Method and apparatus for constructing point cloud correspondence relationship

Abstract

This application provides a method for constructing point cloud correspondences. The method includes: extracting feature points from the original point cloud and the target point cloud by fusing local and global feature values ​​to form an original point set and a target point set; constructing sparse correspondences between the feature points in the original point set and the feature points in the target point set; filtering the sparse correspondences through rigid registration using the sparse correspondences to obtain feature point pairs with correspondences between the original point set and the target point set; performing local matching of nearest neighbor points on the point cloud for the feature point pairs with correspondences; obtaining the correspondences of all points between the original point cloud and the target point cloud; and forming a point cloud correspondence. This application solves the problem that traditional methods can only construct correspondences between point clouds with small deformations and have low computational efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer application, in particular to a method and device for constructing point cloud correspondence. BACKGROUND

[0002] In recent years, with the development of three-dimensional technology, the cost of obtaining point cloud data is reduced, and the processing of such data, such as point cloud registration, semantic segmentation, feature extraction, and point cloud completion, has begun to attract widespread attention. Constructing point cloud correspondence, i.e., the mapping relationship between the original point cloud and the target point cloud, is a very critical step in point cloud registration. The accuracy of the correspondence will affect the results and algorithm efficiency of the subsequent registration algorithm. The current construction of the correspondence can be divided into two types: constructing point-to-point correspondence and constructing region-to-region correspondence.

[0003] The point-to-point correspondence is the mapping relationship between the points of the original point cloud and the target point cloud. The construction of the point-to-point correspondence is only applicable to data with little deformation, and the original point cloud and the target point cloud can be strictly aligned. Moreover, the point cloud needs to be calculated for each point, and there is no existing correspondence for reference, so the calculation efficiency is low. On the other hand, the construction of the point-to-point correspondence is basically based on the local features of each point. For point clouds with deformation, the local features of the corresponding points change greatly, which will lead to the construction of incorrect correspondence between points, so the efficiency is low and the accuracy is not high. The construction of the region-to-region correspondence is to divide the original point cloud and the target point cloud according to a certain method to obtain different parts, and then construct the correspondence between the parts, but it cannot provide the correspondence of each point accurately.

[0004] Therefore, it is currently an urgent problem to efficiently and accurately construct the correspondence of point clouds with large deformation. SUMMARY

[0005] One purpose of the present application is to solve the problem of being unable to efficiently and accurately construct the correspondence of point clouds with large deformation.

[0006] According to an aspect of an embodiment of the present application, a method for constructing point cloud correspondence is disclosed, which comprises:

[0007] extracting feature points in the original point cloud and the target point cloud as an original point set and a target point set through local and global feature value fusion;

[0008] constructing a sparse correspondence between the original point set feature points and the target point set feature points;

[0009] performing rigid registration on the sparse correspondence to screen the sparse correspondence to obtain a pair of feature points between the original point set and the target point set having a correspondence.

[0010] The corresponding feature point pairs with the corresponding relationship are subjected to a near-neighbor point local matching on the point cloud, to obtain a corresponding relationship of all points between the original point cloud and the target point cloud, and the corresponding relationship of all points forms a point cloud corresponding relationship.

[0011] According to an aspect of the embodiment of the present application, a method for extracting feature points in the original point cloud and the target point cloud into an original point set and a target point set by fusing local and global feature values is disclosed, comprising:

[0012] The local feature values of each point in the original point cloud and the target point cloud are calculated respectively, to obtain the local feature values of each point in the original point cloud and the target point cloud;

[0013] The global feature values of each point in the original point cloud and the target point cloud are calculated according to the local feature values of each point in the original point cloud and the target point cloud;

[0014] The fusion feature values of each point in the original point cloud and the target point cloud are calculated according to the global feature values and the local feature values of each point in the original point cloud and the target point cloud;

[0015] The points with the maximum fusion feature values in the original point cloud and the target point cloud are selected as feature points respectively, and other points within the set range of the feature points are removed, and new feature points are repeatedly selected until the number of feature points in the original point cloud and the target point cloud reaches a set number, to obtain the original point set and the target point set.

[0016] According to an aspect of the embodiment of the present application, a method for calculating the local feature values of each point in the original point cloud and the target point cloud is disclosed, comprising:

[0017] The local feature values describing the position relationship of each point are obtained by calculating the position relationship of each point and all points in the set region around the point;

[0018] The size relationship between the local feature value of each point and a preset local feature threshold value is judged, if greater than or equal to the local feature threshold value, the local feature value is kept unchanged, if less than, the local feature value is reduced, to obtain the local feature value of each point in the original point cloud and the target point cloud.

[0019] According to an aspect of the embodiment of the present application, a method for calculating the global feature values of each point in the original point cloud and the target point cloud according to the local feature values of each point in the original point cloud and the target point cloud is disclosed, comprising:

[0020] The original point cloud and the target point cloud are divided into multiple regions, and the average value of the local feature values of all points in each region is taken as the initial global feature value of each region and the global feature value of each point in the region;

[0021] According to the set high-level threshold and low-level threshold, the region corresponding to the initial global feature value of the region determined is determined as a region type, and the region type includes a feature region and a non-feature region;

[0022] The initial global feature value of the center point of the feature region and the non-feature region is updated to a global feature value according to the region type.

[0023] According to an aspect of an embodiment of the application, a method for updating the initial global feature value of each point in the region to a global feature value according to the region type is disclosed, and the initial global feature value of the center point of the feature region and the non-feature region is the global feature value, including:

[0024] For the adjustment point close to the seed point away from the center of the feature region, the global feature value of the adjustment point is calculated according to the global feature value of the seed point, and the adjustment point refers to the target point cloud and the other points in the original point cloud except the seed point and the non-seed point;

[0025] For the adjustment point close to the non-seed point away from the center of the non-feature region, the global feature value of the adjustment point is calculated according to the global feature value of the non-seed point, and the initial global feature value of the seed point and the non-seed point is the global feature value.

[0026] According to an aspect of an embodiment of the application, a method for calculating the fusion feature value of each point in the target point cloud and the original point cloud according to the global feature value and the local feature value of each point in the target point cloud and the original point cloud is disclosed, including:

[0027] The points in the original point cloud and the target point cloud are sorted in descending order according to the global feature value and the local feature value respectively.

[0028] The points with the global feature value and the local feature value ranking within the front-end set range are retained, and the maximum value of the global feature value and the local feature value is taken as the fusion feature value of the point.

[0029] According to an aspect of an embodiment of the application, a method for calculating the fusion feature value of each point in the target point cloud and the original point cloud according to the global feature value and the local feature value of each point in the target point cloud and the original point cloud is disclosed, and further including:

[0030] For the points with the global feature value or the local feature value ranking within the end set range, the product of the global feature value and the local feature value is taken as the fusion feature value of the point, and the average value of the global feature value and the local feature value is taken as the fusion feature value of the remaining points.

[0031] According to an aspect of the embodiments of the present application, a method for rigid registration between an original point set and a target point set based on sparse correspondence is disclosed, which comprises the following steps:

[0032] Comparing the local features of the two feature points with the sparse correspondence to obtain an initial relationship weight value for describing the correlation between the feature points;

[0033] Ranking the point pairs with the sparse correspondence according to the initial relationship weight value, and selecting the point pairs with the top ranking within a certain range as registration point pairs;

[0034] In the process of rigid registration between the original point cloud and the target point cloud, the spatial distance between the two points is added as reference data for calculating the relationship weight value based on the initial relationship weight value, so as to obtain the relationship weight value of the point pairs corresponding to all sparse correspondence.

[0035] According to the relationship weight value of the point pairs with the sparse correspondence, the point pairs with the same number of feature points in the original point set or the target point set are selected as the feature point pairs with the corresponding relationship, and the relationship weight value of the feature point pairs is greater than that of other point pairs.

[0036] According to an aspect of the embodiments of the present application, a method for performing local matching of the nearest neighbor points of the feature point pairs with the corresponding relationship in the point clouds is disclosed, and the corresponding relationship between all points of the original point cloud and the target point cloud is obtained, and the corresponding relationship between all points forms a point cloud correspondence, which comprises the following steps:

[0037] Calculating the spatial distance and the self-defined distance between the nearest neighbor points of the feature points of the target point cloud and each nearest neighbor point of the feature points with the corresponding relationship in the original point cloud, wherein the nearest neighbor points refer to the non-feature points in the region centered on the feature points;

[0038] Selecting the two points with the smallest sum of the spatial distance and the self-defined distance as the feature point pairs with the corresponding relationship, and repeating the above steps until the nearest neighbor points of the feature points of the target point cloud or the original point cloud have the corresponding relationship.

[0039] According to the obtained feature point pairs with the corresponding relationship, the steps of establishing the feature point pairs are repeated until the corresponding relationship between all points of the original point cloud and the target point cloud is obtained.

[0040] According to an aspect of the embodiments of the present application, a point cloud correspondence construction device is disclosed, which comprises:

[0041] An extraction module is configured to extract the feature points in the original point cloud and the target point cloud as an original point set and a target point set through local and global feature value fusion;

[0042] a constructing module configured to construct sparse correspondence between the original set of feature points and the target set of feature points;

[0043] a screening module configured to screen the sparse correspondence to obtain a pair of feature points with correspondence between the original set of feature points and the target set of feature points through rigid registration of the sparse correspondence;

[0044] a matching module configured to perform local matching of the pair of feature points with correspondence on the point clouds where the pair of feature points is located to obtain correspondence between all points of the original point cloud and the target point cloud, and the correspondence between all points forms point cloud correspondence.

