CN108764157B - Building laser foot point extraction method and system based on normal vector Gaussian distribution - Google Patents

Building laser foot point extraction method and system based on normal vector Gaussian distribution Download PDF

Info

Publication number
CN108764157B
CN108764157B CN201810538480.5A CN201810538480A CN108764157B CN 108764157 B CN108764157 B CN 108764157B CN 201810538480 A CN201810538480 A CN 201810538480A CN 108764157 B CN108764157 B CN 108764157B
Authority
CN
China
Prior art keywords
point
histogram
adjacent
normal vector
building
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201810538480.5A
Other languages
Chinese (zh)
Other versions
CN108764157A (en
Inventor
张良
金贵
张谦
周佳雯
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hubei University
Original Assignee
Hubei University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hubei University filed Critical Hubei University
Priority to CN201810538480.5A priority Critical patent/CN108764157B/en
Publication of CN108764157A publication Critical patent/CN108764157A/en
Application granted granted Critical
Publication of CN108764157B publication Critical patent/CN108764157B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/176—Urban or other man-made structures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

The invention relates to a building laser point cloud segmentation method and system, belongs to the technical field of building automatic extraction, and particularly relates to a building laser foot point extraction method and system based on normal vector Gaussian distribution. The invention clusters and segments the acquired airborne LiDAR point cloud data of the building area in a data-driven manner, extracts the building foot points by counting the normal vector Gaussian distribution of each segment, not only can accurately extract the building laser foot points, but also obviously improves the robustness and efficiency in a complex building structure.

