Point cloud registration method and system based on color enhanced geometric information

Abstract

The invention provides a point cloud registration method and system based on color enhanced geometric information, relates to the technical field of computer vision processing, and solves the problem that an existing point cloud registration mode is poor in registration effect and easy to lose efficacy in a weak feature scene. The method comprises the following steps: acquiring key points of a source point cloud and a target point cloud, using an average color value as a color information reference, and acquiring a local color geometric feature CGD of each key point; the global distance distribution feature similarity between corresponding key points in the source point cloud and the target point cloud is preliminarily determined, then correct corresponding matching point pairs are obtained, and a transfer quantity matrix equation is constructed; and obtaining a transfer quantity after solving, constructing a transfer matrix based on the transfer quantity, updating the target point cloud, and completing registration of the source point cloud and the target point cloud after multiple iterations until the transfer quantity converges. According to the method, colors, local geometric information and global distribution features are combined, and a good registration result can be obtained even under the condition that the geometric features are weak.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision processing, and is applied to a point cloud registration process for processing defects of hydraulic structures, in particular to a point cloud registration method and system based on color-enhanced geometric information. BACKGROUND

[0002] With the development of three-dimensional point cloud technology, point clouds have been applied in many fields, such as deformation monitoring of various buildings, smart cities, etc. Point cloud registration is a basic work in point cloud data processing. As a preparation work for applying point clouds, the quality of the registration result has a great influence on subsequent data analysis and application. Existing registration methods mainly include two categories: one is the traditional method, mainly including variants of the iterative closest point (ICP) and variants of the normal distribution transformation registration (NDT); the other is the method using deep learning. Both methods have their own advantages and disadvantages: the traditional method has fast running time and high efficiency, but it needs to set parameters, and the experience value has an influence on the registration result; the deep learning method does not need to preset parameters, and the end-to-end learning is more convenient, but deep learning often needs to down-sample the point cloud, which will lose a large amount of feature information of the point cloud, so the registration effect of the deep learning method in the case of large scenes and weak features is often not as good as that of the traditional method.

[0003] ICP algorithm and its variants, which iteratively register two point clouds based on the nearest key points between them until convergence. There are generally several ways in the prior art: using a hybrid regularization term to represent the registration residual, constructing a hybrid sparse registration function, and combining an alternating multiplier method to optimize the double-loop structure, which can weaken the influence of outliers to some extent; using principal component analysis to obtain the main direction of two point clouds, and then roughly aligning the main direction to provide a good initial condition for the ICP algorithm, which enables ICP to handle point clouds with worse initial positions; using principal component analysis to extract feature points from the fast feature point histogram, and using the feature points as the input of the four-point fast robust algorithm to coarsely match the point cloud, and then using ICP to refine the registration result of the coarsely registered point cloud; in order to solve the problem of slow iteration speed of ICP algorithm, feature points are extracted according to the size relationship between local surface variation factor and average variation factor, which can adaptively extract feature points, and then combined with the principle of random sample consensus to achieve coarse registration, and then using the coarse registration result combined with the feature points to obtain the final registration result by ICP; using the characteristics of pseudo-Huber loss function that is not sensitive to noise and outliers, combined with RGB-D information, to form an ICP variant algorithm that can suppress the influence of outliers on the registration result; using an improved ICP algorithm based on local geometric features, selecting representative feature points through geometric features, and selecting the correct nearest point through geometric feature similarity; in order to solve the problem of abnormal value influence and low recall rate in ICP correspondence, a graph node reliability is designed to obtain the global maximum consensus set. Based on the above methods, the ICP algorithm variants can obtain accurate registration results for most scenes, but have problems such as sensitivity to noise, low efficiency, low recall rate, and dependence on good initial position.

[0004] NDT algorithm and its variants, which is represented by a set of Gaussian distributions with different probability density functions (PDFs), and is based on the expected transformation between the reference point cloud and the source point cloud. There are generally several ways in the prior art: combining sample consistency and normal distribution exchange to improve the accuracy of NDT registration to some extent; NDT algorithm based on three-dimensional shape context features can improve registration efficiency; decompose the seven-parameter transformation problem into three sub-problems, use random sample consensus algorithm to estimate scale and translation parameters, and introduce gradient optimization strategy to solve rotation problem; combine NDT and ICP algorithms, NDT is responsible for rough estimation of the transformation matrix, and then use ICP to refine the point cloud; divide the point cloud using k-means clustering and optimize at multiple cluster sizes to improve convergence speed. Based on the above methods, the variant algorithms of NDT are mostly not affected by the initial position of the point cloud, but are affected by the grid size, and the registration accuracy of NDT algorithm is generally low.