[0045] In the embodiments of the present application, for a given original point cloud and target point cloud, first, feature points in the original point cloud and the target point cloud are extracted as an original set of feature points and a target set of feature points through local and global feature fusion, sparse correspondence is constructed between the original set of feature points and the target set of feature points, then the sparse correspondence is screened to obtain a pair of feature points with correspondence between the original set of feature points and the target set of feature points through rigid registration of the sparse correspondence, and the original point cloud and the target point cloud are extracted at a representative pair of feature points, because these pairs of feature points usually exist in point cloud regions with high similarity, i.e., with small deformation, thereby reducing the difficulty of screening feature points, reducing the calculation time and improving the efficiency, finally, local matching of the pair of feature points with correspondence on the point clouds where the pair of feature points is located is performed to obtain correspondence between all points of the original point cloud and the target point cloud, and the correspondence between all points forms point cloud correspondence.

[0046] Based on the pair of feature points with correspondence, correspondence is constructed for the neighboring points of the pair of feature points, without directly finding the correspondence between points and points, further reducing the calculation time and improving the efficiency, and since the embodiments of the present application do not construct correspondence based on the similarity of each region, even if the deformation degree of the original point cloud and the target point cloud is large, the point-to-point correspondence can be accurately and quickly constructed, and the accuracy of the correspondence between point clouds is also guaranteed. In summary, the present application can solve the problem of inefficient and inaccurate construction of correspondence between point clouds with deformation.

[0047] Other features and advantages of the present application will become apparent from the following detailed description, or will be learned by practice of the present application.

[0048] It should be understood that the foregoing general description and the following detailed description are only exemplary and are not limiting of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0049] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description when taken in conjunction with the accompanying drawings in which:

[0050] Figure 1A flow chart of a method for constructing a point cloud correspondence relationship according to an embodiment of the present application is shown.

[0051] Figure 2 A flow chart of extracting feature points in the original point cloud and the target point cloud as the original point set and the target point set by local and global feature value fusion according to an embodiment of the present application is shown.

[0052] Figure 3 A flow chart of calculating local feature values of each point in the original point cloud and the target point cloud according to an embodiment of the present application is shown.

[0053] Figure 4 A flow chart of calculating global feature values of each point in the target point cloud and the original point cloud according to local feature values of each point in the target point cloud and the original point cloud according to an embodiment of the present application is shown.

[0054] Figure 5 A schematic diagram of segmenting the point cloud using the super voxel point cloud segmentation method according to an embodiment of the present application is shown.

[0055] Figure 6 A flow chart of updating the initial global feature values calculated for each point in the region to the global feature values according to the region type, and the initial global feature values of the center points of the feature region and the non-feature region being the initial global feature values according to an embodiment of the present application is shown.

[0056] Figure 7 A flow chart of calculating the fusion feature values of each point in the target point cloud and the original point cloud according to the global feature values and the local feature values of each point in the target point cloud and the original point cloud according to an embodiment of the present application is shown.

[0057] Figure 8 A flow chart of screening sparse correspondence relationships to obtain feature point pairs having correspondence relationships between the original point set and the target point set by rigid registration of sparse correspondence relationships according to an embodiment of the present application is shown.

[0058] Figure 9 A flow chart of performing local matching of the near neighbor points on the point cloud of the feature point pairs having correspondence relationships to obtain the correspondence relationships of all points between the original point cloud and the target point cloud, and the correspondence relationships of all points forming the point cloud correspondence relationship according to an embodiment of the present application is shown.

[0059] Figure 10 A schematic diagram of the tr_reg_000 and tr_reg_001 models of the data set according to an embodiment of the present application is shown.

[0060] Figure 11A diagram showing raw spatial relative position changes of an original point cloud (arm bending) and a target point cloud (arm stretching) is shown according to an embodiment of the present application.

[0061] Figure 12 A diagram showing feature value of a target point cloud is shown according to an embodiment of the present application.

[0062] Figure 13 A diagram showing feature value of an original point cloud is shown according to an embodiment of the present application.

[0063] Figure 14 A diagram showing selection result of feature points is shown according to an embodiment of the present application.

[0064] Figure 15 A diagram showing correspondence relationship calculation result is shown according to an embodiment of the present application.

[0065] Figure 16 A diagram showing comparison of effects between a traditional algorithm and the method of the present application is shown according to an embodiment of the present application.

[0066] Figure 17 A diagram showing final correspondence relationship calculation is shown according to an embodiment of the present application.

[0067] Figure 18 A diagram showing final correspondence relationship calculation result of calculation data is shown according to an embodiment of the present application.

[0068] Figure 19 A diagram showing a device for constructing a point cloud correspondence relationship is shown according to an embodiment of the present disclosure.

[0069] Figure 20 A diagram showing a hardware structure of a device for constructing a point cloud correspondence relationship is shown according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0070] Example implementations are now described with reference to the drawings. Example implementations can take form in various components and arrangements of components, and in various steps and arrangements of steps. The following description is in the context of a point cloud correspondence relationship construction device. However, the scope of the application is not limited to the examples described herein.

[0071] Moreover, the described features, structures, or characteristics can be combined in any suitable manner in one or more example embodiments. In the following description, numerous specific details are provided to give a thorough understanding of example embodiments. One skilled in the relevant art will recognize, however, that the technology can be practiced without one or more of the specific details, or with other methods, components, materials, and so forth. In other instances, well-known structures, methods, implementations, or operations are not shown or described in detail to avoid obscuring aspects of the application.

[0072] Some of the block diagrams in the drawings are of functional entities that do not necessarily have to correspond to physically or logically independent entities. These functional entities can be implemented in software, or in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0073] In an embodiment of the present application, the corresponding relationship between the point clouds is accurately and quickly constructed mainly through the following process.

[0074] First, the fusion feature values of each point in the original point cloud and the target point cloud are calculated, and the feature points are selected according to the size of the fusion feature values. The process is mainly divided into three stages: (1) calculating the local feature values of each point in the original point cloud and the target point cloud; (2) calculating the global feature values of each point in the original point cloud and the target point cloud based on the local feature values; (3) fusing the local and global feature values to obtain the fusion feature values of each point in the point cloud.

[0075] After obtaining the fusion feature values of each point in the original point cloud and the target point cloud, the points with the largest fusion feature values are selected in the original point cloud and the target point cloud respectively, and the points and other points within the set range of the points are excluded from participating in the selection of subsequent feature points. The above process is repeated to select the feature points until the number of feature points in the original point cloud and the target point cloud reaches the set number, and the original point set and the target point set with the same number of feature points are obtained.

[0076] Secondly, the correspondence is calculated, and the feature point pair is obtained. The process is divided into two stages, (1) the sparse correspondence relationship is constructed by matching each other between the feature points of the original point cloud and the target point cloud, and the sparse correspondence relationship set including the sparse correspondence relationship between any point in the original point cloud and any point in the target point cloud is obtained. (2) After obtaining the sparse correspondence relationship set, the similarity of the local features of each point pair with sparse correspondence is calculated as the initial relationship weight of the two points. The initial relationship weight is used for the first screening, and the correct sparse correspondence relationship is obtained. The registration point pair with the same number of feature points in the original point set or the target point set is obtained, and the initial relationship weight of the registration point pair is greater than the initial relationship weight of other point pairs. Since the first screening may still have errors, the second screening is used to improve the accuracy of the screening. In the process of rigid registration of the original point cloud and the target point cloud, the spatial distance of the two points is added as the reference data for calculating the relationship weight based on the initial relationship weight, the relationship weight of the point pair corresponding to all sparse correspondence relationships is obtained, the second screening is performed according to the relationship weight, the point pair with the same number of feature points in the original point set or the target point set is taken as the feature point pair with the correspondence relationship, and the relationship weight of the feature point pair is greater than the relationship weight of other point pairs.

[0077] Finally, the correspondence relationship between each point in the original point cloud and the target point cloud is constructed, the feature point pair with the correspondence relationship is matched with the local points in the point cloud, and the correspondence relationship between all points in the original point cloud and the target point cloud is obtained. The correspondence relationship between all points forms the point cloud correspondence relationship.

[0078] The specific implementation process is as follows. The spatial distance and the self-defined distance between the near neighbor point of the target point cloud feature point and each near neighbor point of the feature point with the correspondence relationship in the original point cloud are calculated. The near neighbor point refers to the non-feature point in the region centered on the feature point. The two points with the smallest sum of the spatial distance and the self-defined distance are screened as the feature point pair with the correspondence relationship, and the near neighbor point of the target point cloud feature point has the correspondence relationship. According to the obtained feature point pair with the correspondence relationship, the step of establishing the feature point pair is repeated until the correspondence relationship between all points in the original point cloud and the target point cloud is obtained.

[0079] Referring to Figure 1 , Figure 1 A flowchart of a point cloud correspondence relationship construction method according to an embodiment of the present application is shown. The embodiment of the present application provides a point cloud correspondence relationship construction method, which comprises the following steps:

[0080] In step S110, the feature points in the original point cloud and the target point cloud are extracted as the original point set and the target point set through local and global feature value fusion.

[0081] In step S120, the sparse correspondence relationship is constructed between the feature points of the original point set and the target point set.

[0082] Step S130, the sparse correspondence is filtered by sparse correspondence registration to obtain the feature point pairs having the correspondence between the origin set and the target point set;

[0083] Step S140, the feature point pairs having the correspondence are subjected to the local matching of the near points on the point cloud to obtain the correspondence of all points between the origin point cloud and the target point cloud, and the correspondence of all points forms the point cloud correspondence.

[0084] The four steps are described in detail below.

[0085] In step S110, to obtain the origin set and the target set respectively including the feature points in the origin point cloud and the target point cloud, firstly, the fusion feature values of the points in the origin point cloud and the target point cloud are obtained, and secondly, the points in the target point cloud and the origin point cloud are screened according to the fusion feature values of the points in the target point cloud and the origin point cloud to select some points meeting the set requirements as the feature points to form the target point set and the origin set. It should be noted that, in order to more accurately establish the correspondence between the origin point cloud and the target point cloud, the number of the feature points in the target point set is consistent with the number of the feature points in the origin set.

[0086] The process of obtaining the fusion features of the points in the target point cloud and the origin point cloud is a process of distinguishing the feature points and the non-feature points by calculating the feature values in a certain way, and the specific process mainly includes three stages, as follows.

