Three-dimensional point cloud registration method based on statistical spatial distance distribution

Abstract

The invention discloses a three-dimensional point cloud registration method based on statistical spatial distance distribution, and the method comprises the following steps: extracting edge points and angular points in a source point cloud and a target point cloud through employing an HARRIS angular point detection algorithm, and generating a source point cloud key point set and a target point cloud key point set; counting the relative distribution frequency of the spatial distance between each point in the two key point sets and other key points in the field, and generating a high-dimensional feature vector; constructing spatial distance distribution feature descriptors by the spatial distance feature elements; according to the cosine similarity of the feature descriptors, points with the similarity degree of the feature descriptors larger than a set threshold value in the key points of the target point cloud are searched, and an initial corresponding point pair set is obtained; an initial spatial transformation matrix between corresponding point sets is solved by using an SVD algorithm through RANSAC, and coarse registration of the source point cloud and the target point cloud is realized; and after a relatively good initial corresponding pose is obtained, performing fine calibration on a registration result through an iterative closest point ICP algorithm. The method has the advantages of low operation cost, high speed, high precision and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional reconstruction of graphics, and in particular to a three-dimensional point cloud registration method based on statistical space distance distribution. Background Art

[0002] With the rapid development of high-precision scanning equipment such as laser radar and 3D cameras, processing point cloud information of target objects has become a widely used technical means in the fields of three-dimensional reconstruction, visual recognition and positioning. When using point cloud data obtained by laser scanning technology for three-dimensional reconstruction, many challenges are often faced. These point cloud data may be messy, lack obvious features, have holes, or even incomplete information. It is not only difficult to perform three-dimensional reconstruction based on such data, but also the accuracy of the reconstructed model is often not high, and the error is relatively large, which is difficult to meet the needs of practical applications. Therefore, before performing three-dimensional reconstruction, it is very necessary to perform a series of preprocessing work on the scanned point cloud data. These preprocessing work includes but is not limited to filtering, segmentation, registration, etc. Through these processing steps, the quality of the point cloud can be effectively improved, the features can be enhanced, and a more complete and accurate point cloud model can be obtained. Such a point cloud model can not only be used for more accurate measurement, but also lay a solid foundation for further three-dimensional reconstruction.

[0003] Registration is an important part of point cloud processing. The purpose of point cloud registration is to find the spatial pose transformation relationship between two or more frames of point clouds of the same object from different perspectives and calculate the spatial pose transformation matrix. In the field of robotics applications, the spatial pose transformation matrix obtained after registration is an important basis for the robot's terminal path planning; in the field of autonomous driving, registration enables the transformation of multiple frames of point clouds to the same coordinate system, providing a basis for operations such as map construction and ground segmentation. At present, point cloud registration technology has become a key technology in the fields of robotics, aerospace, 3D reconstruction, cultural relics restoration, autonomous driving, etc., and has been widely studied and applied.

[0004] Most existing point cloud registration methods need to calculate multiple feature dimensions such as normals, curvatures, and angles to improve the robustness of feature descriptors and thus improve the accuracy of registration results, but this requires a large time cost. The classic ICP algorithm and some of its variants often use global point cloud data, which requires a large number of iterations to ensure convergence and still requires a high time cost. It is not suitable for application scenarios where some scale differences are not high but have high real-time requirements. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a three-dimensional point cloud registration method based on statistical spatial distance distribution with low computational cost, high speed and high accuracy.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a three-dimensional point cloud registration method based on statistical spatial distance distribution, comprising the following steps:

[0007] The HARRIS corner detection algorithm is used to extract edge points and corner points in the source point cloud and the target point cloud, and generate the key point set of the source point cloud and the key point set of the target point cloud;

[0008] For each point in the two key point sets, the relative distribution frequency of its spatial distance to other key points in the field is counted to generate a high-dimensional feature vector;

[0009] Constructing a spatial distance distribution feature descriptor from spatial distance feature elements;

[0010] The cosine similarity of the feature descriptors is used to find the points in the target point cloud whose feature descriptor similarity is greater than the set threshold, and the initial corresponding point pair set is obtained.