Description

Building laser foot point extraction method and system based on normal vector Gaussian distribution
Technical Field
The invention relates to a building laser point cloud segmentation method and system, belongs to the technical field of building automatic extraction, and particularly relates to a building laser foot point extraction method and system based on normal vector Gaussian distribution.
Background
Airborne laser radar (LiDAR) is an active aerial remote sensing earth observation system, is an emerging technology which is developed by western countries in the early nineties And is put into commercial application, And integrates a laser range finder, a Global Positioning System (GPS) And an Inertial Measurement Unit (IMU). The technology makes a major breakthrough in the aspect of real-time acquisition of three-dimensional spatial information, and provides a brand-new technical means for acquiring geospatial information with high spatial-temporal resolution.
Automated extraction of buildings from aerial images or lidar data is a prerequisite for many geographic information system applications, such as disaster management, virtual reality, city planning, and emergency response. In which, airborne LiDAR point cloud data becomes the main data source for building extraction due to its characteristics of high efficiency, density and high accuracy.
However, due to the complexity of the building structure and the diversity of the three-dimensional scene composition, how to automatically extract the three-dimensional building with high efficiency and high precision is still a difficulty. In addition, the point cloud data of the airborne laser radar has the defects of point density feasibility, non-connectivity and the like, so that the feature selection with remarkable distinguishing degree is difficult to perform in the building extraction process.
Disclosure of Invention
The invention mainly solves the technical problems that in the prior art, a three-dimensional building is difficult to extract automatically with high precision and the characteristic selection without remarkable distinguishing degree is not available, and provides a building laser foot point extraction method and system based on normal vector Gaussian distribution.
The technical problem of the invention is mainly solved by the following technical scheme:
a building laser foot point extraction method based on normal vector Gaussian distribution comprises the following steps:
a descriptor calculation step, namely organizing laser radar point cloud data by using a Kd tree structure, and calculating a Fast Point Feature Histogram (FPFH) descriptor of a query point;
a distance clustering step, namely calculating the distance of adjacent points in a Fast Point Feature Histogram (FPFH) space by using a histogram cross kernel, and performing minimum distance clustering;
a segmentation section obtaining step, namely constructing an adjacency graph of the clustering clusters according to the adjacency relation, carrying out similarity estimation on adjacent clustering clusters based on deterministic minimum covariance determinant estimation, and carrying out region merging and segmentation on the clustering clusters based on the most similar merging principle;
a target extraction step, wherein the projection lengths of the point normal vectors of the segmentation segments on X, Y and Z axes are counted to obtain a normal vector histogram; and performing Gaussian fitting on the histogram data by using a Gaussian function, calculating a certainty coefficient R-square of the fitting function, and automatically extracting the building foot points according to a set threshold value.
In at least one embodiment of the present invention, the descriptor calculating step includes:
point cloud reorganization substep, reorganizing the tissue point cloud data by using a kd-Tree structure, and acquiring a query point P based on a proximity algorithm KNNqAll sample points in the K neighborhood sphere and K nearest neighbor points of each sample point;
and a normal calculation sub-step, namely calculating surface normals of all points in the neighborhood, and performing consistency reorientation on all normals by using the existing viewpoint information:
deviation calculation substep for the query point PqSample point of (2) and its K neighboring point pair { ps,ptDarboux transformation is carried out on (s is not equal to t), and a point p is calculated through a transformed mu v omega coordinate systemsAnd ptAnd their normal nsAnd ntThe deviation between them, represented by the triad (α, φ, θ):
Figure GDA0001788167550000031
Figure GDA0001788167550000032
Figure GDA0001788167550000033
counting the obtained triples to obtain a histogram as a Simplified Point Feature Histogram (SPFH);
influence region calculation sub-step of reassigning the query point PqThe fast feature point histogram FPFH value is weighted by the simplified point feature histogram SPFH based on the following equation, where the fast feature point histogram FPFH calculates the area of influence:
Figure GDA0001788167550000034
wherein, ω iskRepresenting a query point pqAnd neighborhood point pkThe distance in a given metric space.
In at least one embodiment of the present invention, the distance clustering step includes:
adjacent point pair obtaining sub-step, by set resolution RseedAcquiring a clustering seed point; k-Tree based acquisition of adjacent neighborhood pairs of seed points { (t)p,t1),(tp,t2),...,(tp,tk) }; wherein p is a seed point, k is kd-Tree;
an overlap calculation sub-step of counting pairs of adjacent points (t)p,tq) The fast feature point histogram FPFH descriptor of (1) calculates the degree of overlap L (H (t) of the data distribution of the local surface between the pair of points using the histogram cross kernel based on the following formulap),H(tq))；
Figure GDA0001788167550000035
Wherein the content of the first and second substances,
Figure GDA0001788167550000036
and
