Large-scale plane point cloud registration method

Abstract

The invention relates to the field of point cloud registration, in particular to a large-scale plane point cloud registration method, which mainly utilizes a boundary extraction algorithm and a normal vector included angle extraction key point to optimize plane point cloud registration, and specifically comprises the following steps: acquiring a set of input point clouds, and performing boundary extraction; combining a point cloud set after boundary extraction with a normal vector included angle to extract overall feature points; then, carrying out SHOT descriptor description on the extracted feature points, and inputting the described feature points into an RANSAC algorithm for iteration to carry out coarse registration; and inputting the boundary point cloud subjected to coarse registration into an ICP algorithm framework to obtain fine registration and calculate a rotation matrix and a translation matrix. According to the method, large-scale plane point cloud registration and the ICP algorithm are combined to realize rapid and accurate point cloud registration, the ICP algorithm is optimized, the problem that the ICP algorithm is easy to fall into a local optimal solution is solved, the performance of the method is remarkably improved, and the registration precision and efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional reconstruction, and in particular to a method for large-scale planar point cloud registration. Background Art

[0002] Point cloud registration, as a core technology in the field of computer vision and 3D reconstruction, aims to align point cloud data sets captured from different positions or perspectives into a unified coordinate system, thereby building a more complete and consistent 3D model. With the rapid development of high-precision sensors such as 3D laser scanners and RGB-D cameras, point cloud data has become the main way to describe the 3D world. However, limited by the field of view of the sensor, a single scan can often only cover a part of the surface of the target object. Therefore, point cloud registration technology came into being, which can effectively fuse point cloud data obtained from multiple scans to generate a complete 3D scene model.

[0003] The technical system of point cloud registration covers multiple aspects such as feature extraction, transformation estimation and optimization algorithms. In the feature extraction stage, descriptive features such as surface normals and key points are extracted from each point cloud. These features play a key role in the subsequent point cloud alignment process. The transformation estimation stage is responsible for calculating the spatial transformation relationship between different point clouds, such as translation and rotation, so that the corresponding points can overlap as much as possible. The optimization algorithm is used to further reduce the spatial position differences between point clouds and improve the accuracy of registration.

[0004] And point cloud registration technology has a wide range of application value in many fields. In the field of autonomous driving, it is used to build high-precision three-dimensional maps, achieve high-precision positioning and posture estimation of vehicles, and provide accurate environmental perception capabilities for autonomous driving systems. In the field of robot navigation, the three-dimensional environment model constructed by point cloud registration helps robots to make path planning and navigation decisions. In addition, in the field of three-dimensional reconstruction, point cloud registration technology provides key technical support for generating complete three-dimensional scenes, such as the construction of building information models (BIM) and the monitoring of underground mining spaces.

[0005] Point cloud registration needs to be performed in a real-time environment. The processing of large-scale point clouds often fails to meet real-time requirements. However, when performing point cloud registration on large-scale planes, point cloud data may contain a lot of noise and redundant information. This results in a decrease in the robustness and accuracy of the registration process. Large-scale planar point clouds often have different scale features. How to effectively extract and match these multi-scale features is a key issue. Traditional feature extraction and matching algorithms may fail or be inefficient when processing large-scale point clouds.

[0006] Therefore, in response to the above problems, we propose a large-scale planar point cloud registration method to solve the above problems. Summary of the invention

[0007] 1. Technical issues to be resolved

[0008] In view of the shortcomings of existing algorithms for planar point cloud registration, the present invention provides a method for large-scale planar point cloud registration, which solves the problems raised in the above background technology.

[0009] (II) Technical solution

[0010] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions:

[0011] A method for large-scale planar point cloud registration, comprising the following steps:

[0012] S1: By using the optimized boundary extraction algorithm, a statistical outlier removal filter is introduced to reduce the point cloud noise and extract the boundary point cloud;

[0013] S2: Extract feature points by normal vector angle: Construct a KD-Tree for the extracted boundary point cloud model, calculate the average angle between the normal vector of each point and the normal vector of its neighboring points, and extract feature points based on the set angle threshold;

[0014] S3: SHOT descriptor describes feature points: constructs a normalized histogram and describes feature points by calculating the geometric distribution characteristics of the point cloud around the constructed local reference frame;

[0015] S4: Determination of the correspondence relationship of point cloud registration: The SHOT descriptor is input into RANSAC for coarse registration, and the result calculated by the RANSAC algorithm and the point cloud are input into the ICP algorithm. Through iterative calculation, ICP registration can find the optimal rigid body transformation matrix between the source point cloud and the target point cloud to complete the registration.