[0005] Deep learning has developed rapidly in recent years, and its application in point cloud registration is also increasing. There are usually several ways in the prior art: using icosahedron projection to describe features, divided into two parts of rotation invariance and rotation equivariance; extracting features and predicting RPM (Robust Point Matching) parameters by deep learning to realize registration; designing a full convolutional network to match corresponding points; proposing a large-scale outdoor scene point cloud efficient hierarchical network to extract key points and descriptors for registration; proposing a partial-to-partial registration network to find reliable overlapping representations from partially overlapping areas. Based on the above methods, deep learning often needs to sample point clouds to a small number, which makes it difficult to apply in large scenes, so deep learning methods are often applied in scenes with obvious features.

[0006] In summary, most of the existing registration algorithms use point position relationships to extract deep information to obtain corresponding relationships, but this method is easily invalid in weak feature scenes, and the current point cloud acquisition method often contains RGB color information; therefore, how to combine color, local geometric information and global distribution features to ensure good registration results in the case of weak geometric features has become a research point for those skilled in the art. SUMMARY

[0007] The purpose of the present application is to solve the problem that the existing point cloud registration method has poor registration effect and is easily invalid in weak feature scenes, therefore a point cloud registration method and system based on color enhanced geometric information are proposed. The present application uses local color geometric information to express key point features and selects candidate points, and then selects the corresponding points through the distribution relationship of the candidate points. The present application combines color, local geometric information and global distribution features, so that good registration results can be obtained even in the case of weak geometric features.

[0008] The present application adopts the following technical solutions to achieve the purpose: A point cloud registration method based on color enhanced geometric information, the method comprising the following steps: S1, determining the source point cloud and the target point cloud to be registered, and respectively acquiring the key points of the source point cloud and the target point cloud; S2, using the average color value as a color information reference, acquiring the local color geometric features CGD of the key points in the source point cloud and the target point cloud; S3, according to the CGD of the key points of the source point cloud and the target point cloud, preliminarily determining the global distance distribution feature similarity between the corresponding key points in the source point cloud and the target point cloud; S4, according to the preliminarily determined global distance distribution feature similarity, acquiring the correct corresponding matching point pairs between the source point cloud and the target point cloud, and constructing a transition matrix equation based on the corresponding matching point pairs; S5, solving the transition matrix equation to obtain the transition, constructing the transition matrix based on the transition, updating the target point cloud through the transition matrix, completing the registration of the source point cloud and the target point cloud.

[0009] Preferably, after step S5, the method further comprises: S6, after updating the target point cloud based on the transition matrix, iteratively performing the process of steps S1 to S5 until the transition obtained by solving the transition matrix equation in step S5 converges to below a preset threshold, updating the corresponding target point cloud based on the transition and stopping iteration, completing the registration of the source point cloud and the target point cloud.

[0010] Specifically, in step S1, by using the ISS key point detection algorithm, the inherent geometric properties of the source point cloud and / or the target point cloud are used to identify the local geometric feature difference of each point and its neighborhood points, and to determine whether each point can become a key point and to obtain it.

[0011] Further, the point cloud registration problem between the source point cloud and the target point cloud to be registered is expressed as follows:

[0012] In the formula, is a standard orthogonal matrix, represents that it belongs to a special orthogonal group in three-dimensional space; is a translation vector; is the total number of key points in the source point cloud, and is equal to the total number of key points in the target point cloud; is the th key point in the source point cloud, is the th key point in the target point cloud; is the key point corresponding point pair set between the source point cloud and the target point cloud.

[0013] Further, after obtaining the key points of the source point cloud and the target point cloud and forming the key point corresponding point pair set , by using the local color geometric feature CGD and the global distance distribution feature similarity, the noise point pairs are removed from the key point corresponding point pair set , the correct corresponding matching point pairs between the source point cloud and the target point cloud are obtained, and the SVD decomposition method is used to solve the transition matrix equation to obtain the transition , as follows:

[0014]

[0015] In the formula, is the Identification parameter for whether the key point corresponding point pair is a noise point pair; is the set of key point corresponding pairs The corresponding matching point pair set obtained after removing the noise point pairs.

