CN-122023519-A - Point cloud matching method based on point-to-feature descriptors

CN122023519ACN 122023519 ACN122023519 ACN 122023519ACN-122023519-A

Abstract

The invention discloses a point cloud matching method based on a point-to-feature descriptor, which comprises the following steps of adding a target point cloud, solving a normal vector of the target point cloud, carrying out weighted mean smoothing filtering, step 2, screening the normal vector direction of the target point cloud based on the calculated normal vector, step 3, sampling the target point cloud, calculating a sampling point-to-feature value, carrying out neighborhood diffusion, step 4, loading the scene point cloud, carrying out least square fitting calculation on a local plane of a neighborhood point, extracting a plane, adding an isolated point denoising algorithm, screening, selecting the number k of neighbors and a distance threshold d_threshold as denoising parameters, step 5, screening the normal vector of the scene point cloud, setting the normal vector as Z-axis reverse filtering, step 6, combining the distance and the normal vector included angle by the scene point cloud sampling algorithm, defaulting the included angle threshold to 30 degrees, carrying out external setting, step 7, carrying out selective secondary matching in two radiuses when the point-to-feature matching, and carrying out neighborhood expansion strategy voting on the pose after the matching is completed.

Inventors

LIU JIAKUI
WANG TANGMENG
XU ZHIFEI

Assignees

深丝智能科技(江苏)有限公司

Dates

Publication Date: 20260512
Application Date: 20260114

Claims (10)

1. The point cloud matching method based on the point-to-feature descriptors is characterized by comprising the following steps of: Step 1, adding a target point cloud, solving normal vectors of the target point cloud, and carrying out weighted average smoothing filtering on the normal vectors of the target point cloud; Step 2, screening the normal vector direction of the target point cloud based on the normal vector calculated in the target point cloud; Step 3, sampling the target point cloud, calculating a point pair characteristic value of a target point cloud sampling point, and carrying out neighborhood diffusion on the point pair characteristic value; Step 4, loading scene point clouds, performing normal vector calculation on the scene point clouds, adopting a neighborhood point local plane least square fitting method for normal vector of each point, extracting planes by using equations of local planes, adding an isolated point denoising algorithm in the process of calculating the normal vector of the scene point clouds to screen, and selecting neighbor number k and distance threshold d_threshold as parameters of isolated point denoising; Step 5, screening normal vectors of the scene point cloud according to the normal vector direction, and setting the normal vectors as Z-axis reverse filtering; step 6, sampling a sampling point algorithm of a scene point cloud by combining a distance and a normal vector included angle, wherein the threshold angle of the included angle defaults to 30 degrees and is used as an externally settable parameter; Step 7, selectively performing secondary matching in two radiuses of a small radius and a large radius in the characteristic matching process of point cloud scene point cloud and model point cloud point pairs, wherein specific values of the small radius and the large radius are set according to the size of a model; and 8, after the point-to-point feature matching is completed, voting the pose by adopting a neighborhood expansion strategy, and setting the length of the neighborhood to be 3.
2. The point cloud matching method based on the point-to-feature descriptors according to claim 1, wherein in the step 1, the filtering method is adopted to filter by using a neighborhood weighted average value of normal vector points, weights of the neighborhood points in the neighborhood are distributed according to distances from the neighborhood points to the normal vector points, and the closer the distances are, the greater the weights are; wherein the normal vector after filtering at any point p in the point cloud is expressed by the following formula: ; Wherein, the Is the normal vector before the point p filtering, Is the normal vector of the neighborhood point, Is the weight of the self point and can be set as a fixed constant, such as setting Weighting of neighborhood points Sigma may be set at 3 times the resolution.
3. The point cloud matching method based on point-to-feature descriptors according to claim 1, wherein in the step 2, Z-axis back filtering is set, and if the normal vector of the valid feature points in the target point cloud is at an angle smaller than 90 degrees with respect to the Z-axis unit direction vector (0, 1), such feature points are valid, and not valid.
4. The method according to claim 1, wherein in the step 3, the distance and the included angle between the two normal vectors are used for sampling, and the included angle threshold angle includes 30 °, 45 ° or 180 °.
5. The point cloud matching method based on the point pair feature descriptor according to claim 4, wherein the neighborhood size is set to 3, the point pair feature is stored with hash table { p, { index, angle }, wherein key p in the hash table represents the point pair feature, and the value { index, angle } represents the index of the first point in all the point pairs belonging to a point pair feature p and all the sets of rotation angles of the point pairs.
6. The method according to claim 1, wherein in the step 4, the point cloud point p is used to calculate the algorithm vector when finding K points nearest to the point p, and the distances di from K points to the point p are calculated together when finding K points, and if the distances di from the point i to the point p in the K points are smaller than the set d_threshold, the normal vector calculated by the point p is assigned as an invalid value, and the point p is not considered by the subsequent sampling points.
7. A point cloud matching method based on point-to-feature descriptors according to claim 3, wherein the back filtering in step 5 is the same as in step 2.
8. The method according to claim 4, wherein the sampling algorithm in the point cloud of the scene in step 6 is the same as that in step 3.
9. The point cloud matching method based on the point-to-feature descriptor according to claim 1, wherein in the step 7, the point-to-feature matching of the scene point cloud to the target point cloud is performed within two neighborhood ranges of a small radius and a large radius, the two neighborhood ranges are represented by R1 and R2, R1 is smaller than R2, wherein r1= (qMedian + qMin)/2, r2= qObj; qMin and qMedian are the smallest side length and the middle side length of three different side lengths of the positive bounding box of the target point cloud, and qObj is the largest diagonal length of the target point cloud bounding box.
10. The method according to claim 5, wherein in the step 8, when the point pairs in the scene are matched to the point pair features on the target point cloud, voting 1 is added to the corresponding matching point index and rotation angle { index; The voting of the pose adopts the neighborhood expansion, the voting is also carried out on the value of angle plus 1 on the basis of voting of angle, da represents the discretization resolution of angle, and the resolution is set to be 12 degrees, because the pose is mainly determined by angle, the voting result after the neighborhood expansion is reduced by the noise interference result.

