WO2019242174A1

WO2019242174A1 - Method for automatically detecting building structure and generating 3d model based on laser radar

Info

Publication number: WO2019242174A1
Application number: PCT/CN2018/110825
Authority: WO
Inventors: 张平; 郑蓝翔; 李方; 杜广龙
Original assignee: 华南理工大学
Priority date: 2018-06-21
Filing date: 2018-10-18
Publication date: 2019-12-26
Also published as: CN109509256A; CN109509256B

Abstract

A method for automatically detecting a building structure and generating a 3D model based on a laser radar comprises the following steps: acquiring three-dimensional data of a building structure; performing matching and fusion operations on multiple point clouds; performing pre-processing of the point clouds; extracting a region growing plane; generating an editable 3D model of the building structure; and calibrating the editable 3D model. The invention can effectively remove complex obstacles, thereby building a highly precise editable 3D model of a building structure.

Description

Automatic measurement of building structure based on lidar and 3D model generation method

Technical field

The invention relates to the technical field of three-dimensional building models, in particular to a method based on a lidar sensor, which is applied to the detection of indoor and outdoor building structures and the generation of a 3D editable model.

Background technique

Three-dimensional building model is an important content of urban planning and digital city construction. The 3D building model can be simply divided into an outdoor 3D building model and an indoor 3D building model. Through the outdoor building model, the building can be simulated and maintained and evaluated online, which facilitates the generation of high-precision 3D maps. By simulating and decorating the interior model, and feeling the decoration effect in advance, it is helpful for the construction of a digital and modern city.

At present, the commonly used building reconstruction methods include vision reconstruction based on depth camera and point cloud reconstruction based on lidar. The domestic work on the reconstruction of lidar buildings mainly includes: "Swing Lidar-based Indoor 3D Point Cloud Map Generation System and Method" by Shanghai Jiaotong University (public number: CN106199626A), and Heilongjiang Institute of Technology's "Based on Lidar and Quadcopter Multi-resolution indoor three-dimensional scene reconstruction device and method "(publication number: CN104503339A)," A depth camera-based 3D modeling method and device for large-scale scenes "by Hangzhou Guangpai Intelligent Technology Co., Ltd. (public number: CN106997614A )Wait. Although the existing lidar building reconstruction schemes have already achieved good results, it is difficult to filter out complex indoor and outdoor scene interference and obtain intuitive and editable building structure models.

Summary of the Invention

The purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art, and provide a method for automatic measurement of building structures and 3D building model generation based on lidar, which can effectively filter out complex obstacles and build a highly accurate editable 3D building. The structure model avoids artificially measuring the size of the indoor building structure and drawing the indoor structure drawing, saving time and effort.

The object of the present invention is achieved by the following technical solutions:

A method for automatically measuring and generating a 3D model of a building structure based on lidar includes the following steps:

S1. Acquisition of three-dimensional data of building structure: acquisition of three-dimensional point cloud data through two-dimensional structural movement and two-dimensional laser measurement;

S2. Multi-point cloud matching fusion: Matching and fusion of three-dimensional measurement points measured at different positions of the same building structure through matching algorithms to form a complete and dense three-dimensional point cloud data set of building structures;

S3. Point cloud preprocessing: remove interference points and abnormal points due to the accuracy of the lidar sensor and measurement noise;

S4. Extraction of regional growth planes: Each point in the point cloud is clustered according to its characteristics, and planes are extracted from different categories;

S5. Editable building structure 3D model generation: merge and stitch the extracted planes, filter out various obstacles, and generate a 3D model of the building structure.

Preferably, in step S1, a two-dimensional laser radar is carried by a swinging mechanism or a rotating mechanism for reciprocating or circular motion, and the three-dimensional point cloud data of the building structure is obtained by swinging or rotating.

Preferably, step S2 specifically includes:

S2-1. Standard point cloud information conversion: Lidar measurement data is polar coordinate data with lidar as the coordinate origin, that is, distance information ρ and angle information θ from obstacles, and convert it into standard point cloud information, that is, Space rectangular coordinate system data (x, y, z) with the origin as the lidar;

S2-2. Point cloud matching: The nearest point iterative algorithm is used to perform data matching on multiple measured point clouds. By finding the corresponding point pairs between the source point cloud and the target point cloud, a rotation translation matrix is constructed based on the corresponding point pairs. , And use the obtained matrix to transform the source point cloud into the coordinate system of the target point cloud, and estimate the error function between the source point cloud and the target point cloud after the transformation. If the error function value is greater than the threshold, iteratively perform the above operations until it is satisfied Given error requirements;

S2-3. Point cloud fusion.