Figure GDA0001788167550000037
respectively representing the values of two adjacent points corresponding to the ith dimension of the FPFH descriptor;
and a category determination substep of setting the category label of the point whose degree of overlap satisfies a predetermined condition to coincide with the seed point.
In at least one embodiment of the present invention, the segment acquisition step,
an adjacency graph obtaining sub-step, namely establishing an adjacency graph G (M, E) by utilizing the three-dimensional space domain relation, wherein a node M represents the centroid coordinate of a clustering cluster, and an edge E represents the similarity of adjacent clustering clusters in a binary form;
a similarity calculation substep, which is used for obtaining the similarity between adjacent clustering clusters based on the DetMCD robust estimation;
and an adjacent edge merging sub-step, namely merging similar adjacent edges based on the adjacent image G to obtain a group of point cloud segmentation segments.
In at least one embodiment of the present invention, the target extracting step specifically includes the following sub-steps:
a normal vector obtaining sub-step, namely calculating and obtaining a normal vector of each point based on the segmentation segment obtained after segmentation and the steady estimated local point;
a histogram obtaining sub-step, namely counting the projection lengths of the normal vector on an X axis, a Y axis and a Z axis to obtain a normal vector histogram;
a building extraction sub-step, fitting the statistical histogram data by using a Gaussian function to obtain a Gaussian fitting function; and extracting the building according to the determined coefficient R-square of the fitting function.
A building laser foot point extraction system based on normal vector Gaussian distribution comprises:
the descriptor calculation module is used for organizing laser radar point cloud data by utilizing a Kd tree structure and calculating a Fast Point Feature Histogram (FPFH) descriptor of the query point;
the distance clustering module is used for calculating the distance between adjacent points in the Fast Point Feature Histogram (FPFH) space by using a histogram cross kernel and carrying out minimum distance clustering;
the segmentation section acquisition module is used for constructing an adjacency graph of the clustering clusters according to the adjacency relation, carrying out similarity estimation on the adjacent clustering clusters based on deterministic minimum covariance determinant estimation, and carrying out region merging and segmentation on the clustering clusters based on the most similar merging principle;
the target extraction module is used for counting the projection lengths of the point normal vectors of the segmentation segments on X, Y and Z axes to obtain a normal vector histogram; and performing Gaussian fitting on the histogram data by using a Gaussian function, calculating a certainty coefficient R-square of the fitting function, and automatically extracting the building foot points according to a set threshold value.
In at least one embodiment of the invention, the descriptor computation module comprises:
a point cloud reorganizing unit, which reorganizes the tissue point cloud data by using a kd-Tree structure and obtains a query point P based on a proximity algorithm KNNqAll sample points in the K neighborhood sphere and K nearest neighbor points of each sample point;
and the normal calculation unit is used for calculating surface normals of all points in the neighborhood and carrying out consistency reorientation on all normals by utilizing the existing viewpoint information:
a deviation calculation unit for calculating the query point PqSample point of (2) and its K neighboring point pair { ps,ptDarboux transformation is carried out on (s is not equal to t), and a point p is calculated through a transformed mu v omega coordinate systemsAnd ptAnd to themNormal n tosAnd ntThe deviation between them, represented by the triad (α, φ, θ):
Figure GDA0001788167550000051
Figure GDA0001788167550000052
Figure GDA0001788167550000053
counting the obtained triples to obtain a histogram as a Simplified Point Feature Histogram (SPFH);
area of influence calculation unit, reassigning query point PqThe fast feature point histogram FPFH value is weighted by the simplified point feature histogram SPFH based on the following equation, where the fast feature point histogram FPFH calculates the area of influence:
Figure GDA0001788167550000054
wherein, ω iskRepresenting a query point pqAnd neighborhood point pkThe distance in a given metric space.
In at least one embodiment of the present invention, the distance clustering module comprises:
an adjacent point pair acquisition unit for acquiring the adjacent point pairs with a set resolution RseedAcquiring a clustering seed point; k-Tree based acquisition of adjacent neighborhood pairs of seed points { (t)p,t1),(tp,t2),...,(tp,tk) }; wherein p is a seed point, and k is the number of adjacent points in the kd-Tree;
an overlap degree calculating unit for counting the adjacent point pairs (t)p,tq) The fast feature point histogram FPFH descriptor of (1) calculating the number of local surfaces between the pair of points using histogram cross-kernels based on the following equationAccording to the degree of overlap L (H (t) of the distributionp),H(tq))；
Figure GDA0001788167550000061
Wherein the content of the first and second substances,
Figure GDA0001788167550000062
and
Figure GDA0001788167550000063
respectively representing the values of two adjacent points corresponding to the ith dimension of the FPFH descriptor;
and a category determination unit configured to set a category label of a point having an overlapping degree satisfying a predetermined condition to be identical to the seed point.
In at least one embodiment of the invention, a segment acquisition module,
