Three-dimensional point cloud registration method and system based on adaptive neighborhood feature extraction

Abstract

The invention discloses a three-dimensional point cloud registration method and system based on adaptive neighborhood feature extraction. Firstly, a source point cloud and a target point cloud are obtained and preprocessed; dividing initial neighborhoods for points in the source point cloud and the target point cloud according to the density; dividing a fine neighborhood on the basis of the initial neighborhood according to geometric features of the points; then, feature point extraction is carried out on the source point cloud and the target point cloud based on the fine neighborhood; finally, performing coarse registration on the point cloud to obtain an initial transformation matrix; and performing fine registration by using an improved ICP algorithm. The system comprises a data acquisition module, a preprocessing module, a self-adaptive neighborhood calculation module, a feature point extraction module and a matching and optimizing module. According to the method, the size of the neighborhood is adaptively adjusted according to the density and the geometric features, an appropriate neighborhood is found for each point, more representative feature points are extracted in the fine neighborhood, the method can better adapt to local density and geometric feature changes of point cloud data, and the precision and robustness of point cloud registration are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of point cloud data processing, and in particular, to a three-dimensional point cloud registration method and system based on adaptive neighborhood feature extraction. Background Art

[0002] With the rapid development of three-dimensional scanning technology, point cloud data has been widely used in the fields of computer vision, medical imaging, robot navigation, and reverse engineering. Currently, due to the limitations of three-dimensional scanning technology, it is impossible to directly obtain the complete point cloud data of the target object, and it is necessary to align and merge the point cloud data obtained from multiple different perspectives or at different times through point cloud registration. Therefore, point cloud registration, as a key step in point cloud data processing, its accuracy and efficiency directly affect the effect of subsequent reconstruction.

[0003] Point cloud registration is mainly divided into two types: global point cloud registration and point cloud registration based on feature point extraction and matching. The most representative global search algorithm is the Iterative Closest Point algorithm (ICP), which solves the transformation matrix by seeking the closest neighbor points between the source point cloud and the target point cloud. Although the ICP algorithm is simple and easy to implement, it has the disadvantage of being easily trapped in local optima. The point cloud registration based on feature point extraction and matching solves the problem that the global registration method is sensitive to the initial position of the point cloud data. This method is mostly divided into two stages: rough registration and fine registration. The initial position of the point cloud obtained by rough registration directly affects the accuracy of fine registration.

[0004] In the current point cloud registration methods, the extraction of point cloud feature points is mostly completed in a fixed neighborhood or K-nearest neighbors. The neighborhood range of each point will affect the extraction effect of local geometric features. For point clouds with uneven density, a fixed radius may result in too many or too few points in the neighborhood, affecting the accuracy of feature extraction. Although the KNN algorithm can ensure that the number of points in the neighborhood is consistent, in regions with large curvature changes, a fixed K value may not accurately reflect the local geometric features, resulting in inaccurate description of feature points and thus affecting the registration accuracy. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the technical problem to be solved by the present invention is to provide a three-dimensional point cloud registration method and system based on adaptive neighborhood feature extraction.

[0006] The present invention adopts the following technical solutions to solve the above technical problems:

[0007] A three-dimensional point cloud registration method based on adaptive neighborhood feature extraction, characterized by including the following steps:

[0008] The first step: Obtain the source point cloud and the target point cloud, and preprocess the source point cloud and the target point cloud;

[0009] Step 2: Divide the initial neighborhood for each point in the source point cloud and the target point cloud according to the density;

[0010] For any point p in the source point cloud i , calculate the local density around the point according to the number of neighborhood points within a fixed radius. The calculation formula is:

[0011]

[0012] In the formula, d i represents the local density around point p i , N i represents the number of neighborhood points within a fixed radius, and V and S represent the volume and area of the neighborhood;

[0013] Divide the initial neighborhood according to the local density distribution. Then, the number of neighborhood points included in the initial neighborhood is calculated by the following formula:

[0014]

[0015] In the formula, k i represents the number of neighborhood points included in the initial neighborhood of point p i , d med represents the median of the local density, α is the density adjustment factor, and k max and k min are the upper and lower limits of the number of neighborhood points respectively;