[0087] In the process of obtaining the fusion features of the points in the target point cloud and the origin point cloud, three steps are mainly included, which are as follows.

[0088] The first step is to calculate the local feature values of the points in the origin point cloud and the target point cloud. The local feature value is a feature value representing the spatial position relationship of the points in a set range around each point with the point, and the greater the feature value, the more special the spatial position relationship of the point with the points in the set range around the point, thereby providing a calculation basis for the extraction of the feature points.

[0089] The second step is to calculate the global feature values of the points in the origin point cloud and the target point cloud based on the local feature values. The global feature value is a feature value representing the region to which each point in the origin point cloud and the target point cloud belongs, and the greater the global feature value, the more obvious the feature of the region and the more special the region. The target point cloud and the origin point cloud are divided into several regions, the feature values of the regions are calculated, and then the feature values of the regions are introduced into the points in the target point cloud and the origin point cloud.

[0090] In the embodiments of the present application, the initial global feature value of each region is the average of the local feature values of all points in the region. After the initial global feature values of each region are calculated, the points whose initial global feature values meet the set condition are selected according to the initial global feature values of the points; then, the global feature values of all points in the target point cloud and the original point cloud are optimized according to the points meeting the set condition, the initial global feature values of the points with initial global feature values higher than the set condition are increased to be the global feature values, and the initial global feature values of the points with initial global feature values lower than the set condition are reduced to be the global feature values. In this way, the global feature values of the points with high initial global feature values are more prominent, and the difference with the points with low initial global feature values is more obvious, which is more conducive to the identification of feature points.

[0091] The third step is to fuse the local and global feature values to obtain the fusion feature value of each point of the point cloud. The fusion feature value is used to measure the final special degree of each point in the point cloud where the point is located. The greater the fusion feature value, the higher the obvious degree of the final feature of the point, and the more special the point is. After obtaining the fusion feature values of the points in the original point cloud and the target point cloud, the points with the maximum fusion feature value are selected in the original point cloud and the target point cloud respectively, and the selected points and other points within a set range of the selected points are removed to continue the selection of feature points, that is, the feature points are repeatedly selected in the remaining points according to the above method until the number of feature points in the original point cloud and the target point cloud reaches a set number, and the original point set and the target point set are obtained.

[0092] Please refer to Figure 2 , Figure 2 A flowchart for extracting feature points in the original point cloud and the target point cloud as the original point set and the target point set by fusing the local and global feature values is shown according to an embodiment of the present application. The embodiment of the present application provides a step S110 for extracting feature points in the original point cloud and the target point cloud as the original point set and the target point set by fusing the local and global feature values, which includes:

[0093] Step S111, local feature values of the points in the original point cloud and the target point cloud are calculated respectively to obtain the local feature values of the points in the original point cloud and the target point cloud;

[0094] Step S112, global feature values of the points in the original point cloud and the target point cloud are calculated according to the local feature values of the points in the original point cloud and the target point cloud;

[0095] Step S113, fusion feature values of the points in the original point cloud and the target point cloud are calculated according to the global feature values and the local feature values of the points in the original point cloud and the target point cloud;

[0096] Step S114, the points with the maximum fusion feature value in the original point cloud and the target point cloud are selected as feature points respectively, and the selected points and other points within a set range of the selected points are removed, and the selection is repeated until the number of feature points in the original point cloud and the target point cloud reaches a set number, and the original point set and the target point set are obtained.

[0097] The four steps are described in detail as follows.

[0098] In step S111, the local feature value of each point is calculated according to the spatial position relationship between each point in the original point cloud and the target point cloud and other points in the set range around the point, that is, the local feature value of each point quantifies the spatial position relationship characteristics of each point and other points in the set range around the point in a numerical manner, to represent the obviousness of the characteristics of the point. For example, the relative position relationship of the coordinates, the density of the surrounding points, and the like. The greater the feature value, the more obvious the spatial position relationship of the point and the points in the set range around the point, which provides a calculation basis for the extraction of the feature points described below.

[0099] Referring to Figure 3 , Figure 3 A flowchart for calculating the local feature value of each point in the original point cloud and the target point cloud according to an embodiment of the present application is shown. The embodiment of the present application provides a step S111 of calculating the local feature value of each point in the original point cloud and the target point cloud, including:

[0100] In step S1111, the initial local feature value of the position relationship of each point is calculated according to the position relationship of all points in the set region around the point, and the initial local feature value of the position relationship of each point is obtained.

[0101] In step S1112, the size relationship between the initial local feature value of each point and the preset local feature threshold value is judged. If it is greater than or equal to the local feature threshold value, it is kept unchanged, and if it is less than the local feature threshold value, the local feature value of the point is reduced, and the local feature value of each point in the target point cloud and the original point cloud is obtained.

[0102] The two steps are described in detail as follows.

[0103] In step S1111, the spatial position relationship of all points in the set region around each point in the target point cloud and the original point cloud is calculated, and the spatial position relationship is calculated by the spatial coordinate information of each point in the target point cloud and the original point. For example, the spatial position of other points within a set radius centered on each point is calculated with the point. When calculating the spatial position relationship between two points, the distance between the two points, the existing vector, and the density of the points in the set region around the point are calculated according to the number of points.

[0104] In an embodiment of the present application, the step of calculating the local feature value of each point in the original point cloud and the target point cloud is as follows:

[0105] The calculation of the local feature value uses FPFH (Fast Point Feature Histogram), which provides information that is invariant to translation and rotation for point clouds, is robust to sampling density and noise points, and can better describe the local information of the point cloud. In actual calculation, the FPFH feature is a vector of length N. Assuming that point p i ∈P, P is a point cloud object for which the local feature is calculated, FPFH(p i ) refers to the FPFH feature descriptor (Fast Point Feature Descriptor) of point p i , which refers to a way of describing the feature of the point. The local feature value D(p i ) feature descriptor of point p i is calculated according to the following formula:

[0106]

[0107] where R refers to the neighbor point used for calculation of p i , ||p i -p j || refers to the distance between two points. χ 2 (p i ,p j ) is a user-defined distance between two points, which represents the spatial position relationship between the two points, such as the coordinate position relationship. The definition is as follows:

[0108]

[0109] In step S1112, since the local feature values of the points of the original point cloud and the target point cloud are not significantly different, the subsequent calculation of the global feature values and the fusion feature values of the points may not be significantly different. Therefore, by executing step S1112, the local feature values of the points with weak local feature values are reduced, so that the comparison between the points with weak local feature values and the points with high local feature values is more obvious, and the feature values of the points of the target point cloud and the original point cloud are more obviously different in size in the subsequent calculation of the global feature values and the fusion feature values, which is more convenient for selecting feature points and more obvious when viewing the point cloud feature distribution image.

[0110] In an embodiment of the present application, a local feature value threshold is set, and the size relationship between the local feature value of each point and the preset local feature threshold is determined. If the local feature value is greater than or equal to the local feature threshold, it remains unchanged, and if the local feature value is less than the local feature threshold, the local feature value is reduced, thereby obtaining the local feature values of the points in the target point cloud and the original point cloud.

[0111] According to an embodiment of the present application, the following algorithm is used to achieve the effect of suppressing the points with small local feature values. A piecewise suppression function is used to prevent too many points from being marked. It should be noted that the local feature value of a point p is in the range of 0 to 1, and the local feature value of the final point p i is: i

[0112]

[0113] h is the local feature threshold, which is calculated according to the maximum local feature value in each point, such as 80% or 81% of the maximum local feature value, or any value in the range of 0% to 100% of the maximum local feature value, according to the calculation requirement.

[0114] Further, according to another embodiment of the present application, the points in the original point cloud and the target point cloud are filtered to obtain points with small local feature values, and the method for reducing the local feature values of the obtained points further includes:

[0115] The points in the original point cloud and the target point cloud are sorted according to the local feature values, and the points within a certain range at the end of the ranking are taken to perform the action of reducing the local feature values. The action of reducing the local feature values includes multiplying the local feature value by itself (the local feature value is greater than 0 and less than 1).

[0116] In step S112, the original point cloud and the target point cloud are divided into several regions, and the initial global feature values are calculated for each region. The initial global feature value of a region is the average of the local feature values of all points in the region. Then, the initial global feature value of the region is taken as the initial global feature value of each point in the region, and the seed points and non-seed points are determined according to the initial global feature values of the regions.

[0117] According to the seed points and non-seed points, the global feature values of other points except the seed points and non-seed points are calculated. The global feature values of points with high final increased global feature values are increased, and the global feature values of points with low final increased global feature values are decreased. In this way, the global feature values of points with high global feature values are more prominent, which is more conducive to the identification of feature points.

[0118] Referring to Figure 4 , Figure 4 FIG. 7 shows a flowchart of calculating the global feature values of the points in the original point cloud and the target point cloud according to the local feature values of the points in the original point cloud and the target point cloud, according to an embodiment of the present application. The embodiment of the present application provides a step S112 of calculating the global feature values of the points in the original point cloud and the target point cloud according to the local feature values of the points in the original point cloud and the target point cloud, which includes:

[0119] ​Step S1121, the target point cloud and the original point cloud are segmented into multiple regions, and the average value of the local feature values of all points in each region is used as the initial global feature value of each region, and also as the global feature value of each point in the region;

[0120] Step S1122, the region corresponding to the initial global feature value is determined according to the set high-level threshold and low-level threshold, and the region type includes a feature region and a non-feature region.

[0121] Step S1123, the initial global feature value of each point in the region is updated to the global feature value according to the region type, and the initial global feature value of the center point of the feature region and the non-feature region is the initial global feature value.

[0122] The three steps will be described in detail below.

[0123] In order to obtain the global feature value of the target point cloud and the original point cloud, the target point cloud and the original point cloud need to be segmented into several regions first, and the initial global feature value of each region is calculated, then the seed points and non-seed points in the target point cloud and the original point cloud are found according to the initial global feature value, and finally the global feature value of each point other than the seed points and the non-seed points is calculated according to the positional relationship of each point.