[0011] The SVD algorithm is used through RANSAC to obtain the initial spatial transformation matrix between the corresponding point sets, and the rough registration of the source point cloud and the target point cloud is achieved;

[0012] After obtaining a good initial corresponding pose, the registration result is finely calibrated by iterative closest point ICP algorithm.

[0013] The beneficial effect of adopting the above technical solution is that: the method uses the spatial distance distribution state between the key points of the point cloud to construct the feature descriptor of the key points, without calculating the normal, curvature and other information, reducing the amount of calculation and saving a lot of operation time. For rigid objects, the spatial distance distribution between their corners and cusps will not change due to rotation or change in viewing angle. When calculating the spatial distance feature vector, only the distance to the key points in the neighborhood is considered, which reduces the impact of the low overlap of the point cloud due to partial occlusion or defects on the registration results, and ensures that the algorithm has good robustness. During coarse registration, RANSAC is used to select points with better effects to calculate the spatial transformation matrix, filtering out the influence of some invalid corresponding points on the registration accuracy. During fine registration, the final corresponding point set is used for ICP iteration, which not only ensures the reliability of the accuracy of the present invention and expands the applicable scenarios, but also reduces a large amount of point data and further reduces the operation time. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0015] Figure 1 is a flow chart of the method described in an embodiment of the present invention;

[0016] Figure 2 A schematic diagram of constructing a key point feature descriptor in an embodiment of the present invention;

[0017] Figure 3 This is a schematic diagram of the initial corresponding point pair generation result in an embodiment of the present invention;

[0018] Figure 4 Schematic diagram of the registration result in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the drawings 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.

[0020] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0021] Embodiment 1

[0022] like Figure 1 As shown, the embodiment of the present invention discloses a three-dimensional point cloud registration method based on statistical space distance distribution, comprising the following steps:

[0023] S1: Use the HARRIS corner detection algorithm to extract edge points and corner points in the source point cloud and the target point cloud, and generate the key point set of the source point cloud and the key point set of the target point cloud;

[0024] S2: For each point in the two key point sets, the relative distribution frequency of its spatial distance to other key points in the field is counted to generate a high-dimensional feature vector;

[0025] S3: Construct spatial distance distribution feature descriptor from spatial distance feature elements;

[0026] S4: Using the cosine similarity of the feature descriptors, find the points in the target point cloud whose feature descriptor similarity is greater than the set threshold, and obtain the initial corresponding point pair set.

[0027] S5: Using SVD algorithm through RANSAC to obtain the initial spatial transformation matrix between corresponding point sets, and realize the rough registration of source point cloud and target point cloud;

[0028] S6: After obtaining a good initial corresponding pose, the registration result is finely calibrated by iterative closest point ICP algorithm.

[0029] The step S2 calculates the spatial distance feature vector in the neighborhood for each key point, specifically comprising the following steps:

[0030] Step S2.1, constructing KD trees for the key point sets of the source point cloud and the target point cloud respectively, and establishing K nearest neighbor relationships and radius neighborhood relationships for each key point in the key point set by establishing KD trees;

[0031] Step S2.2, set a suitable neighborhood range, and calculate the Euclidean distance between each key point and other key points in its neighborhood:

[0032]

[0033] Among them, d AB is the key point P A and a key point P in the neighborhood B The spatial Euclidean distance between A ,y A ,z A ) is the key point P A The three-dimensional space coordinates, (x B ,y B ,z B ) is the key point P B The three-dimensional space coordinates of .

[0034] Step S2.3, with key point P A For example, for each key point, the spatial distance feature vector is generated by the Euclidean distance between it and other key points in the neighborhood:

[0035] Df A =[d A1 ,d A2 ,d A3 ,……,d A(n-1) ] (2)

[0036] Among them, Df A The key point P A The spatial distance feature vector in the neighborhood contains P A There are n key points in total. Among them, Figure 2 Construct a schematic for keypoint feature descriptors.

[0037] The method of constructing a spatial distance distribution feature descriptor from the spatial distance feature elements in step S3 comprises the following steps:

[0038] Step S3.1, according to the neighborhood range, select a suitable interval and construct a one-dimensional feature histogram with the same number of intervals for each key point.

[0039] Step S3.2, on the one-dimensional feature histogram, traverse each feature element in the spatial distance feature vector of the key point, and calculate the distribution frequency of the feature element in each interval.

