge-Kutta method: , to adaptively adjust the step size . The tracing stops when one of the following conditions is met: 1. the boundary of the point cloud is reached; 2. the gradient magnitude falls below a threshold ; 3. the tracing length exceeds the maximum allowed length .

[0111] Step 13. For each streamline, a set of ordered points is generated , and a spline fitting is performed on each curve to obtain a smooth parametric curve representation . From each smooth curve, a discrete boundary point sequence is resampled at equal intervals , ensuring that the distance between adjacent points is approximately the average point distance. For each boundary point, its curve ID, parameter position, and gradient direction at that point are recorded as an estimate of the boundary normal.

[0112] Step 14. Based on the gradient positivity of the tracing starting point and the curve morphology, the boundary segments obtained in step 13 are divided into three categories:

[0113] External boundary: open or closed curves generated by positive gradient seed points, enclosing low-entropy regions;

[0114] Internal boundary: closed curves generated by negative gradient seed points, corresponding to interior spaces such as courtyards and patios;

[0115] Ambiguous boundary: short curves that do not show a clear closing trend, possibly due to noise or complex over-regions.

[0116] Based on the persistent homology-based topological feature importance evaluation scheme, the topologically meaningful boundary structures that persist at multiple scales are identified from a large number of boundary segments, including the following steps:

[0117] Step 15. Obtain the boundary point set obtained in step 13, and construct the of the boundary point set as the radius parameter increases from 0 to , the birth time of the generator of the homology group of each dimension and the death time :

[0118] 0-dimensional features: appearance and merging of connected components;

[0119] 1-dimensional features: generation and filling of loop structures;

[0120] 2-dimensional features: formation and filling of voids.

[0121] For boundary detection, this embodiment focuses on 0-dimensional features (boundary endpoints) and 1-dimensional features (boundary loops). The persistence of each topological feature is calculated : .

[0122] Step 16, set the persistence threshold where, and are the mean and standard deviation of the persistence of all 1-dimensional features, usually taken as 1.0. Significant 1-dimensional topological features that satisfy , i.e. "long-lived" loop structures, are filtered out, and these significant topological features are mapped back to the original boundary segments, i.e. identify those boundary curves that form early (small ) and disappear late (large ) in the filtering process. Each boundary segment is assigned a topological importance score :

[0123] ;

[0124] where, is the persistence of the associated topological feature, is the maximum persistence, is the curve length, is the maximum curve length.

[0125] Based on the closed-loop boundary loop construction, candidate building region division is performed. Including the following steps:

[0126] Step 17, establish a boundary segment connection graph: if the end points of two boundary segments and are less than the connection threshold , usually taken as 2 times the average point distance, 0.3m in this embodiment, and the gradient direction angle at the end point is less than (as ), they are considered to be connectable. Based on the constructed connection graph, a depth-first search is used to find closed loops:

[0127] 0 Prioritize the segments with high topological importance score;

[0128] 2. Allow geometric bridging at gaps: if the distance between two endpoints is and the middle region has no significant gradient points, then connect them with a straight segment. Output a set of closed boundary loops , each represented by an ordered sequence of boundary points.

[0129] Step 18, for each closed loop , compute its geometric and topological verification indicators:

[0130] 1. Area : the area of the loop in horizontal projection;

[0131] 2. Circular degree : , where is the perimeter;

[0132] 3. Internal average entropy : the average structural entropy of all points inside the loop

[0133] 4. Boundary gradient consistency : the average cosine similarity of gradient directions of adjacent boundary points on the loop.

[0134] Set thresholds to filter out non-building loops, with the condition:

[0135] ;

[0136] where is the minimum area, is the average entropy of all non-ground points globally, is the minimum circular degree, which is , in this implementation.

[0137] Step 19, each verified boundary loop defines a candidate building region . At the same time, the inclusion and intersection relationship between loops needs to be processed:

[0138] Inclusion relationship processing: if a loop is completely contained in another loop , establish a parent-child hierarchical relationship, marked as inner boundary, marked as outer boundary;

[0139] Intersection handling: If two rings intersect, calculate the intersection angle and area overlap ratio. If the overlap area exceeds 30% of the smaller ring's area and the intersection angle is close to perpendicular, it may be a connected building structure, and the two rings will be merged into a complex boundary.

