Multi-view three-dimensional point cloud data coarse registration method based on branch and bound

Abstract

The invention provides a multi-view three-dimensional point cloud data coarse registration method based on branch and bound. The method comprises the following steps: 1, extracting three-dimensional point cloud data feature points to be registered and obtaining a point set; step 2, processing the point set obtained in the step 1 to preliminarily obtain matching points in the three-dimensional point cloud data; 3, taking the matching point set preliminarily obtained in the step 2 as input data, further screening feature points, and accurately obtaining matching points in the three-dimensional point cloud data; and step 4, taking the accurate matching point set obtained in the step 3 as input data, and carrying out camera pose parameter estimation, so that compared with the existing three-dimensional point cloud data registration method, the Chebyshev distance used by the method is stricter, and the method has the characteristics of high speed and high accuracy, and can be applied to the field of three-dimensional point cloud data registration. According to the invention, coarse registration can be carried out more quickly and more accurately, so that more computing resources and time are saved in the subsequent accurate registration process.

Description

Technical Field

[0001] The present invention belongs to the technical field of three-dimensional point cloud data registration, and particularly relates to a rough registration method for multi-view three-dimensional point cloud data based on branch and bound. Background Art

[0002] In the current wave of technological development, three-dimensional reconstruction technology, with its powerful spatial modeling ability, has been widely applied in such cutting-edge fields as AR / VR immersive interaction, robot intelligent navigation, digital library / museum cultural heritage digitization, and unmanned autonomous driving environment perception. Among them, multi-view three-dimensional reconstruction, as a key direction of technological development, has become a research focus jointly concerned by the academic and industrial communities. And three-dimensional point cloud data registration technology, as the core link of multi-view three-dimensional reconstruction, its importance is mainly reflected in the following three dimensions:

[0003] First of all, from the perspective of the entire three-dimensional reconstruction process, point cloud data registration is a key hub connecting the upstream and downstream. In the data acquisition stage, the three-dimensional point cloud data collected by devices such as lidar and depth cameras is often scattered in different coordinate systems; while in the data fusion stage, it is necessary to integrate these scattered point cloud data into a unified three-dimensional model. Point cloud data registration, as the intermediate link connecting data acquisition and fusion, its accuracy and efficiency directly determine the final quality of three-dimensional reconstruction.

[0004] Secondly, in the application scenario of multi-sensor collaborative acquisition, due to the differences in the installation positions and viewing angles of each depth sensor, the obtained two-dimensional image information needs to be converted into three-dimensional point cloud data. In this process, establishing an accurate coordinate transformation relationship becomes a prerequisite for achieving high-precision three-dimensional point cloud data fusion. Only through precise point cloud registration can the point cloud data in different coordinate systems be unified into the same spatial reference system, so as to construct a three-dimensional reconstruction model that truly reflects the shape of the target object.

[0005] Finally, the application value of three-dimensional point cloud data registration technology is not limited to the three-dimensional reconstruction field, but is also the core technical support for augmented reality (AR) virtual-real fusion and robot simultaneous localization and mapping (SLAM) algorithms. In AR applications, accurate point cloud registration can achieve seamless fusion of virtual objects and real scenes; in the robot SLAM system, through point cloud registration technology, an environmental map can be constructed in real time and the position of the robot itself can be determined, providing a basic guarantee for autonomous navigation.

[0006] From a technical implementation perspective, the 3D point cloud data registration process is generally divided into two stages: rough registration and fine registration. As a preprocessing step for fine registration, rough registration mainly uses a fast matching algorithm to preliminarily align the 3D point cloud data from different viewpoints, reducing the spatial position difference between the data. This process can significantly reduce the computational complexity required for subsequent fine registration, effectively save computational resources, and greatly improve the real-time performance and operating efficiency of the 3D reconstruction system, thus providing strong support for the practical application of multi-viewpoint 3D reconstruction technology.

[0007] Therefore, the research on the rough registration technology of 3D point cloud data has important practical significance and theoretical value. Summary of the Invention

[0008] The present invention proposes a rough registration method for multi-viewpoint 3D point cloud data based on branch and bound, which realizes the process of roughly registering the 3D point cloud data from two different viewpoints.