[0016] Furthermore, in S1, optimizing the boundary extraction algorithm includes:

[0017] The boundary extraction method is based on each point p i Neighborhood, Neighborhood is defined to include the center p i The set of k nearest points of , where k = 30, a represents the number of neighborhood points, creates a 3×3 dimensional covariance matrix C to estimate the surface normal, and calculates the eigenvector ν and eigenvalue λ by eigendecomposition;

[0018]

[0019] Cv j =λ j v j ,j∈{1,2,3}

[0020] In pi In the local coordinate system of i and We can get With u, w dot product, due to p i a May be distributed on the plane up i w, so there is an approximate formula: θ a ∈(-π,π), can be calculated as:

[0021]

[0022] Calculate θ for each k closest point a , sort it in ascending order to obtain the set {θ 1 ,θ 2 ,θ 3 ,…,θ k}, then use this formula to get the difference between two adjacent angles: Δθ t =θ t+1 -θ t And t∈{1,2,…,k-1}, it can be converted to: Δθ t =2π-θ t +θ 1 , and t = k. The maximum angle difference is: If the maximum angle difference is greater than the threshold α, then point p i Can be identified as a boundary point cloud.

[0023] Furthermore, in S2, extracting feature points by using normal vector angles includes:

[0024] Compute eigenvectors and eigenvalues ​​ν j and λ j , the formula is:

[0025]

[0026] Cv j =λ j v j ,j∈{1,2,3}

[0027] in a∈(1,k) is p i The a-th k closest point of Yes i The centroid of all k closest points, ν j and λ j They are the jth eigenvector and eigenvalue of the covariance matrix C, and then we calculate the normal vector n based on i , obtain unit orthogonal vectors u and w to construct a local coordinate system;

[0028] Then calculate each point p i The normal vector n i and its neighboring point q j Normal vector n j The angle θ between ij :

[0029]

[0030] Find the average angle between the normal vector of each point and the normal vectors of its neighboring points Then according to the set normal vector angle threshold θ th Extract the overall feature points of the boundary point cloud.

[0031] Furthermore, in said S3, the SHOT descriptor includes:

[0032] Given a point p in a point cloud i , whose normal vector is n i , the normal vector of the neighborhood point is n j , the local reference system Z axis is z r , the SHOT descriptor can be expressed as:

[0033] SHOT(p i )=[H 1 ,H 2 ,……,H NRF ]

[0034] cosθ qi =z r n j

[0035] Where NRF refers to the number of neighborhood points for calculating the SHOT descriptor. For each neighborhood point q j , which can be calculated with p i The relative position ΔP = q j -p i , through the polar radius r = ||ΔP|| of the relative position ΔP and the polar angle θ = atan2(Δp y ,Δp x ); then use the point's normal vector n i and the normal vector n of the neighboring point J To calculate the angle difference between the polar angles α = cos -1 (η i ·n j ), divide the polar angle and angle difference into several intervals, and count the number of neighborhood points in each interval to obtain the histogram H k , where k = 1, 2,…, NRF.

[0036] Furthermore, in S4, the determination of the point cloud registration correspondence relationship includes:

[0037] After calculating the rough transformation between the point cloud pairs through the RANSAC and SVD algorithms, the roughly aligned point clouds are submitted to the ICP framework for fine alignment. The final transformation matrix M f It can be expressed as:

[0038]

[0039] Among them, P i , Q i Represent the points in the source point cloud and the target point cloud respectively, R and T represent the parameters of rotation and translation respectively, and R and T are calculated by the following objective function:

[0040]

[0041] (III) Beneficial effects

[0042] Compared with the prior art, the present invention provides a large-scale planar point cloud registration method and system, which has the following beneficial effects:

[0043] The present invention effectively reduces point cloud noise and extracts accurate boundary point clouds by using an optimized boundary extraction algorithm and introducing a statistical outlier removal filter, thereby solving the problem of the influence of noise on registration accuracy.

[0044] The present invention adopts a method of extracting feature points by using the normal vector angle, constructs a KD-Tree and calculates the average normal vector angle, thereby being able to accurately extract feature points and ensure the accuracy and representativeness of the feature points, thereby improving the reliability of registration.

