CN111862111A - Point cloud registration algorithm based on region segmentation and fusion - Google Patents

Abstract

The invention belongs to the technical field of three-dimensional point cloud registration, and particularly relates to a point cloud registration algorithm based on region segmentation and fusion. The two point cloud data with different visual angles are respectively divided into two groups of data according to regional properties, one group of data is target point cloud data, the other group of data is background point cloud data, the different groups of data are respectively registered based on an ICP (inductively coupled plasma) algorithm, corresponding fusion is carried out according to different weights, the optimal point cloud rotation and translation matrix is worked out, the overall registration effect can be improved, and meanwhile the robustness of the algorithm is improved.

Description

Point cloud registration algorithm based on region segmentation and fusion

The technical field is as follows:

the invention belongs to the technical field of three-dimensional point cloud registration, and particularly relates to a point cloud registration algorithm based on region segmentation and fusion.

Background art:

point cloud registration is an important research direction of computer vision and image graphics, and aims to splice local three-dimensional point cloud data of multiple visual angles into a complete three-dimensional point cloud model. When the three-dimensional target of the actual scene is scanned, the effect of point cloud registration is poor under the influence of background data such as ground and wall surfaces, low quality of collected point cloud and defects of a registration algorithm.

At present, the most widely applied point cloud registration algorithm is an iterative closest point algorithm (ICP algorithm), which is simple and intuitive in calculation and high in registration accuracy, but is easy to fall into local optimization under the condition of poor initial value.

The invention content is as follows:

the invention aims to solve the technical problem that when a three-dimensional target in an actual scene is scanned, due to the fact that the point cloud data quality is low and the point cloud data with a complex structure formed by a ground wall surface, an actual target object and the like exist, the traditional ICP algorithm has the defect of large error during registration, and the point cloud registration effect is poor.

In order to solve the problems, the invention provides an improved algorithm based on classical ICP registration, which is characterized in that two different sets of point cloud data are respectively divided into two sets of data, one set is target point cloud data, the other set is background point cloud data, the different sets of data are registered based on the ICP algorithm, corresponding fusion is carried out according to different weights, the optimal point cloud rotation and translation matrix is solved, the integral registration effect is improved, and meanwhile, the robustness of the algorithm is increased.

In order to achieve the above object, the present invention is implemented by the following technical solution, and a point cloud registration algorithm based on region segmentation and fusion, including the following steps:

(1) collecting scene point cloud data: obtaining two initial point cloud data with different visual angles;

(2) area division: and dividing each initial point cloud data into a target area and a background area. Performing region segmentation on each point cloud data based on a region growing method, wherein the segmented regions are as follows:

R∈{R₀,R₁,..R_N-1n is the number of zones;

according to a threshold value th_lowAnd th_highThe region R is divided into regions. Number of points in the region_lowth_high]The regions within the range constitute a target region; the number of the region points is (0, th)_low) Or (th)_high, + ∞) constitutes the background region. Therein, threshold th_lowAnd th_highAnd determining according to the point cloud scale. Respectively carrying out region division on the two initial point clouds according to the method to obtain a group of target regions (the corresponding point clouds are Cloud)_ob1And Cloud_ob2) A set of background regions (corresponding point clouds are Cloud)_bg1And Cloud_bg2)。

(3) Area registration: the target region and the background region are registered separately according to a conventional ICP algorithm. The method comprises the following specific steps: solving a corresponding rotation and translation matrix of the region, wherein the rotation and translation matrix of the target region is Mat _objectSplit it into a rotation matrix R_obAnd a translation vector V_obThe background area rotation-translation matrix is TMat_backgroundSplit it into a rotation matrix R_bgAnd a translation vector V_bg。

(4) Setting weight: setting different weights for the split rotation matrix and translation vector according to registration errors of different regions: the weight of the translation component of the target region rotation-translation matrix is set as W_vThe translation component of the background region has a weight of 1-W_v(ii) a Similarly, the weight of the rotation component of the target region rotation-translation matrix is set to W_rThe weight of the rotation matrix of the background area is 1-W_r。W_v＝W_rW is calculated from the following equation:

wherein，_obA registration error representing the target region is determined,_bgrepresenting registration errors of background regions.

(5) Final translation vector V_finalDirectly solving by a translation vector calculation formula:

V_final＝V_ob*W_v，+V_bg*(1-W_v)

wherein, V_finalTranslation vectors, V, representing the final two initial point clouds_obTranslation vector, V, representing the target region_bgA translation vector representing a background region.

(6) And (3) solving a final rotation matrix, namely converting the final rotation matrix into a corresponding quaternion for calculation, and fusing:

R_ob→q_ob；R_bg→q_bg，q_final＝q_ob*W_r+q_bg*(1-W_r)

then q_finalNormalization, converting quaternion to rotation matrix, q_final→R_final。

Wherein R is_obA rotation matrix representing the target area, q_obRepresenting quaternions, R, corresponding to the rotation matrix of the target area _bgRotation matrix representing background region, q_bgRepresenting the quaternion, q, corresponding to the rotation matrix of the background region_finalQuaternion, R, corresponding to the rotation matrix representing the final registration of the two initial point clouds_finalA rotation matrix representing the final registration of the two initial point clouds.

(7) Weighted translation vector V_finalAnd a rotation matrix R_finalRotation and translation matrix TMat combined into two initial point clouds and finally registered_final。

The invention has the beneficial effects that:

according to the invention, two different point cloud data are respectively divided into two groups of data, one group of data is target area data to be registered, the other group of data is background area data, different groups of data are registered based on an ICP algorithm respectively, corresponding fusion is carried out according to different weights, and the optimal point cloud rotation translation matrix is worked out, so that the registration effect can be improved, and meanwhile, the robustness of the algorithm is increased.

Drawings

FIG. 1 is a flow chart of the split-zone registration algorithm of the present invention;

fig. 2 is a scene point cloud data acquisition diagram of embodiment 1, wherein the target area is a cabinet, a flowerpot, etc., and the background area is a ground, a wall, etc.;

FIG. 3 is a right side view of the scene of embodiment 1;

FIG. 4 is a left side view of the scene of embodiment 1;

FIG. 5 is a graph of composite ICP for example 1;

fig. 6 is a diagram of a conventional ICP of embodiment 1.

The specific implementation mode is as follows:

in order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are clearly and completely described below, and it is obvious that the described embodiments are a part of the embodiments of the present invention, but not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Example 1:

a point cloud registration algorithm based on region segmentation fusion comprises the following steps:

(1) collecting scene point cloud data: as shown in fig. 1, two point cloud data with different viewing angles are obtained;

(2) area division:

and dividing each point cloud data into a target area and a background area. Performing region segmentation on each point cloud data based on a region growing method, wherein the segmented regions are as follows:

R∈{R₀,R₁,..R_N-1n is the number of zones;

according to a threshold value th_lowAnd th_highThe region R is divided into regions. Number of points in the region_lowth_high]The regions within the range constitute a target region; the number of the region points is (0, th)_low) Or (th)_high, + ∞) constitutes the background region. Therein, threshold th _lowAnd th_highAnd determining according to the point cloud scale. Respectively carrying out region division on the two initial point clouds according to the method to obtain a group of target regions (the corresponding point clouds are Cloud)_ob1And Cloud_ob2) A set of background regions (corresponding point clouds are Cloud)_bg1And Cloud_bg2)。

(3) Area registration: the target region and the background region are registered separately according to a conventional ICP algorithm. The method comprises the following specific steps: solving a corresponding rotation and translation matrix of the area, namely a rotation and translation matrix Mat of the target area_objectSplit into a rotation matrix R_obAnd a translation vector V_obBackground area rotation translation matrix TMat_backgroundSplit into a rotation matrix R_bgAnd a translation vector V_bg。

Area registration: each region was registered according to the conventional ICP algorithm: the steps of the registration algorithm are as follows:

1. registering two target area point clouds through traditional ICP algorithm_ob1And Cloud_ob2Obtaining a target area rotation translation matrix Mat_objectAnd split into rotation matrices R_obAnd a translation vector V_ob；

2. Registering two background area point clouds through traditional ICP algorithm_bg1And Cloud_bg2Obtaining a background area rotation translation matrix TMat_backgroundAnd split into rotation matrices R_bgAnd a translation vector V_bg；

(4) Setting weight: setting different weights for the split rotation matrix and translation vector according to registration errors of different regions: the weight of the translation component of the target rotational-translation matrix is set to W _vThe translation component of the background region has a weight of 1-W_v(ii) a Similarly, the weight of the rotation component of the target rototranslation matrix is set to W_rThe weight of the rotation matrix of the background area is 1-W_r。W_v＝W_rW is calculated from the following equation:

wherein,_oba registration error representing the target region is determined,_bgrepresenting registration errors of background regions.

(5) Final translation vector V_finalDirectly solving by a translation vector calculation formula:

V_final＝V_ob*W_v+V_bg*(1-W_v)

(6) and (3) solving a final rotation matrix, namely converting the final rotation matrix into a corresponding quaternion for calculation, and fusing:

R_ob→q_ob；R_bg→q_bg，q_final＝q_ob*W_r+q_bg*(1-W_r)

then q_finalNormalization processing, converting quaternion into rotation matrix, q_final→R_final。

(7) Weighted translation vector V_finalAnd a rotation matrix R_finalRotation and translation matrix TMat combined into two initial point clouds and finally registered_final。

As can be seen from FIGS. 2-5, when the invention is applied to actual scene for actual measurement point cloud data registration, the point is used

Cloud data quality is not high, meanwhile, complex structure point cloud data composed of a ground wall surface, an actual target object and the like exist, the defect of large error occurs in the traditional ICP algorithm during registration, the algorithm provided by the text can well realize registration of two scene data, and the accuracy is higher than that of the traditional ICP algorithm.

The algorithm is realized on a PC of an Intel Xeon E33.31GHz CPU, an 8G memory and a 64-bit Window 10 operating system. The algorithm is applied to the registration of the actual complex scene point cloud data. The scene data comprises backgrounds such as the ground and the wall, targets such as cabinets and flowerpots, and a large amount of noise exists in the collected point cloud, so that the registration difficulty is increased.