[0140] Output the final candidate building region geometry. Each region has the following attributes: boundary ring, area, average entropy, and hierarchical relationship (i.e., a list of parent and child regions).

[0141] The process involves constructing a multimodal feature field for the candidate building region, performing feature clustering based on a Gaussian mixture model and density peaks to obtain a high-quality seed point set, including:

[0142] For each point within the candidate building area, calculate the local point density, normalized relative elevation, and final structural entropy. Perform robust statistical calculations and winsorization on each feature dimension to establish a mapping from the three-dimensional coordinate space to the feature space, and obtain the feature point set distribution.

[0143] The optimal number of components is determined by the Bayesian information criterion and a Gaussian mixture model is fitted. The local density and minimum distance of each point are calculated in the feature space. The posterior probability of the Gaussian mixture model is fused with the density peak information. Based on Mahalanobis distance, similar clusters are merged to obtain the high-quality seed point set.

[0144] In one specific embodiment, the construction and normalization of a multimodal feature field for the candidate building region includes the following steps:

[0145] Step 20, for candidate building areas Each point inside Calculate three core features:

[0146] Local point density : Among them, the search radius Set to 0.8m;

[0147] Normalized relative elevation ,in, and Candidate building areas The lowest and highest elevations of all points within the area are calculated. For areas including the inner boundary, a watershed algorithm is used to identify the main body of the roof, and the elevation range is calculated using only the roof points.

[0148] Final structural entropy The result calculated in step 6 is sufficient.

[0149] Step 21: For each feature dimension calculated in Step 20, calculate the candidate building region. Robust statistics within the inner region, including:

[0150] Median First quartile Third quartile ;

[0151] Interquartile range ;

[0152] Apply winsorization based on quantiles:

[0153] ;

[0154] Interval normalization on the processed feature values:

[0155] ;

[0156] Three normalized features are obtained: .

[0157] Step 22, for all points within the candidate building region , establish a mapping from the three-dimensional coordinate space to the feature space:

[0158] ;

[0159] ;

[0160] Record the point set distribution of the candidate building region in the feature space :

[0161] .

[0162] Feature clustering based on Gaussian mixture model and density peak value in the feature space, including the following steps:

[0163] Step 23, for the feature point set , use the Bayesian information criterion to determine the optimal number of clusters :

[0164]

[0165] where is the likelihood value of the K-component GMM, is the number of points, is the total number of parameters. Use the expectation maximization algorithm to fit ​GMM of components, each component corresponds to a latent semantic class, such as main roof, accessory structure, residual vegetation, noise, etc. And calculate the posterior probability of each point belonging to each component, get soft assignment matrix.

[0166] Step 24, calculate the local density of each point in the feature space And the minimum distance :

[0167] ;

[0168] ;

[0169] Wherein, Take 5% of the diameter of the feature space. Calculate the decision value , select the local maximum point as the candidate cluster center, fuse the posterior probability of GMM and the density peak value information, and correct the cluster center position.

[0170] Step 25, calculate the Mahalanobis distance between each cluster:

[0171]

[0172] Wherein, The global covariance matrix of the feature space. And based on the Mahalanobis distance, merge similar clusters, that is, when Merge into the same class, output the final clustering set Each cluster Includes feature space center , covariance matrix , point set in three-dimensional space .

[0173] From the cluster cluster, the discriminative screening of the building core seed point is carried out, and the high-quality seed point set of the main body of the building is extracted, including the following steps:

[0174] Step 26, for each cluster cluster , calculate the following discriminative features:

[0175] Building core index : , wherein, The mean value of the cluster in three feature dimensions, respectively, and the weight .

[0176] Spatial compactness : , wherein, The three-dimensional convex hull volume of the cluster point cloud, The surface area, high spatial compactness represents compact body structure.