[0016] Step 3: Divide the fine neighborhood based on the geometric characteristics of the points on the basis of the initial neighborhood;

[0017] Use the PCA method to analyze the initial neighborhood, calculate the covariance matrix, perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues λ1, λ2, and λ3, where λ1 ≥ λ2 ≥ λ3; calculate the curvature at the point according to the eigenvalues:

[0018]

[0019] In the formula, δ i represents the curvature at point p i ;

[0020] The third principal component of PCA is the normal vector of the point. Calculate the included angle between the normal vector of point p i and the normal vectors of each neighborhood point; if the included angle of the normal vectors is greater than the normal vector included angle threshold, it is considered that the normal directions of the two points are inconsistent. If the proportion of the inconsistent normal directions between point p i and the neighborhood points exceeds the set threshold, the neighborhood needs to be reduced; if point p iIf the curvature at a certain point is greater than the curvature threshold, the neighborhood needs to be shrunk; by iteratively and adaptively adjusting the number of points in the neighborhood, the neighborhood is shrunk so that the proportion of inconsistent normal directions is less than or equal to the set threshold and the curvature at the point is less than or equal to the curvature threshold, obtaining the fine neighborhood of point p i of the fine neighborhood;

[0021] Step 4: Extract feature points from the source point cloud and the target point cloud based on the fine neighborhood;

[0022] Calculate the covariance matrix within the fine neighborhood, decompose the covariance matrix to obtain eigenvalues; according to the eigenvalues of the covariance matrix, calculate the average curvature at the point through Equation (6), and take the points with an average curvature greater than the local curvature threshold as feature points;

[0023]

[0024] In the formula, represents the average curvature at point p i where r is the neighborhood radius and c is an empirical constant;

[0025] Step 5: Coarsely register the point cloud to obtain the initial transformation matrix; use the improved ICP algorithm for fine registration.

[0026] Furthermore, the local curvature threshold is calculated by the following formula:

[0027]

[0028] In the formula, T i represents the local curvature threshold at point p i where σ δ represents the mean and standard deviation of the average curvatures at all points within the fine neighborhood, and ρ represents the adjustment factor.

[0029] Furthermore, the improved ICP algorithm introduces a K-D tree in the search stage, and at the same time, in the stage of matching point pairs, if the included angle between the normal vectors of the matching point pairs is greater than the set threshold, then the matching point pairs are removed.

[0030] The present invention also provides a three-dimensional point cloud registration system for implementing the above method, which is characterized in that it includes a data acquisition module, a preprocessing module, an adaptive neighborhood calculation module, a feature point extraction module, and a matching and optimization module;

[0031] Data acquisition module: Acquire the source point cloud and the target point cloud;

[0032] Preprocessing module: Denoise and downsample the source point cloud and the target point cloud;

[0033] Adaptive neighborhood calculation module: Adaptively obtain the fine neighborhood according to density and geometric features;

[0034] Feature point extraction module: Extract feature points in the fine neighborhood;

[0035] Matching and optimization module: Used for initial registration and fine registration between feature points, and optimize the registration.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] 1. Divide the initial neighborhood based on density, and further adjust the initial neighborhood according to the geometric features of the points to obtain a fine neighborhood; adaptively adjust the neighborhood size according to density and geometric features, find a suitable neighborhood for each point, extract more representative feature points in the fine neighborhood, can better adapt to the changes of the local density and geometric features (including curvature and normal vector) of the point cloud data, improve the accuracy of feature point description, thereby improving the accuracy and robustness of point cloud registration, and at the same time solve the problem that the existing algorithms select fewer key feature points, are fast but have low registration accuracy.

[0038] 2. Introduce a K-D tree on the basis of the traditional ICP algorithm to speed up the registration speed; in the matching stage, eliminate the matching point pairs with large differences in normal vector directions through the normal vector angle constraint, realize the effective screening of matching point pairs, reduce the operation complexity, and achieve a good registration effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Is a flowchart of the method of the present invention;

[0040] Figure 2 Is a structural diagram of the system of the present invention;

[0041] Figure 3 Is a comparison diagram of the registration results of the method of the present invention and the ICP algorithm. DETAILED DESCRIPTION OF THE INVENTION

[0042] The following gives specific embodiments in conjunction with the drawings. The specific embodiments are only used to introduce the technical solutions of the present invention in detail, and do not limit the protection scope of this application.