[0009] The technical solution of the present invention is implemented as follows: A rough registration method for multi-viewpoint 3D point cloud data based on branch and bound, including:

[0010] The first step is to extract feature points from the multi-viewpoint 3D point cloud data to be registered to obtain an initial feature point set;

[0011] The second step is to remove the points in the flat area from the initial feature point set according to the feature distance threshold to obtain a matching candidate point set;

[0012] The third step is to further screen based on the similarity of multi-dimensional geometric feature vectors with the matching candidate point set as the input to accurately obtain a set of matching point pairs;

[0013] The fourth step is to use the accurate set of matching point pairs to estimate the initial pose parameters of the camera by the quaternion method and calculate the rotation matrix and translation vector.

[0014] As a preferred implementation manner, in the first step, the feature point extraction includes:

[0015] 2.1 Input two pieces of 3D point cloud data P and Q to be registered;

[0016] 2.2 Calculate the distance from each adjacent point in the K-neighborhood of each point to the tangent plane of the point;

[0017] 2.3 Retain the adjacent points with a distance greater than the preset threshold and remove the adjacent points with a distance less than the threshold;

[0018] 2.4 Take the average value of the distances of each adjacent point as the feature distance of the point. If the feature distance is greater than the preset threshold, the point is marked as a feature point to form an initial feature point set.

[0019] As a preferred embodiment, the specific operation method in the feature point extraction step is as follows: For two point clouds to be registered, if the entire point cloud is searched for matching point pairs, it will consume a large amount of time and there will also be a large number of incorrect matching point pairs. In order to perform fast and accurate registration, it is necessary to preprocess the point cloud to obtain a feature point set. By observing the tangent plane of the points, it is easy to find that if the distance from the neighborhood point set of a local area of the point cloud to the tangent plane is small, it means that this area is relatively flat and the feature is not obvious; on the contrary, if the distance from the neighborhood point set to the tangent plane is large, it indicates that the area has large fluctuations and the feature is more obvious. Based on the above observations, define a certain point P in the three-dimensional point cloud data i The average distance from the points within its K-neighborhood to the tangent plane of this point is the feature distance g i As the basis for determining whether point P i is a feature point, the calculation formula is:

[0020]

[0021] where d ij refers to the distance from a certain point in the neighborhood of point P i to the tangent plane of point P i , and g i is the average distance from the neighborhood points of point P i to the tangent plane. According to this definition, the point with a larger feature distance indicates that the area has large fluctuations. Select a threshold σ1, remove the flat points in the three-dimensional point cloud data, and retain the points in the three-dimensional point cloud data where g i > σ1. For any point P n among the retained points, if it satisfies,

[0022] g(P n ) = max[g(P n1 ), g(P n2 ),..., g(P nk )]

[0023] then P n is taken as a feature point, where g(P n1 ), g(P n2 ),..., g(P nk ) are the feature distances of the K-nearest neighbors of point P n . Assume that the two three-dimensional point cloud data are P and Q respectively, where P is the target point set and Q is the reference point set. Feature extraction is performed on the two three-dimensional point cloud data respectively, and the feature point set of P is P t = {P t1 , P t2 ,..., P tm}, and the feature point set of Q is Q t = {Q t1 , Qt2 , …, Q tn}, where m and n are the numbers of feature points of P and Q respectively.

[0024] As a preferred embodiment, the step of obtaining matching points in the three-dimensional point cloud data in the second step includes: ① calculating four feature quantities of each point in the two point clouds respectively; ② obtaining a matching point set according to the principle of the same or similar feature quantities. In the second step, the point set obtained after the feature points in the first step are extracted is used as the input data of the second step. In the second step, the following 4 feature description operators are calculated for each point in the point set obtained in the first step. According to the 4 solved feature description operators, a feature vector point set is calculated. The present invention selects four basic geometric features as the feature description operators to initially search for matching point pairs in the feature point set. Using multiple geometric feature information can not only describe the neighborhood feature information more carefully and accurately, but also avoid the appearance of a large number of incorrect matching point pairs. Taking the target point set as an example, the four geometric features are described as follows:

[0025] As a preferred embodiment, the first feature quantity is the feature distance of the K-nearest neighbors of each point P t in the point set P ti . The calculation formula is as follows:

[0026] f1(P ti ) = g(P ti );

[0027] The second feature quantity is the centroid O(P t ) of the K-nearest neighbors of each point P ti calculated according to the point set P ti . The distance value between this point and its K-nearest neighbor centroid O(P ti ) is calculated as follows:

[0028] f2(P ti ) = P ti - O(P ti ); p

[0029] The third feature quantity is the degree of change of the normal vector of each point calculated according to the point set P t , that is, the geometric mean of the cosine of the angle between the normal vector of this point and the normal vectors of its K-nearest neighbors. The calculation formula is as follows:

[0030]

[0031] The fourth feature quantity is the cosine value of the angle between the normal vector n t of each point in the point set P i and the line connecting this point and the centroid of its K-nearest neighbors. The calculation formula is as follows:

[0032] f4(P ti ) = cos < n i , [O(P ti ) - P ti >

[0033] By solving the above four characteristic quantities, the eigenvectors of P ti and Q ti are obtained, and the eigenvector point sets L1 and L2 of P t and Q t are obtained respectively.

[0034] As a preferred implementation manner, the method for obtaining the matching point set according to the principle of the same or similar characteristic quantities is as follows: For each point in the point set P t , search for the matching point in the point set Q t . If P ti and Q ti are a pair of matching points, they should have the same or at least similar geometric characteristics, so the eigenvectors should also be the same or similar; in the step of initially obtaining the matching feature point set, the Chebyshev distance in the eigenvector space should be used as the comparison criterion, and when the following conditions are met at the same time, it can be considered as a preliminary matching point:

[0035]

[0036]

[0037] The σ i parameters in the above four conditions are all set to 0.01, and the preliminary feature matching point set can be obtained according to the above conditions. Since there may be multiple feature-similar regions in the two point clouds, in order to avoid a large number of wrong matching point pairs, multiple feature correspondence relationships are comprehensively searched for matching points, and a reasonable threshold is selected to establish the preliminary matching point set and denoted as:

[0038] W = {(mi1, m i2 ) | m i1 ∈P t , m i2 ∈Q t , i = 1, 2, 3,..., N}.

[0039] As a preferred implementation manner, in the third step, the preliminary obtained matching point set in the second step will be used as the input data for further screening of the feature points. After the following screening, the matching points in the point cloud will be accurately obtained as the output, and then the external parameters of the camera will be calculated. The operation method for accurately obtaining the matching points in the three-dimensional point cloud data in the third step is as follows: Set a new threshold σ6 = 0.01, and for each matching point pair (m i1 , mi2 ) ∈ W, calculate the number N of point pairs in W that meet the distance constraint conditions m , if a point pair in W satisfies:

[0040]

[0041] then this point pair is considered a qualified point pair. If most of the point pairs in W meet the conditions, then (mi1, m i2 ) is a correct point pair; otherwise, it is incorrect and should be excluded. Set another threshold σ7 = 0.8. If the N i1 , m i2 ) calculated m satisfies:

[0042] N m ≥ σ7 * N

[0043] it can be considered a correct matching point pair; otherwise, it is considered incorrect and the point pair is excluded from the point set to obtain the final matching point set:

[0044] W = {(m i1 , m i2 ) | m i1 ∈ P t , m i2 ∈ Q t , i = 1, 2, 3,..., N}

[0045] where N' is the number of matching point pairs after excluding incorrect matching point pairs.

[0046] As a preferred implementation manner, in the fourth step, the accurate matching point set obtained in the third step is used as input data to roughly calculate the camera pose parameter matrix, and the rotation matrix and translation vector parameters are obtained.

[0047] As a preferred implementation manner, the calculation of the initial camera pose parameters adopts the quaternion method, and the calculation methods of the camera rotation matrix and translation vector are as follows: According to the result in the third step, the point m i1 in the target three-dimensional point cloud data is denoted as m' i1 after rotation and translation transformation.