Preferably, step S3 specifically includes:

S3-1. Remove outliers: For each point in the point cloud, calculate the average distance from all neighboring points. Points with an average distance outside the standard range can be defined as outliers and can be removed from the data set. ;

S3-2. Remove the point cloud curvature and point normal anomaly: after obtaining the normal vector of each point in the point cloud, remove the point where the normal vector does not exist.

Specifically, in S3-2, the surface normal is approximately inferred from the point cloud data set, and the normal vector of a point in the point cloud is approximated to estimate a tangent plane normal of the point cloud surface, which can be converted into a least square method The plane fitting estimation problem is to analyze the eigenvectors and eigenvalues of a covariance matrix. This covariance matrix is created from the nearest neighbors of the query point. After obtaining the normal vector of each point in the point cloud, the normal vector is not removed. The point of existence.

Preferably, step S4 specifically includes:

S4-1. Regional growth clustering: Use the regional growth algorithm to cluster point clouds, including:

1) Suppose a point is a seed point, and the points around the seed are compared with the seed;

2) Whether the normal directions are close enough;

3) Whether the curvature is small enough;

4) If both 2) and 3) are satisfied, this point can be used as a seed;

5) If only 2) is satisfied, classify without seeding;

6) Starting from a certain seed, which satisfies the conditions 2), 3), the "sub-seed" no longer appears, then the aggregation is completed;

7) Limit the size of the class;

S4-2, RANSAC plane extraction: The RANSAC algorithm plane extraction is performed on the extracted classes, and the parameters of the plane model are estimated from a set of clusters containing outliers in an iterative manner, and the border and area of the plane are calculated.

Preferably, step S5 specifically includes:

S5-1. Remove the clusters whose extracted area is smaller than a certain threshold α and the number of planes is larger than the threshold ε;

S5-2. Merge similar planes, and merge planes whose plane parameter variance is less than a certain threshold β into the same plane, and modify their borders and areas;

S5-3. Calculate the intersection point of two adjacent planes, extend each plane to the intersection to obtain a 3D model of the building structure, and save the 3D model as a general editable graphic file format.

Preferably, the editable building structure 3D model generation step further includes an editable 3D model calibration step, calibrating the dimensions of the lidar and the generated 3D model, and automatically adjusting the model size.

Specifically, the distance information and angle information measured by the lidar are calibrated to remove the zero offset.

Specifically, the size calibration will compare the generated model size with the actual size, fine-tune the proportion of the model, and get the final 3D model of the building structure.

Compared with the prior art, the present invention has the following advantages and beneficial effects:

The present invention is dedicated to generating a high-precision editable three-dimensional building model. The present invention can use two-dimensional lidar instead of expensive three-dimensional lidar to obtain the point cloud information of the building structure, and finally obtain a high-precision editable building structure 3D model after a series of processing. The invention can filter out most obstacles and automatically generate an editable building 3D model. The invention supports the measurement and generation of building structures of various indoor and outdoor structures, and can effectively reduce the time and material consumption of manual measurement and drawing of a three-dimensional room model.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a general flowchart of the embodiment scheme.

FIG. 2 is a multi-point cloud matching fusion flowchart according to the embodiment.

FIG. 3 is a flowchart of point cloud preprocessing according to the embodiment.

FIG. 4 is a flowchart of extracting a region growth plane according to the embodiment.

FIG. 5 is a flowchart of region growth clustering in the embodiment.

FIG. 6 is a flowchart of generating a 3D model of an editable building structure according to an embodiment.

FIG. 7 is a flowchart of complex cluster filtering in the embodiment.

FIG. 8 is an editable 3D model calibration flowchart of the embodiment.

FIG. 9 is a schematic diagram of a rotating mechanism according to the embodiment.

FIG. 10 is a schematic diagram of a rocking mechanism according to an embodiment.

detailed description

The present invention is described in further detail below with reference to the embodiments and the accompanying drawings, but the embodiments of the present invention are not limited thereto.