[0045] The present invention uses SHOT descriptors to describe feature points, constructs a normalized histogram, describes the geometric distribution characteristics of point clouds around a local reference frame, and improves the accuracy and consistency of feature point descriptions; in the registration process, the SHOT descriptor is input into RANSAC for coarse registration, and then combined with the ICP algorithm for precise registration, and the optimal rigid body transformation matrix is ​​found through iterative calculation, which solves the problem of easy errors in the registration process of large-scale point cloud data and obtains more accurate and stable registration results.

[0046] This method is suitable for point cloud registration tasks in various complex environments, and can be applied to indoor 3D reconstruction, urban modeling, unmanned driving, industrial inspection and other fields, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flowchart of the working process in an embodiment of the present invention;

[0048] Figure 2It is a schematic diagram of a model in an embodiment of the present invention;

[0049] Figure 3 Schematic diagram of normal vector angle in an embodiment of the present invention;

[0050] Figure 4 Schematic diagram of a SHOT descriptor in an embodiment of the present invention;

[0051] Figure 5 A schematic diagram comparing the boundary extraction algorithm results in the embodiment of the present invention with the SOR boundary extraction algorithm; DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] Example 1

[0054] like Figure 1-5 As shown, a large-scale planar point cloud registration method proposed in one embodiment of the present invention includes the following specific steps S1-S4:

[0055] S1: Use the optimized boundary extraction algorithm to process the point cloud data. First, reduce the point cloud noise by introducing a statistical outlier removal filter, and then extract the boundary point cloud. In this step, the optimized boundary extraction algorithm is specifically operated as follows:

[0056] First, select a point in the point cloud data as the center point p i , and based on p i Select 30 nearest neighbor points from the neighborhood of Using these neighborhood points, a 3×3 dimensional covariance matrix C is created, and the eigenvector ν and eigenvalue λ are calculated by eigendecomposition.

[0057]

[0058] Cv j =λ j v j ,j∈{1,2,3}

[0059] In p i In the local coordinate system, by calculating p i and The dot product of the connection vector of and the vectors u and w in the local coordinate system gives θ a. The θ of each neighborhood point is calculated by the following formula a values, and sort these values ​​to form a set {θ 1 ,θ 2 ,θ 3 ,…,θ k}:

[0060]

[0061] Then, the difference Δθ between adjacent angles is calculated t :

[0062] Δθ t =θ t+1 -θ t

[0063] If the maximum angle difference is: If the maximum angle difference is greater than the maximum value of the set threshold α, then point p i Can be identified as a boundary point cloud.

[0064] S2: Further extract feature points from the extracted boundary point cloud. First, construct a KD-Tree structure for each boundary point and calculate the average angle between the normal vector of each point and the normal vector of its neighboring points. The specific operations are as follows:

[0065] The covariance matrix C is calculated in the same way as in S1, and the eigenvector ν and eigenvalue λ are calculated again by eigendecomposition. These eigenvectors are used to construct p i The local coordinate system of i and its neighboring point q j Normal vector n j The angle θ between ij :

[0066]

[0067] Next, calculate the average angle between the normal vector of each point and the normal vector of its neighboring points And extract the feature points in the boundary point cloud according to the preset normal vector angle threshold.

[0068] See also Figure 3 , which is a schematic diagram of the regional normal vector of this embodiment.

[0069] S3: Use SHOT descriptors to describe the extracted feature points. Construct a normalized histogram and describe each feature point by calculating the geometric distribution characteristics of the point cloud in the local reference frame. The specific operations are as follows:

[0070] Given a point p in a point cloud i , whose normal vector is n i , the normal vector of the neighborhood point is nj , the local reference system Z axis is z r , calculate the neighborhood point q j Relative to p i The relative position ΔP = q j -p i , and calculate the polar radius r and polar angle θ based on ΔP:

[0071] cosθ qi =z r n j

[0072] SHOT(p i )=[H 1 ,H 2 ,……,H NRF ]

[0073] Divide the polar angle and angle difference into several intervals, count the number of neighborhood points in each interval, and get the histogram H k .

[0074] See also Figure 4 , which is a schematic diagram of the SHOT descriptor structure of this embodiment.

[0075] S4: Determine the corresponding relationship of point cloud registration. The SHOT descriptor is input into the RANSAC algorithm for rough registration, and the result is input into the ICP algorithm together with the point cloud. Finally, the optimal rigid body transformation matrix M between the source point cloud and the target point cloud is found through iterative calculation. f :

[0076]

[0077] Among them, P i , Q i Represent the points in the source point cloud and the target point cloud respectively, R and T represent the parameters of rotation and translation respectively, and are optimized by the following objective function:

[0078]

[0079] Apply the calculated R and T results to the source point cloud to complete the precise registration of the source point cloud and the target point cloud.