[0043] The present invention provides a three-dimensional point cloud registration method based on adaptive neighborhood feature extraction, including the following steps:

[0044] The first step: Use a three-dimensional laser scanning device to obtain a source point cloud and a target point cloud; preprocess the source point cloud and the target point cloud;

[0045] The preprocessing includes filtering and downsampling; the source point cloud and the target point cloud are filtered by normal-based bilateral filtering to remove outliers. The normal-based bilateral filtering combines the geometric distance between a point and its neighboring points and the similarity of normal directions, retaining geometric features while removing noise; voxel downsampling is used to downsample the filtered source point cloud and target point cloud, which can effectively retain the overall structure and geometric features of the point cloud while significantly reducing the amount of point cloud data.

[0046] Step 2: Divide the initial neighborhoods for each point in the source point cloud and the target point cloud according to the density;

[0047] For any point p in the source point cloud i , according to the number of neighboring points N within a fixed radius i , the number of neighboring points per unit volume or area is the local density around the point. Then, the local density around point p i is calculated by the following formula:

[0048]

[0049] In the formula, d i represents the local density around point p i , and V and S represent the volume and area of the neighborhood;

[0050] According to the local density distribution, the median of the local density is used as the threshold. According to the proportional relationship between the local density and the threshold, the initial neighborhood radius is adaptively and dynamically adjusted. In the area with a large local density, the initial neighborhood radius is small to avoid noise interference; in the area with a small local density, the initial neighborhood radius is large to ensure that each point has a sufficient number of neighboring points for calculation or other analysis tasks. The number of neighboring points included in the initial neighborhood is calculated by the following formula:

[0051]

[0052] In the formula, k i represents the number of neighboring points included in the initial neighborhood of point p i , d med represents the median of the local density, α is the density adjustment factor, and k max and k min are the upper and lower limits of the number of neighboring points respectively;

[0053] The neighborhood scaling amplitude is controlled by parameters, and at the same time, the stability of the neighborhood scale is ensured by the upper and lower limits of the number of neighboring points. Similarly, the initial neighborhoods of the remaining points in the source point cloud and each point in the target point cloud are obtained.

[0054] Step 3: Adjust the initial neighborhood according to the geometric features of the points to obtain a refined neighborhood;

[0055] The principal component analysis (PCA) is used to analyze the spatial distribution of points within the initial neighborhood of each point, and the covariance matrix C is calculated through Equation (3);

[0056]

[0057] where p j represents the neighborhood points of point p i , and p represents the centroid of all neighborhood points;

[0058] The curvature reflects the degree of bending of the local surface; the covariance matrix is eigen-decomposed to obtain the eigenvalues λ1, λ2, and λ3, where λ1 ≥ λ2 ≥ λ3; in PCA, the smaller the variance ratio corresponding to the smallest eigenvalue, the smaller the curvature and the smoother the neighborhood; conversely, the larger the variance ratio, the larger the curvature and the less smooth the neighborhood; the curvature at a point is calculated by the following formula:

[0059]

[0060] where δ i represents the curvature at point p i ;

[0061] The normal vector reflects the direction of the minimum change of the local surface, and the third principal component of PCA is the normal vector of the point; calculate the angle between the normal vector of point p i and the normal vectors of each neighborhood point. Assuming that the normal vectors of point p i and neighborhood point p j are respectively then the angle θ between the normal vectors of point p i and p j is calculated by the following formula;

[0062]

[0063] Taking the normal direction consistency and curvature smoothness tests as evaluation indicators for adaptively adjusting the neighborhood size to avoid overfitting problems caused by a single evaluation criterion; the normal direction consistency is measured by the angle between normal vectors. If the angle between normal vectors is greater than the normal vector angle threshold, it is considered that the normal directions of two points are inconsistent. If the proportion of inconsistent normal directions between point p i and neighborhood points exceeds the set threshold, the neighborhood needs to be reduced; the curvature smoothness is measured by the curvature magnitude. If the curvature is greater than the curvature threshold and the curvature smoothness test fails, the neighborhood needs to be reduced; by iteratively and adaptively adjusting the number of points within the neighborhood to reduce the neighborhood so that the proportion of inconsistent normal directions is less than or equal to the set threshold and the curvature at the point is less than or equal to the curvature threshold, thereby obtaining the fine neighborhood of point p i ; similarly, a suitable fine neighborhood is found for each point to ensure a high degree of consistency in the normal direction and curvature of the points within the neighborhood.