[0048] After adopting the above technical solutions, the beneficial effects of the present invention are:

[0049] Compared with the existing three-dimensional point cloud data registration methods, the method of the present invention uses the Chebyshev distance more strictly, has the characteristics of fast speed and high accuracy. In the application of the rough registration method, it can perform rough registration more quickly and accurately, making the subsequent accurate registration process more time - and computing - resource - saving. Description of the Drawings

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0051] Figure 1 It is a flow schematic diagram of the present invention. Specific embodiments

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0053] As Figure 1 shown, a rough registration method for multi-viewpoint three-dimensional point cloud data based on branch and bound includes:

[0054] The first step is to extract feature points from the multi-viewpoint three-dimensional point cloud data to be registered to obtain an initial feature point set;

[0055] The second step is to remove the points in the flat area from the initial feature point set according to the feature distance threshold to obtain a matching candidate point set;

[0056] The third step is to further screen based on the similarity of multi-dimensional geometric feature vectors with the matching candidate point set as the input to accurately obtain a set of matching point pairs;

[0057] The fourth step is to use the set of accurate matching point pairs to estimate the initial pose parameters of the camera by the quaternion method and calculate the rotation matrix and translation vector.

[0058] In the first step, the feature point extraction includes:

[0059] 2.1 Input two pieces of three-dimensional point cloud data P and Q to be registered;

[0060] 2.2 Calculate the distance from each neighbor point to the tangent plane of the point within the K-neighborhood of each point;

[0061] 2.3 Retain the neighbor points with a distance greater than the preset threshold and remove the neighbor points with a distance less than the threshold;

[0062] 2.4 Take the average value of the distances of each neighbor point as the feature distance of the point. If the feature distance is greater than the preset threshold, the point is recorded as a feature point to form an initial feature point set.

[0063] The specific operation method in the feature point extraction step is as follows: For the two point clouds to be registered, if the entire point cloud is searched for matching point pairs, it will consume a large amount of time and there will also be a large number of incorrect matching point pairs. In order to perform fast and accurate registration, it is necessary to preprocess the point cloud to obtain a feature point set. By observing the tangent plane of the points, it is easy to find that if the distance from the neighborhood point set of the local area of the point cloud to the tangent plane is small, it means that this area is relatively flat and the feature is not obvious; on the contrary, if the distance from the neighborhood point set to the tangent plane is large, it indicates that the area has large fluctuations and the feature is relatively obvious. Based on the above observations, a certain point P in the three-dimensional point cloud data is defined i The average distance from the points within the K-neighborhood of this point to the tangent plane of this point is the feature distance g i As the basis for determining whether point P i is a feature point, the calculation formula is:

[0064]

[0065] where d ij refers to the distance from a certain point in the neighborhood of point P i to the tangent plane of point P i , and g i is the average distance from the neighborhood points of point P i to the tangent plane of the point. According to this definition, the greater the feature distance of a point, the greater the fluctuations in this area. Select the threshold σ1, remove the flat points in the three-dimensional point cloud data, and retain the points in the three-dimensional point cloud data where g i >σ1. For any point P n among the retained points, if it satisfies,

[0066] g(P n ) = max[g(P n1 ), g(P n2 ), …, g(P nk )]

[0067] then P n is taken as the feature point, where g(P n1 ), g(P n2 ),..., g(P nk ) are the feature distances of the K-nearest neighbors of point P n . Assume that the two three-dimensional point cloud data are P and Q respectively, where P is the target point set and Q is the reference point set. Feature extraction is performed on the two three-dimensional point cloud data respectively, and the feature point set of P is P t = {P t1 , P t2 , …, P tm}, and the feature point set of Q is Q t = {Q t1 , Qt2 , …, Q tn} where m and n are the numbers of feature points of P and Q respectively.