[0177] planar dominance : , is the eigenvalue of the clustered point cloud covariance matrix, close to 1 indicates strong planar characteristics.

[0178] Step 27, set the constraint condition to filter out non-building cluster, the retention condition is:

[0179] ;

[0180] Establish a soft scoring system:

[0181] ;

[0182] wherein, is the three-dimensional spatial distribution entropy of the clustered points. According to , arrange in descending order, select the first as the building candidate cluster, .

[0183] Step 28, check the spatial relationship between the candidate cluster, if the two cluster in three-dimensional space distance is less than (such as 2m), and the elevation difference is less than (such as 1m), consider merging the two cluster, and calculate the comprehensive score after merging, get the final cluster .

[0184] For the building seed point set, construct a multi-level seed point structure, including the following steps:

[0185] Step 29, for each cluster determined in step 28, calculate the local building confidence of each point in the cluster : , wherein denotes Mahalanobis distance, is the normalized feature vector, is the cluster feature space center. Select the first N points with the highest confidence as the core seed point set . is the exponential function.

[0186] Step 30, around each core seed point set , in three-dimensional space, take each core point as the center, and select the points that meet the following conditions in the sphere with radius (taken as 2m in this embodiment) as the extended seed point:

[0187] ;

[0188] in, The standard deviation of the cluster entropy value. and The mean and standard deviation of the three-dimensional locations of the clusters are given. Merging all expanded points yields the expanded seed point set. Calculate the growth priority of the expansion point. :

[0189]

[0190] in, The distance to the nearest core point. The maximum distance to the nearest core point. The angle between the normal vector and the dominant plane of the cluster is denoted as .

[0191] Step 31, for each candidate building area Constructing a hierarchical seed point set :

[0192] .

[0193] The process involves seed activation based on the high-quality seed point set, absorption decision-making for candidate points based on a structural convergence dynamic evaluation method, and directional growth based on the structural entropy field to obtain the building point cloud aggregation region, including:

[0194] A weighted undirected graph is constructed based on the high-quality seed point set, and confidence propagation iteration is performed to obtain the activation sequence. An initial growth region is created starting from the seed point with the highest confidence. The baseline value of the region structure entropy and the dominant direction of the region normal vector are calculated. The active boundary set, candidate point buffer, growth history and dynamic convergence threshold are initialized.

[0195] Starting from each point in the active boundary set, search for unlabeled candidate points. For each unlabeled candidate point, calculate the local structure compatibility, regional entropy change prediction, and topological continuity index. Construct an absorption decision state vector, combine it with a pre-trained policy network and rule thresholds to make absorption decisions, and perform batch absorption to update the growing region.

[0196] In one specific embodiment, starting from high-quality seed points in the high-quality seed point set, multi-level seed activation and growth initialization are performed based on confidence propagation, including the following steps:

[0197] Step 32, for candidate building areas seed point set Construct a weighted undirected graph ,node Includes core seed point and expanding seed points ,side Connection space distance is less than connection radius ( Seed point pair with average point distance Based on three-dimensional space distance and feature similarity:

[0198] ;

[0199] Where, is the normalized feature vector of node .

[0200] Step 33, initialize the confidence of each seed point

[0201] , execute confidence propagation iteration (usually 5 times),

[0202] , where is the damping coefficient, is the field point of node . According to the final confidence , all seed points are sorted in descending order, and the activation sequence is obtained.

[0203] Step 34, create an initial growth region starting from the seed point with the highest confidence, calculate the region structure entropy baseline value of : At the same time, calculate the dominant direction of the region normal vector , obtained by the first principal component of the region point cloud PCA. Initialize the region growth state parameters, including:

[0204] Active boundary set : initially as the seed point set;

[0205] Candidate point buffer : initially empty;

[0206] Growth history record : record the change of region entropy of each iteration;

[0207] Dynamic convergence threshold : initially set to .