Example 1

This method is mainly applied to automatic detection of building structures and 3D model generation. The general flow chart is shown in Figure 1. The two-dimensional lidar is fixed on the swing mechanism (as shown in Figure 10) or the rotation mechanism (as shown in Figure 9) to obtain the three-dimensional point cloud information of the building structure. After the information, multi-point cloud matching fusion is performed on the point cloud data to obtain high-density and complete building structure information. After preprocessing the point cloud, the area growth algorithm is used to cluster the point cloud, remove complex clusters, and perform plane extraction for each class after clustering. A 3D model of the building structure is generated by merging and splicing the planes. Finally, the model is modified by dimensional calibration to finally generate a high-precision 3D model of the building structure. The method flow can be divided into the following six parts: three-dimensional data acquisition of building structure, multi-point cloud matching fusion, point cloud preprocessing, area growth plane extraction, editable building structure 3D model generation, editable 3D model calibration. The specific implementation of each part is as follows:

1. 3D data acquisition of building structure

The two-dimensional lidar is combined with a rotating mechanism or a swinging mechanism, and the two-dimensional lidar can scan the three-dimensional data of the room through periodic reciprocating motion or uniform speed rotation. As shown in Figure 9, the rotating mechanism carries the lidar with the center of the rotating table as The origin makes a uniform circular motion; as shown in FIG. 10, the swing mechanism carries a lidar for reciprocating motion. The data measured by the lidar is sent to the PC through the serial port, and the data transmission follows the serial port RS232 protocol.

2.Multipoint cloud matching fusion

Lidar measurement data is polar coordinate data with the lidar as the origin, that is, distance information ρ and angle information θ from the obstacle, so it needs to be converted into standard point cloud information, that is, the space rectangular coordinates with the lidar as the origin System data (x, y, z).

The lidar emits a laser beam to detect the position of the target during work. There is a certain angular interval between adjacent laser beams. Therefore, the lidar scan will show a feature of denser close-up points and sparse far-off points, so a single scan The obtained point cloud will appear locally dense and sparse. Multiple 3D point clouds of different building structures can be obtained by scanning at different positions multiple times. Multiple point clouds are fused by the point cloud matching algorithm to obtain higher density and more complete structure information.

Because the point cloud of a single scan of the lidar will exhibit the characteristics of uneven density, which affects the generation of subsequent models, multiple measurements need to be performed at different locations in the same building structure. The nearest point iteration algorithm (Iterative Closest Point, ICP) is used to perform data matching on multiple point clouds measured. ICP is characterized by obtaining corresponding point pairs between the source point cloud and the target point cloud, and based on the corresponding points. The rotation and translation matrix is constructed, and the source point cloud is transformed into the coordinate system of the target point cloud by using the obtained matrix. The error function between the source point cloud and the target point cloud is estimated after the transformation. If the error function value is greater than the threshold, iterate The above operation is performed until the given error requirement is satisfied. After matching and fusion, the density of the point cloud is more uniform than that of a single point cloud, and the number of points in the point cloud also increases significantly. Multipoint cloud matching fusion flowchart is shown in Figure 2.

3.Point cloud preprocessing

Since the measurement of the lidar will be affected by the accuracy and introduce noise errors, it is necessary to preprocess the fused point cloud. The preprocessing flowchart is shown in Figure 3. Due to the existence of noise in the results of the lidar scan, the point cloud matching fusion cannot completely match the fusion, so a small amount of useless points are mixed in the obtained point cloud. By removing outliers in the point cloud, calculating the curvature and normal vector of each point in the point cloud, and removing points without the normal vector, the obtained point cloud is pre-processed.

For each point in the point cloud, calculate its average distance from all nearby points. Assume that the result is a Gaussian distribution whose shape is determined by the mean and standard deviation. Points whose average distance is outside the standard range (defined by the global distance average and variance) can be defined as outliers and can be collected from the data set. Remove it.

The invention directly infers the surface normal from the point cloud data set directly. Calculating the normal vector of a point in a point cloud approximates the tangent plane normal of the point cloud surface, so it can be converted into a least squares plane fitting estimation problem, that is, analyzing the eigenvectors and eigenvalues of a covariance matrix (Or PCA-principal component analysis), this covariance matrix is created from the nearest neighbors of the query point. After obtaining the normal vector of each point in the point cloud, remove the points where the normal vector does not exist.