the adjacency graph obtaining unit is used for establishing an adjacency graph G (M, E) by utilizing the three-dimensional space domain relation, wherein the node M represents the centroid coordinate of the clustering cluster, and the edge E represents the similarity of the adjacent clustering clusters in a binary form;
the similarity calculation unit is used for obtaining the similarity between adjacent clustering clusters based on the DetMCD robust estimation;
and the adjacent edge merging unit merges similar adjacent edges based on the adjacent image G to obtain a group of point cloud segmentation sections.
In at least one embodiment of the present invention, the target extraction module specifically includes the following units:
the normal vector acquisition unit is used for calculating and acquiring a normal vector of each point based on the segmentation segment obtained after segmentation and the steady estimated local point;
the histogram acquisition unit is used for counting the projection lengths of the normal vector on an X axis, a Y axis and a Z axis to obtain a normal vector histogram;
the building extraction unit is used for fitting the statistical histogram data by utilizing a Gaussian function to obtain a Gaussian fitting function; and extracting the building according to the determined coefficient R-square of the fitting function.
Therefore, the invention has the following advantages:
1. aiming at the defects that the selected features are difficult to calculate and have wide applicability in the existing building extraction method, the invention analyzes the distribution of the normal vectors of the point sets in the segmentation sections, provides the attribute features of the normal vector Gaussian distribution with strong robustness and significance, and extracts the building foot points based on the features, thereby not only reducing the setting of artificial parameters and improving the automation degree of the method, but also enhancing the robustness of the method in a complex building structure due to the robust estimation of the features.
2. Compared with other building extraction algorithms, the building foot point extraction based on the normal vector Gaussian distribution takes the airborne laser radar data as the only data source, and extraction errors caused by registration errors among multi-source data and inconsistent data features are avoided. Meanwhile, the invention carries out the automatic extraction of the foot points in a data-driven mode, does not need to select a model in advance and can adapt to various building structures.
Drawings
FIG. 1: a flow chart of a building laser foot point automatic extraction method based on normal vector Gaussian distribution;
FIG. 2: darboux coordinate transformation diagram;
FIG. 3: an influence area schematic diagram of FPFH calculation;
FIG. 4: a histogram intersection function schematic;
FIG. 5: based on the DetMCD robust estimation segmentation result graph;
FIGS. 6(a) - (c) are X, Y, Z axial normal lengths for vegetation, respectively, (d) - (f) are X, Y, Z axial normal lengths for buildings, respectively;
fig. 7(a) - (c) are diagrams of X, Y, Z normal axis gaussian distributions of vegetation, respectively, and (d) - (f) are diagrams of X, Y, Z normal axis gaussian distributions of buildings, respectively.
Detailed Description
The technical scheme of the invention is further specifically described by the following embodiments and the accompanying drawings.
Example (b):
a building laser foot point automatic extraction method based on normal vector Gaussian distribution improves the extraction precision and robustness of building laser foot points through the significance characteristics of robust estimation, and is characterized by comprising the following steps:
preparing and installing an airborne laser radar system, and generating airborne laser radar data;
step (2) airborne LiDAR data clustering is carried out by using a fast point feature histogram descriptor;
step (3) similarity region segmentation based on the DetMCD robust estimation;
and (4) extracting the building by using the Gaussian distribution of the normal vectors of the point sets of the segments.
In the step (1), the preparation and installation of the airborne laser radar system and the generation of the airborne laser radar data comprise the following steps:
carrying a set of LiDAR system on an aviation carrier, wherein the LiDAR system comprises an Inertial Measurement Unit (IMU), a Differential Global Positioning System (DGPS), a laser scanning ranging system and an imaging device;
step (b), carrying out aviation flight on the building area according to the formulated flight scheme;
and (c) generating a theoretical model according to the airborne laser radar data to obtain airborne laser radar point cloud data in the three-dimensional scene of the building.
Wherein, in the step (2), the airborne LiDAR data clustering by using the fast point feature histogram descriptor comprises the following steps:
constructing a Kd-Tree structure of Lidar point cloud data;
acquiring a nearest neighbor point set of the query point based on the KNN;
step (c) calculating a fast point feature histogram descriptor (FPFH) of the query point according to the nearest neighbor point set obtained in the step (b);
step (d) calculating the FPFH spatial distance between adjacent points based on the histogram cross kernel;
and (e) clustering according to the FPFH space distance obtained in the step (d) to obtain a cluster.
In the step (3), the similarity region segmentation based on the detemcd robust estimation includes the following steps:
step (a) establishing an adjacency graph according to the adjacency relation for the clustering clusters obtained in step (2);
step (b) calculating the similarity between adjacent clustering clusters based on the DetMCD according to the adjacency graph in the step (a);
and (c) carrying out region segmentation according to the similarity between the clustering clusters.