[0208] Based on the dynamic evaluation method of structure convergence, the absorption decision of candidate points is made, including the following steps:

[0209] Step 35, starting from each point in the active boundary set , in the radius (first analysis scale) and (unlabeled candidate points are searched in the third analysis scale). For each candidate point , its absorption evaluation features are calculated:

[0210] local structure consistency : where, is the local normal vector of the candidate point;

[0211] regional entropy change prediction : ;

[0212] topological continuity index : the minimum connection distance between the candidate point and the active boundary point is calculated , and the k-nearest neighbor overlap rate between the candidate point and the points in the region is counted , .

[0213] Step 36, construct the state vector of absorption decision:

[0214] ;

[0215] Calculate the absorption probability using the pre-trained Q-learning policy network: The network is trained in the offline stage through a large number of building point cloud samples, and the reward function is designed as:

[0216] ;

[0217] Make a final decision combined with the rule threshold, and the absorption condition is:

[0218] ;

[0219] Step 37, batch optimization is performed on the candidate point set that passes the preliminary screening:

[0220] Construct a candidate point conflict graph: if the absorption spheres (radius ) of two candidate points intersect, it is considered to be in conflict.

[0221] Solve the maximum weight independent set problem, select the non-conflict subset with the maximum total absorption benefit:

[0222] ;

[0223] where the weight . For all conflict subsets, and The sum is not greater than 1. Perform batch absorption, update the growing region and regional attributes.

[0224] Directional growth based on entropy gradient field guidance and adaptive parameter adjustment with history feedback, comprising the following steps:

[0225] Step 38, calculate the current growth region Average entropy gradient direction of boundary points:

[0226] ;

[0227] Establish a growth direction preference model, including:

[0228] Primary growth direction: along , the direction of fastest entropy reduction;

[0229] Secondary growth direction: expand along the current region plane direction;

[0230] Forbidden direction: along , the direction of increasing entropy.

[0231] Assign a growth direction weight to each boundary point:

[0232] ;

[0233] Where, is the preset preferred direction.

[0234] Step 39, monitor the convergence indicators during growth, including:

[0235] Entropy change trend: the average entropy change rate of the region in the last k iterations,

[0236] ;

[0237] Flatness retention: the consistency of the region point cloud normal vector,

[0238] ;

[0239] Where, is the eigenvalue of the region point cloud covariance matrix.

[0240] Then:

[0241] If (the entropy value rises), tighten the convergence threshold: ;

[0242] If , increase the normal vector constraint weight;

[0243] If the growth speed is too slow (more than 50 iterations), expand the search radius .

[0244] Construct a multi-condition convergence criterion to intelligently terminate growth, and calculate the comprehensive quality index of the final growth area, and construct the topological relationship diagram between regions. Including the following steps:

[0245] Step 40, establish a convergence criterion index system, and determine the convergence of growth when one of the following conditions is met:

[0246] ① Structural entropy convergence condition: , the relative change of the last 10 iterations is less than 1%;

[0247] ② Boundary stability condition: Calculate the Hausdorff distance change of the boundary point set:

[0248] ;

[0249] Take 2 times the average point distance.

[0250] ③ Candidate point exhaustion condition: , the ratio of candidate buffer area to active boundary point number is less than 0.1.

[0251] ④ Maximum iteration protection: , in this embodiment .

[0252] Step 41, calculate the comprehensive quality index of the final growth area:

[0253] ① Structural consistency score : , where is the standard deviation of the point entropy value in the region;

[0254] ② Geometric integrity score : , the ratio of projected area to the area of the smallest circumscribed rectangle , measures the compactness of the region;

[0255] ③ Boundary clarity score : .

[0256] Based on the above scores, the growth results are classified:

[0257] High-quality area: all scores >0.7, directly output;

[0258] Optimization area: the score of a certain item is between 0.5 and 0.7, and subsequent optimization is required;

[0259] Problem area: the score of a certain item is <0.5, marked as failed.

