Point cloud registration method guided by similarity of overlapping regions

Abstract

In order to solve the problem that an existing point cloud registration method is difficult to realize accurate registration of measurement point clouds with the characteristics of extremely low overlap, lack of features, noise, different density and the like, the invention provides a point cloud registration method guided by similarity of overlapped regions. According to the method, the overlapping areas of the measurement point clouds of the adjacent visual angles are firstly extracted, and the sizes of the overlapping areas are consistent through scale comparison adjustment, so that the influence of initial roughness caused by inherent geometric errors of a measurement system on inaccurate extraction of the overlapping areas can be reduced; the problem that the shapes of overlapping areas are inconsistent due to noise and different densities can be solved through bidirectional interaction, then probability registration is carried out on the overlapping areas with the sizes and the shapes adjusted to be consistent, transformation from a source point cloud subjected to coarse registration to a target point cloud is obtained, finally, the obtained transformation acts on measurement point clouds of adjacent visual angles, and finally the measurement point clouds of the adjacent visual angles are obtained. Robust registration of measurement point clouds with the problems of extremely low overlapping, feature lacking, noise and different densities is achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of image data processing. Background Art

[0002] With the advantages of flexibility, high efficiency, good feasibility and dense measurement points, optical measurement is expected to solve the problem of fast and accurate measurement of complex industrial products. However, due to the limitation of the sensor's measurement range, it is difficult for a single sensor to obtain a complete point cloud of the measurement target in a single measurement. Therefore, it is often necessary to integrate one or more sensors on a mechanical device that can perform multi-view measurement, such as a machine tool, a robot or a multi-axis platform, to form a measurement system that can obtain a complete point cloud of the measurement target. Although the measurement system is calibrated before measurement, due to the inherent geometric error of the measurement system, the measurement point clouds obtained by the measurement system at different viewing angles are difficult to align perfectly after being unified to a common coordinate system (rough registration), resulting in inevitable large stitching errors, which affect the accuracy of optical measurement. Therefore, it is usually ensured that the measurement ranges of adjacent viewing angles have a certain overlap, so that the measurement point clouds of adjacent viewing angles have a certain overlapping area. At this time, after the measurement point clouds of each viewing angle are unified to a common coordinate system, point cloud registration (fine registration) is performed on the measurement point clouds of adjacent viewing angles to reduce the stitching error.

[0003] At present, the mainstream point cloud registration methods are divided into three categories: distance-based, probability-based and feature-based. Among the distance-based registration methods, the iterative closest point registration algorithm (ICP) and its variants are widely used in various fields due to their simplicity and speed, but their binary evaluation method makes them sensitive to the initial pose, noise, and density of the point cloud, and they are prone to fall into local optimality; the probability-based registration method is represented by the coherent point drift algorithm and the GMMReg algorithm. The model representation of such methods can effectively deal with point clouds with noise and density, but the registration accuracy needs to be further improved; the feature-based method extracts point cloud features to guide the corresponding allocation, but when the point cloud features are not significant, such methods are not always effective. In recent years, driven by deep learning technology, a large number of point cloud registration networks have been born, aiming to extract deep features to guide registration through powerful abstract information learning capabilities, or enhance noise recognition capabilities to improve robustness, but the generalization ability of such methods is affected by the quality of training data, and it is difficult and costly to obtain high-quality training data.

[0004] In general, the above point cloud registration methods perform well in specific scenarios. However, in the field of aero-engines, in pursuit of higher service performance, aero-engine components are developing in the direction of thin-walled and complex technology. These structural characteristics lead to significantly low overlap (overlap rate is far less than 50%) of their measured point clouds compared to those of other industrial products, and they also have the characteristics of lack of features, noise, and uneven density. For such measured point clouds, the above point cloud registration methods are difficult to establish effective correspondence between the point clouds to be registered, and cannot complete the task of precise point cloud registration.

[0005] Therefore, it is of great significance to carry out research on the registration method of measurement point clouds with extremely low overlap, lack of features, noise, and varying density, which can provide important technical support for the accurate measurement of thin-walled complex components of aircraft engines. Summary of the invention

[0006] In order to solve the technical problem that existing point cloud registration methods are difficult to achieve accurate registration of measurement point clouds with extremely low overlap, lack of features, noise, and uneven density, the present invention proposes a point cloud registration method guided by overlapping area similarity.