[0124] In step S1121, the original point cloud and the target point cloud are first segmented into multiple regions, and the average value of the local feature values of all points in each region is used as the initial global feature value of each region, and also as the initial global feature value of each point in the region.

[0125] In an embodiment of the present application, the original point cloud and the target point cloud are segmented using the super voxel segmentation method (VCCS), please refer to Figure 5 , Figure 5 A schematic diagram of segmenting a point cloud using a super voxel point cloud segmentation method according to an embodiment of the present application is shown. The super voxel point cloud segmentation method (VCCS) is a region growing algorithm, and the implementation process includes:

[0126] I. Construct a voxel grid of the target point cloud, where a voxel refers to a structural unit formed by a certain number of point clouds, and the voxel grid refers to a representation of the target point cloud in units of voxels.

[0127] II. Construct initial voxel seed points on the voxel grid according to the distance between the preset initial voxel seed points, and the distance between the uniformly constructed initial voxel seed points is much larger than the size of the voxel grid.

[0128] III. Randomly select an initial voxel seed point, and set a radius, query the number of voxels in the region centered on the initial voxel seed point, if the number is greater than the set threshold, keep the initial voxel seed point as the voxel seed point, if the number is less than the set threshold, delete the initial voxel seed point.

[0129] IV. Take the voxel seed point as the starting point, search for the neighboring voxels. According to the voxel feature information, calculate the correlation with the neighboring voxels, divide the voxels with greater correlation with the voxel seed point than other voxel seed points into the region centered on the voxel seed point, until each voxel in the voxel grid has a belonging region, successfully segmenting the target point cloud into several regions.

[0130] The calculation method of the initial global feature value is as follows:

[0131] Get the region set C of the target point cloud. Let a region c i ∈C, FPFH(c i ) represents the FPFH descriptor (fast point feature descriptor) of the region c i , which is defined as the FPFH average value of the points in the region. The initial global feature value G(c i ) of the region c i is:

[0132]

[0133] Where ||c i -c j || refers to the distance between the center points of the two regions,

[0134] χ 2 (p i ,p j ) is a user-defined distance between two points, which is defined as follows:

[0135]

[0136] Here N refers to the entire region set.

[0137] After obtaining the initial global feature value of each region, introduce the initial global feature value of each region into each point in the region, that is, the initial global feature value of each region is also the initial local feature value of each point in the region.

[0138] Similarly, segment the original point cloud according to the above segmentation method.

[0139] In step S1122, in order to obtain the seed points and non-seed points in the original point cloud, firstly, the seed points and non-seed points are obtained by screening according to the initial global feature values of each region in the original point cloud and the target point cloud, and then the seed points and non-seed points are obtained by screening according to the high-level threshold and the low-level threshold.

[0140] In an embodiment of the present application, the high-level threshold and the low-level threshold are set, and then the regions are screened according to the initial global feature values of the regions, and the size relationship of the initial global feature values of the regions relative to the high-level threshold and the low-level threshold is determined, the initial global feature value of the feature region is greater than the high-level threshold, and the initial global feature value of the non-feature region is less than the low-level threshold.

[0141] According to an embodiment of the present application, the threshold is set as follows:

[0142] The thresholds h1 and h2 are set to select the seed points and non-seed points. The region with the initial global feature value greater than h1 is set as a feature region, and the center point thereof is set as a seed point. The region with the initial global feature value less than h2 is set as a non-feature region, and the center point thereof is set as a non-seed point. Wherein, h1 and h2 are defined as:

[0143]

[0144] h2 = mean(G)

[0145] Δ = max(G) - mean(G)

[0146] Referring to Figure 6 , Figure 6 A flow chart for updating the initial global feature values of the points in the regions to global feature values according to the region types, and the initial global feature values of the center points of the feature regions and the non-feature regions are the initial global feature values, is shown according to an embodiment of the present application. The embodiment of the present application provides a step S1123 for updating the initial global feature values of the points in the regions to global feature values according to the region types, and the initial global feature values of the center points of the feature regions and the non-feature regions are the initial global feature values, which comprises:

[0147] Step S201, for the adjustment points close to the seed points of the center of the feature region, the global feature values of the adjustment points are calculated according to the global feature values of the seed points, and the adjustment points refer to the other points in the target point cloud and the original point cloud except the seed points and the non-seed points;

[0148] Step S202, for the adjustment points close to the non-seed points of the center of the non-feature region, the global feature values of the adjustment points are calculated according to the global feature values of the non-seed points, and the initial global feature values of the seed points and the non-seed points are the global feature values.

[0149] The two steps are described in detail below.

[0150] If the global feature value of an adjustment point is calculated, the distance between the adjustment point and the surrounding seed points and non-seed points is also calculated, and it is determined whether the adjustment point is closer to the seed points or the non-seed points before the following steps are performed. The adjustment point refers to a point other than the seed points and the non-seed points in the target point cloud and the original point cloud.

[0151] In step S201, the center point of the feature region is taken as a seed point, the initial global feature value of the seed point is updated as the global feature value, and the global feature value of an adjustment point closer to the seed point is calculated according to the global feature value of the seed point.

[0152] In an embodiment of the present application, for the adjustment point p i , the global feature value G(p i ) of the point p i is:

[0153] G(p i ) = exp(-G(c i )*||p i -c i ||*a)

[0154] ||p i -c i || represents the distance between the point p i and the nearest seed point.

[0155] In step S201, the center point of the non-feature region is taken as a non-seed point, the initial global feature value of the non-seed point is updated as the global feature value, and the global feature value of an adjustment point closer to the non-seed point is calculated according to the global feature value of the non-seed point.

[0156] In an embodiment of the present application, for the adjustment p i , the global feature value G(p i ) of the point p i is:

[0157] G(p i ) = exp(-G(c i )*||p i -c i ||*b)

[0158] ||p i -c i || represents the distance between the point p i and the nearest non-seed point.

[0159] It should be noted that in the formulae in step S201 and step S202, a > b.

[0160] In step S113, the local feature value and the global feature value of each point are substituted into a pre-set calculation method to calculate a fusion feature value of each point.

[0161] Please refer to Figure 7 , Figure 7 A flow chart is shown for calculating a fusion feature value of each point in the target point cloud and the original point cloud according to the global feature value and the local feature value of each point in the target point cloud and the original point cloud according to one embodiment of the present application. The embodiment of the present application provides step S113 for calculating a fusion feature value of each point in the target point cloud and the original point cloud according to the global feature value and the local feature value of each point in the target point cloud and the original point cloud, which includes:

[0162] In step S1131, each point in the original point cloud and the target point cloud is sorted according to the global feature value and the local feature value respectively.

[0163] In step S1132, the points with the global feature value and the local feature value ranked within the front end setting range are screened out according to the sorting of each point in the original point cloud and the target point cloud in step S1131, and then the local feature value is compared with the global feature value, and the larger one is taken as the fusion feature value of the point. It is further explained that the front end setting range of the global feature value ranking and the front end setting range of the local feature value ranking can be consistent or inconsistent, for example, the points with the global feature value and the local feature value ranked in the front 20% or the points with the global feature value ranked in the front 20% and the local feature value ranked in the front 10%.

[0164] The two steps are described in detail below.

[0165] In step S1131, each point in the original point cloud and the target point cloud is sorted according to the global feature value and the local feature value respectively.

[0166] In step S1132, the points with the global feature value and the local feature value ranked within the front end setting range are screened out according to the sorting of each point in the original point cloud and the target point cloud in step S1131, and then the local feature value is compared with the global feature value, and the larger one is taken as the fusion feature value of the point. It is further explained that the front end setting range of the global feature value ranking and the front end setting range of the local feature value ranking can be consistent or inconsistent, for example, the points with the global feature value and the local feature value ranked in the front 20% or the points with the global feature value ranked in the front 20% and the local feature value ranked in the front 10%.

[0167] It is further explained that in the embodiment of the present application, for the points with the global feature value or the local feature value ranked within the end setting range, the product of the global feature value and the local feature value is taken as the fusion feature value of the point. Similarly, the end setting range of the global feature value ranking and the end setting range of the local feature value ranking can be consistent or inconsistent, for example, the points with the global feature value and the local feature value ranked in the back 30% or the points with the global feature value ranked in the back 15% and the local feature value ranked in the back 20%.

[0168] It is also needed to be explained that in an embodiment of the present application, for the points whose global evaluation value or local feature value ranking is neither at the end of the set range nor at the front of the set range, the average value of the global evaluation value and the local feature value is taken as the fusion feature value.

[0169] In step S114, taking the selection of the feature points of the target point cloud as an example, according to the fusion feature values of the points in the target point cloud, the points are sorted, the point with the largest fusion feature value is selected as a feature point, and the feature point and other points within a set range around the feature point are removed and do not participate in the selection of the feature point in the next time; then the point with the largest fusion feature value in all the remaining points is selected as a feature point again, and the feature point and other points within a set range around the feature point are removed and do not participate in the selection of the feature point in the next time, the step of obtaining the feature point is repeated until the number of the feature points reaches a set number, and the target point set is obtained. Similarly, for the selection of the feature points in the original point cloud, the same method is used.

[0170] In an embodiment of the present application, the original point set and the target point set are obtained according to the fusion feature values of the points in the original point cloud and the target point cloud as follows:

[0171] The fusion feature values of the points are sorted, the points whose feature value rankings are within a set range at the front are selected as a candidate feature point set, and the feature points are selected from the candidate feature point set for the construction of the sparse correspondence relationship.

[0172] Taking the selection of the feature points of the target point cloud as an example, first, the target point set k T is set, and the access queue flag of the candidate feature point set of the target point cloud is set, the length of the flag queue is the length of the candidate feature point set of the target point cloud, and the initial value of each candidate feature point is 0, which is used to indicate that the point has not been accessed, and 1 indicates that the point has been accessed. The number T of the target feature points to be obtained is set artificially, and the target point set k T is obtained. The target feature point selection algorithm steps are as follows:

[0173] (1) The point with the highest feature value in the candidate feature point set of the target point cloud and the access queue parameter not being 1 is inquired, the point is added to the target point set k T , and the access queue parameter of the point is set to 1;

[0174] (2) The access queue parameters of the points within a set range around the target feature point and belonging to the candidate feature point set of the target point cloud are set to 1;

[0175] (3) If the number of elements of the feature point set k T is equal to T, or the access queue parameters are all 1, stop, otherwise loop to step 1.