[0260] Step 42, each successfully generated growth area is

[0261] The optimal fusion of the building point cloud aggregation area is performed based on the minimum description length metric method, the three-dimensional point cloud building object after fusion is converted into a two-dimensional vector contour, and quality evaluation is performed on the two-dimensional vector contour to obtain a structured building expression, including:

[0262] The optimal fusion of the building point cloud aggregation area is performed based on the minimum description length metric method between regions, the total description length is defined as the weighted sum of the geometric description length and the structure entropy description length, the merging benefit and merging confidence evaluation of the candidate merging pair are calculated, the region merging meeting the conditions is performed, and the final building object set is obtained;

[0263] The contour polygon extraction is performed on the final building object set based on the adaptive alpha-shape contour extraction, the inner angle size is calculated by detecting the contour angle, the inner angle close to 90 degrees is adjusted to 90 degrees, the two-dimensional vector contour is obtained, the quality evaluation is performed by calculating the geometric fidelity and the area consistency, and the structured building expression is generated.

[0264] The optimal fusion of the building point cloud aggregation area is performed based on the minimum description length metric method between regions, the total description length is defined as the weighted sum of the geometric description length and the structure entropy description length, the merging benefit and merging confidence evaluation of the candidate merging pair are calculated, the region merging meeting the conditions is performed, and the final building object set is obtained, including:

[0265] The geometric description length includes boundary complexity and internal structure complexity, the structure entropy description length is calculated based on the region point entropy value distribution covariance and the KL divergence of the region distribution and the global building point distribution, the region pairs with effective adjacent distance less than a preset value are reserved based on the spatial adjacency relationship, and the region merging is performed on the regions with merging benefit less than a preset value and merging confidence greater than a preset value.

[0266] In a specific embodiment, the optimal fusion of the same building is performed on the growth region based on the minimum description length metric method between regions. Including the following steps:

[0267] Step 43, for a growth region , the total description length is defined as the weighted sum of two components:

[0268] ;

[0269] Wherein, is an adjustable weight,

[0270] The geometric description length : , ]= [0.6, 0.4]. Boundary complexity where, is the projected boundary perimeter, is the equi-area circle perimeter, is the ratio of convex hull area to actual area, measuring the concave-convex property of the boundary. Internal structure complexity where, is the local normal vector entropy, measuring the surface direction variation, measuring the elevation fluctuation.

[0271] Structure entropy description length , the first term is based on the covariance of the entropy value distribution of region points , following the information entropy formula of Gaussian distribution; the second term is the KL divergence between the region distribution and the global building point distribution , penalizing the region deviating from the typical building characteristics.

[0272] Step 44, based on spatial adjacency relationship, calculate the effective adjacency distance of all region pairs :

[0273] ;

[0274] where, is the average radius of the region horizontal projection. Retain region pairs. For each candidate merging pair , calculate the merging benefit:

[0275] ;

[0276] Introduce merging confidence evaluation:

[0277] ;

[0278] The confidence evaluation comprehensively considers the significance of description length reduction and the balance of region size.

[0279] Step 45, iteratively perform candidate merging pairs, select the best merging pair that satisfies and , and perform merging. When there are no more merging pairs that satisfy the conditions, or the maximum merging rounds are reached (such as 10 rounds), output the final building object set.

[0280] In this embodiment, the adaptive alpha-shape contour is selected for initial extraction, the three-dimensional point cloud building obtained in steps 43-45 is subjected to contour polygon extraction, and by detecting contour angles and calculating interior angle sizes, for interior angles close to 90 degrees, directly adjusting to 90 degrees, thereby ensuring the accuracy of building angles, and obtaining the final contour .

[0281] Quality assessment and topological consistency verification are performed on the final contour and a structured building representation is generated, including the following steps:

[0282] Step 46, for each building object, the final contour is calculated respectively for its geometric fidelity and area consistency .

[0283] wherein, is the shortest distance from the point to the final contour , is the average distance of the point cloud. The value close to 1 indicates that the contour can well cover the point cloud.

[0284] wherein, is the convex hull area of the point cloud.

[0285] Step 47, output the final structured building representation for subsequent use.