[0040] Step S3.3, with key point P A As an example, according to the distribution frequency, a spatial distance distribution feature descriptor is constructed:

[0041] F A =[f A1 ,f A2 ,f A3 ,……,f Am ] (3)

[0042] Among them, F A The key point P A The spatial distance distribution feature descriptor of , m is the number of intervals of the one-dimensional histogram, and:

[0043]

[0044] Among them, R is the set neighborhood radius (neighborhood range), +1 prevents the descriptor subscript from overflowing when R is not an integer, and size is the size of each interval of the feature histogram.

[0045] Furthermore, the step S3.2 calculates the distribution frequency of the key point feature elements in each interval of the feature histogram, and uses relative frequency to eliminate the impact caused by the large difference in the data volume of the two point clouds. For each key point, first calculate the number of feature elements distributed in each interval of the feature histogram to obtain the absolute frequency. A For example:

[0046]

[0047] Among them, N a is the key point P A The absolute frequency of all feature elements in the feature vector distributed in the ath interval of the feature histogram, min is the minimum boundary value of the feature histogram, and size is the size of each interval of the feature histogram.

[0048] Then convert the absolute frequency into relative frequency:

[0049]

[0050] f Aa The key point P A The ath feature in the feature descriptor.

[0051] The method of finding corresponding key points in step S4 and obtaining corresponding point pairs comprises the following steps:

[0052] Step S4.1, judging whether the two key points are the corresponding point pair is based on the cosine similarity of the feature descriptors of the two key points, with the source point cloud key point P S and the target point cloud key point P t Take as an example, first calculate the cosine similarity of the two feature descriptors:

[0053]

[0054] F s 、F t are the feature descriptors of the two key points, f si 、f ti are the features in the feature descriptors of the two key points respectively, and m is the number of intervals of the above feature histogram;

[0055] Step S4.2, setting a cosine similarity threshold t. If the cosine similarity of the above two point feature descriptors is greater than t, the two points are determined to be an initial corresponding point pair.

[0056] Step S4.3, traverse all key points of the target point cloud, find their corresponding points in the source point cloud, store all corresponding point pairs, and obtain the initial corresponding point pair set, where Figure 3 Schematic diagram of the generated results for the initial corresponding point pairs.

[0057] Furthermore, the specific method of the rough registration in step S5 includes the following steps:

[0058] Using the initial corresponding point set, set the iteration upper limit, and adopt the RANSAC algorithm idea to perform the following iterations: randomly select three pairs of points from the corresponding point set, and use the SVD method to calculate the spatial transformation matrix; calculate the spatial distance error between the unselected corresponding point pairs under the action of the above spatial transformation matrix. If it is less than the set threshold, it is recorded as an inlier pair, and the number of inlier pairs is counted. After the iteration is completed, select the spatial transformation matrix with the largest number of inlier pairs after transformation as the initial spatial transformation matrix, and store the three pairs of points and the inlier pairs at this time as the final corresponding point pair set.

[0059] Furthermore, the step S6 is a precise registration, using the final corresponding point pair set, after ICP calibration, to obtain the final spatial transformation matrix. Figure 4 This is a schematic diagram of the registration results.

[0060] Embodiment 2

[0061] Point cloud preprocessing: The open source data set in the Stanford University 3D scan repository is selected as the implementation object to preprocess the original point cloud. In this embodiment, voxel filtering (Voxel Grid Filter) is used to downsample the original point cloud, and Radius Outlier Removal (Radius Outlier Removal) is used to smooth the original point cloud.

[0062] The embodiment of the present invention discloses a three-dimensional point cloud registration method based on statistical space distance distribution, the method comprising the following steps:

[0063] S1, uses HARRIS corner detection algorithm to sample key points.

[0064] S2, for each key point, calculate the spatial distance feature vector within the neighborhood.

[0065] S2.1, create a KD tree object, import the key point sets of the source point cloud and the target point cloud respectively, build a KD tree for the key point sets of the source point cloud and the target point cloud, and establish K nearest neighbor relationships and radius neighborhood relationships.

[0066] S2.2, set the neighborhood radius R, and use the above formula (1) to calculate the spatial Euclidean distance between each key point and other key points in its radius neighborhood.