[0176] Similarly, for the selection of the feature points in the original point cloud, the same method is used to obtain the original point set.

[0177] In step S120, after the origin point set and the target point set are obtained in step S110, sparse correspondence relations between the feature points belonging to different point sets are constructed, so that any feature point in the target point set has a sparse correspondence relation with any feature point in the origin point set. For example, there are 3 feature points in the origin point set and 3 feature points in the target point set, and the feature points in the origin point set are 1, 2 and 3 respectively, and the feature points in the target point set are a, b and c respectively. Then 9 sparse correspondence relations should be constructed, which are 1a, 1b, 1c, 2a, 2b, 2c, 3a, 3b and 3c respectively. It should be clear that each feature point should have only one corresponding feature point, that is, among the 9 sparse correspondence relations, there should be 3 correct sparse correspondence relations as the correspondence relations, and the other 6 sparse correspondence relations are all incorrect. The 9 sparse correspondence relations are screened by step S130 to obtain 3 correct sparse correspondence relations as the correspondence relations.

[0178] In step S130, according to the sparse correspondence relations between the feature points in the origin point set and the feature points in the target point set obtained in step S120, there are correct sparse correspondence relations and incorrect sparse correspondence relations. The initial relation weight between two feature points is calculated according to the sparse correspondence relation between the two feature points. The higher the initial relation weight is, the greater the possibility that the sparse correspondence relation between the two feature points is correct. According to the size of the relation weight between the feature points, the sparse correspondence relations are screened for the first time. The sparse correspondence relations are sorted according to the size of the relation weight, and the point pairs in the front end of the set range are obtained as the registration point pairs. According to the sparse correspondence relations of the registration point pairs, the origin point cloud and the target point cloud are rigidly registered. In the registration process, the relation weight of all point pairs with sparse correspondence relations is calculated. According to the relation weight between the obtained point pairs, the sparse correspondence relations are screened for the second time, and the same number of feature points in the origin point set or the target point set are obtained again. These point pairs have correct sparse correspondence relations, that is, they are feature point pairs with correspondence relations.

[0179] Please refer to Figure 8 , Figure 8 A flow chart of rigid registration by sparse correspondence relation to screen sparse correspondence relations to obtain feature point pairs with correspondence relations between an origin point set and a target point set according to an embodiment of the present application is shown. The embodiment of the present application provides step S130 of rigid registration by sparse correspondence relation to screen sparse correspondence relations to obtain feature point pairs with correspondence relations between an origin point set and a target point set, which includes:

[0180] In step S131, the initial relation weight describing the correlation between the feature points by a numerical value is obtained by comparing the local features of the two feature points with the sparse correspondence relation.

[0181] Step S132, ranking the point pairs with sparse correspondence according to the initial relationship weight, and screening the point pairs with the top ranking within a certain range as the registration point pairs;

[0182] Step S133, in the process of rigid registration of the original point cloud and the target point cloud, adding the spatial distance between the two points based on the initial relationship weight as the reference data for calculating the relationship weight, to obtain the relationship weight of the point pairs corresponding to all sparse correspondence;

[0183] Step S134, according to the relationship weight of the point pairs with sparse correspondence, taking the point pairs with the same number of feature points in the original point set or the target point set as the feature point pairs with correspondence, and the relationship weight of the feature point pairs is greater than the relationship weight of other point pairs.

[0184] The above four steps will be described in detail below.

[0185] In step S131, in order to more intuitively judge the correctness and error of sparse correspondence, the local features of the two feature points are compared, such as the spatial position relationship of the two feature points and their neighbors, the feature vectors of the two feature points, etc. The initial relationship weight representing the correct probability of the sparse correspondence of the two feature points is obtained.

[0186] In an embodiment of the present application, the local feature of the feature point is also the spatial position relationship of the feature point and its surrounding neighbor points. The spatial position relationship of the two feature points and their surrounding neighbor points with the same sparse correspondence is compared, and the initial relationship weight representing the similarity degree of the spatial position relationship of the two feature points and their respective neighbor points is obtained according to the spatial position relationship of the two feature points and their respective neighbor points. That is, the greater the initial relationship weight, the higher the similarity degree of the spatial position relationship of the two feature points and their respective neighbor points, and the greater the correct probability of the sparse correspondence of the two feature points.

[0187] Further, in an embodiment of the present application, the way to obtain the initial relationship weight of the two feature points with sparse correspondence according to their local features is as follows:

[0188] For any point p i , its correlation coefficient factor(p i ) is calculated as follows: 1, using the point p i and its neighbor points to construct a covariance matrix C i ; 2, solving the eigenvalues λ1, λ2, λ3 of C i , where λ1> λ2> λ3; factor(p i ) represents the geometric features of the region formed by the point p i and its neighbor points. Finally, factor(p iThe definition is as follows:

[0189]

[0190] Suppose a pair of correspondences (s) in a sparse correspondence relation. j ,t i ), s j and t i These are a point on the original point cloud and a point on the target point cloud, respectively, and the weight w of this pair's correspondence. ij for:

[0191] w ij =exp(-factor_dis)+exp(-FPFH_dis)

[0192] Where factor_dis is point s j and t i The absolute value is the difference between the correlation coefficients. FPFH_dis is also χ². 2 (p i ,p j () is a custom distance between two points.

[0193]

[0194] The initial relation weights consider not only the point p i The geometric characteristics of the region formed by the point p and its nearest neighbors were also considered. i The spatial relationship between the feature points and their nearest neighbors makes the initial relationship weight more representative of the correlation between the two feature points, resulting in a higher accuracy in screening point pairs with sparse correspondences based on the initial relationship weight.

[0195] In step S132, based on the initial relation weights of the sparse correspondence point pairs obtained in step S131, the sparse correspondences are sorted. Since the number of correct sparse correspondences is fixed (i.e., the number of feature points in the origin or target set), point pairs with the same number of feature points as the origin or target set should ideally be selected as registration point pairs. However, because it is impossible to accurately filter out correct sparse correspondences at this point, the acquisition range is expanded to obtain point pairs within the predefined range of the ranking front as registration point pairs. The initial relation weight of the registration point pair is greater than the initial relation weight of other point pairs. The first screening of sparse correspondences is completed using the initial relation weights.

[0196] In step S133, according to the registration point pair obtained in step S132, the target point cloud and the original point cloud are rigidly registered, which means that the shape and size of the target point cloud and the original point cloud do not change, and only the spatial position and attitude change. This process can be imagined as a translation and rotation of the original point cloud and the target point cloud to maximize the coincidence of the two.

[0197] In the process of coincidence of the original point cloud and the target point cloud, several iterations are required to complete the rigid registration between the original point cloud and the target point cloud. The distance between the two feature points with correct sparse correspondence is definitely getting smaller and smaller in each iteration. After each iteration is completed, the intermediate relationship weight of all point pairs is recalculated according to the obtained spatial distance of the two feature points with sparse correspondence and the initial relationship weight, sorted and screened, and the top-ranked point pairs within the set range are reselected as the registration point pairs. Then, the next iteration of rigid registration is continued according to the sparse correspondence of the registration point pairs. Until the original point cloud and the target point cloud are rigidly registered, the last obtained intermediate relationship weight is updated as the relationship weight. The spatial distance of the two points is added to the initial relationship weight as the reference data for calculating the intermediate relationship weight. Finally, the relationship weight of the point pairs corresponding to all sparse correspondence is obtained. At this time, the greater the relationship weight of the two feature points, the higher the similarity of the spatial position relationship between the two feature points and their respective neighboring points, the smaller the spatial distance of the two feature points in the process of rigid registration of the original point cloud and the target point cloud, and the greater the possibility of correct sparse correspondence between the two feature points.

[0198] By continuously updating the registration point pairs for rigid registration, the calculation amount of registration is greatly reduced, the efficiency is higher, and the selection of feature point pairs is more accurate through continuous screening.

[0199] In an embodiment of the present application, the spatial distance of the two points is added to the initial relationship weight as the reference data for calculating the relationship weight. The method for calculating the relationship weight of the point pairs corresponding to all sparse correspondence is as follows:

[0200] After one rigid registration, the weight of all sparse correspondences is recalculated. The calculation formula is as follows:

[0201]

[0202] is a value set by man, which is set to 0.5 here. Then, the weight of each correspondence is reordered from large to small, and the top 5 correspondences with the largest weight are still selected as the new data for the next rigid registration. The above process is repeated until the result of the rigid registration no longer changes or the set number of iterations is reached, and the relationship weight of all point pairs is obtained.

[0203] In step S134, according to the relationship weight of the point pair with sparse correspondence obtained in step S133, a second screening is performed on all point pairs with sparse correspondence,

[0204] Among all point pairs with sparse correspondence, point pairs with the same number of feature points in the original point set or the target point set are selected as feature point pairs, and the feature point pairs with correspondence are obtained, and it is further illustrated that the relationship weight of the feature point pair is greater than the relationship weight of other point pairs. The two screenings of sparse correspondence ensure the correctness of the sparse correspondence to the greatest extent, and greatly improve the accuracy of the correspondence between the original point cloud and the target point cloud.

[0205] In step S140, after the feature point pairs obtained in step S130, the correspondence between each point in the original point cloud and each point in the target point cloud is constructed, and the specific process is as follows: taking the feature point belonging to the original point cloud as the center, the non-feature points within the set range around the feature point of the original point cloud are set as the near neighbor points of the feature point, and similarly, taking the feature point of the target point cloud corresponding to the feature point of the original point cloud as the center, the non-feature points within the set range around the feature point of the target point cloud are set as the near neighbor points of the feature point. The sparse correspondence between the near neighbor points of the feature point of the original point cloud and the near neighbor points of the feature point of the target point cloud is constructed one by one, and then the method of steps S120 and S130 is used to obtain new feature point pairs. Repeat the step of matching the near neighbor points between the feature point pairs until the correspondence between all points in the original point cloud and the target point cloud is obtained, and the correspondence between all points forms the point cloud correspondence.