In the step (4), the building extraction by using the Gaussian distribution of the normal vectors of the point sets of the segments comprises the following steps:
step (a) calculating and obtaining a normal vector of each point based on the segmentation segment obtained after segmentation in step (3) and the steady estimated local point;
counting the projection lengths of the normal vector on an X axis, a Y axis and a Z axis to obtain a normal vector histogram;
fitting the statistical histogram data by using a Gaussian function to obtain a Gaussian fitting function;
and (d) extracting the building according to the determined coefficient R-square of the fitting function.
The present embodiment will be further described with reference to the accompanying drawings.
As shown in fig. 1, an automatic building laser foot point extraction method based on normal vector gaussian distribution includes the following steps:
step 1, carrying a set of LiDAR system on an aviation carrier, wherein the LiDAR system comprises an Inertial Measurement Unit (IMU), a Differential Global Positioning System (DGPS), a laser scanning ranging system and an imaging device, and acquiring airborne laser radar point cloud data of a building scene;
step 2, organizing the laser radar data in the step 1 by using a Kd-Tree structure, and calculating an FPFH descriptor of the query point based on KNN;
step 3, calculating the distance of adjacent points in the FPFH space by using a histogram cross kernel to perform region growing clustering according to the feature descriptors calculated in the step 2;
step 4, constructing an adjacency graph of the clustering clusters according to the adjacency relation, carrying out similarity estimation on the adjacent clustering clusters based on the DetMCD, and carrying out segmentation based on the most similar combination principle;
step 5, counting the projection lengths of the point normal vectors of the segmentation segments obtained in the step 4 on X, Y and Z axes to obtain a normal vector histogram;
and 6, performing Gaussian fitting on the histogram data by using a Gaussian function, calculating a certainty coefficient R-square of the fitting function, and automatically extracting the building foot points according to a set threshold value.
In step 2, the method for calculating the FPFH descriptor of the query point based on KNN includes:
step 2.1, reorganizing the point cloud data obtained in the step 1 by using a kd-Tree structure, and processing the point data in steps 2.2-2.6;
step 2.2, obtaining a query point P based on K-NNqAll sample points in the K neighborhood sphere and K nearest neighbor points of each sample point;
step 2.3, surface normals of all points in the neighborhood are calculated, typically estimated using Principal Component Analysis (PCA):
npi＝V0,λ0≤λ1≤λ2 (1)
wherein, V0Representing the minimum eigenvalue lambda0The feature vector of (2);
step 2.4, once the normals of all the points are obtained, carrying out consistent redirection on all the normals by using the existing viewpoint information:
Figure GDA0001788167550000101
wherein n ispiRepresents a point piV is the viewpoint.
Step 2.5, for the query point PqSample point of (2) and its K neighboring point pair { ps,ptDarboux transformation (as shown in FIG. 2) is performed on (s ≠ t), and a point p is calculated by the transformed μ ν ω coordinate systemsAnd ptAnd their normal nsAnd ntThe deviation between them, represented by the triad (α, φ, θ):
Figure GDA0001788167550000111
Figure GDA0001788167550000112
Figure GDA0001788167550000113
step 2.6, performing statistics on the triplets obtained in step 2.5 to obtain a Histogram, which is called as a Simplified Point Feature Histogram (SPFH);
step 2.7, reassign query Point PqSPFH is weighted according to the FPFH value (as shown in equation (6)), wherein the FPFH calculation has an impact area as shown in fig. 3:
Figure GDA0001788167550000114
wherein, ω iskRepresenting a query point PqAnd neighborhood point PkThe distance in a given metric space.
In step 3, the method for calculating and clustering the spatial distance of the FPFH of the neighboring points by using the histogram cross kernel includes:
step 3.1, by setting the resolution RseedAcquiring clustered seed points, and performing the treatment of steps 3.2-3.6 on each seed point P;
step 3.2, acquiring adjacent near point pairs { (t) based on kd-Treep,t1),(tp,t2),...,(tp,tk) -3.6 for each point pair;
step 3.3, to the adjacent point pairs (t)p,tq) Is statistically based on the FPFH descriptorThe degree of overlap of the data distribution is calculated in the histogram cross-kernel (as shown in fig. 4), and is expressed by equation (7):
Figure GDA0001788167550000115
wherein the content of the first and second substances,
Figure GDA0001788167550000116
and
Figure GDA0001788167550000117
respectively representing the values of two adjacent points corresponding to the ith dimension of the FPFH descriptor;
step 3.4, since the greater the degree of overlap, the greater the degree of similarity between two points, the smaller the spatial distance, passing through the intersection function L (H (t)p),H(tq) The reciprocal of) is expressed as:
DFPFH(tp,tq)＝1/L(H(tp),H(tq)) (8)
step 3.5, when D is calculatedFPFH(tp,tq) If the value is less than the threshold value epsilon, setting the label of the point to be consistent with the seed point, and performing step 3.6; otherwise, performing step 3.7;
step 3.6, setting the newly added point as a seed point, and repeating the step 3.2-3.5;
and 3.7, not setting the labels of the adjacent points, and terminating the clustering process when no new labels of the points need to be set.
In step 4, the method for performing similarity estimation and data segmentation on neighboring clusters based on DetMCD includes:
step 4.1, establishing an adjacency graph G (M, E) by utilizing a three-dimensional space domain relation according to the clustering cluster obtained in the step 3, wherein a node M represents a centroid coordinate of the clustering cluster, an edge E represents the similarity of adjacent clustering clusters in a binary form, and steps 4.2-4.6 are carried out on adjacent nodes;