Description

Point cloud matching method based on point-to-feature descriptors Technical Field The invention relates to the field of computer technology application, in particular to a point cloud matching method based on point-to-feature descriptors. Background The pose estimation technology of the target object in the scene point cloud has wide application in the fields of robot guidance, navigation, measurement and the like. Among the methods for realizing pose estimation of a target object, there are currently two main methods: The first method is to combine images and point clouds, train the images and the point clouds shot under different visual angles to form training data with marked pose, train to form a network, and utilize the network to identify in the identifying process. This approach is deployed in the industry for too long and has limited accuracy, relies on data sets, and therefore limits the scope of application. The second method is point cloud matching, most of the existing algorithms for multiple point cloud matching directly utilize feature descriptors established by the point cloud for matching, and the method has great limitation in the aspects of resisting the noise interference of the point cloud and matching speed under the condition of not optimizing. In view of the foregoing, it is desirable to provide a point cloud matching method based on point-to-feature descriptors. Disclosure of Invention In order to solve the technical problems, the invention provides a point cloud matching method based on a point pair feature descriptor, which adopts the matching of small neighborhood radius in the matching process, reduces the radius and the number of point pairs searched by each point, greatly improves the point cloud matching speed, and increases the robustness of the point cloud matching. The technical scheme of the invention is that the point cloud matching method based on the point-to-feature descriptors comprises the following steps: Step 1, adding a target point cloud, solving normal vectors of the target point cloud, and carrying out weighted average smoothing filtering on the normal vectors of the target point cloud; Step 2, screening the normal vector direction of the target point cloud based on the normal vector calculated in the target point cloud; Step 3, sampling the target point cloud, calculating a point pair characteristic value of a target point cloud sampling point, and carrying out neighborhood diffusion on the point pair characteristic value; Step 4, loading scene point clouds, performing normal vector calculation on the scene point clouds, adopting a neighborhood point local plane least square fitting method for normal vector of each point, extracting planes by using equations of local planes, adding an isolated point denoising algorithm in the process of calculating the normal vector of the scene point clouds to screen, and selecting neighbor number k and distance threshold d_threshold as parameters of isolated point denoising; Step 5, screening normal vectors of the scene point cloud according to the normal vector direction, and setting the normal vectors as Z-axis reverse filtering; step 6, sampling a sampling point algorithm of a scene point cloud by combining a distance and a normal vector included angle, wherein the threshold angle of the included angle defaults to 30 degrees and is used as an externally settable parameter; Step 7, selectively performing secondary matching in two radiuses of a small radius and a large radius in the characteristic matching process of point cloud scene point cloud and model point cloud point pairs, wherein specific values of the small radius and the large radius are set according to the size of a model; and 8, after the point-to-point feature matching is completed, voting the pose by adopting a neighborhood expansion strategy, and setting the length of the neighborhood to be 3. Further, in the step 1, the filtering method is used for filtering by using a neighborhood weighted average value of normal vector points, and weights of the neighborhood points in the neighborhood are distributed according to the distance from the neighborhood points to the normal vector points, and the closer the distance is, the larger the weights are; wherein the normal vector after filtering at any point p in the point cloud is expressed by the following formula: Wherein, the Is the normal vector before the point p filtering,Is the normal vector of the neighborhood point,Is the weight of the self point and can be set as a fixed constant, such as settingWeighting of neighborhood pointsSigma may be set at 3 times the resolution. Further, in the step 2, a Z-axis back filtering is set, and if the normal vector of the effective feature point in the target point cloud is an included angle smaller than 90 degrees with the Z-axis unit direction vector (0, 1), such feature point is effective, otherwise, it is ineffective. Further, in the step 3, the distance and the incl