[0016] Furthermore, for any key point in the source point cloud , when obtaining its local color geometric features CGD, the key points Convert RGB color information into one-dimensional features , at the key point of Neighborhood Search and key points The maximum and minimum points of the absolute value difference of the RGB color information are determined, and the two points are determined to be the key points. The angle between the vectors ; Then, the key point Local color geometric features The calculation formula is as follows:

[0017] Where, For key points of Neighborhood Any point in the target point cloud; For any key point in the target point cloud , using any key point in the source point cloud In the same way, obtain its local color geometric feature CGD and record it as ; When the key points in the source point cloud Local color geometric features and key points in the target point cloud Local color geometric features After obtaining all, determine the key points The closest key point , calculate the similarity of local features between the two , as follows:

[0018] When the local feature similarity is calculated When the similarity of the representation is greater than the preset similarity threshold, the key point With key points As corresponding candidate point pairs; for the key points in the source point cloud and key points in the target point cloud After obtaining multiple sets of corresponding candidate point pairs, they are stored in the corresponding candidate point pair set In the following formula:

[0019] Where, represents the point set containing all source candidate points, Represents the point set containing all target candidate points; and The candidate points in are one-to-one corresponding.

[0020] Furthermore, for the point set Key points in , determine its point set in Neighborhood is ; The key points with it Neighborhood The Euclidean distances of the points in the compose key points Wide-area geometric distribution characteristics of , expressed as the following formula:

[0021] Where, Indicates key points The Euclidean distance to its nearest point, Indicates key points With its The Euclidean distance between neighboring points; then the corresponding candidate point set Correspondence between candidate points in the point set Center and key points The target candidate points corresponding to each neighborhood point , these points and key points Corresponding key points Wide-area geometric distribution characteristics of It is expressed as the following formula:

[0022] Similarly, Indicates key points The Euclidean distance to its nearest point, Indicates key points With its The Euclidean distance between neighboring points; by calculating the wide-area geometric distribution characteristics and The similarity method is used to determine the key points With key points Is the corresponding relationship correct? If it is correct, the key points With key points As a correct corresponding matching point pair between the source point cloud and the target point cloud; after completing the corresponding candidate point pair set After judging the corresponding relationship of the plurality of corresponding candidate point pairs, the plurality of corresponding matching point pairs are obtained, i.e. the corresponding matching point pair set is formed .

[0023] Preferably, the global geometric distribution feature is calculated And The similarity is calculated by using the overall score method, and the calculation formula of the overall score is as follows:

[0024] In the formula, denotes the distance error tolerance, and the calculation formula is as follows:

[0025] If the overall score is greater than a preset overall score threshold, it is judged that the corresponding relationship of the key point and the key point is correct.

[0026] Specifically, in each round of iterative updating process, the target point cloud and the corresponding matching point pair set are updated until the calculated transition amount satisfies the following formula:

[0027] In the formula, denotes the transition amount calculated in the i-th round of iteration ; after satisfying the formula, the corresponding target point cloud is updated, and the iteration is stopped.

[0028] The application also provides a point cloud registration system based on color-enhanced geometric information, which comprises a memory, a processor and a computer program stored in the memory, and the processor executes the computer program to realize the steps of the aforementioned point cloud registration method based on color-enhanced geometric information.

[0029] In summary, due to the adoption of the technical scheme, the application has the following advantages: ​The application effectively improves the distinguishability and stability of features in the point cloud registration process by introducing color information for multi-dimensional feature extraction of point cloud data, and solves the point cloud registration problem in various feature scenes under the combination of local and global features. The local descriptor applied, i.e. the local color geometric feature, contains color and geometric information at the same time, candidate corresponding points are found through the local descriptor, and then noise is eliminated through the global distribution of the candidate corresponding points. The application effectively improves the accuracy of point cloud registration by adding color information and global distribution.