[0064] Step 4: Extract feature points from the source point cloud and the target point cloud;

[0065] Recalculate the covariance matrix within the fine neighborhood, decompose the covariance matrix to obtain the eigenvalues; according to the eigenvalues of the covariance matrix, calculate the average curvature at each point through the following formula:

[0066]

[0067] In the formula, represents the average curvature at point p i where r is the neighborhood radius and c is an empirical constant; the eigenvectors ν1 and ν2 corresponding to the two larger eigenvalues λ1 and λ2 of the covariance matrix respectively correspond to the minimum curvature direction and the maximum curvature direction.

[0068] Compare the average curvature at the point with the local curvature threshold, and take the points with an average curvature greater than the local curvature threshold as feature points to achieve feature point extraction; among them, the local curvature threshold is dynamically adjusted according to the mean value and the standard deviation σ δ of the average curvature at all points within the fine neighborhood, and the calculation formula is:

[0069]

[0070] In the formula, T i represents the local curvature threshold at point p i where ρ represents a regulation factor used to control the strictness of the threshold.

[0071] Step 5: Use the sampling consistency method to achieve rough registration between point clouds and obtain the initial transformation matrix; use the improved ICP algorithm to achieve fine registration between point clouds. The improved ICP algorithm is based on the traditional ICP algorithm and introduces a K-D tree composed of all neighborhood points to accelerate the nearest neighbor search, so as to improve the efficiency and accuracy of fine registration; at the same time, add a normal vector angle constraint in the matching point pair stage to eliminate the matching point pairs with large differences in normal vector directions; when the angle between the normal vectors of the matching point pairs is greater than the set threshold, then eliminate the matching point pair, thereby optimizing the registration result.

[0072] The present invention also provides a three-dimensional point cloud registration system based on adaptive neighborhood feature extraction, including a data acquisition module, a preprocessing module, an adaptive neighborhood calculation module, a feature point extraction module, and a matching and optimization module;

[0073] Data acquisition module: Obtain the source point cloud and the target point cloud through devices such as 3D scanners, lidars, and depth cameras;

[0074] Preprocessing module: includes a denoising unit and a downsampling unit; the denoising unit uses methods such as statistical filtering and radius filtering to denoise the source point cloud and the target point cloud; the downsampling unit uses the voxel grid downsampling method to downsample the denoised point cloud;

[0075] Adaptive neighborhood calculation module: Dynamically adjusts the neighborhood size of each point according to the local density and geometric features of the point cloud data for adaptive neighborhood calculation;

[0076] Feature point extraction module: Based on the results of adaptive neighborhood calculation, extracts the feature points in the point cloud and uses geometric feature descriptors to describe the feature points;

[0077] Matching and optimization module: Calculates the similarity between feature points to match the feature points and uses the matching results for initial registration; optimizes the initial registration results, uses an improved ICP algorithm to achieve fine registration, and further improves the registration accuracy; evaluates the error of the registration results. If the error meets the requirements, outputs the final registration results; otherwise, performs re-registration.

[0078] To verify the effectiveness of the method of the present invention, a comparative experiment was carried out using the publicly available dataset Bunny. The experimental results are shown in Figure 3 . The experimental results show that the registration accuracy of the traditional ICP algorithm is 4.01×10 -3 m, while the registration accuracy of the method of the present invention is 1.47×10 -3 m. The traditional ICP algorithm usually uses a fixed neighborhood or global unified parameters and is difficult to adapt to the changes in the local density and geometric structure of the point cloud. After initially dividing the neighborhood by density, the present invention further dynamically adjusts the neighborhood size in combination with geometric features such as curvature and normal vector, so that the neighborhood of each point can more accurately reflect the local surface characteristics (such as edges, planes, etc.), enabling more representative feature points to be extracted at parts such as the ears and legs where the curvature of Bunny changes significantly, thus providing a more accurate initial correspondence for registration. Therefore, the registration effect of the method of the present invention is better than that of the traditional ICP algorithm.

[0079] Matters not described in the present invention are applicable to the prior art.

Claims

1. A three-dimensional point cloud registration method based on adaptive neighborhood feature extraction, characterized in that: The following steps are involved: Step 1: Obtain the source point cloud and the target point cloud, and pre-process the source point cloud and the target point cloud; Step 2: Divide the initial neighborhood for each point in the source point cloud and the target point cloud according to the density; For any point p in the source point cloud i , the local density around the point is calculated according to the number of neighborhood points within a fixed radius, and the calculation formula is: Where, d i Represents point p i The local density around, N i represents the number of neighborhood points within a fixed radius, V and S represent the volume and area of ​​the neighborhood; The initial neighborhood is divided according to the local density distribution, and the number of neighborhood points contained in the initial neighborhood is calculated by the following formula: In the formula, k i Represents point p i The number of neighborhood points contained in the initial neighborhood of med represents the median of local density, α is the density adjustment factor, k max and k min are the upper and lower limits of the number of neighborhood points respectively; Step 3: Divide the fine neighborhood based on the initial neighborhood according to the geometric characteristics of the points; The PCA method is used to analyze the initial neighborhood, calculate the covariance matrix, perform eigendecomposition on the covariance matrix, and obtain the eigenvalues ​​λ1, λ2, and λ3, λ1≥λ2≥λ3; the curvature at the point is calculated based on the eigenvalues: In the formula, δ i Represents point p i The curvature at The third principal component of PCA is the normal vector of the point. Calculate the point p i The angle between the normal vectors of each neighboring point; if the normal vector angle is greater than the normal vector angle threshold, it is considered that the normal directions of the two points are inconsistent. i If the inconsistency ratio of the normal direction with the neighboring points exceeds the set threshold, the neighborhood needs to be reduced; if point p i If the curvature at is greater than the curvature threshold, the neighborhood needs to be reduced; By iteratively and adaptively adjusting the number of points in the neighborhood and shrinking the neighborhood, the normal direction inconsistency ratio is less than or equal to the set threshold and the curvature at the point is less than or equal to the curvature threshold. i The fine neighborhood of Step 4: Extract feature points of source point cloud and target point cloud based on fine neighborhood; Calculate the covariance matrix in the fine neighborhood, decompose the covariance matrix, and obtain the eigenvalues; According to the eigenvalue of the covariance matrix, the average curvature at the point is calculated by formula (6), and the point whose average curvature is greater than the local curvature threshold is taken as the feature point; In the formula, Represents point p i The average curvature at , r is the neighborhood radius, c is an empirical constant; Step 5: Roughly register the point cloud to obtain the initial transformation matrix; The improved ICP algorithm is used for precise registration.

2. The three-dimensional point cloud registration method based on adaptive neighborhood feature extraction according to claim 1, characterized in that: The local curvature threshold is calculated as follows: Where, T i Represents point p i The local curvature threshold at σ δ represents the mean and standard deviation of the average curvature at all points in the fine neighborhood, and ρ represents the adjustment factor.

3. The three-dimensional point cloud registration method based on adaptive neighborhood feature extraction according to claim 1 or 2, characterized in that: The improved ICP algorithm introduces the KD tree in the search stage. At the same time, in the matching point pair stage, if the normal vector angle of the matching point pair is greater than the set threshold, the matching point pair is eliminated.

4. A 3D point cloud registration system for implementing the method of claim 1, characterized in that: It includes data acquisition module, preprocessing module, adaptive neighborhood calculation module, feature point extraction module, matching and optimization module; Data acquisition module: obtain source point cloud and target point cloud; Preprocessing module: denoising and downsampling the source point cloud and target point cloud; Adaptive neighborhood calculation module: Adaptively obtains fine neighborhood based on density and geometric features; Feature point extraction module: extract feature points in a fine neighborhood; Matching and optimization module: used for initial and precise registration between feature points, and optimizes the registration.

Images