[0206] Please refer to Figure 9 , Figure 9 A flow chart for performing local matching of near neighbor points on the point cloud of the feature point pairs with correspondence to obtain the correspondence between all points in the original point cloud and the target point cloud, and the correspondence between all points forms the point cloud correspondence according to an embodiment of the present application is shown. The embodiment of the present application provides step S140 of performing local matching of near neighbor points on the point cloud of the feature point pairs with correspondence to obtain the correspondence between all points in the original point cloud and the target point cloud, and the correspondence between all points forms the point cloud correspondence, which comprises:

[0207] Step S141, the spatial distance and the self-defined distance of each near neighbor point of the target point cloud feature point and the near neighbor point of the feature point with correspondence in the original point cloud are calculated, and the near neighbor point refers to the non-feature point within the set region with the feature point as the center;

[0208] Step S142, screen the two points with the smallest sum of the spatial distance and the self-defined distance, and the two points are the feature point pairs with correspondence, until the near neighbor points of the target point cloud feature point or the original point cloud feature point all have correspondence;

[0209] Step S143, according to the obtained corresponding feature point pair, repeat the step of establishing the feature point pair until the corresponding relationship of all points between the original point cloud and the target point cloud is obtained.

[0210] The following three steps are described in detail.

[0211] In step S141, the target point cloud feature points and the original point cloud feature points in a feature point pair are determined respectively. The near neighbor points are non-feature points within a certain range centered on a feature point. Then the spatial distance between a near neighbor point of the target point cloud feature point and each near neighbor point of the original point cloud feature point is randomly selected and compared with the self-defined distance.

[0212] Further, at this time, the original point cloud and the target point cloud have completed the rigid registration, so the smaller the spatial distance, the greater the possibility that the near neighbor points of the target point cloud feature point and the near neighbor points of the original point cloud feature point have a corresponding relationship. When the original point cloud and the target point cloud are consistent in size and shape, the near neighbor point pairs with corresponding relationship can be selected only by calculating the spatial distance. However, when the original point cloud and the target point cloud have deformation, the near neighbor point pairs with corresponding relationship cannot be selected only by the spatial distance, so the self-defined distance is used for screening. The self-defined distance is a manifestation of the difference between the feature descriptors of the two near neighbor points. The specific calculation method is as follows:

[0213]

[0214] As can be seen from the above, the smaller the spatial distance and the self-defined distance, the greater the possibility that the two near neighbor points have a corresponding relationship. By this way of screening, the corresponding relationship of the two point clouds without deformation can be constructed, and the corresponding relationship of the two point clouds with deformation can also be constructed.

[0215] It should be noted that the above "randomly selecting the spatial distance between a near neighbor point of the target point cloud feature point and each near neighbor point of the original point cloud feature point and the self-defined distance" can also be "randomly selecting the spatial distance between a near neighbor point of the original point cloud feature point and each near neighbor point of the target point cloud feature point and the self-defined distance".

[0216] In step S142, according to the spatial distance and the self-defined distance of each near neighbor point pair obtained in step S141, the two points with the smallest sum of the spatial distance and the self-defined distance are the feature point pair with corresponding relationship. Repeat steps S141 and S142 until each near neighbor point of the target point cloud feature point has a corresponding relationship with a near neighbor point of the original point cloud feature point, forming a new feature point pair. Or until each near neighbor point of the original point cloud feature point has a corresponding relationship with a near neighbor point of the target point cloud feature point, forming a new feature point pair.

[0217] In step S143, according to the obtained new feature point pairs, the above-mentioned manner of obtaining feature point pairs is repeated to obtain the correspondence of all points between the original point cloud and the target point cloud.

[0218] The following describes a point cloud correspondence construction method of the application in terms of the construction of the correspondence between the target point cloud and the original point cloud.

[0219] The input point cloud data, i.e., the original point cloud and the target point cloud, is subjected to global and local feature fusion to extract a feature point set, and an initial sparse correspondence is constructed between the feature point set. Then, the sparse correspondence is preliminarily screened according to the local feature (FPFH is used here) and the correlation coefficient, followed by rigid registration, and then screening based on the self-defined distance and the spatial distance to obtain the correspondence and the feature point pairs. The correspondence is calculated through the near-neighbor point matching manner. The specific manner is as follows:

[0220] The first step is to screen the points of the original point cloud and the target point cloud according to the fusion feature values of the original point cloud and the target point cloud to obtain feature points.

[0221] The feature value calculation steps for each point are as follows: 1. local feature value calculation, 2. global feature value calculation, and 3. local and global feature value fusion to obtain the final feature value of each point of the point cloud. Then, the feature points are selected based on the obtained feature values.

[0222] The local feature value calculation uses FPFH (Fast Point Feature Histogram), which is a feature represented by the spatial relationship of the point and its surrounding neighbor points within a specified range. The local feature values of the points in the target point cloud and the original point cloud are obtained and optimized.

[0223] There are two ways to optimize the local features of the points in the original point cloud and the target point cloud.

[0224] First, let h be the local feature threshold. For the local feature value greater than or equal to the local feature threshold, the local feature value remains unchanged; for the local feature value less than the local feature threshold, the local feature value is reduced to a new local feature value.

[0225] Second, after sorting the local feature values from large to small, for the points ranked after 20%, the local feature value is reduced as a new local feature value, and the local feature values of the remaining points remain unchanged.

[0226] The calculation of the global feature value is to divide the point cloud into multiple regions using the super-voxel method, take the average of the local feature values of each point in each region as the initial global feature value of the region, introduce the initial global feature value of each pixel region into each point of the region, set threshold values h1 and h2, and select feature seed points and non-feature seed points. The region with an initial global feature value greater than h1 is set as a feature region, and the center point thereof is set as a feature seed point. The region with an initial global feature value less than h2 is set as a non-feature region, and the center point thereof is set as a non-feature seed point.

[0227] For feature seed points and other points other than feature seed points, which are adjustment points, the distances of the adjustment points from the feature seed points or the non-feature seed points are compared to determine whether the adjustment points are close to the feature seed points or the non-feature seed points, and then different formulas are used to calculate the global feature values of the adjustment points. The final effect is that the adjustment points close to the feature seed points increase the initial global feature value to the global feature value, and the adjustment points close to the non-feature seed points decrease the initial global feature value to the global feature value. The feature points and the non-feature points update the initial global feature value to the global feature value.

[0228] Finally, the global feature value and the local feature value of each point are integrated. The global feature value and the local feature value of each point are sorted from large to small. For points with global feature values and local feature values ranked in the top 20%, the maximum value is retained. For points with local feature values or local feature values ranked in the last 20%, the product of the two values is taken. For the remaining points, the average of the two values is taken.

[0229] The second step is feature point selection and corresponding relationship construction.

[0230] According to the fusion feature values of the points of the original point cloud and the target point cloud, the fusion feature values of the points are sorted, and points with feature values ranked in a set range at the front end are selected as a candidate feature point set, and a feature point is selected from the point set for construction of a sparse corresponding relationship.

[0231] Taking the selection of the target point cloud feature as an example, first, set the target point set k T , and the access queue flag of the target point cloud candidate feature point set, the length of the flag queue is the length of the target point cloud candidate feature point set, and the initial value of each candidate feature point is 0, which indicates that the point has not been accessed, and 1 indicates that the point has been accessed. The number of target point cloud feature points to be obtained T is set artificially, and the target point set k T is obtained. The target feature point selection algorithm steps are as follows:

[0232] (1) Query the point with the highest feature value in the current target point cloud candidate feature point set and the access queue parameter not being 1, and add the point to the target point set k T, the access queue parameter of the point is set to 1;

[0233] (2) Set the access queue parameter of the point belonging to the target point cloud candidate feature point point set within the range to 1, with the target point cloud feature point as the center;

[0234] (3) If the number of elements of the feature point point set k T is equal to T, or the access queue parameter is 1, stop, otherwise loop to step 1.

[0235] Similarly, for the selection of feature points in the original point cloud, the same method is used. Obtain the feature point set ks of the original point cloud, where k T and ks are set to 6. Then build a corresponding relationship between the feature point point sets of the original point cloud and the target point cloud, that is, there are k T (36) sparse correspondences.

[0236] After obtaining the sparse correspondences, the incorrect sparse correspondences need to be deleted, and here the local feature value of the point and the correlation coefficient are used to calculate an initial relationship weight for each point pair with sparse correspondence. Then sort the weight of each sparse correspondence from large to small, and take the point pairs corresponding to the top 20% of the sparse correspondence as the registration point pairs as the initial data of the coarse rigid registration. Then based on the sparse correspondence of the registration point pair, the original point cloud and the target point cloud are coarsely rigidly registered, and in the process of continuous iteration and coincidence, the intermediate relationship weight is repeatedly calculated, and the spatial distance is added as the data basis for the calculation of the intermediate relationship weight. Update the intermediate relationship weight of each point pair to the relationship weight of each point pair, and finally obtain the correct sparse correspondence as the corresponding relationship and the point pair with the corresponding relationship as the feature point pair.

[0237] Thirdly, the final relationship between the original point cloud and the target point cloud is constructed.

[0238] A corresponding relationship of the near neighbor points is used to construct the corresponding relationship of the remaining points. The near neighbor points of each point (s j , t i ) in the corresponding relationship are searched, and the near neighbor point sets of s j and t i are obtained, where s j and t i are the feature points of the original point cloud and the target point cloud with the corresponding relationship. Then the search range of the corresponding point of the near neighbor point of s j is limited in the near neighbor point set of t i , and the new corresponding relationship obtained is used as the basis for the next corresponding relationship calculation. In this way, it can effectively reduce the search space and improve the accuracy. The steps of the algorithm are as follows:

[0239] A correspondence set H is set, the correspondence obtained is put into the set H, and access queues S H and T H are set, the lengths are initialized as the number of points of the original point cloud and the target point cloud respectively, and are used to represent whether the points in the original point cloud and the target point cloud have corresponding points.