[0030] In the application, the local descriptor calculates the local geometric feature through the distribution of local color information, which enhances the robustness of the geometric feature. In order to eliminate the corresponding ambiguity caused by the similarity of local features between different key points, the global feature is obtained by distributing a large range of candidate points, and the outlier candidate corresponding relationship is removed. Experiments have verified that the registration effect of the method of the application in different feature complexity scenes is better than that of the existing geometric feature method, and since the global distribution feature is added, the application is also better than the existing color point cloud registration method. BRIEF DESCRIPTION OF DRAWINGS

[0031] The embodiments and technical solutions of the application are further described in detail by the following drawings, specifically including 9 drawings, as follows: Figure 1 The schematic diagram of the overall process of the method of the application is briefly described; Figure 2 The detailed schematic diagram of the process of the method of the application is described; Figure 3 The schematic diagram of the registration algorithm process corresponding to the method of the application is described; Figure 4 The schematic diagram of 3Dmatch living room point cloud data in the experimental verification is described; Figure 5 The schematic diagram of 3Dmatch office point cloud data in the experimental verification is described; Figure 6 The schematic diagram of weak geometric feature point cloud data obtained by HD100 in the experimental verification is described; Figure 7 The schematic diagram of the registration result of 3Dmatch living room point cloud data in the experimental verification is described; Figure 8 The schematic diagram of the registration result of 3Dmatch office point cloud data in the experimental verification is described; Figure 9 The schematic diagram of the registration result of HD100 point cloud data in the experimental verification is described. DETAILED DESCRIPTION

[0032] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings can be arranged and designed in various different configurations.

[0033] Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative labor are within the scope of protection of the present application.

[0034] Embodiment 1 A point cloud registration method based on color enhancement of geometric information, Figure 1 The overall flow of the method is shown, which can be referred to simultaneously, and the key steps can be summarized as follows: S1, determining the source point cloud and the target point cloud to be registered, and respectively acquiring the key points of the source point cloud and the target point cloud; S2, using the average color value as the color information reference, acquiring the local color geometric features CGD of the key points in the source point cloud and the target point cloud; S3, preliminarily determining the global distance distribution feature similarity between the corresponding key points in the source point cloud and the target point cloud according to the CGD of the key points of the source point cloud and the target point cloud; S4, acquiring the correct corresponding matching point pair between the source point cloud and the target point cloud according to the preliminarily determined global distance distribution feature similarity, and constructing a transition quantity matrix equation based on the corresponding matching point pair; S5, solving the transition quantity matrix equation to obtain a transition quantity, constructing a transition matrix based on the transition quantity, and updating the target point cloud through the transition matrix; S6, after updating the target point cloud based on the transition matrix, iteratively performing the process of steps S1 to S5 until the transition quantity obtained by solving the transition quantity matrix equation in step S5 converges to be lower than a preset threshold, updating the corresponding target point cloud based on the transition quantity and stopping iteration, and completing the registration of the source point cloud and the target point cloud.

[0035] The flowchart of the method can also be referred to Figure 2The embodiment will be described in detail according to the contents of each part of the method. Firstly, the whole embodiment can be introduced that the ISS key point detection algorithm uses the inherent geometric properties of the point cloud, that is, to identify the local geometric feature difference of each point and its neighborhood point to determine whether the point can be a key point; the algorithm has repeatability and robustness, and is therefore suitable for key point detection in the point cloud registration of the embodiment. Subsequently, since the information weight provided by the RGB three channels in the point cloud is the same, the embodiment therefore uses the average color value as the color information reference to obtain the local color geometric feature CGD (Color Geometry Descriptor) of the key point.

[0036] The embodiment preliminarily determines the corresponding matching point pair through the CGD similarity between the nearest key points in the source point cloud and the target point cloud. Since the CGD fuses color information to compensate for the description difficulty in the weak geometric feature scene, it also has good description ability in the weak geometric feature. The global distance distribution feature similarity between the corresponding key points preliminarily determined by the CGD is used to further obtain the correct corresponding matching point pair. The transition matrix equation is constructed through the corresponding matching point pair, and the transition amount is obtained by using SVD decomposition to solve the matrix equation. The transition amount is used to construct the transition matrix to update the target point cloud, and then the process of CGD feature searching, global feature screening, and SVD decomposition solving matrix is repeated until the transition amount converges to a small value, and the registration is completed.

[0037] The following is a separate introduction of the contents of each part of the method.