[0007] The present invention adopts the following technical solution:

[0008] The point cloud registration method guided by overlap region similarity is special in that it includes the following steps:

[0009] Step 1: Define the perspective and The measured point clouds obtained under the perspectives are respectively the target point cloud and the source point cloud, and the source point cloud is roughly registered with the target point cloud to obtain the roughly registered source point cloud;

[0010] Step 2: Based on the endpoints of the target point cloud and the endpoints of the roughly registered source point cloud, a nearest neighbor search algorithm is used to extract the first source overlapping region from the roughly registered source point cloud and the first target overlapping region from the target point cloud;

[0011] Step 3: Overlapping region similarity-guided registration;

[0012] Step 3.1: Based on scale comparison, the scales of the first source overlapping region and the first target overlapping region are adjusted to obtain a second source overlapping region and a second target overlapping region with the same scales;

[0013] Step 3.2: Through two-way interaction, the shapes of the second source overlapping region and the second target overlapping region are adjusted to obtain a third source overlapping region and a third target overlapping region with the same scale and shape;

[0014] Step 3.3: finely register the third source overlap region to the third target overlap region through a probability-based registration algorithm, and obtain a transformation from the roughly registered source point cloud to the target point cloud;

[0015] Step 4: Point cloud precise registration;

[0016] The transformation is applied to the roughly registered source point cloud, so that the roughly registered source point cloud is precisely registered to the target point cloud, and the precisely registered source point cloud is obtained, thus completing the point cloud registration.

[0017] Furthermore, the nearest neighbor search algorithm in step 2 is a Kd-tree algorithm.

[0018] Furthermore, step 3.1 is specifically as follows:

[0019] Step 3.1.1: Solve the directed bounding boxes of the first source overlapping area and the first target overlapping area respectively, and calculate the scale difference between the first source overlapping area and the first target overlapping area based on the directed bounding boxes using the following formula: :

[0020] ;

[0021] Where a, b, and c are the weights of the length, width, and height of the directed bounding box when calculating the scale difference, and the value ranges of a, b, and c are all 0 to 1, and a+b+c=1; d1, d2, and d3 are the length, width, and height of the directed bounding box of the first source overlapping area, respectively; d4, d5, and d6 are the length, width, and height of the directed bounding box of the first target overlapping area, respectively;

[0022] Step 3.1.2: The scale difference The scale difference threshold is set In comparison, if , delete the end point of the overlapping area of ​​the first source; if , delete the starting point of the first target overlapping area;

[0023] Step 3.1.3: Repeat steps 3.1.1-3.1.2 until the scale difference is satisfy When , a second source overlapping region and a second target overlapping region with the same scale are obtained;

[0024] Said The value range is 0.001-0.01mm.

[0025] Furthermore, step 3.2 is specifically as follows:

[0026] Step 3.2.1: Construct the closest point set;

[0027] By using a nearest neighbor search algorithm, the nearest points corresponding to the points in the second source overlapping area are searched in the second target overlapping area, and a nearest point set of the second source overlapping area is constructed;

[0028] By using a nearest neighbor search algorithm, the nearest points corresponding to the points in the second target overlapping area are searched in the second source overlapping area, and a nearest point set of the second target overlapping area is constructed;

[0029] Step 3.2.2: Solve the source intermediate point cloud and the target intermediate point cloud;

[0030] Each point in the overlapping area of ​​the second source corresponds to each point in the nearest point set of the overlapping area of ​​the second source to form a point pair, which together constitutes Set point pairs and solve The midpoints of each pair of points in the group of points, which constitute the source midpoint cloud; is the number of points in the point cloud of the second source overlap area; the number of points in the source middle point cloud is equal to the number of points in the point cloud of the second source overlap area;

[0031] Each point in the overlapping area of ​​the second target corresponds to each point in the nearest point set of the overlapping area of ​​the second target to form a point pair, which together constitutes Set point pairs and solve The midpoints of each point pair in the group of point pairs, these midpoints constitute the target intermediate point cloud; is the number of points in the point cloud of the overlapping area of ​​the second target; the number of points in the target middle point cloud is equal to the number of points in the point cloud of the overlapping area of ​​the second target;