[0067] S2.3, using the spatial Euclidean distance calculated above to generate a spatial distance feature vector for each key point, as shown in the above formula (2).

[0068] S3, constructs a spatial distance distribution feature descriptor from spatial distance feature elements.

[0069] S3.1, set the appropriate feature histogram interval size size and construct a one-dimensional feature histogram for each key point.

[0070] S3.2, traverse each feature element in the spatial distance feature vector of each key point (i.e., the distance to other points in the neighborhood), and use the above formula (5) to calculate the absolute distribution frequency of the feature element in the feature histogram interval.

[0071] S3.3, use the above formula (6) to normalize the absolute distribution frequency, use the above formula (6) to convert it into a relative distribution frequency, and use this to construct a feature descriptor, such as the above formula (3).

[0072] S4, find the corresponding key points and obtain the corresponding point pair set.

[0073] S4.1, set the feature descriptor similarity threshold t, for each key point of the source point cloud, search in the target point cloud key point set, and use the above formula (7) to calculate the similarity of the feature descriptors of the two key points.

[0074] S4.2, create an initial corresponding point pair container. If the similarity of the feature descriptors of two points is greater than t, save these two points as the initial corresponding point pair and delete them from the key point set to prevent repeated matching.

[0075] S4.3, after completing the above operations on the key point set, an initial corresponding point pair set is obtained.

[0076] S5, rough registration. The initial corresponding point pairs use the RANSAC algorithm idea, randomly select three pairs of points from the initial corresponding point pair set, and use the SVD method to calculate the spatial transformation matrix; calculate the spatial distance error of the unselected corresponding point pairs under the action of the above spatial transformation matrix. If it is less than the set threshold, it is recorded as an inlier pair, and the number of inlier pairs is counted. After the iteration is completed, the spatial transformation matrix with the largest number of inlier pairs after the transformation is selected as the initial spatial transformation matrix, and the three pairs of points and the inlier pairs at this time are stored as the final corresponding point pair set.

[0077] S6, fine registration. Use the ICP algorithm to fine-tune the registration results, create an ICP object, import the above final corresponding point pair set, and obtain the final spatial transformation matrix

[0078] The present invention uses the Stanford rabbit data file in the Stanford University 3D scan repository to conduct experiments with the armadillo data file. The RMSE (root mean square error) after the Stanford rabbit bun000 and bun045 are registered is 8.60859e-05, and the registration time is 1473ms. The RMSE after the armadillo ArmadilloStand_30 and ArmadilloStand_60 are registered is 6.48353e-04, and the registration time is 1170ms. According to the registration experiment of Stanford rabbit point clouds with different initial poses conducted by Wang Xinmiao et al. of Liaoning University of Engineering and Technology in "Point Cloud Registration Algorithm Based on Statistical Local Feature Description and Matching", the traditional ICP algorithm has an RMSE of 8.85048e-05 after the Stanford rabbit is registered when the initial pose is good, but the registration time takes 36.087s. While achieving the same order of magnitude registration accuracy, this method saves about 95.91% in time.

[0079] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A three-dimensional point cloud registration method based on statistical spatial distance distribution, characterized in that The steps include: The HARRIS corner detection algorithm is used to extract edge points and corner points in the source point cloud and the target point cloud, and generate the key point set of the source point cloud and the key point set of the target point cloud; For each point in the two key point sets, the relative distribution frequency of its spatial distance to other key points in the field is counted to generate a high-dimensional feature vector; Constructing a spatial distance distribution feature descriptor from spatial distance feature elements; The cosine similarity of the feature descriptors is used to find the points in the target point cloud whose feature descriptor similarity is greater than the set threshold, and the initial corresponding point pair set is obtained; The SVD algorithm is used through RANSAC to obtain the initial spatial transformation matrix between the corresponding point sets, and the rough registration of the source point cloud and the target point cloud is achieved; After obtaining a good initial corresponding pose, the registration result is finely calibrated by iterative closest point ICP algorithm.