The scene is shown in fig. 2, and the algorithm proposed herein is used to collect scene point cloud data from different view angles, as shown in fig. 3

As shown in fig. 4, region division and region registration are performed; then calculating a final translation vector and a final rotation matrix according to different weights; and finally, combining the weighted translation vectors and the weighted rotation matrixes to obtain a rotation translation transformation matrix of the two point clouds, so that the point clouds can be transformed to realize the registration of the point clouds. Clearly, the algorithm herein is feasible.

To verify the actual performance of the registration algorithm described herein, registration was performed on the point cloud data of fig. 3 and 4 using the present algorithm and the conventional ICP algorithm, respectively, with the following results:

as can be seen from the above table, the algorithm herein has a higher registration accuracy. Compared with the traditional ICP algorithm, the speed is reduced by about 4%, and the precision is improved by about 48%. In low-quality multi-view point cloud data with a large number of complex structures, a traditional ICP algorithm adopts a global registration strategy, so that the traditional ICP algorithm is easily influenced by factors such as background, noise, coverage rate and the like, and a large registration error is caused; the target and background regions provided by the method are respectively registered and then weighted and fused, the invariance of local regions and the difference of the importance degrees of the target and background regions are considered, and the overall error is reduced by a grouping registration and re-fusion mode. Therefore, the algorithm based on region segmentation fusion can realize accurate registration of point cloud.

Claims (5)

1. A point cloud registration algorithm based on region segmentation and fusion is characterized by comprising the following steps:

(1) collecting scene point cloud data: obtaining two initial point cloud data with different visual angles;

(2) area division: dividing each initial point cloud data into a target area and a background area;

(3) area registration: the target area and the background area are respectively registered according to a traditional ICP algorithm;

(4) setting weight: setting different weights for the split rotation matrix and translation vector according to registration errors of different areas;

(5) final translation vector V_finalDirectly solving by a translation vector calculation formula:

V_final＝V_ob*W_v+V_bg*(1-W_v)

wherein, V_finalTranslation vectors, V, representing the final two initial point clouds_obTranslation vector, V, representing the target region_bgA translation vector representing a background region.

(6) And (3) solving a final rotation matrix, namely converting the final rotation matrix into a corresponding quaternion for calculation, and fusing:

R_ob→q_ob；R_bg→q_bg，q_final＝q_ob*W_r+q_bg*(1-W_r)

then q_finalNormalization, converting quaternion to rotation matrix, q_final→R_final。

Wherein R is_obA rotation matrix representing the target area, q_obRepresenting quaternions, R, corresponding to the rotation matrix of the target area_bgRotation matrix representing background region, q_bgRepresenting the quaternion, q, corresponding to the rotation matrix of the background region _finalQuaternion, R, corresponding to the rotation matrix representing the final registration of the two initial point clouds_finalA rotation matrix representing the final registration of the two initial point clouds.

(7) Weighted translation vector V_finalAnd a rotation matrix R_finalRotation and translation matrix TMat combined into two initial point clouds and finally registered_final。

2. The region segmentation fusion based point cloud registration algorithm of claim 1, wherein: and (3) carrying out region segmentation on each point cloud data by region segmentation in the step (2) based on a region growing method, wherein the segmented regions are as follows:

R∈{R₀,R₁,..R_N-1n is the number of zones;

according to a threshold value th_lowAnd th_highDividing the region R into regions with the number of points in [ th ]_lowth_high]The regions within the range constitute a target region; the number of the region points is (0, th)_low) Or (th)_high, + ∞) constitutes a background region; therein, threshold th_lowAnd th_highDetermining according to the point cloud scale; respectively carrying out region division on the two initial point clouds according to the method to obtain a group of target regions (the corresponding point clouds are Cloud)_ob1And Cloud_ob2) A set of background regions (corresponding point clouds are Cloud)_bg1And Cloud_bg2)。

3. The region segmentation fusion based point cloud registration algorithm of claim 1, wherein: the specific steps of the step (3) comprise: solving a corresponding rotation and translation matrix of the region, wherein the rotation and translation matrix of the target region is Mat _objectSplit it into a rotation matrix R_obAnd a translation vector V_obThe background area rotation-translation matrix is TMat_backgroundSplit it into a rotation matrix R_bgAnd a translation vector V_bg。

4. The region segmentation fusion based point cloud registration algorithm of claim 1, wherein: step (4) specifically, according to the registration errors of different areas, the split rotation matrix and the split translation vector are set to have different weights: the weight of the translation component of the target region rotation-translation matrix is set as W_vThe translation component of the background region has a weight of 1-W_v(ii) a Similarly, the weight of the rotation component of the target region rotation-translation matrix is set to W_rThe weight of the rotation matrix of the background area is 1-W_r。

5. The region segmentation fusion based point cloud registration algorithm of claim 4, wherein: w_v＝W_rW is calculated from the following equation:

wherein,_oba registration error representing the target region is determined,_bgrepresenting registration errors of background regions.

Images