[0038] I. Registration problem The most basic point cloud registration problem can be expressed as follows:

[0039] In the formula, is a standard orthogonal matrix, represents that it belongs to the special orthogonal group in three-dimensional space; is a translation vector; is the total number of key points in the source point cloud, and is equal to the total number of key points in the target point cloud; is the th key point in the source point cloud, is the th key point in the target point cloud; is the corresponding point pair set of the key points between the source point cloud and the target point cloud.

[0040] The registration problem is to find an optimal solution, so that the distance sum of the points in the transformed corresponding point pair set is the smallest; in this process is unknown and needs to be estimated; however, it may contain incorrect corresponding point pairs, which will cause registration errors. Usually it consists of the closest points in the source point cloud and the target point cloud. The noise point pairs are removed based on .

[0041] In this embodiment, after the source point cloud and the target point cloud are processed by the ISS key point detection algorithm, the detected key points form a key point corresponding point pair set. Then, the noise is eliminated by local color geometric features and global geometric features to obtain the correct corresponding matching point pairs between the source point cloud and the target point cloud. After solving the transfer matrix equation using the SVD decomposition method, the transfer amount is obtained. , as follows:

[0042]

[0043] Where, For the Identification parameter for whether the key point corresponding point pair is a noise point pair; is the set of key point corresponding pairs The corresponding matching point pair set obtained after removing the noise point pairs.

[0044] 2. Extraction of local color geometric features CGD The color features of a colored point cloud are primarily localized, manifesting as color differences and spatial distribution of the points. Excellent color features should possess the following characteristics: sufficient differentiation between the color features of different points to ensure the accuracy of subsequent matching; a simple extraction process with low feature extraction complexity. Therefore, this embodiment proposes a local color geometric feature (CGD) as a color geometry descriptor. CGD can effectively describe the local information of key points, is computationally simple, consumes little computing power, and shortens the time required for registration.

[0045] In this embodiment, for any key point in the source point cloud , when obtaining its local color geometric features CGD, the key points Convert RGB color information into one-dimensional features , at the key point of Neighborhood Search and key points The maximum and minimum points of the absolute value difference of the RGB color information are determined, and the two points are determined to be the key points. The angle between the vectors ; Then, the key point Local color geometric features The calculation formula is as follows:

[0046] In the formula, is a key point . neighborhood of any point. For any key point in the target point cloud, the local color geometric feature CGD is obtained in the same way as any key point in the source point cloud, and is recorded as .

[0047] When the local color geometric feature of the key point in the source point cloud and the local color geometric feature of the key point in the target point cloud are obtained, the distance closest key point corresponding to the key point is determined, and the local feature similarity of the two is calculated, as follows:

[0048] When the local feature similarity calculated is greater than 0.7, it is considered that the local features of the two points are similar enough, and the two points are taken as a corresponding candidate point pair. Subsequently, in order to ensure that the two points correspond correctly, it is also necessary to ensure the feature similarity between them on a larger scale; here, for the key point in the source point cloud and the key point in the target point cloud, after a plurality of corresponding candidate point pairs are obtained, they are stored in the corresponding candidate point pair set , as follows:

[0049] In the formula, represents a point set containing all source candidate points, represents a point set containing all target candidate points; and are one-to-one corresponding candidate points.

[0050] III. Global geometric similarity judgment In order to eliminate the geometric ambiguity of the candidate point pair in a larger range, the embodiment is screened by the candidate point distance, which determines whether the candidate point can become a corresponding point according to the distance distribution difference of the candidate point in the source point cloud and the target point cloud.

[0051] For the key point in the point set , its distance in the point set in Neighborhood is , since the neighborhood selection is in Based on the source point cloud instead of the entire ,therefore The range is wider and can reflect the global distribution of candidate points. with it Neighborhood The Euclidean distances of the points in the compose key points Wide-area geometric distribution characteristics of , expressed as the following formula:

[0052] Where, Indicates key points The Euclidean distance to its nearest point, Indicates key points With its The Euclidean distance between neighboring points; then the corresponding candidate point set Correspondence between candidate points in the point set Center and key points The target candidate points corresponding to each neighborhood point , these points and key points Corresponding key points Wide-area geometric distribution characteristics of It is expressed as the following formula:

[0053] Similarly, Indicates key points The Euclidean distance to its nearest point, Indicates key points With its The Euclidean distance between neighboring points; by calculating the wide-area geometric distribution characteristics and The similarity method is used to determine the key points With key points If the similarity between the two is high, it means that the correspondence is reliable and the probability that they are the same mapping in the real world is higher. With key points As the correct corresponding matching point pair between the source point cloud and the target point cloud; after completing the corresponding candidate point pair set After judging the correspondence relationship of multiple sets of corresponding candidate point pairs, multiple sets of corresponding matching point pairs are obtained, that is, a corresponding matching point pair set is formed. .