[0286] The application also provides a building extraction system based on multi-dimensional point cloud structure entropy phase detection, the system comprising:

[0287] A data preprocessing and structure entropy field construction module is used to denoise the original point cloud data and obtain non-ground points by using a progressive triangular net encryption filtering algorithm, and a structure entropy field is constructed by weighted fusion of geometric structure entropy, echo distribution entropy and elevation consistency entropy;

[0288] A gradient field calculation and boundary identification module is used to calculate the gradient field of the non-ground points based on the structure entropy field, identify the boundary structure based on gradient flow tracking and persistent homology, and construct a candidate building area based on a closed loop boundary ring;

[0289] A multi-modal feature clustering module is used to construct a multi-modal feature field for the candidate building area, perform feature clustering based on Gaussian mixture model and density peak value, and obtain a high-quality seed point set;

[0290] A seed activation and region growing module is used to perform seed activation based on the high-quality seed point set, perform absorption decision on candidate points based on a structure convergence dynamic evaluation method, and perform directional growth based on the structure entropy field, to obtain a building point cloud aggregation area;

[0291] The region fusion and contour extraction module is configured to perform optimal fusion on the building point cloud aggregated region based on a minimum description length metric method, convert the fused three-dimensional point cloud building object into a two-dimensional vector contour, and perform quality evaluation on the two-dimensional vector contour to obtain a structured building expression.

[0292] Specifically, the building extraction system based on multi-dimensional point cloud structure entropy phase transition detection in the embodiment realizes a full-automatic processing flow from original point cloud data to a structured building expression through cooperative work of five functional modules, and realizes efficient intelligent processing of large-scale point cloud data and automatic extraction of high-precision buildings by combining structure entropy field construction, gradient field boundary identification, multi-modal feature clustering, seed activation region growth and region fusion contour extraction.

[0293] The above merely describes preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A building extraction method based on multi-dimensional point cloud structure entropy phase transition detection, characterized in that, The method comprises the following steps: The original point cloud data is denoised and a progressive triangulation encryption filtering algorithm is used to obtain non-ground points, a structure entropy field is constructed by weighted fusion of geometric structure entropy, echo distribution entropy and height consistency entropy; The structure entropy field quantifies the structure order of the local neighborhood of the point cloud by fusing the three dimensions of geometric anisotropy, echo distribution complexity and height fluctuation consistency; Gradient field calculation is performed on the non-ground points based on the structure entropy field, boundary structures are identified based on gradient flow tracking and persistent homology, and a candidate building area is constructed based on a closed-loop boundary ring; A multi-modal feature field is constructed for the candidate building area, feature clustering is performed based on a Gaussian mixture model and a density peak value, and a high-quality seed point set is obtained; Seed activation is performed based on the high-quality seed point set, absorption decision is made on candidate points based on a structure convergence dynamic evaluation method, directional growth is performed based on the structure entropy field, and a building point cloud aggregation area is obtained. The optimal fusion of the building point cloud aggregation area is performed based on a minimum description length metric method, the fused three-dimensional point cloud building object is converted into a two-dimensional vector contour, the quality of the two-dimensional vector contour is evaluated, and a structured building expression is obtained.

2. The building extraction method based on multi-dimensional point cloud structure entropy phase transition detection according to claim 1, characterized in that, The gradient field calculation on the non-ground points based on the structure entropy field, the identification of boundary structures based on gradient flow tracking and persistent homology, and the construction of a candidate building area based on a closed-loop boundary ring comprise: Gradient field calculation and significance analysis of the three-dimensional structure entropy field of the non-ground points are performed, a scalar entropy value field of the point neighborhood is fitted by using a moving least squares method, local linear fitting coefficients are obtained by solving a weighted least squares problem, an entropy gradient vector and a gradient amplitude at each point are calculated, and a significant gradient point set that meets the gradient amplitude condition is screened out; The significant gradient point set is taken as a seed point, bidirectional streamline tracking is performed along the gradient direction, a boundary segment is obtained by using a fourth-order Runge-Kutta method to numerically integrate the streamline equation, significant boundary structures in the boundary segment are identified based on a topological feature importance evaluation scheme of persistent homology, and the candidate building area is constructed based on a closed-loop boundary ring.