4.Regional growth plane extraction

The points in the point cloud are clustered according to their characteristics, and planes are extracted from different categories. The points in the point cloud are distributed on multiple objects and planes, and sub-plane extraction is needed to obtain structures such as walls and floors in the building structure, as shown in Figure 4. Due to incomplete matching and fusion of the pre-processed point cloud, the phenomenon of multiple wall surfaces will be caused, that is, one plane is divided into multiple planes, so the point cloud is extracted to classify the point cloud. Here, the region growth algorithm is used to cluster the point cloud, and the plane is extracted after clustering. The area growth algorithm determines whether a point belongs to this class by calculating the normal and curvature obtained. The algorithm flow is as follows, as shown in Figure 5:

2) Whether the normal directions are close enough;

3) Whether the curvature is small enough;

4) If 2) and 3) are satisfied, this point can be used as a seed;

5) If only 2) is satisfied, classify without seeding;

7) Limit the size of the class.

The RANSAC (Random Sample) Consensus algorithm plane extraction is performed on the extracted classes, and the parameters of the plane model are estimated from a set of clusters containing outliers (points not on the same plane) in an iterative manner, and the plane is calculated Border and area.

5.Editable building structure 3D model generation

Merging and splicing the extracted planes, filtering out various obstacles, and generating a 3D model of the building structure, as shown in FIG. 6. Calculate the area of the extracted sub-planes and delete planes smaller than a certain threshold. The approximate planes are merged, and the planes whose adjacent areas are larger than a certain threshold are calculated. The outermost planes are selected as the adjacent walls, and finally all the walls are connected to obtain their 3D models.

The planar features of building structures are generally simple, while the planar features of obstacles (tables, stools, trees) are generally more complex. Therefore, the clusters with areas smaller than a certain threshold α and larger than the threshold ε are removed from the cluster. As shown in Figure 7. The planes with similar spatial positions in the remaining planes and the variance of the plane parameters less than a certain threshold β are merged into the same plane, and the frame and area thereof are modified. Calculate the intersection point between the plane with an area larger than a certain threshold χ and the adjacent plane in the remaining plane, and expand the plane to obtain the final 3D model.

6.Editable 3D model calibration

The size of the lidar and the generated model is calibrated, and the model size is automatically adjusted. A lidar sensor is a radar system that emits a laser beam to detect the target position, velocity, and other characteristic quantities. Its measurement accuracy is affected by the working environment light and service life, so it needs to be calibrated before use. After the model is generated, the size of the 3D model can be calibrated, and the 3D model can be automatically adjusted according to the size ratio of each face of the generated 3D model. The 3D model meets the format requirements of common editable graphics files, such as "STEP", "STL", "VRML" and other formats, which can be modified in a variety of editable software.

Lidar measurement errors are generally zero offset errors, that is, the measurement distance is increased or decreased by a fixed value compared to the actual distance, and the measurement angle is increased or decreased by a fixed value compared to the actual distance. Therefore, before using lidar to measure data, record the distance from the lidar to an obstacle and the angle between the obstruction and the positive direction of the lidar, and compare the error offset with the actual value to obtain an error offset. Add the offset to the data. Measure the actual size of the room, compare the corresponding side in the model with the input size, and modify the length of each side in the model in proportion to generate a 3D model of the final building structure. The generated 3D model of the building structure is saved in accordance with the general editable graphic file format requirements, so as to facilitate modification and editing of the model.

The above embodiment is a preferred embodiment of the present invention, but the embodiment of the present invention is not limited by the above embodiment. Any other changes, modifications, substitutions, combinations, and modifications made without departing from the spirit and principle of the present invention, Simplified, all should be equivalent replacement methods, and all are included in the protection scope of the present invention.

Claims

The method of automatic measurement and 3D model generation of building structures based on lidar is characterized in that it includes the following steps:

S1. Acquisition of three-dimensional data of building structure: acquisition of three-dimensional point cloud data through two-dimensional structural movement and two-dimensional laser measurement;

S2. Multi-point cloud matching fusion: Matching and fusion of three-dimensional measurement points measured at different positions of the same building structure through matching algorithms to form a complete and dense three-dimensional point cloud data set of building structures;

S3. Point cloud preprocessing: remove interference points and abnormal points due to the accuracy of the lidar sensor and measurement noise;

S4. Extraction of regional growth planes: Each point in the point cloud is clustered according to its characteristics, and planes are extracted from different categories;