[0054] In this embodiment, the wide-area geometric distribution characteristics are calculated and The overall score method is used when the similarity is The calculation formula is as follows:

[0055] Where, Represents the distance error tolerance, which is calculated as follows:

[0056] If the total score Greater than , then determine the key points With key points The corresponding relationship is correct.

[0057] 4. Point Cloud Registration In each round of iterative update, the target point cloud and the corresponding matching point pair set are updated. , until the calculated transfer amount Satisfy the following formula:

[0058] Where, Indicates the The transfer amount calculated during the round iteration ; After satisfying this formula, the corresponding target point cloud is updated and the iteration is stopped, thereby completing the registration of the source point cloud and the target point cloud.

[0059] The overall registration process of the method can be described as follows when implemented in a computer system: Figure 3 The algorithm process shown. The above method of this embodiment can therefore be formed into a point cloud registration system based on color enhancement geometric information, which includes a memory, a processor, and a computer program stored in the memory; when the processor executes the computer program, the steps of the above-mentioned point cloud registration method based on color enhancement geometric information are implemented.

[0060] Example 2 Based on Example 1, this example demonstrates the registration performance of the present method in weak geometric scenes through experimental verification. This example uses a 3Dmatch dataset, a point cloud dataset with tens to hundreds of thousands of points per scan, and a spacing of approximately 1 mm between points. This example also uses an HD100 binocular structured light camera to acquire point cloud data in scenes with weak geometric features.

[0061] The experimental verification process of the embodiment runs on a Windows 11 system, uses an Intel(R) i9-12900H CUP, and has a memory of 16 GB of RAM. The algorithm program corresponding to the method is written in a Visual C++ development environment and is compiled and run by using PCL 1.11.1 + Visual Studio 2019.

[0062] The embodiment also compares several existing advanced registration algorithms. The first is a variant of the ICP algorithm with multiple geometric features, GFO-ICP; the second is CPCR, which uses color information and geometric information to find the best correspondence by projection relationship; and the third is GROR, which uses fast point feature histogram (FPFH) to select corresponding matches and uses a correspondence graph to remove abnormal correspondences. As shown in Table 1, multiple sets of parameters are set for each algorithm, and the best parameters on each data are used as the parameters for comparing each algorithm in the embodiment.

[0063] Table 1 shows the parameter setting table of each algorithm

[0064] The embodiment evaluates the registration results from the program running time, registration rotation error, and translation error. The rotation error and the translation error are calculated as follows:

[0065]

[0066] wherein, is the real rotation matrix; is the estimated rotation matrix; is the real translation vector; is the estimated translation vector.

[0067] Figure 4 and Figure 5 are the living room and office data of 3Dmatch, respectively. These two sets of data have relatively obvious geometric features, and the color information is also rich, so the embodiment selects these two sets of data as experimental data. Figure 6 HD100 is weak geometric feature point cloud data collected, which is used to verify that the algorithm of the embodiment can still use color information to estimate correct correspondence under weak geometric features.

[0068] Figure 7Fig. 2 and Table 2 are respectively the registration result images and evaluation indexes of 3Dmatch living room. The average rotation error of the method of the present application on this data is 0.39 deg, the average translation error is 42.1 mm, and the average running time is 2.2 s. The average rotation error of CPCR is 0.42 deg, the average translation error is 49.1 mm, and the average running time is 2.5 s. The average rotation error of GROR is 0.49 deg, the average translation error is 55.7 mm, and the average running time is 2.8 s. The average rotation error of GFO-ICP is 0.48 deg, the average translation error is 58.7 mm, and the average running time is 2.4 s.

[0069] Table 2 Evaluation table of registration results of different algorithms for 3Dmatch living room point cloud data

[0070] Figure 8 Fig. 3 and Table 3 are respectively the registration result images and evaluation indexes of 3Dmatch office. The average rotation error of the method of the present application on this data is 0.36 deg, the average translation error is 46.4 mm, and the average running time is 2.2 s. The average rotation error of CPCR is 0.46 deg, the average translation error is 62.2 mm, and the average running time is 3.0 s. The average rotation error of GROR is 0.43 deg, the average translation error is 53.6 mm, and the average running time is 3.1 s. The average rotation error of GFO-ICP is 0.52 deg, the average translation error is 62.4 mm, and the average running time is 2.3 s.

[0071] Table 3 Evaluation table of registration results of different algorithms for 3Dmatch office point cloud data

[0072] In the registration experiment of 3Dmatch data, the indexes of the method of the present application are the best. Among the compared algorithms, CPCR has the best effect, and the main reason is that it also considers the color information of the point cloud, so that the corresponding matching accuracy is increased. GROR considers the global distribution characteristics of key points, and GFO-ICP considers the fusion of multiple local geometric features, while the geometric features of 3Dmatch data set are obvious, so adding global features makes the corresponding matching more reliable, so the registration accuracy of GROR algorithm is slightly higher than that of GFO-ICP.

[0073] Figure 9Table 4 is the registration result and evaluation index of the point cloud data of weak geometric feature scene collected by HD100. The average rotation error of the method of the application on this data is 0.36 deg, the average translation error is 25.5 mm, and the average running time is 3.0 s. The average rotation error of CPCR is 0.44 deg, the average translation error is 29.5 mm, and the average running time is 3.3 s. The average rotation error of GROR is 1.41 deg, the average translation error is 53.5 mm, and the average running time is 5.4 s. The average rotation error of GFO-ICP is 0.59 deg, the average translation error is 39.3 mm, and the average running time is 2.5 s.

[0074] Table 4 is the registration result and evaluation index of the point cloud data of weak geometric feature scene collected by HD100. The average rotation error of the method of the application on this data is 0.36 deg, the average translation error is 25.5 mm, and the average running time is 3.0 s. The average rotation error of CPCR is 0.44 deg, the average translation error is 29.5 mm, and the average running time is 3.3 s. The average rotation error of GROR is 1.41 deg, the average translation error is 53.5 mm, and the average running time is 5.4 s. The average rotation error of GFO-ICP is 0.59 deg, the average translation error is 39.3 mm, and the average running time is 2.5 s.

[0075] The point cloud data collected by HD100 contains a large number of planar structures, resulting in a lack of obvious geometric features in the data, which makes the algorithms such as GROR and GFO-ICP that rely only on point distribution and geometric features for corresponding matching to produce a large number of incorrect correspondences, increasing the registration error. CPCR considers color information in the point cloud but ignores the global distribution of key points and only considers local feature information.

[0076] The method of the application uses color information to make up for the lack of information in the weak geometric feature scene, and uses global distribution features to further ensure the correctness of the corresponding matching, so that it can still obtain stable registration results in the weak geometric feature scene.

Claims

1. A point cloud registration method based on color-enhanced geometric information, characterized in that: The method comprises the following steps: S1. Determine the source point cloud and target point cloud to be registered, and obtain the key points of the source point cloud and target point cloud respectively; S2, using the average color value as a color information reference, obtain the local color geometric features CGD of the key points in the source point cloud and the target point cloud; S3, based on the CGD of the key points of the source point cloud and the target point cloud, preliminarily determine the similarity of the global distance distribution features between the corresponding key points in the source point cloud and the target point cloud; S4. According to the preliminarily determined global distance distribution feature similarity, the correct corresponding matching point pairs between the source point cloud and the target point cloud are obtained, and a transfer matrix equation is constructed based on the corresponding matching point pairs; S5. Solve the transfer matrix equation to obtain the transfer amount, construct a transfer matrix based on the transfer amount, update the target point cloud through the transfer matrix, and complete the alignment of the source point cloud and the target point cloud.

2. The point cloud registration method according to claim 1, wherein: After step S5, the method further includes: S6. After updating the target point cloud based on the transfer matrix, iterate the process from step S1 to step S5 until the transfer amount obtained by solving the transfer amount matrix equation in step S5 converges to below the preset threshold, update the corresponding target point cloud based on the transfer amount and stop the iteration to complete the alignment of the source point cloud and the target point cloud.

3. The point cloud registration method according to claim 2, wherein: In step S1, the ISS key point detection algorithm is used to identify the differences in local geometric features between each point and its neighboring points using the inherent geometric properties of the source point cloud and / or target point cloud, and to determine whether each point can become a key point and obtain it.