3. The building extraction method based on multi-dimensional point cloud structure entropy phase transition detection according to claim 2, characterized in that, The gradient field calculation and significance analysis of the three-dimensional structure entropy field of the non-ground points, the fitting of the scalar entropy value field of the point neighborhood by using the moving least squares method, the obtaining of the local linear fitting coefficients by solving the weighted least squares problem, the calculation of the entropy gradient vector and the gradient amplitude at each point, and the screening out of the significant gradient point set that meets the gradient amplitude condition comprise: The dot product of the gradient direction and the local normal vector is calculated to determine the positive and negative gradients, the positive gradient of the positive and negative gradients indicates that the building points point to the environment corresponding to the outer boundary of the building, the negative gradient of the positive and negative gradients indicates that the environment points to the building interior corresponding to the hole or courtyard boundary, the statistics of the global gradient amplitude distribution is calculated, and the significant gradient point set that meets the condition of the gradient amplitude being greater than the mean value plus the standard deviation multiple is screened out.

4. The building extraction method based on multi-dimensional point cloud structure entropy phase transition detection according to claim 2, characterized in that, The significant gradient point set is taken as a seed point, bidirectional streamline tracking is performed along a gradient direction, a fourth-order Runge-Kutta method numerical integral streamline equation is used to obtain a boundary segment, a topological feature importance evaluation scheme based on persistent homology is used to identify a significant boundary structure in the boundary segment, and a candidate building region is constructed based on a closed-loop boundary ring, including: The positive gradient points are integrated along the positive gradient direction to simulate the entropy increasing path from the building interior to the exterior, the negative gradient points are integrated along the negative gradient direction to simulate the entropy decreasing path from the exterior to the building interior, an ordered point set is generated on each streamline, and spline fitting is performed to obtain a smooth parameterized curve representation, and a α The complex is calculated, the birth time and death time of the generators of the homology groups of each dimension are calculated, the topological features satisfying the persistence threshold are screened, and the corresponding boundary segments are extracted according to the streamline point set corresponding to the topological features.

5. The building extraction method based on multi-dimensional point cloud structure entropy phase transition detection according to claim 1, characterized in that, The building point cloud aggregation region is optimally fused based on a minimum description length metric method, a three-dimensional point cloud building object after fusion is converted into a two-dimensional vector contour, and quality evaluation is performed on the two-dimensional vector contour to obtain a structured building expression, including: The building point cloud aggregation region is optimally fused based on an inter-region minimum description length metric method, a total description length is defined as a weighted sum of a geometric description length and a structure entropy description length, a merging benefit and a merging confidence of a candidate merging pair are calculated, a region merging that meets a condition is performed, and a final building object set is obtained; A contour polygon is extracted from the final building object set based on an adaptive alpha-shape contour extraction, an inner angle size is calculated by detecting a contour angle, an inner angle close to 90 degrees is adjusted to 90 degrees, the two-dimensional vector contour is obtained, and the quality evaluation is performed by calculating a geometric fidelity and an area consistency to generate the structured building expression.

6. The building extraction method based on multi-dimensional point cloud structure entropy phase transition detection according to claim 5, characterized in that, The building point cloud aggregation region is optimally fused based on an inter-region minimum description length metric method, a total description length is defined as a weighted sum of a geometric description length and a structure entropy description length, a merging benefit and a merging confidence of a candidate merging pair are calculated, a region merging that meets a condition is performed, and a final building object set is obtained, including: The geometric description length includes a boundary complexity and an internal structure complexity, the structure entropy description length is calculated based on a region point entropy value distribution covariance and a KL divergence of a region distribution and a global building point distribution, a region pair with an effective adjacent distance less than a preset value is retained based on a spatial adjacency relationship, and a region merging is performed on a region with a merging benefit less than a preset value and a merging confidence greater than a preset value.