[0032] Step 3.2.3: Centroid transformation and point cloud fusion;

[0033] Centroid transformation: Based on the difference between the centroid coordinates of the source middle point cloud and the centroid coordinates of the overlapping area of ​​the second source, the centroid transformation is performed on each point in the source middle point cloud other than the starting point and the end point to eliminate the difference in posture between the source middle point cloud and the overlapping area of ​​the second source; Based on the difference between the centroid coordinates of the target middle point cloud and the centroid coordinates of the overlapping area of ​​the second target, the centroid transformation is performed on each point in the target middle point cloud other than the starting point and the end point to eliminate the difference in posture between the target middle point cloud and the overlapping area of ​​the second target;

[0034] Point cloud fusion: replace the coordinates of the starting point and the end point of the source middle point cloud with the coordinates of the starting point and the end point of the second source overlapping area, thereby adding the pose information of the second source overlapping area to the source middle point cloud; replace the coordinates of the starting point and the end point of the target middle point cloud with the coordinates of the starting point and the end point of the second target overlapping area, thereby adding the pose information of the second target overlapping area to the target middle point cloud;

[0035] After centroid transformation and point cloud fusion, the source middle point cloud is transformed to the second source overlapping area, and the target middle point cloud is transformed to the second target overlapping area, so as to obtain a third source overlapping area and a third target overlapping area with the same scale and shape.

[0036] Furthermore, the nearest neighbor search algorithm described in step 3.2.1 is a Kd-tree algorithm.

[0037] Furthermore, in step 3.2.2:

[0038] The midpoints of the point pairs are calculated according to the following formula:

[0039] ;

[0040] ;

[0041] ;

[0042] in, , , for Group point alignment The coordinates of the midpoint of the pair of points, , , is the first The coordinates of the points, , , The closest point in the overlapping area of ​​the second source is the The coordinates of the points corresponding to the points;

[0043] The midpoints of the point pairs are calculated according to the following formula:

[0044] ;

[0045] ;

[0046] ;

[0047] in, , , for Group point alignment The coordinates of the midpoint of the pair of points, , , The second target overlap area The coordinates of the points, , , The closest point in the overlapping area of ​​the second target is the first point in the overlapping area of ​​the second target. The coordinates of the point corresponding to the point.

[0048] Furthermore, the method of centroid transformation in step 3.2.3 is specifically as follows:

[0049] The points other than the starting point and the end point in the source intermediate point cloud are recorded as , n is 2, 3, ..., , point The centroid transformation is performed as follows:

[0050] ;

[0051] ;

[0052] ;

[0053] in, , , are the coordinates of the centroid of the source middle point cloud, , , are the coordinates of the centroid of the overlapping region of the second source;

[0054] Each point other than the starting point and the end point in the target intermediate point cloud is recorded as ,m is 2,3,…, , point The centroid transformation is performed as follows:

[0055] ;

[0056] ;

[0057] ;

[0058] in, , , is the coordinate of the centroid of the target middle point cloud, , , is the coordinate of the centroid of the overlapping area of ​​the second target.

[0059] Furthermore, the probability-based registration algorithm described in step 3.3 is a coherent point drift algorithm.

[0060] The beneficial effects of the present invention are as follows:

[0061] 1. Since the measured point cloud features of thin-walled complex components in aviation are not significant, the present invention adopts an overlap-based registration strategy, that is, extracting the overlapping area of ​​the adjacent perspective measurement point clouds after rough registration to obtain a first source overlapping area and a first target overlapping area, and performing similarity-guided registration on the point clouds of the first source overlapping area and the first target overlapping area to obtain a transformation from the source point cloud after rough registration to the target point cloud, and then applying the obtained transformation to the source point cloud after rough registration, thereby improving the ability of the present invention to cope with extremely low overlap point cloud registration; secondly, the present invention adjusts the scale of the first source overlapping area and the first target overlapping area by scale comparison to obtain a second source overlapping area and a second target overlapping area with consistent scales. domain, reducing the impact of initial roughness caused by the inherent geometric error of the measurement system on inaccurate extraction of overlapping areas; thirdly, in order to address the problem of inconsistent shapes of overlapping areas caused by noise and different densities, the present invention adopts two-way interaction to adjust the shapes of the second source overlapping area and the second target overlapping area, and obtains the third source overlapping area and the third target overlapping area that are consistent in scale and shape; finally, the probability-based registration algorithm is used to guide the precise registration of the third source overlapping area and the third target overlapping area that are consistent in scale and shape, so as to obtain the transformation from the source point cloud after coarse registration to the target point cloud, and finally realizes the robust registration of measurement point clouds with extremely low overlap, lack of features, noise, and different density problems.

[0062] 2. It has been verified in practice that the registration success rate and registration error of the present invention are superior to those of the existing mainstream registration methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 The figure is a flow chart of the method of the present invention.

[0064] Figure 2 It is the target point cloud and the source point cloud after rough registration in the embodiment of the present invention.

[0065] Figure 3 It is the first source overlapping area and the first target overlapping area in the embodiment of the present invention.

[0066] Figure 4 The second source overlapping area and the second target overlapping area are obtained after the scale is adjusted through scale comparison in the embodiment of the present invention.

[0067] Figure 5 The third source overlapping area and the third target overlapping area are obtained after the shapes are adjusted through bidirectional interaction in the embodiment of the present invention.

[0068] Figure 6 It is the target point cloud and the source point cloud after precise registration in the embodiment of the present invention.

[0069] Description of reference numerals:

[0070] 1-target point cloud; 2-source point cloud after coarse alignment; 3-starting point of source point cloud after coarse alignment; 4-end point of target point cloud; 5-first closest point; 6-second closest point; 7-first source overlapping area; 8-first target overlapping area; 9-second source overlapping area; 10-second target overlapping area; 11-third source overlapping area; 12-third target overlapping area; 13-source point cloud after fine alignment. DETAILED DESCRIPTION

[0071] In the following, in conjunction with the accompanying drawings, the point cloud registration method of the present invention is described in detail by taking the measurement point cloud (the measurement point cloud is a two-dimensional point cloud) obtained by the line laser multi-view measurement system of a thin-walled complex component in an aircraft engine as an example.

[0072] For the convenience of description, the first The measured point cloud (i.e. The measured point cloud obtained from the first viewing angle is recorded as the source point cloud. The measured point cloud (i.e. The measured point cloud obtained from a viewing angle is recorded as the target point cloud, and the index of each point in the source point cloud and the target point cloud increases in the clockwise direction.

[0073] Reference Figure 1 The steps of registering the source point cloud to the target point cloud using the method of the present invention are as follows:

[0074] Step 1: Rough registration of point cloud;

[0075] After the calibration of the line laser multi-view measurement system is completed, the transformation relationship between the coordinate systems in the line laser multi-view measurement system is known. Therefore, the source point cloud can be transformed based on the transformation relationship between the coordinate systems in the line laser multi-view measurement system, and the source point cloud can be roughly aligned to the target point cloud 1 to obtain the roughly aligned source point cloud 2. Figure 2 .

[0076] Step 2: Based on the endpoints of the roughly aligned source point cloud and the endpoints of the target point cloud, the overlapping areas are extracted from both the roughly aligned source point cloud and the target point cloud using the nearest neighbor search algorithm;

[0077] In multi-view measurement, point cloud acquisition is sequential, and the end points and starting points of the target point cloud 1 and the source point cloud 2 after rough registration are known;

[0078] By using the nearest neighbor search algorithm (such as the Kd-tree algorithm), the nearest point corresponding to the end point 4 of the target point cloud is searched in the roughly registered source point cloud 2, which is recorded as the second nearest point 6. The point cloud between the starting point 3 of the roughly registered source point cloud and the second nearest point 6 is extracted to obtain the first source overlapping area 7 (see Appendix Figure 3), denoted as , N represents the first source overlapping area The number of points in the point cloud.

[0079] By using the nearest neighbor search algorithm (such as the Kd-tree algorithm), the nearest point corresponding to the starting point 3 of the source point cloud after rough registration is searched in the target point cloud 1, which is recorded as the first nearest point 5. The point cloud between the first nearest point 5 and the end point 4 of the target point cloud is extracted to obtain the first target overlapping area 8 (see Appendix Figure 3 ), denoted as , M represents the first target overlapping area The number of points in the point cloud;

[0080] Step 3: Overlapping region similarity-guided registration;

[0081] Step 3.1: Based on scale comparison, the scales of the first source overlapping region and the first target overlapping region are adjusted to obtain a second source overlapping region and a second target overlapping region with the same scales;

[0082] Step 3.1.1 calculates the scale difference between the first source overlap region and the first target overlap region;

[0083] Solving the first source overlap region by principal component analysis The directed bounding box of Overlapping area with the first target The directed bounding box of , a directed bounding box The length, width, and height are recorded as d1, d2, and d3 respectively, and the directed bounding box The length, width and height are recorded as d4, d5 and d6 respectively, and the overlapping area of ​​the first source is calculated using the following formula Overlapping area with the first target The scale difference :

[0084] ;

[0085] In the formula, a, b, and c are the weights of the length, width, and height of the directed bounding box when calculating the scale difference, respectively. The value ranges of a, b, and c are all 0 to 1, and a+b+c=1. In this embodiment, a=0.5, b=0.5, and c=0. In addition, since the measurement point cloud obtained by the line laser multi-view measurement system is a two-dimensional point cloud, the directed bounding box here is a two-dimensional bounding box. At this time, d3 and d6 are both 0.

[0086] Step 3.1.2: adjusting the scale of the first source overlapping area and / or the first target overlapping area;

[0087] The scale difference calculated in step 3.1.1 is The scale difference threshold is set For comparison, The value range is 0.001-0.01mm:

[0088] like , delete the first source overlapping area End point, after deletion, the first source overlapping area The current point cloud point count is reduced by 1;

[0089] like , delete the first target overlapping area The starting point of the first target overlap area after deletion The current point cloud point count is reduced by 1.

[0090] Step 3.1.3 Repeat steps 3.1.1-3.1.2 until the calculated scale difference satisfy When , the second source overlapping region 9 and the second target overlapping region 10 with the same scale are obtained, as shown in Figure 4 shown.

[0091] The second source overlap region 9 is denoted as :

[0092] ;

[0093] in, The second source overlap area Middle Points, The second source overlap area The number of points in the point cloud, , The overlapping area of ​​the first source The number of times scaling was performed;

[0094] The second target overlapping area 10 is recorded as :

[0095] ;

[0096] in, Overlapping area for the second target Middle Points, Overlapping area for the second target The number of points in the point cloud, , Overlapping area for the first target The number of times to perform rescaling.

[0097] Step 3.2: adjusting the shape of the second source overlapping region and the second target overlapping region through bidirectional interaction to obtain a third source overlapping region and a third target overlapping region with the same scale and shape;

[0098] Step 3.2.1: Construct the closest point set;

[0099] Use the nearest neighbor search algorithm (such as Kd-tree algorithm) to find the overlapping area of ​​the second target Find the second source overlapping area The nearest point corresponding to each point in the source is defined as the nearest point set of the overlapping area of ​​the second source. :

[0100] ;

[0101] Use a nearest neighbor search algorithm (such as Kd-tree algorithm) to find the overlapping area of ​​the second source Find the second target overlapping area The nearest point corresponding to each point in the target is defined as the nearest point set of the overlapping area of ​​the second target. :

[0102] .

[0103] Step 3.2.2: Solve the source intermediate point cloud and the target intermediate point cloud;

[0104] The second source overlaps the area The first The points are represented as , the closest point set of the overlapping area of ​​the second source The first The points are represented as , due to the overlap region of the second source The closest point set The number of points is the same and each point corresponds to each other to form a point pair, which constitutes Set point pairs and solve The midpoints of each point pair in the group of point pairs constitute the source intermediate point cloud, denoted as ;

[0105] in, The source middle point cloud The Points, ; , ;

[0106] Overlapping area of ​​the second target The first The points are represented as , the closest point set of the overlapping area of ​​the second target The first The points are represented as , due to the overlap area of ​​the second target The closest point set The number of points is the same and each point corresponds to each other to form a point pair, which constitutes Set point pairs and solve The midpoints of each point pair in the group of point pairs constitute the target intermediate point cloud, denoted as ;

[0107] in, The target middle point cloud The Points, ; , .

[0108] Step 3.2.3: By means of centroid transformation and point cloud fusion, the source middle point cloud is transformed to the second source overlapping region, and the target middle point cloud is transformed to the second target overlapping region, so as to obtain a third source overlapping region and a third target overlapping region with the same scale and shape;

[0109] Step 3.2.3.1 Source intermediate point cloud Transform to the second source overlap area ;

[0110] For the source intermediate point cloud , whose position overlaps with the second source There is a difference, we use the difference in the center of mass coordinates of the source middle point cloud The points outside the starting point and the end point are transformed by centroid to eliminate the difference in posture between them; through point cloud fusion, the source middle point cloud The starting point and end point are set as the second source overlapping area The starting and ending points of the source point cloud Add a second source overlap area The posture information of .

[0111] The second source overlaps the area Point cloud between the source The centroids are recorded as:

[0112] ;

[0113] ;

[0114] Source intermediate point cloud Points within The specific steps for centroid transformation and point cloud fusion are as follows:

[0115] When n=1, point , is the source middle point cloud The starting point in the middle of the source point cloud The coordinate value of the starting point in is set as the second source overlapping area The coordinate value of the starting point, that is:

[0116] ;

[0117] ;

[0118] When 1 <n< When n is 2, 3, ..., Time, point The source middle point cloud For all points other than the starting point and the end point, according to the centroid With the centroid , for the source middle point cloud The points other than the starting point and the end point are processed as follows:

[0119] ;

[0120] ;

[0121] when Time, point , is the source middle point cloud The end point in the source middle point cloud The coordinates of the end point in the second source overlap area are set The coordinate value of the end point, that is:

[0122] ;

[0123] .

[0124] After the above centroid transformation and point cloud fusion processing, the source intermediate point cloud can be Transform to the second source overlap area , and obtain the third source overlapping region 11, as shown in the attached Figure 5 shown.

[0125] Step 3.2.3.2 Target middle point cloud Transform to the second target overlapping area ;

[0126] For the target intermediate point cloud , whose position overlaps with the second target area There is a difference, we use the difference between the centroid coordinates of the two to calculate the target intermediate point cloud The points outside the starting point and the end point are transformed by centroid to eliminate the difference in posture between them; through point cloud fusion, the target middle point cloud The starting point and end point of the second target overlap area are set as The starting and ending points of the target point cloud Add a second target overlap area The posture information of .

[0127] Overlapping area of ​​the second target Point cloud between the target and The centroids are recorded as:

[0128] ;

[0129] ;

[0130] For the target middle point cloud Points within The specific steps for centroid transformation and point cloud fusion are as follows:

[0131] When m=1, point , is the target middle point cloud The starting point in the target middle point cloud The coordinate value of the starting point in is set as the second target overlapping area The coordinate value of the starting point, that is:

[0132] ;

[0133] ;

[0134] When 1 <m< , that is, m is 2, 3, ..., Time, point The target middle point cloud For all points other than the starting point and the end point, according to the centroid With the centroid , for the target middle point cloud The points other than the starting point and the end point are processed as follows:

[0135] ;

[0136] ;

[0137] When m= Time, point , is the target middle point cloud The end point in the target middle point cloud The coordinate value of the end point in is set as the second target overlapping area The coordinate value of the end point, that is:

[0138] ;

[0139] .

[0140] After the above centroid transformation and point cloud fusion processing, the target middle point cloud can be Transform to the second target overlapping area , and obtain the third target overlapping area 12, as shown in the attached Figure 5 shown.

[0141] The third source overlapping region 11 and the third target overlapping region 12 are consistent in size and shape.

[0142] The above steps 3.2.3.1 and 3.2.3.2 can be interchanged in order.

[0143] Step 3.3: Obtain the transformation from the coarsely registered source point cloud to the target point cloud through probabilistic registration;

[0144] By using a probability-based registration algorithm (a known method, such as the coherent point drift algorithm), the third source overlapping area 11 is precisely registered to the third target overlapping area 12, and the transformation from the roughly registered source point cloud 2 to the target point cloud 1 can be obtained. The transformation is a homogeneous matrix that includes a rotation matrix and a translation vector.

[0145] Step 4: Point cloud precise registration;

[0146] Apply the transformation obtained in step 3.3 to the roughly registered source point cloud 2, so that the roughly registered source point cloud 2 is precisely registered to the target point cloud 1, and the precisely registered source point cloud 13 is obtained. At this point, the point cloud registration is completed, as shown in the attached figure. Figure 6 shown.

[0147] The above is an explanation of the present invention using a two-dimensional point cloud as an example. The registration method for a three-dimensional point cloud is the same as that for the two-dimensional point cloud, the only difference being that the z-coordinate value of the coordinates of the points involved is not equal to 0.

[0148] Technical effect verification:

[0149] The registration effects of the present invention and four existing mainstream registration methods were compared in 28 test data sets, and the success rates of the five methods in completing the registration task and the proportions of achieving the lowest registration error in the 28 test data sets were statistically analyzed, as shown in the following table.

[0150]

[0151] The results show that the success rate of the proposed method in completing the registration task is 28 / 28, and the ratio of achieving the lowest registration error is 25 / 28, which is significantly better than the other four mainstream registration methods.

Claims

1. Overlapping region similarity-guided point cloud registration method, characterized in that: The following steps are involved: Step 1: Define the perspective and The measured point clouds obtained under the perspectives are respectively the target point cloud and the source point cloud, and the source point cloud is roughly registered with the target point cloud to obtain the roughly registered source point cloud; Step 2: Based on the endpoints of the target point cloud and the endpoints of the roughly registered source point cloud, a nearest neighbor search algorithm is used to extract the first source overlapping region from the roughly registered source point cloud and the first target overlapping region from the target point cloud; Step 3: Overlapping region similarity-guided registration; Step 3.1: Based on scale comparison, the scales of the first source overlapping region and the first target overlapping region are adjusted to obtain a second source overlapping region and a second target overlapping region with the same scales; Step 3.2: Through two-way interaction, the shapes of the second source overlapping region and the second target overlapping region are adjusted to obtain a third source overlapping region and a third target overlapping region with the same scale and shape; Step 3.3: finely register the third source overlap region to the third target overlap region through a probability-based registration algorithm, and obtain a transformation from the roughly registered source point cloud to the target point cloud; Step 4: Point cloud precise registration; The transformation is applied to the roughly registered source point cloud, so that the roughly registered source point cloud is precisely registered to the target point cloud, and the precisely registered source point cloud is obtained, thus completing the point cloud registration.

2. The overlapping region similarity guided point cloud registration method according to claim 1, characterized in that: The nearest neighbor search algorithm in step 2 is the Kd-tree algorithm.

3. The overlapping region similarity guided point cloud registration method according to claim 2, characterized in that: Step 3.1 is as follows: Step 3.1.1: Solve the directed bounding boxes of the first source overlapping area and the first target overlapping area respectively, and calculate the scale difference between the first source overlapping area and the first target overlapping area based on the directed bounding boxes using the following formula: : ； Where a, b, and c are the weights of the length, width, and height of the directed bounding box when calculating the scale difference. The value ranges of a, b, and c are all 0 to 1, and a+b+c=1; d1, d2, and d3 are the length, width, and height of the directed bounding box of the first source overlapping area; d4, d5, and d6 are the length, width, and height of the directed bounding box of the first target overlapping area; Step 3.1.2: Difference The scale difference threshold is set In comparison, if , delete the end point of the overlapping area of ​​the first source; like , delete the starting point of the first target overlapping area; Step 3.1.3: Repeat steps 3.1.1-3.1.2 until the scale difference is satisfy When , a second source overlapping region and a second target overlapping region with the same scale are obtained; Said The value range is 0.001-0.01mm.

4. The overlapping region similarity guided point cloud registration method according to claim 3, characterized in that: Step 3.2 is as follows: Step 3.2.1: Construct the closest point set; By using a nearest neighbor search algorithm, the nearest points corresponding to the points in the second source overlapping area are searched in the second target overlapping area, and a nearest point set of the second source overlapping area is constructed; By using a nearest neighbor search algorithm, the nearest points corresponding to the points in the second target overlapping area are searched in the second source overlapping area, and a nearest point set of the second target overlapping area is constructed; Step 3.2.2: Solve the source intermediate point cloud and the target intermediate point cloud; Each point in the overlapping area of ​​the second source corresponds to each point in the nearest point set of the overlapping area of ​​the second source to form a point pair, which together constitutes Set point pairs and solve The midpoints of each pair of points in the group of points, which constitute the source midpoint cloud; is the number of points in the point cloud of the second source overlap area; the number of points in the source middle point cloud is equal to the number of points in the point cloud of the second source overlap area; Each point in the overlapping area of ​​the second target corresponds to each point in the nearest point set of the overlapping area of ​​the second target to form a point pair, which together constitutes Set point pairs and solve The midpoints of each point pair in the group of point pairs, these midpoints constitute the target intermediate point cloud; is the number of points in the point cloud of the overlapping area of ​​the second target; the number of points in the target middle point cloud is equal to the number of points in the point cloud of the overlapping area of ​​the second target; Step 3.2.3: Centroid transformation and point cloud fusion; Centroid transformation: Based on the difference between the centroid coordinates of the source middle point cloud and the centroid coordinates of the overlapping area of ​​the second source, the centroid transformation is performed on each point other than the starting point and the end point in the source middle point cloud to eliminate the posture difference between the source middle point cloud and the overlapping area of ​​the second source; Based on the difference between the centroid coordinates of the target middle point cloud and the centroid coordinates of the overlapping area of ​​the second target, centroid transformation is performed on each point other than the starting point and the end point in the target middle point cloud to eliminate the posture difference between the target middle point cloud and the overlapping area of ​​the second target; Point cloud fusion: Replace the coordinates of the starting point and the end point of the source middle point cloud with the coordinates of the starting point and the end point of the second source overlapping area, thereby adding the pose information of the second source overlapping area to the source middle point cloud; The coordinates of the start point and the end point of the target middle point cloud are replaced by the coordinates of the start point and the end point of the second target overlapping area, thereby adding the position and posture information of the second target overlapping area to the target middle point cloud; After centroid transformation and point cloud fusion, the source middle point cloud is transformed to the second source overlapping area, and the target middle point cloud is transformed to the second target overlapping area, so as to obtain a third source overlapping area and a third target overlapping area with the same scale and shape.

5. The overlapping region similarity guided point cloud registration method according to claim 4, characterized in that: The nearest neighbor search algorithm described in step 3.2.1 is the Kd-tree algorithm.

6. The overlapping region similarity guided point cloud registration method according to claim 4 or 5, characterized in that: In step 3.2.2: The midpoints of the point pairs are calculated as follows: ； ； ； in, , , for Group point alignment The coordinates of the midpoint of the pair of points, , , is the first The coordinates of the points, , , The closest point in the overlapping area of ​​the second source is the The coordinates of the points corresponding to the points; The midpoints of the point pairs are calculated as follows: ； ； ； in, , , for Group point alignment The coordinates of the midpoint of the pair of points, , , The second target overlap area The coordinates of the points, , , The closest point in the overlapping area of ​​the second target is the first point in the overlapping area of ​​the second target. The coordinates of the point corresponding to the point.

7. The overlapping region similarity-guided point cloud registration method according to claim 6, characterized in that: The method of centroid transformation in step 3.2.3 is as follows: The points other than the starting point and the end point in the source intermediate point cloud are recorded as , n is 2, 3, ..., , point The centroid transformation is performed as follows: ； ； ； in, , , are the coordinates of the centroid of the source middle point cloud, , , are the coordinates of the centroid of the overlapping region of the second source; Each point other than the starting point and the end point in the target intermediate point cloud is recorded as ,m is 2,3,…, , point The centroid transformation is performed as follows: ； ； ； in, , , is the coordinate of the centroid of the target middle point cloud, , , is the coordinate of the centroid of the overlapping area of ​​the second target.

8. The overlapping region similarity guided point cloud registration method according to claim 7, characterized in that: The probability-based registration algorithm described in step 3.3 is the coherence point drift algorithm.

Images