[0068] The steps of obtaining the matching points in the three-dimensional point cloud data in the second step include: ① calculating four feature quantities of each point in the two point clouds respectively; ② obtaining the matching point set according to the principle of the same or similar feature quantities. In the second step, the point set obtained after extracting the feature points in the first step is used as the input data of the second step. In the second step, the following 4 feature description operators are calculated for each point in the point set obtained in the first step. According to the 4 solved feature description operators, the feature vector point set is calculated. The present invention selects four basic geometric features as the feature description operators to initially search for the matching point pairs in the feature point set. Using multiple geometric feature information can not only describe the neighborhood feature information more carefully and accurately, but also avoid the appearance of a large number of wrong matching point pairs. Taking the target point set as an example, the four geometric features are described as follows:

[0069] , the first feature quantity is the feature distance of the K-nearest neighbor points of each point P t in the point set P ti , and the calculation formula is as follows:

[0070] f1(P ti ) = g(P ti );

[0071] The second feature quantity is the centroid O(P t ) of the K-nearest neighbor points of each point P ti calculated according to the point set P ti . The distance value between this point and the centroid O(P ti ) of its K-nearest neighbor points, and the calculation formula is as follows:

[0072] f2(P ti ) = P ti - O(P ti ); p

[0073] The third feature quantity is the degree of change of the normal vector of each point calculated according to the point set P t , that is, the geometric mean of the cosine of the angle between the normal vector of this point and the normal vectors of the K-nearest neighbor points. The calculation formula is as follows:

[0074]

[0075] The fourth feature quantity is the cosine value of the angle between the normal vector n t of each point in the point set P i and the line connecting this point and the centroid of its K-nearest neighbor points. The calculation formula is as follows:

[0076] f4(Pti ) = cos <n i , [O(P ti ) - P ti >

[0077] By solving the above four characteristic quantities, the eigenvectors of P ti and Q ti are obtained, and the eigenvector point sets L1 and L2 of P t and Q t are obtained respectively.

[0078] The method for obtaining the matching point set according to the principle of the same or similar characteristic quantities is as follows: For each point in the point set P t , search for the matching point in the point set Q t . If P ti and Q ti are a pair of matching points, they should have the same or at least similar geometric characteristics. Therefore, the eigenvectors should also be the same or similar. In the step of initially obtaining the matching feature point set, the Chebyshev distance in the eigenvector space should be used as the comparison criterion. When the following conditions are met at the same time, they can be considered as initially matching points:

[0079]

[0080] The σ i parameters in the above four conditions are all set to 0.01. According to the above conditions, the initial feature matching point set can be obtained. Since there may be multiple feature-similar regions in the two point clouds, in order to avoid a large number of false matching point pairs, multiple feature correspondence relationships are comprehensively searched for matching points, and a reasonable threshold is selected to establish the initial matching point set and denoted as:

[0081] W = {(m i1 , m i2 ) | m i1 ∈P t , m i2 ∈Q t , i = 1, 2, 3,..., N}.

[0082] In the third step, the initially obtained matching point set in the second step will be used as the input data for further screening of the feature points. After the following screening, the matching points in the point cloud will be accurately obtained as the output, and then the external parameters of the camera will be calculated. The operation method for accurately obtaining the matching points in the three-dimensional point cloud data in the third step is as follows: Set a new threshold σ6 = 0.01, and calculate the number N i1 of point pairs in W that meet the distance constraint condition for each matching point pair (m i2 ) ∈ W obtained in the second step. If a point pair in W satisfies: m If

[0083]

[0084] Then it is considered that this point is a qualified point pair. If most of the point pairs in W meet the conditions, then (m i1 , m i2 ) is a correct point pair; otherwise, it is incorrect and should be excluded. Set another threshold σ7 = 0.8. If the N i1 , m i2 ) calculated for the point pair satisfies: m Satisfied:

[0085] N m ≥σ7*N

[0086] It can be considered as a correct matching point pair; otherwise, it is considered incorrect and the point pair is excluded from the point set to obtain the final matching point set:

[0087] W = {(m i1 , m i2 )|m i1 ∈P t , m i2 ∈Q t , i = 1, 2, 3,..., N}

[0088] where N' is the number of matching point pairs after excluding incorrect matching point pairs.