[0240] The correspondence set H is traversed, and a pair of correspondences (s j ,t i )∈H is set, s j and t i are a feature point on the original point cloud and the target point cloud respectively, and the two have a correspondence. A neighbor point set s j _n without a correspondence of s j is obtained by searching, a neighbor point set t i _n without a correspondence of t i is obtained, a sum of a spatial distance and a custom distance between each point of s j _n and t i _n is calculated, and a point with a minimum sum of the spatial distance and the custom distance corresponding to each point is taken as a pair of feature points with a correspondence, and the correspondence is added to the set H. When S H or T H represents that all points have a correspondence, the traversal is stopped, and a final correspondence set of the original point cloud and the target point cloud is obtained. The construction of the correspondence of the original point cloud and the target point cloud is completed.

[0241] The correspondence constructed by the original point cloud and the target point cloud is explained, and the effectiveness of the method of the application is explained.

[0242] First of all, it is explained that the isometric distortion D isoThe error in the correspondence is characterized by isometric distortion. For example, three feature points a, b, and c are randomly selected on the target point cloud, and corresponding feature points A, B, and C are found on the origin point cloud. The term "geodesic" comes from geodesy. Geodesic originally refers to the shortest path between two points on the Earth's surface. Therefore, Geodesic Distance initially referred to the shortest distance between two points on the Earth's surface, but it has been applied to arbitrary shapes, referring to the distance along the shortest path on an object's surface. According to the geodesic principle, line segments AB, AC, and BC are theoretically equal to line segments ab, ac, and bc, respectively. That is, the values ​​of AB / ab, AC / ac, and BC / bc are all 1, and the error is zero. However, since absolute precision is impossible in calculations, the ratios are generally less than 1; the closer to 1, the more precise the calculation. However, for point clouds with deformation, the corresponding geodesics are not equal. Therefore, when using equidistant distortion to represent the accuracy requirements, the requirements should be reduced according to the degree of deformation. For example, in the embodiments of this application, the accuracy requirement is 0.8 or higher. If it is less than 0.8, the correspondence between point clouds should be reconstructed.

[0243] Geodesic distance refers to the shortest distance between two points on a curved surface.

[0244]

[0245] Where, d iso (s j ,t i ) is an individual correspondence (s j ,t i Contribution to overall isometric distortion:

[0246]

[0247] g(s j ,s m () is the geodesic distance between two points.

[0248] Experiments were conducted using models tr_reg_000 to tr_reg_003 from the FAUST dataset and mesh020 to mesh022 from the SCAPE dataset. An example was provided using data tr_reg_000 as the target point cloud and tr_reg_001 as the origin point cloud. Since the original data was already aligned, [example data would be inserted here]. Figure 10 As shown, Figure 10 The diagram illustrates the initial spatial positions of the tr_reg_000 and tr_reg_001 models of a dataset according to an embodiment of this application.

[0249] To verify the effectiveness of the method, the relative spatial positions of the model in the data set are completely shuffled manually, as shown in Figure 11 Figure 11 shows a schematic diagram of the original relative spatial position change of the original point cloud (arm bending) and the target point cloud (arm stretching) according to an embodiment of the present application. The left of the figure is the original relative position, and the right is the changed relative position. The feature value calculation results of the target point cloud and the original point cloud are shown in Figure 12 and Figure 13 . Figure 12 shows a feature value schematic diagram of the target point cloud according to an embodiment of the present application. From left to right, they are the local feature value, global feature value, and fusion feature value result schematic diagram of the point cloud. Figure 13 shows a feature value schematic diagram of the original point cloud according to an embodiment of the present application. From left to right, they are the local feature value, global feature value, and fusion feature value result schematic diagram of the point cloud. The feature points are selected in the gray (light color) area in the figure.

[0250] Figure 14 shows a feature point selection result schematic diagram according to an embodiment of the present application. The left is the feature point of the original point cloud selected, and the right is the feature point of the target point cloud selected. The final sparse correspondence relationship constructed between the two feature points is shown in Figure 15 . Figure 15 shows a correspondence relationship calculation result schematic diagram according to an embodiment of the present application. The rigid registration result based on the sparse correspondence relationship is shown in Figure 16 , Figure 16 shows a comparison schematic diagram of the effect of the traditional algorithm and the method of the present application according to an embodiment of the present application. The right is the calculation result of the traditional ICP algorithm, and the left is the calculation result of the algorithm proposed in the present application. It can be seen that when the initial relative spatial position difference is large, the algorithm implementation of the present application can give a better alignment result.

[0251] The correspondence relationship calculation process of the remaining points of the point cloud is shown in Figure 17 . Figure 17 shows a final correspondence relationship calculation schematic diagram according to an embodiment of the present application. The leftmost is the correspondence relationship that has been calculated, the middle is the new correspondence relationship constructed through the matching of the near neighbor points between the correspondence relationships. The rightmost shows part of the final calculated correspondence relationship result.

[0252] Please refer to Figure 18 and Table 1, Figure 18 ​Fig. 1 shows a schematic diagram of final correspondence calculation results of calculation data according to one embodiment of the present application. Taking tr_reg_000 of the FAUS dataset as the original point cloud and tr_reg_001-tr_reg_003 as the target point cloud, the correspondence calculation results are shown as a, b and c respectively. Taking mesh020 of the SCAPE dataset as the original point cloud and mesh021 and mesh022 as the target point cloud, the correspondence calculation results are shown as d and e respectively.

[0253] Fig. 2 shows final correspondence calculation results of calculation data. Figure 18 The correspondence evaluation results of the above several groups of models are shown in Table 1. The results prove that the method of the present application has good effect on the correspondence calculation of human point clouds with not very large deformation.

[0254] Table 1: Evaluation data of point cloud correspondence of each correspondence number

[0255]

[0256] From the above process, it can be seen that only point clouds with coordinate data are needed to automatically realize the construction of the correspondence between point clouds, and the efficiency and accuracy are good. No additional information is needed, and the method of significant detection through local and global feature fusion can stably obtain feature regions with high consistency between the original point cloud and the target point cloud, thereby providing good conditions for obtaining more reliable sparse correspondence in the subsequent process.

[0257] Referring to Figure 19 , Figure 19 Fig. 3 shows a schematic diagram of a point cloud correspondence construction device according to one embodiment of the present disclosure. The embodiment of the present disclosure provides a point cloud correspondence construction device, which comprises:

[0258] The extraction module 610 is configured to extract feature points in the original point cloud and the target point cloud as an original point set and a target point set through local and global feature value fusion;

[0259] The construction module 620 is configured to construct a sparse correspondence between the original point set feature points and the target point set feature points;

[0260] The screening module 630 is configured to screen the sparse correspondence to obtain feature point pairs with correspondence between the original point set and the target point set through rigid registration of the sparse correspondence;

[0261] The matching module 640 is configured to perform local matching of near neighbor points on the point cloud of the feature point pairs with correspondence to obtain the correspondence of all points between the original point cloud and the target point cloud, and the correspondence of all points forms the point cloud correspondence.

[0262] The point cloud correspondence construction apparatus according to the embodiments of the present application can be implemented by Figure 20 The point cloud correspondence construction apparatus according to the embodiments of the present application can be implemented by Figure 20 The point cloud correspondence construction apparatus according to the embodiments of the present application can be implemented by Figure 20 The point cloud correspondence construction apparatus shown is one example and should not be taken as limiting the functionality or use of embodiments of the present application.

[0263] As shown in Figure 20 , the point cloud correspondence construction apparatus is in the form of a general computing device. The components of the point cloud correspondence construction apparatus can include, but are not limited to, the at least one processing unit 810 described above, the at least one storage unit 820 described above, and a bus 830 that connects the different system components including the storage unit 820 and the processing unit 810.

[0264] The storage unit stores program code that can be executed by the processing unit 810, so that the processing unit 810 performs the steps according to various exemplary embodiments of the present application described in the description of the exemplary methods described above in this specification. For example, the processing unit 810 can perform each of the steps as shown in Figure 1

[0265] The storage unit 820 can include a readable medium in the form of a volatile storage unit, such as a random access memory (RAM) 8201 and / or a cache memory 8202, and can further include a read-only memory (ROM) 8203.

[0266] The storage unit 820 can further include program / utility 8204 having a set of the program modules 8205, such as an operating system, one or more application programs, other program modules, and program data, and each of these examples, or some combination thereof, can include implementation of a network environment.

[0267] The bus 830 can be representative of one or more of several types of bus structures, including a storage unit bus or storage unit controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of a variety of bus structures.

[0268] ​The point cloud correspondence construction device can also communicate with one or more external devices 700 (e.g., a keyboard, a pointing device, a Bluetooth device, etc.), with one or more devices that enable a user to interact with the point cloud correspondence construction device, and / or with any devices (e.g., a router, a modem, etc.) that enable the point cloud correspondence construction device to communicate with one or more other computing devices. Such communication can occur via Input / Output (I / O) interface 850. Also, the point cloud correspondence construction device can communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or the public network, such as the Internet) via network adapter 860. As illustrated, network adapter 860 communicates with other modules of the point cloud camera 12 via bus 830. It should be appreciated that other hardware and / or software modules can be used in conjunction with the point cloud correspondence construction device, as are known in the art, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data archival storage systems, etc.

[0269] From the above description of the embodiments, those skilled in the art will easily understand that the example embodiments described herein can be implemented by software, or by software in combination with necessary hardware. Therefore, the technical solution according to the embodiments of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a U disk, a mobile hard disk, etc.) or a network, and includes a number of instructions to make a computing device (which can be a personal computer, a server, a terminal device, or a network device, etc.) execute the method according to the embodiments of the present application.