[0080] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for large-scale planar point cloud registration, characterized in that: The method comprises the following steps: S1: By using the optimized boundary extraction algorithm, a statistical outlier removal filter is introduced to reduce the point cloud noise and extract the boundary point cloud; S2: Extract feature points by normal vector angle: Construct a KD-Tree for the extracted boundary point cloud model, calculate the average angle between the normal vector of each point and the normal vector of its neighboring points, and extract feature points based on the set angle threshold; S3: SHOT descriptor describes feature points: constructs a normalized histogram and describes feature points by calculating the geometric distribution characteristics of the point cloud around the constructed local reference frame; S4: Determination of the correspondence relationship of point cloud registration: The SHOT descriptor is input into RANSAC for coarse registration, and the result calculated by the RANSAC algorithm and the point cloud are input into the ICP algorithm. Through iterative calculation, ICP registration can find the optimal rigid body transformation matrix between the source point cloud and the target point cloud to complete the registration.

2. A method for large-scale planar point cloud registration according to claim 1, characterized in that: In S1, optimizing the boundary extraction algorithm includes: The boundary extraction method is based on each point p i Neighborhood, Neighborhood is defined to include the center p i The set of k nearest points of , where k = 30, a represents the number of neighborhood points, creates a 3×3 dimensional covariance matrix C to estimate the surface normal, and calculates the eigenvector ν and eigenvalue λ by eigendecomposition; Cv j =λ j v j ,j∈{1,2,3} In p i In the local coordinate system of i and We can get With u, w dot product, due to p i a May be distributed on the plane up i w, so there is an approximate formula: θ a ∈(-π,π), can be calculated as: Calculate θ for each k closest point a , sort it in ascending order to obtain the set {θ1,θ2,θ3,…,θ k }, then use this formula to get the difference between two adjacent angles: Δθ t =θ t+1 -θ t And t∈{1,2,…,k-1}, it can be converted to: Δθ t =2π-θ t +θ1, and t=k. The maximum angle difference is: If the maximum angle difference is greater than the threshold α, then point p i Can be identified as a boundary point cloud.

3. The method for large-scale planar point cloud registration according to claim 1, characterized in that: In S2, the feature point is extracted by the normal vector angle, including: Compute eigenvectors and eigenvalues ​​ν j and λ j , the formula is: Cv j =λ j v j ,j∈{1,2,3} in Yes i The a-th k closest point of Yes i The centroid of all k closest points, ν j and λ j They are the jth eigenvector and eigenvalue of the covariance matrix C, and then we calculate the normal vector n based on i , obtain unit orthogonal vectors u and w to construct a local coordinate system; Then calculate each point p i The normal vector n i and its neighboring point q j Normal vector n j The angle θ between ij : Find the average angle between the normal vector of each point and the normal vectors of its neighboring points Then according to the set normal vector angle threshold θ th Extract the overall feature points of the boundary point cloud.

4. The method for large-scale planar point cloud registration according to claim 1, characterized in that: In the S3, the SHOT descriptor includes: Given a point p in a point cloud i , whose normal vector is n i , the normal vector of the neighborhood point is n j , the local reference system Z axis is z r , the SHOT descriptor can be expressed as: SHOT(p i )=[H1,H2,……,H NRF ] cosθ qi =z r n j Where NRF refers to the number of neighborhood points for calculating the SHOT descriptor. For each neighborhood point q j , which can be calculated with p i The relative position ΔP = q j -p i , through the polar radius r = ||ΔP|| of the relative position ΔP and the polar angle θ = atan2(Δp y ,Δp x ); then use the point's normal vector n i and the normal vector n of the neighboring point J To calculate the angle difference between the polar angles α = cos -1 (η i ·n j ), divide the polar angle and angle difference into several intervals, and count the number of neighborhood points in each interval to obtain the histogram H k , where k = 1, 2,…, NRF.

5. The method for large-scale planar point cloud registration according to claim 1, characterized in that: In S4, the determination of the corresponding relationship of the point cloud registration includes: After calculating the rough transformation between the point cloud pairs through the RANSAC and SVD algorithms, the roughly aligned point clouds are submitted to the ICP framework for fine alignment. The final transformation matrix M f It can be expressed as: Among them, P i , Q i Represent the points in the source point cloud and the target point cloud respectively, R and T represent the parameters of rotation and translation respectively, and R and T are calculated by the following objective function:

Images