[0089] In the fourth step, the accurate matching point set obtained in the third step is used as input data to roughly calculate the camera pose parameter matrix, and the rotation matrix and translation vector parameters are obtained. The calculation of the initial camera pose parameters uses the quaternion method. The calculation methods of the camera rotation matrix and translation vector are as follows: According to the result in the third step, the point m i1 in the target three-dimensional point cloud data is transformed through rotation and translation, and the obtained point is denoted as m' i1 :

[0090] m' i1 = R*m i1 + T.

[0091] The feature operator proposed by the present invention: the feature distance of the K-nearest points of each point P t in the point set P; the feature quantity is the distance value between each point P ti in the point set P and the centroid O(P t ) of its K-nearest points, calculated based on each point P ti in the point set P; according to the point set P ti ti t ​​For each point in, calculate the degree of change of the normal vector of this point, that is, the geometric mean of the cosine of the angle between the normal vector of this point and the normal vectors of its K-nearest neighbor points; it is based on the point set P t For each point in, with the normal vector n of this point i and the cosine value of the angle between the line connecting this point and the centroid of its K-nearest neighbor points.

[0092] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by the terms "longitudinal", "transverse", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of the present invention. In the description of the present invention, unless otherwise specified and defined, it should be noted that the terms "installation", "connection", "connection" should be understood in a broad sense. For example, it can be a mechanical connection or an electrical connection, or it can be the communication inside two elements. It can be directly connected or indirectly connected through an intermediate medium. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific situations.

[0093] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A rough registration method for multi-viewpoint 3D point cloud data based on branch and bound, characterized in that, Including: The first step is to extract feature points from the multi-viewpoint 3D point cloud data to be registered, obtaining an initial feature point set. The second step is to remove the points in the flat area from the initial feature point set according to the feature distance threshold to obtain a matching candidate point set. The third step is to further screen based on the similarity of multi-dimensional geometric feature vectors with the matching candidate point set as the input to accurately obtain a set of matching point pairs. The fourth step is to use the accurately matched point pair set to estimate the initial pose parameters of the camera by the quaternion method and calculate the rotation matrix and translation vector.

2. A rough registration method for multi-viewpoint three-dimensional point cloud data based on branch and bound, characterized in that In the first step, the feature point extraction includes: 2.1 Input two pieces of 3D point cloud data P and Q to be registered. 2.2 Calculate the distance from each neighbor point in the K-neighborhood of each point to the tangent plane of the point. 2.3 Keep the neighbor points whose distance is greater than the preset threshold and remove the neighbor points whose distance is less than the threshold. 2.4 Take the average value of the distances of each neighbor point as the feature distance of the point. If the feature distance is greater than the preset threshold, the point is marked as a feature point, forming an initial feature point set.

3. A rough registration method for multi-viewpoint 3D point cloud data based on branch and bound, as claimed in claim 2, wherein Define a certain point P in the three-dimensional point cloud data i At this point, the average distance from the points within its K-neighborhood to the tangent plane of this point is the characteristic distance g i As the basis for determining whether the point P i is a feature point, the calculation formula is: where d ij the designated point P i the distance from a certain point in the neighborhood to P i to the tangent plane of the point, and g i is the average distance from the neighborhood points of point P i to the tangent plane of the point. According to this definition, the larger the characteristic distance of a point, the greater the undulation change in the area. Select a threshold σ1 to remove the flat points in the three-dimensional point cloud data of the points and retain the points in the three-dimensional point cloud data where g i > σ1. For any point P n among the retained points, if it satisfies, g(P n ) = max[g(P n1 ), g(P n2 ), ..., g(P nk )] Then take P n as the feature point, where g(P n1 ), g(P n2 ),..., g(P nk ) are the feature distances of the K-nearest points of point P n . Assume that the two pieces of 3D point cloud data are P and Q respectively, where P is the target point set and Q is the reference point set. Feature extraction is performed on the two pieces of 3D point cloud data respectively, and the feature point set of P is P t = {P t1 , P t2 ,..., P tm}, and the feature point set of Q is Q t = {Q t1 , Q t2 , …, Q tn}, where m and n are the numbers of feature points of P and Q respectively.

4. A rough registration method for multi-viewpoint three-dimensional point cloud data based on branch and bound, characterized in that, The steps to obtain matching points in the 3D point cloud data in the second step include: ① Calculate four feature quantities of each point in the two point clouds respectively; ② Obtain a set of matching points according to the principle of the same or similar feature quantities.

5. A rough registration method for multi-viewpoint three-dimensional point cloud data based on branch and bound, characterized in that, The first type of the feature quantity is the feature distance of the K-nearest neighbors of each point P in the point set P, and the calculation formula is as follows: t in the point set P ti The calculation formula is as follows: f1(P ti ) = g(P ti ); The second characteristic quantity is based on the point set P t For each point P ti in it, the centroid O(P ti ) of the K-nearest neighbors of this point is calculated. Taking the distance value between this point and the centroid O(P ti ) of its K-nearest neighbors, the calculation formula is as follows: f2(P ti ) = P ti - O(P ti ); p The third characteristic quantity is based on each point in the point set P t to calculate the degree of change of the normal vector of this point, that is, the geometric mean of the cosine of the angle between the normal vector of this point and the normal vector of the K-nearest neighbor points. The calculation formula is as follows: The fourth characteristic quantity is based on each point in the point set P t and the cosine value of the angle between the normal vector n i of this point and the line connecting this point and the centroid of its K-nearest neighbor points, and the calculation formula is as follows: f4(P ti ) = cos < n i , [O(P ti ) - P ti > The feature quantity P is obtained by solving the above four feature quantities ti and Q ti feature vectors. The feature vector sets L1 and L2 of P t and Q t are obtained respectively 6. A rough registration method for multi-viewpoint three-dimensional point cloud data based on branch and bound, characterized in that The method for obtaining the matching point set according to the principle of the same or similar feature quantities is as follows: For each point in the point set P t search for the matching point in the point set Q t . If P ti and Q ti are a pair of matching points, they should have the same or at least similar geometric features, so the feature vectors should also be the same or similar; in the step of initially obtaining the matching feature point set, the Chebyshev distance in the feature vector space should be used as the comparison criterion, and when the following conditions are met at the same time, it can be considered as a preliminary matching point: σ among the above four conditions i The parameters are all set to 0.

01. According to the above conditions, a preliminary set of feature matching points can be obtained and denoted as: W = {(m i1 , m i2 ) | m i1 ∈ P t , m i2 ∈ Q t , i = 1, 2, 3, …, N}.

7. A rough registration method for multi-viewpoint 3D point cloud data based on branch and bound, characterized in that, The operation method for accurately obtaining the matching points in the three-dimensional point cloud data in the third step is as follows: Set a new threshold σ6 = 0.01, and for each matching point pair (m i1 , m i2 ) ∈ W obtained in the second step, calculate the number N m of point pairs in W that meet the distance constraint condition. If a point pair in W satisfies: Then it is considered that this point is a point pair that meets the conditions. If most of the point pairs in W meet the conditions, then (m i1 , m i2 ) is a correct point pair; otherwise, it is incorrect and should be excluded. Set another threshold σ7 = 0.

8. If the N i1 , m i2 ) calculated for the point pair (m m satisfies: N m ≥σ7 * N It can be considered as a correct matching point pair, otherwise it is considered incorrect and the point pair is removed from the point set to obtain the final set of matching points: W = {(m i1 , m i2 ) | m i1 ∈ P t , m i2 ∈ Q t , i = 1, 2, 3,..., N‘} where N' is the number of matching point pairs after removing the incorrect matching point pairs.

8. A rough registration method for multi-viewpoint 3D point cloud data based on branch and bound, characterized in that In the fourth step, the accurately matched point set obtained in the third step is used as the input data to roughly calculate the camera pose parameter matrix, obtaining the rotation matrix and translation vector parameters.

9. A rough registration method for multi-viewpoint 3D point cloud data based on branch and bound, characterized in that The calculation of the initial parameters of the camera pose adopts the quaternion method, and the calculation methods of the camera rotation matrix and the translation vector are as follows: According to the result in the third step, the point m in the target three-dimensional point cloud data i1 after rotation and translation transformation is denoted as m'. i1 .

Images