7. The building extraction method based on multi-dimensional point cloud structure entropy phase transition detection according to claim 1, characterized in that, The original point cloud data is denoised, and a progressive triangular net encryption filtering algorithm is used to obtain non-ground points, a structure entropy field is constructed by weighted fusion of a geometric structure entropy, a echo distribution entropy, and a height consistency entropy, including: A statistical outlier removal method is used on the original point cloud data to eliminate sensor noise and abnormal points, a voxel grid method is used to regularize and resample the point cloud to obtain resampled point cloud; A progressive triangular net encryption filtering algorithm is used to separate the resampled point cloud to obtain the non-ground points; A multi-scale adaptive neighborhood is constructed according to the characteristic scale of the building component, a final representative entropy value of each point is generated by weighted geometric mean fusion of the geometric structure entropy, the echo distribution entropy, and the height consistency entropy, and the structure entropy field is constructed based on a local terrain complexity to dynamically determine the credibility weight of the entropy value of each scale.

8. The building extraction method based on multi-dimensional point cloud structure entropy phase transition detection according to claim 1, characterized in that, The multi-modal feature field is constructed for the candidate building region, feature clustering is performed based on a Gaussian mixture model and a density peak value, and a high-quality seed point set is obtained, including: The local point density, normalized relative height and final structure entropy are calculated for each point in the candidate building area, and the robust statistical quantity calculation and winsorization processing are performed on each feature dimension to establish the mapping from the three-dimensional coordinate space to the feature space, and the feature point set distribution is obtained; The optimal component number is determined by the Bayesian information criterion to fit the Gaussian mixture model, the local density and minimum distance of each point in the feature space are calculated, the posterior probability of the Gaussian mixture model is fused with the density peak value information, and the similar clustering clusters are merged based on the Mahalanobis distance, and the high-quality seed point set is obtained.

9. The building extraction method based on multi-dimensional point cloud structure entropy phase transition detection according to claim 1, characterized in that, The seed activation is performed based on the high-quality seed point set, the candidate points are absorbed based on the dynamic evaluation method of structure convergence, and the directional growth is performed based on the structure entropy field, and the building point cloud aggregation area is obtained, including: A weighted undirected graph is constructed based on the high-quality seed point set, and the confidence propagation iteration is performed to obtain an activation sequence, an initial growth area is created from the seed point with the highest confidence, the regional structure entropy reference value and the dominant direction of the regional normal vector are calculated, and the active boundary set, the candidate point buffer area, the growth history record and the dynamic convergence threshold are initialized; Each unmarked candidate point is searched from each point in the active boundary set, the local structure compatibility, the regional entropy change prediction and the topological continuity index are calculated for each unmarked candidate point, the absorption decision state vector is constructed, the absorption decision is made in combination with the pre-trained strategy network and the rule threshold, and the batch absorption updates the growth area.

10. A building extraction system based on multi-dimensional point cloud structure entropy phase transition detection, configured to perform a building extraction method based on multi-dimensional point cloud structure entropy phase transition detection according to any one of claims 1-9, characterized in that, The system comprises: A data preprocessing and structure entropy field construction module is configured to denoise the original point cloud data and obtain non-ground points by using a progressive triangular mesh encryption filtering algorithm, and to construct a structure entropy field by weighted fusion of geometric structure entropy, echo distribution entropy and elevation consistency entropy; A gradient field calculation and boundary identification module is configured to calculate a gradient field for the non-ground points based on the structure entropy field, identify boundary structures based on gradient flow tracking and persistent homology, and construct a candidate building area based on a closed-loop boundary ring; A multi-modal feature clustering module is configured to construct a multi-modal feature field for the candidate building area, perform feature clustering based on a Gaussian mixture model and a density peak value, and obtain a high-quality seed point set; A seed activation and region growing module is configured to perform seed activation based on the high-quality seed point set, make absorption decisions for candidate points based on a dynamic evaluation method of structure convergence, and perform directional growth based on the structure entropy field, and obtain a building point cloud aggregation area. A region fusion and contour extraction module is configured to perform optimal fusion on the building point cloud aggregation area based on a minimum description length metric method, convert the fused three-dimensional point cloud building object into a two-dimensional vector contour, perform quality evaluation on the two-dimensional vector contour, and obtain a structured building representation.

Images