step 4.2, robust estimation based on DetMCDCounting to obtain a cluster point set QAAnd QBMean vector of
Figure GDA0001788167550000121
Sum covariance matrix
Figure GDA0001788167550000122
Step 4.3, calculating the Mahalanobis distance (shown as the formula (9)) of each point in the clustering cluster, eliminating gross errors, and respectively obtaining and obtaining the local point IAAnd IB：
Figure GDA0001788167550000123
Wherein the content of the first and second substances,
Figure GDA0001788167550000124
is χ with degree of freedom p and quantile alpha2Distribution, a is usually 0.975;
step 4.4, obtaining the common local point I by using a formulaAB,IAB：
Figure GDA0001788167550000125
Step 4.5, calculating the ratio M of the common local interior points:
Figure GDA0001788167550000131
wherein card () is used to return the number of collection elements;
and 4.6, assigning a value to the edge E by utilizing the ratio of the common local interior points:
Figure GDA0001788167550000132
wherein 1 represents similarity, and 0 represents no similarity;
and 4.7, based on the adjacency graph G, merging the adjacent edges E when the adjacent edges E are 1, and not merging when the adjacent edges E are 0, so as to obtain a group of segmentation segments (as shown in fig. 5).
In step 5, the projection lengths of the point normal vectors of the segments obtained in step 4 on the X, Y and Z axes are counted, and a method for obtaining a normal vector histogram includes:
step 5.1, respectively carrying out steps 5.2-5.4 on each segmentation segment obtained in step 4;
and 5.2, based on the robust estimated local interior point obtained in the step 4, performing normal vector solution by using principal component analysis (as shown in a formula (13)):
Figure GDA0001788167550000133
step 5.3, counting the normal vector
Figure GDA0001788167550000134
The projection lengths in the X-axis, Y-axis and Z-axis directions respectively are obtained to obtain a set { X }i,yi,zi}i＝1,2,...,n；
Step 5.4, respectively mixing xi,yi,ziStatistics were performed, divided into ten sets of data and plotted as histograms (as shown in fig. 6(a) - (f)).
In step 6, the method for performing gaussian fitting on the histogram data by using the gaussian function, calculating the certainty coefficient R-square of the fitting function, and automatically extracting the building foot points according to the set threshold value comprises the following steps:
step 6.1, acquiring data of the statistical histogram in the step 5, and processing each data in the steps 6.2-6.11;
step 6.2, assuming that the data distribution satisfies a unimodal gaussian distribution function:
f(x)＝A·exp(-(x-μ)2/(2σ2)) (14)
wherein, the parameters (A, mu, sigma) to be estimated are respectively peak value, peak value position and half width information of the Gaussian curve.
Step 6.3, setting radius delta k of a letter domain, performing Taylor series expansion on the objective function, then taking a quadratic function as an approximate value of the objective function, and performing the steps 6.4-6.5 by iteration by using a letter domain method;
6.4, selecting proper displacement d in the range of delta k;
step 6.5, calculating a quadratic approximation function q when the kth iterative displacement is dk(d) And to have a small value (as shown in formula (15)):
Figure GDA0001788167550000141
s.t.||d||≤Δk (16)
in the formula (I), the compound is shown in the specification,
Figure GDA0001788167550000145
and HkA Jocabian matrix and a Hessian matrix which are parameters theta ═ A, mu and sigma at the k-th iteration respectively;
and 6.6, judging whether the displacement d can reduce the quadratic approximation function, if so, updating the parameters by using a formula (17), otherwise, not needing:
θj+1＝θj+d (17)
step 6.7, judging whether the iteration termination condition is metkI < epsilon (epsilon is a set minimum value); if yes, performing step 6.8, otherwise, repeating steps 6.4-6.7;
step 6.8, returning parameter values (A, mu, sigma) to obtain a Gaussian model f (x) with optimal fitting, as shown in FIGS. 7(a) - (f);
step 6.9, calculate the fit data f (x) at this timei) And the mean value of the original data
Figure GDA0001788167550000142
SSR of the sum of squares of the differences, and the raw data y (x)i) And mean value
Figure GDA0001788167550000143
The difference is flatRecipe and SST:
Figure GDA0001788167550000144
Figure GDA0001788167550000151
step 6.10, calculating the determination coefficient R-square of the fitting function:
Figure GDA0001788167550000152
step 6.11, determining coefficients R of normal vector Gaussian distribution of the segmentation sections on the X axis, the Y axis and the Z axisx,Ry,RzSatisfies the condition (R)x≥0.9&&Ry≥0.9&&RzNot less than 0.9), the segmentation section belongs to the building and is extracted.
Table 1 is a building foot point extraction accuracy analysis table according to the present embodiment, as shown in the table, to be pixel-based, object-based (area greater than 50 m)2) And evaluating the building foot point extraction accuracy based on four geometrical evaluation methods. As can be seen from the table, the building foot point extraction method provided by the invention has higher extraction precision in the pixel-based evaluation method, the correctness is as high as 95.9%, the accuracy is 100% for basically extracting all the large buildings, and meanwhile, the extracted buildings have higher geometric precision.
TABLE 1 analysis table for extracting precision of building foot point
Figure GDA0001788167550000153
The specific embodiments described herein are merely illustrative of the spirit of the invention. Various modifications or additions may be made to the described embodiments or alternatives may be employed by those skilled in the art without departing from the spirit or ambit of the invention as defined in the appended claims.