[0270] In the example embodiments of the present application, a computer program medium is also provided, which stores computer readable instructions, and when the computer readable instructions are executed by a processor of a computer, the computer executes the method described in the method embodiment part.

[0271] According to one embodiment of the present application, a program product for implementing the method in the method embodiment is also provided, which can be in the form of a portable compact disc read-only memory (CD-ROM) and includes program codes, and can be run on a terminal device, such as a personal computer. However, the program product of the present application is not limited to this, and in this document, the readable storage medium can be any tangible medium containing or storing a program, which can be used or combined with an instruction execution system, device or apparatus.

[0272] The program product can employ any combination of one or more computer-readable media. The computer-readable media can be a computer-readable storage medium or a computer-readable signal medium. The computer-readable storage medium can be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer-readable storage medium include the following: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0273] The computer-readable signal medium can include a computer-readable storage medium that is configured to store and deliver a computer-readable program code. The computer-readable program code can be propagated as a computer-readable signal medium.

[0274] The program code embodied on the computer-readable media can be transmitted using any appropriate medium, including but not limited to wireless, wired, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0275] The program code can be executed by one or more programmable processors, which can be implemented using one or more microprocessors, microcontrollers, digital signal processors, application specific integrated circuits, field programmable gate arrays, programmable logic devices, or the like. The program code can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, C++, or the like, and conventional procedural programming languages, such as the "C" programming language or the like. The program code can execute entirely on the user's computing device, partly on the user's computing device, as a stand-alone software package, partly on the user's computing device and partly on a remote computing device or entirely on the remote computing device or server. In the latter scenario, the remote computing device can be connected to the user's computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computing device, such as through the Internet using an Internet Service Provider. The program code can also be downloaded to the user's computing device from the remote computing device or server through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computing device, such as through the Internet using an Internet Service Provider.

[0276] It should be noted that, although the above detailed description refers to several modules or units of the device for action execution, this division is not mandatory. Indeed, according to an embodiment of the present application, the features and functionalities of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functionalities of one module or unit described above can be further divided into several modules or units embodied by.

[0277] Furthermore, although individual steps of the methods in the present application are described in a particular order in the drawings, this is not required or implied as to the order in which the steps are performed, nor is it required that all of the steps shown be performed to achieve the desired result. Additionally or alternatively, certain steps can be omitted, combined into a single step, broken into multiple steps, and / or the like.

[0278] Those skilled in the art will readily understand that the example embodiments described herein can be implemented by software and / or by software in combination with the necessary hardware. Thus, the technical solutions according to the embodiments of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash disk, a mobile hard disk, or the like) or a network, and includes a number of instructions to enable a computing device (which can be a personal computer, a server, a mobile terminal, or a network device, etc.) to perform the methods according to the embodiments of the present application.

[0279] Other embodiments of the present application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. It is intended that the present application cover any and all variations of the application that come within the scope of the general inventive concepts described herein and including all such variations as fall within the scope of the claims. The specification and examples are illustrative of the application and are not intended to be limiting.

Claims

1. A method for constructing a point cloud correspondence relationship, characterized in that, The method comprises: According to the position relationship between each point and all points in the set region around it, the local feature value of each point is calculated; The size relationship between the local feature value of each point and the preset local feature threshold value is judged, if greater than or equal to the local feature threshold value, it remains unchanged, if less than, the local feature value is reduced, and the local feature value of each point in the original point cloud and the target point cloud is obtained; According to the local feature value of each point in the original point cloud and the target point cloud, the global feature value of each point in the original point cloud and the target point cloud is calculated; According to the global feature value and the local feature value of each point in the original point cloud and the target point cloud, the fusion feature value of each point in the original point cloud and the target point cloud is obtained; The point with the maximum fusion feature value is selected as a feature point in the original point cloud and the target point cloud respectively, and the feature point and other points in the set range of the feature point are removed, and a new feature point is repeatedly selected until the number of feature points in the original point set and the target point set reaches a set number, and the original point set and the target point set are obtained; A sparse correspondence relationship is constructed between the feature points of the original point set and the target point set; The sparse correspondence relationship is used for rigid registration to obtain the feature point pairs with correspondence relationship between the original point set and the target point set; The feature point pairs with correspondence relationship are subjected to local matching of the near neighbor points on the point cloud to obtain the correspondence relationship between all points in the original point cloud and the target point cloud, and the correspondence relationship between all points forms a point cloud correspondence relationship.

2. The method of claim 1, wherein, According to the local feature value of each point in the original point cloud and the target point cloud, the global feature value of each point in the original point cloud and the target point cloud is calculated, comprising: The original point cloud and the target point cloud are divided into multiple regions, and the average value of the local feature values of all points in the region is used as the initial global feature value of each region and the global feature value of each point in the region; According to the set high-level threshold value and low-level threshold value, the region corresponding to the region type is determined, and the region type comprises a feature region and a non-feature region; The initial global feature value of each point in the region is updated to the global feature value according to the region type, and the initial global feature value of the center point of the feature region and the non-feature region is the initial global feature value.

3. The method of claim 2, wherein, The initial global feature value of each point in the region is updated to the global feature value according to the region type, and the initial global feature value of the center point of the feature region and the non-feature region is the global feature value, comprising: For the adjustment point close to the seed point of the feature region center, the global feature value of the adjustment point is calculated according to the global feature value of the seed point, and the adjustment point refers to other points in the target point cloud and the original point cloud except the seed point and the non-seed point; For the adjustment point close to the non-seed point of the non-feature region center, the global feature value of the adjustment point is calculated according to the global feature value of the non-seed point, and the initial global feature value of the seed point and the non-seed point is the global feature value.

4. The method of claim 2, wherein, According to the global feature value and the local feature value of each point in the original point cloud and the target point cloud, the fusion feature value of each point in the original point cloud and the target point cloud is obtained, comprising: The points in the original point cloud and the target point cloud are sorted from large to small according to the global feature value and the local feature value respectively. The global feature value and the local feature value of the point whose ranking is within the range at the front end are reserved, and the maximum value of the global feature value and the local feature value is taken as the fusion feature value of the point.

5. The method of claim 4, wherein, The global feature value and the local feature value of the point whose ranking is within the range at the front end are reserved, and the maximum value of the global feature value and the local feature value is taken as the fusion feature value of the point. For the point whose ranking of the global feature value or the local feature value is within the range at the end, the product of the global feature value and the local feature value is taken as the fusion feature value of the point, and the average value of the global feature value and the local feature value is taken as the fusion feature value of the remaining points.

6. The method of claim 1, wherein, The sparse correspondence relationship between the original point set and the target point set is screened to obtain the feature point pairs having the correspondence relationship between the original point set and the target point set. The local features of the two feature points having the sparse correspondence relationship are compared to obtain an initial relationship weight value for describing the correlation between the feature points by a numerical value; The point pairs having the sparse correspondence relationship are ranked according to the initial relationship weight value, and the point pairs ranked within the range at the front end are screened as registration point pairs; In the process of rigid registration of the original point cloud and the target point cloud, the spatial distance of the two points is added as reference data for calculating the relationship weight value on the basis of the initial relationship weight value, and the relationship weight value of the point pairs corresponding to all sparse correspondence relationships is obtained. According to the relationship weight value of the point pairs having the sparse correspondence relationship, the point pairs having the same number of feature points in the original point set or the target point set are taken as the feature point pairs having the correspondence relationship, and the relationship weight value of the feature point pairs is greater than the relationship weight value of other point pairs.

7. The method of claim 1, wherein, The feature point pairs having the correspondence relationship are subjected to local matching of the near neighbor points in the point cloud to obtain the correspondence relationship between all points of the original point cloud and the target point cloud, and the correspondence relationship between all points forms a point cloud correspondence relationship. The spatial distance and the self-defined distance between a near neighbor point of a feature point of the target point cloud and each near neighbor point of a feature point having the correspondence relationship in the original point cloud are calculated, the near neighbor point being a non-feature point in a region centered on the feature point; The two points having the minimum sum of the spatial distance and the self-defined distance are screened as the feature point pairs having the correspondence relationship, until the near neighbor points of the feature points of the target point cloud or the original point cloud all have the correspondence relationship. According to the obtained feature point pairs having the correspondence relationship, the step of establishing the feature point pairs is repeated until the correspondence relationship between all points of the original point cloud and the target point cloud is obtained.

8. An apparatus for constructing a point cloud correspondence relationship, characterized in that, The device comprises: The extraction module is configured to: calculate local feature values of each point in the original point cloud and the target point cloud according to a position relationship of each point with all points in a set region around the point; determine a size relationship between the local feature values of each point and a preset local feature threshold value; if the local feature value is greater than or equal to the local feature threshold value, the local feature value is kept unchanged; if the local feature value is less than the local feature threshold value, the local feature value is reduced to obtain the local feature values of each point in the original point cloud and the target point cloud; calculate global feature values of each point in the original point cloud and the target point cloud according to the local feature values of each point in the original point cloud and the target point cloud; calculate fusion feature values of each point in the original point cloud and the target point cloud according to the global feature values and the local feature values of each point in the original point cloud and the target point cloud; select a point with the largest fusion feature value as a feature point in the original point cloud and the target point cloud respectively, and remove the feature point and other points in a set range of the feature point; and repeat the selection of a new feature point until the number of feature points in the original point cloud and the target point cloud reaches a set number, to obtain an original point set and a target point set. The construction module is configured to construct a sparse correspondence relationship between the feature points in the original point set and the feature points in the target point set. The screening module is configured to screen the sparse correspondence relationship to obtain a feature point pair having a correspondence relationship between the original point set and the target point set through rigid registration of the sparse correspondence relationship. The matching module is configured to perform local matching of near neighbor points on a point cloud of the feature point pair having the correspondence relationship, to obtain a correspondence relationship between all points in the original point cloud and the target point cloud, and to form a point cloud correspondence relationship based on the correspondence relationship between all points.

Images