4. The point cloud registration method according to claim 2, wherein: The point cloud registration problem between the source point cloud and the target point cloud to be registered is expressed as follows: Where, is a standard orthogonal matrix, It represents a special orthogonal group in three-dimensional space; is the translation vector; is the total number of key points in the source point cloud and is equal to the total number of key points in the target point cloud; is the first point in the source point cloud A key point, is the first point in the target point cloud key points; is the set of key point correspondences between the source point cloud and the target point cloud.

5. The point cloud registration method according to claim 4, wherein: After obtaining the key points of the source point cloud and the target point cloud and forming a key point corresponding point pair set Then, the local color geometric feature CGD and the global distance distribution feature similarity are used to obtain the corresponding point pair set from the key points. After removing the noise point pairs, the correct corresponding matching point pairs between the source point cloud and the target point cloud are obtained. After solving the transfer matrix equation using the SVD decomposition method, the transfer amount is obtained. , as follows: Where, For the Identification parameter for whether the key point corresponding point pair is a noise point pair; is the set of key point corresponding pairs The corresponding matching point pair set obtained after removing the noise point pairs.

6. The point cloud registration method according to claim 4, wherein: For any key point in the source point cloud , when obtaining its local color geometric features CGD, the key points Convert RGB color information into one-dimensional features , at the key point of Neighborhood Search and key points The maximum and minimum points of the absolute value difference of the RGB color information are determined, and the two points are determined to be the key points. The angle between the vectors ; Then, the key point Local color geometric features The calculation formula is as follows: Where, For key points of Neighborhood Any point in the target point cloud; For any key point in the target point cloud , using any key point in the source point cloud In the same way, obtain its local color geometric feature CGD and record it as ; When the key points in the source point cloud Local color geometric features and key points in the target point cloud Local color geometric features After obtaining all, determine the key points The closest key point , calculate the similarity of local features between the two , as follows: When the local feature similarity is calculated When the similarity of the representation is greater than the preset similarity threshold, the key point With key points As corresponding candidate point pairs; for the key points in the source point cloud and key points in the target point cloud After obtaining multiple sets of corresponding candidate point pairs, they are stored in the corresponding candidate point pair set In the following formula: Where, represents the point set containing all source candidate points, Represents the point set containing all target candidate points; and The candidate points in are one-to-one corresponding.

7. The point cloud registration method according to claim 6, characterized in that: For point sets Key points in , determine its point set in Neighborhood is ; The key points with it Neighborhood The Euclidean distances of the points in the compose key points Wide-area geometric distribution characteristics of , expressed as the following formula: Where, Represents key points The Euclidean distance to its nearest point, Represents key points With its The Euclidean distance between neighboring points; then the corresponding candidate point set Correspondence between candidate points in the point set Center and key points The target candidate points corresponding to each neighborhood point , these points and key points Corresponding key points Wide-area geometric distribution characteristics of It is expressed as the following formula: Similarly, Represents key points The Euclidean distance to its nearest point, Represents key points With its The Euclidean distance between neighboring points; by calculating the wide-area geometric distribution characteristics and The similarity method is used to determine the key points With key points Is the corresponding relationship correct? If it is correct, the key points With key points As the correct corresponding matching point pairs between the source point cloud and the target point cloud; After completing the corresponding candidate point pair set After judging the correspondence relationship of multiple sets of corresponding candidate point pairs, multiple sets of corresponding matching point pairs are obtained, that is, a corresponding matching point pair set is formed. .

8. The point cloud registration method according to claim 7, wherein: Calculate wide-area geometric distribution characteristics and The overall score method is used when the similarity is The calculation formula is as follows: Where, Represents the distance error tolerance, which is calculated as follows: If the total score If the total score is greater than the preset threshold, the key point is determined With key points The corresponding relationship is correct.

9. The point cloud registration method according to claim 5, wherein: In each round of iterative update, the target point cloud and the corresponding matching point pair set are updated. , until the calculated transfer amount Satisfy the following formula: Where, Indicates the The transfer amount calculated during the round iteration ; After satisfying this formula, the corresponding target point cloud is updated and the iteration is stopped.

10. A point cloud registration system based on color-enhanced geometric information, comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the point cloud registration method based on color enhancement geometric information according to any one of claims 1 to 9.

Images