Claims (8)

1. A building laser foot point extraction method based on normal vector Gaussian distribution is characterized by comprising the following steps:
a descriptor calculation step, namely organizing laser radar point cloud data by using a Kd tree structure, and calculating a Fast Point Feature Histogram (FPFH) descriptor of a query point;
a distance clustering step, namely calculating the distance of adjacent points in a Fast Point Feature Histogram (FPFH) space by using a histogram cross kernel, and performing minimum distance clustering;
a segmentation section obtaining step, namely constructing an adjacency graph of the clustering clusters according to the adjacency relation, carrying out similarity estimation on adjacent clustering clusters based on deterministic minimum covariance determinant estimation, and carrying out region merging and segmentation on the clustering clusters based on the most similar merging principle;
a target extraction step, wherein the projection lengths of the point normal vectors of the segmentation segments on X, Y and Z axes are counted to obtain a normal vector histogram; performing Gaussian fitting on the normal vector histogram by using a Gaussian function, calculating a certainty coefficient R-square of the fitting function, and automatically extracting building foot points according to a set threshold;
wherein the distance clustering step comprises:
adjacent point pair obtaining sub-step, by set resolution RseedAcquiring a clustering seed point; k-Tree based acquisition of adjacent neighborhood pairs of seed points { (t)p,t1),(tp,t2),...,(tp,tk) }; wherein p is a seed point, and k is the number of adjacent points in the kd-Tree;
an overlap calculation sub-step of counting pairs of adjacent points (t)p,tq) The fast feature point histogram FPFH descriptor of (1) calculates the degree of overlap L (H (t) of the data distribution of the local surface between the pair of points using the histogram cross kernel based on the following formulap),H(tq))；
Figure FDA0003299084950000021
Wherein the content of the first and second substances,
Figure FDA0003299084950000022
and
Figure FDA0003299084950000023
respectively representing the values of two adjacent points corresponding to the ith dimension of the FPFH descriptor;
and a category determination substep of setting the category label of the point whose degree of overlap satisfies a predetermined condition to coincide with the seed point.
2. The building laser foot point extraction method based on normal vector Gaussian distribution as claimed in claim 1, wherein the descriptor calculation step comprises:
point cloud reorganization substep, reorganizing point cloud data by using a kd-Tree structure, and acquiring a query point P based on a proximity algorithm KNNqAll sample points in the K neighborhood sphere and K nearest neighbor points of each sample point;
and a normal calculation sub-step, namely calculating surface normals of all points in the neighborhood, and performing consistency reorientation on all normals by using the existing viewpoint information:
deviation calculation substep for the query point PqSample point of (2) and its K neighboring point pair { ps,ptDarboux transformation is carried out on (s is not equal to t), and a point p is calculated through a transformed mu v omega coordinate systemsAnd ptAnd their normal nsAnd ntThe deviation between them, represented by the triad (α, φ, θ):
Figure FDA0003299084950000024
Figure FDA0003299084950000025
Figure FDA0003299084950000026
counting the obtained triples to obtain a histogram as a Simplified Point Feature Histogram (SPFH);
influence region calculation sub-step of reassigning the query point PqThe fast feature point histogram FPFH value is weighted by the simplified point feature histogram SPFH based on the following equation, where the fast feature point histogram FPFH calculates the area of influence:
Figure FDA0003299084950000031
wherein, ω iskRepresenting a query point pqAnd neighborhood point pkThe distance in a given metric space.
3. The building laser foot point extraction method based on normal vector Gaussian distribution as claimed in claim 1, characterized by the segment acquisition step,
an adjacency graph obtaining sub-step, namely establishing an adjacency graph G (M, E) by utilizing the three-dimensional space domain relation, wherein a node M represents the centroid coordinate of a clustering cluster, and an edge E represents the similarity of adjacent clustering clusters in a binary form;
a similarity calculation substep, which is used for obtaining the similarity between adjacent clustering clusters based on the DetMCD robust estimation;
and an adjacent edge merging sub-step, namely merging similar adjacent edges based on the adjacent image G to obtain a group of point cloud segmentation segments.
4. The building laser foot point extraction method based on normal vector Gaussian distribution as claimed in claim 1, wherein the target extraction step specifically comprises the following substeps:
a normal vector obtaining sub-step, namely calculating and obtaining a normal vector of each point based on the segmentation segment obtained after segmentation and the steady estimated local point;
a histogram obtaining sub-step, namely counting the projection lengths of the normal vector on an X axis, a Y axis and a Z axis to obtain a normal vector histogram;
a building extraction sub-step, fitting the statistical histogram data by using a Gaussian function to obtain a Gaussian fitting function; and extracting the building according to the determined coefficient R-square of the fitting function.
5. A building laser foot point extraction system based on normal vector Gaussian distribution is characterized by comprising:
the descriptor calculation module is used for organizing laser radar point cloud data by utilizing a Kd tree structure and calculating a Fast Point Feature Histogram (FPFH) descriptor of the query point;
the distance clustering module is used for calculating the distance between adjacent points in the Fast Point Feature Histogram (FPFH) space by using a histogram cross kernel and carrying out minimum distance clustering;
the segmentation section acquisition module is used for constructing an adjacency graph of the clustering clusters according to the adjacency relation, carrying out similarity estimation on the adjacent clustering clusters based on deterministic minimum covariance determinant estimation, and carrying out region merging and segmentation on the clustering clusters based on the most similar merging principle;
the target extraction module is used for counting the projection lengths of the point normal vectors of the segmentation segments on X, Y and Z axes to obtain a normal vector histogram; performing Gaussian fitting on the normal vector histogram by using a Gaussian function, calculating a certainty coefficient R-square of the fitting function, and automatically extracting building foot points according to a set threshold;
wherein the distance clustering module comprises:
an adjacent point pair acquisition unit for acquiring the adjacent point pairs with a set resolution RseedAcquiring a clustering seed point; k-Tree based acquisition of adjacent neighborhood pairs of seed points { (t)p,t1),(tp,t2),...,(tp,tk) }; wherein p is a seed point, and k is the number of adjacent points in the kd-Tree;
an overlap degree calculating unit for counting the adjacent point pairs (t)p,tq) Fast feature point histogram (FPFH) descriptionAnd calculating an overlap degree L (H (t) of data distribution of the local surface between the pair of points by using a histogram cross kernel based on the following equationp),H(tq))；
Figure FDA0003299084950000041
Wherein the content of the first and second substances,
Figure FDA0003299084950000042
and
Figure FDA0003299084950000043
respectively representing the values of two adjacent points corresponding to the ith dimension of the FPFH descriptor;
and a category determination unit configured to set a category label of a point having an overlapping degree satisfying a predetermined condition to be identical to the seed point.
6. The building laser foot point extraction system based on normal vector Gaussian distribution as claimed in claim 5, wherein the descriptor computation module comprises:
a point cloud reorganizing unit for reorganizing the point cloud data by using the kd-Tree structure and acquiring a query point P based on a proximity algorithm KNNqAll sample points in the K neighborhood sphere and K nearest neighbor points of each sample point;
and the normal calculation unit is used for calculating surface normals of all points in the neighborhood and carrying out consistency reorientation on all normals by utilizing the existing viewpoint information:
a deviation calculation unit for calculating the query point PqSample point of (2) and its K neighboring point pair { ps,ptDarboux transformation is carried out on (s is not equal to t), and a point p is calculated through a transformed mu v omega coordinate systemsAnd ptAnd their normal nsAnd ntThe deviation between them, represented by the triad (α, φ, θ):
Figure FDA0003299084950000051
Figure FDA0003299084950000052
Figure FDA0003299084950000053
counting the obtained triples to obtain a histogram as a Simplified Point Feature Histogram (SPFH);
area of influence calculation unit, reassigning query point PqThe fast feature point histogram FPFH value is weighted by the simplified point feature histogram SPFH based on the following equation, where the fast feature point histogram FPFH calculates the area of influence:
Figure FDA0003299084950000054
wherein, ω iskRepresenting a query point pqAnd neighborhood point pkThe distance in a given metric space.
7. The building laser foot point extraction system based on normal vector Gaussian distribution as claimed in claim 5, wherein the segment acquisition module,
the adjacency graph obtaining unit is used for establishing an adjacency graph G (M, E) by utilizing the three-dimensional space domain relation, wherein the node M represents the centroid coordinate of the clustering cluster, and the edge E represents the similarity of the adjacent clustering clusters in a binary form;
the similarity calculation unit is used for obtaining the similarity between adjacent clustering clusters based on the DetMCD robust estimation;
and the adjacent edge merging unit merges similar adjacent edges based on the adjacent image G to obtain a group of point cloud segmentation sections.
8. The building laser foot point extraction system based on normal vector Gaussian distribution as claimed in claim 5, wherein the target extraction module specifically comprises the following units:
the normal vector acquisition unit is used for calculating and acquiring a normal vector of each point based on the segmentation segment obtained after segmentation and the steady estimated local point;
the histogram acquisition unit is used for counting the projection lengths of the normal vector on an X axis, a Y axis and a Z axis to obtain a normal vector histogram;
the building extraction unit is used for fitting the statistical histogram data by utilizing a Gaussian function to obtain a Gaussian fitting function; and extracting the building according to the determined coefficient R-square of the fitting function.
CN201810538480.5A 2018-05-30 2018-05-30 Building laser foot point extraction method and system based on normal vector Gaussian distribution Active CN108764157B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810538480.5A CN108764157B (en) 2018-05-30 2018-05-30 Building laser foot point extraction method and system based on normal vector Gaussian distribution