2. The three-dimensional point cloud registration method based on statistical spatial distance distribution according to claim 1, characterized in that: For each key point, calculating the spatial distance feature vector in the neighborhood includes the following steps: 1) Construct KD trees for the key point sets of the source point cloud and the target point cloud respectively, and establish K nearest neighbor relationships and radius neighborhood relationships for each key point in the key point set by establishing KD trees; 2) Set a suitable neighborhood range and calculate the Euclidean distance between each key point and other key points in its neighborhood: Among them, d SB is the key point P A and a key point P in the neighborhood B The spatial Euclidean distance between A ,y A ,z A ) is the key point P A The three-dimensional space coordinates, (x B ,y B ,z B ) is the key point P B The three-dimensional space coordinates of 3) For the key point P A , and the spatial distance feature vector is generated by the Euclidean distance between it and other key points in the neighborhood: Df A =[d A1 ,d A2 ,d A3 ,……,d A(n-1) ] (2) Among them, Df A is the key point P A The spatial distance feature vector in the neighborhood contains P A There are n key points in total.

3. The three-dimensional point cloud registration method based on statistical spatial distance distribution according to claim 1, characterized in that: Constructing a spatial distance distribution feature descriptor from spatial distance feature elements includes the following steps: 1) According to the neighborhood range, select intervals and construct a one-dimensional feature histogram with the same number of intervals for each key point; 2) On the one-dimensional feature histogram, traverse each feature element in the spatial distance feature vector of the key point and calculate the distribution frequency of the feature element in each interval; 3) According to the distribution frequency, construct a spatial distance distribution feature descriptor: F A =[f A1 ,f A2 ,f A3 ,……,f Am ] (3) Among them, F A is the key point P A The spatial distance distribution feature descriptor of , m is the number of intervals of the one-dimensional histogram and: R is the neighborhood radius, +1 prevents descriptor subscript overflow when R is not an integer, and size is the size of each interval of the feature histogram.

4. The three-dimensional point cloud registration method based on statistical spatial distance distribution according to claim 3, characterized in that: Calculate the distribution frequency of key point feature elements in each interval of the feature histogram, and use relative frequency to eliminate the impact caused by large differences in the amount of data between the two point clouds; For each key point, first calculate the number of feature elements distributed in each interval of the feature histogram and obtain the absolute frequency: Among them, N a is the key point P A The absolute frequency of all characteristic elements in the characteristic vector distributed in the ath interval of the characteristic histogram, d Ab is the key point P A With key point P b The spatial Euclidean distance of , min is the minimum boundary value of the feature histogram, and size is the size of each interval of the feature histogram; Then convert the absolute frequency into relative frequency: f Aa is the key point P A The ath feature in the feature descriptor.

5. The three-dimensional point cloud registration method based on statistical spatial distance distribution according to claim 1, characterized in that: The method for obtaining an initial corresponding point pair set includes the following steps: 1) The basis for judging whether two key points are the corresponding point pair is the cosine similarity of the feature descriptors of the two key points. First, the cosine similarity of the feature descriptors of the two points is calculated: Where P S is the key point of the source point cloud, P t is the key point of the target point cloud; F s 、F t are the feature descriptors of the two key points, f si 、f ti are the features in the feature descriptors of the two key points respectively, and m is the number of intervals of the above feature histogram; 2) Setting a cosine similarity threshold t, if the cosine similarity of the above two point feature descriptors is greater than t, the two points are determined to be an initial corresponding point pair; 3) Traverse all key points of the target point cloud, find their corresponding points in the source point cloud, store all corresponding point pairs, and obtain the initial corresponding point pair set.

6. The three-dimensional point cloud registration method based on statistical spatial distance distribution according to claim 1, characterized in that The coarse registration comprises the following steps: Use the initial corresponding point pair set, set the iteration upper limit, and use the RANSAC method to perform the following iterations: Three pairs of points are randomly selected from the initial corresponding point pair set, and the spatial transformation matrix is ​​calculated using the SVD algorithm; the spatial distance error of the unselected corresponding point pairs under the action of the above spatial transformation matrix is ​​calculated, and if it is less than the set threshold, it is recorded as an inner point pair, and the number of inner point pairs is counted; after the iteration is completed, the spatial transformation matrix with the largest number of inner point pairs after transformation is selected as the initial spatial transformation matrix, and the three pairs of points and the corresponding inner point pairs at this time are stored as the final corresponding point pair set.

Images