S5. Editable building structure 3D model generation: merge and stitch the extracted planes, filter out various obstacles, and generate a 3D model of the building structure.
The method for automatically measuring and generating a 3D model of a building structure based on lidar according to claim 1, wherein in step S1, a two-dimensional lidar is carried by a swinging mechanism or a rotating mechanism to perform a reciprocating or circular motion, and the swinging or rotating Obtain 3D point cloud data of building structure.
The method for automatically measuring and generating a 3D model of a building structure based on lidar according to claim 1, wherein step S2 specifically comprises:

S2-1. Standard point cloud information conversion: Lidar measurement data is polar coordinate data with lidar as the coordinate origin, that is, distance information ρ and angle information θ from obstacles, and convert it into standard point cloud information, that is, Space rectangular coordinate system data (x, y, z) with the origin as the lidar;

S2-2. Point cloud matching: The nearest point iterative algorithm is used to perform data matching on multiple measured point clouds. By finding the corresponding point pairs between the source point cloud and the target point cloud, a rotation translation matrix is constructed based on the corresponding point pairs. , And use the obtained matrix to transform the source point cloud into the coordinate system of the target point cloud, and estimate the error function between the source point cloud and the target point cloud after the transformation. If the error function value is greater than the threshold, iteratively perform the above operations until it is satisfied Given error requirements;

S2-3. Point cloud fusion.
The method for automatically measuring and generating a 3D model of a building structure based on lidar according to claim 1, wherein step S3 specifically comprises:

S3-1. Remove outliers: For each point in the point cloud, calculate the average distance from all neighboring points. Points with an average distance outside the standard range can be defined as outliers and can be removed from the data set. ;

S3-2. Remove the point cloud curvature and point normal anomaly: after obtaining the normal vector of each point in the point cloud, remove the point where the normal vector does not exist.
The method for automatic measurement and 3D model generation of a building structure based on lidar according to claim 4, characterized in that in S3-2, the surface normal is approximately inferred from the point cloud data set, and a normal vector of a point in the point cloud is calculated It is approximated by estimating a tangent normal to the surface of the point cloud, which can be transformed into a least squares plane fitting estimation problem, that is, analyzing the eigenvectors and eigenvalues of a covariance matrix. This covariance matrix Created in the element, after obtaining the normal vector of each point in the point cloud, remove the points where the normal vector does not exist.
The method for automatically measuring and generating a 3D model of a building structure based on lidar according to claim 1, wherein step S4 specifically comprises:

S4-1. Regional growth clustering: Use the regional growth algorithm to cluster point clouds, including: assuming a certain point is a seed point, and the points around the seed are compared with the seed: 1) whether the normal direction is close enough; 2) Whether the curvature is small enough;

If both 1) and 2) are satisfied at the same time, this point can be used as a seed; if only 1) is satisfied, they are classified without seeding; starting from a certain seed, the "sub-seeds" that satisfy the conditions 1) and 2) are no longer When it appears, the aggregation of one class is completed; at the same time, the size of the class needs to be limited;

S4-2, RANSAC plane extraction: The RANSAC algorithm plane extraction is performed on the extracted classes, and the parameters of the plane model are estimated from a set of clusters containing outliers in an iterative manner, and the border and area of the plane are calculated.
The method for automatically measuring and generating a 3D model of a building structure based on lidar according to claim 1, wherein step S5 specifically comprises:

S5-1. Remove the clusters whose extracted area is smaller than a certain threshold α and the number of planes is larger than the threshold ε;

S5-2. Merge similar planes, and merge planes whose plane parameter variance is less than a certain threshold β into the same plane, and modify their borders and areas;

S5-3. Calculate the intersection point of two adjacent planes, extend each plane to the intersection to obtain a 3D model of the building structure, and save the 3D model as a general editable graphic file format.
The method for automatically measuring and generating a 3D model of a building structure based on lidar according to claim 1, further comprising an editable 3D model calibration step after the step of generating the editable 3D model of the building structure, for the lidar and generating a 3D model The size is calibrated, and the model size is automatically adjusted.
The automatic measurement and 3D model generation method for a building structure based on lidar according to claim 8, characterized in that the distance information and angle information measured by the lidar are calibrated to remove the zero offset.
The automatic measurement and 3D model generation method of a building structure based on lidar according to claim 8, characterized in that the dimensional calibration compares the generated model size with the actual size, fine-tunes the model proportions, and obtains the final 3D model of the building structure .