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810538480.5A CN108764157B (en) 2018-05-30 2018-05-30 Building laser foot point extraction method and system based on normal vector Gaussian distribution

Publications (2)

Publication Number Publication Date
CN108764157A CN108764157A (en) 2018-11-06
CN108764157B true CN108764157B (en) 2022-01-14

Family

ID=64003970

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810538480.5A Active CN108764157B (en) 2018-05-30 2018-05-30 Building laser foot point extraction method and system based on normal vector Gaussian distribution

Country Status (1)

Country Link
CN (1) CN108764157B (en)

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109887024A (en) * 2019-02-16 2019-06-14 西南科技大学 A kind of cloud normal estimates new method
CN110097598B (en) * 2019-04-11 2021-09-07 暨南大学 Three-dimensional object pose estimation method based on PVFH (geometric spatial gradient frequency) features
CN110363743B (en) * 2019-06-10 2021-08-10 长安大学 Surface texture separation method based on laser three-dimensional data of asphalt concrete pavement
CN114463338B (en) * 2022-01-07 2024-05-03 武汉大学 Automatic building laser foot point extraction method based on graph cutting and post-processing
CN114463638B (en) * 2022-02-23 2022-09-20 中国科学院空天信息创新研究院 Geometric correction method for airborne interferometric synthetic aperture radar image

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104299260B (en) * 2014-09-10 2017-05-17 西南交通大学 Contact network three-dimensional reconstruction method based on SIFT and LBP point cloud registration
US10259164B2 (en) * 2016-06-22 2019-04-16 Massachusetts Institute Of Technology Methods and apparatus for 3D printing of point cloud data
CN106407925B (en) * 2016-09-09 2019-09-27 厦门大学 Laser scanning point cloud trees extraction method based on local section maximum
CN107123161A (en) * 2017-06-14 2017-09-01 西南交通大学 A kind of the whole network three-dimensional rebuilding method of contact net zero based on NARF and FPFH
CN108022262A (en) * 2017-11-16 2018-05-11 天津大学 A kind of point cloud registration method based on neighborhood of a point center of gravity vector characteristics

Also Published As

Publication number Publication date
CN108764157A (en) 2018-11-06

Similar Documents

Publication Publication Date Title
CN108764157B (en) Building laser foot point extraction method and system based on normal vector Gaussian distribution
Biosca et al. Unsupervised robust planar segmentation of terrestrial laser scanner point clouds based on fuzzy clustering methods
Yang et al. An individual tree segmentation method based on watershed algorithm and three-dimensional spatial distribution analysis from airborne LiDAR point clouds
WO2021143778A1 (en) Positioning method based on laser radar
Wang et al. Modeling indoor spaces using decomposition and reconstruction of structural elements
WO2024077812A1 (en) Single building three-dimensional reconstruction method based on point cloud semantic segmentation and structure fitting
US9942535B2 (en) Method for 3D scene structure modeling and camera registration from single image
CN112526513B (en) Millimeter wave radar environment map construction method and device based on clustering algorithm
Matei et al. Building segmentation for densely built urban regions using aerial lidar data
CN110992341A (en) Segmentation-based airborne LiDAR point cloud building extraction method
CN110599489A (en) Target space positioning method
CN112070769A (en) Layered point cloud segmentation method based on DBSCAN
CN111781608A (en) Moving target detection method and system based on FMCW laser radar
WO2010042466A1 (en) Apparatus and method for classifying point cloud data based on principal axes
CN114526739A (en) Mobile robot indoor repositioning method, computer device and product
CN116258857A (en) Outdoor tree-oriented laser point cloud segmentation and extraction method
Gilani et al. Robust building roof segmentation using airborne point cloud data
Jiang et al. Learned local features for structure from motion of uav images: A comparative evaluation
Lu et al. A lightweight real-time 3D LiDAR SLAM for autonomous vehicles in large-scale urban environment
CN112070787B (en) Aviation three-dimensional point cloud plane segmentation method based on opponent reasoning theory
CN114332172A (en) Improved laser point cloud registration method based on covariance matrix
CN109508674A (en) Airborne lower view isomery image matching method based on region division
Wang et al. Self-calibration of traffic surveillance cameras based on moving vehicle appearance and 3-D vehicle modeling
Sun et al. Automated segmentation of LiDAR point clouds for building rooftop extraction
CN115619977A (en) High-order dangerous rock monitoring method based on airborne laser radar

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant