Robot point cloud feature registration method and system and readable storage medium

Abstract

The invention discloses a robot point cloud feature registration method and system and a readable storage medium, and the method comprises the steps: collecting a scene point cloud, and carrying out the preprocessing of the scene point cloud, and obtaining a target point cloud; performing point cloud clustering segmentation based on the target point cloud and calculating a clustering center of gravity to obtain a feature point cloud; performing feature coarse registration based on the feature point cloud and a preset point cloud template to obtain a matched pose; and based on a preset algorithm, carrying out filtering fine registration on the matching pose to obtain a registration result. According to the method, the problems of rapid and accurate identification and positioning of objects with complex shapes in a robot automation system are solved, faster and more accurate point cloud registration is realized in a 3D disordered grabbing task, and the operation efficiency and reliability of the robot automation system are improved.

Description

Technical Field

[0001] The present invention relates to the field of robot operation technology, and more specifically, to a robot point cloud feature registration method, system and readable storage medium. Background Art

[0002] Point cloud registration is a key technology in 3D unordered grasping tasks. Robots rely on 3D vision technology to accurately obtain the position information of the object to be grasped, thereby ensuring the accuracy and reliability of the grasping operation.

[0003] Among them, point cloud registration is a key technology in this process. Most existing point cloud registration methods require accurate initial position estimation and are easily affected by noise and outliers, which makes them not robust enough in practical applications. Especially when faced with a large number of disorderly stacked objects, how to quickly and accurately find the target object from complex point cloud data and complete the estimation of its three-dimensional posture is a huge challenge. Summary of the Invention

[0004] The purpose of the present invention is to provide a robot point cloud feature registration method, system and readable storage medium, aiming to solve the problem of rapid and accurate recognition and positioning of complex-shaped objects in robotic automation systems.

[0005] A first aspect of the present invention provides a robot point cloud feature registration method, comprising the following steps:

[0006] Collecting scene point clouds and preprocessing the scene point clouds to obtain target point clouds;

[0007] Performing point cloud clustering segmentation based on the target point cloud and calculating the cluster centroid to obtain a feature point cloud;

[0008] Performing rough feature registration based on the feature point cloud and a preset point cloud template to obtain a matching pose;

[0009] The matching postures are filtered and precisely registered based on a preset algorithm to obtain a registration result.

[0010] In this solution, training to obtain the point cloud template specifically includes:

[0011] Obtaining the template point cloud of the scene where the object being measured is located;

[0012] A normal vector for each point in the point cloud is calculated based on the template point cloud;

[0013] The point cloud template is obtained by performing feature extraction and training of a point pair feature descriptor based on the normal vector and the template point cloud, wherein the point cloud template includes a point pair information feature database.

[0014] In this solution, the acquisition of scene point clouds and preprocessing of the scene point clouds to obtain target point clouds specifically include:

[0015] Collecting the scene point cloud based on a preset 3D camera;

[0016] The scene point cloud is subjected to interference removal preprocessing to obtain the target point cloud, and the processing method includes deleting interference points and deleting noise points, wherein,

[0017] Delete points outside the preset detection area, and delete background points within the preset detection area that do not belong to the object being detected;

[0018] The scene point cloud is filtered to remove noise points, wherein the spatial connection between adjacent objects is disconnected after the noise points are removed.

[0019] In this solution, performing point cloud clustering segmentation based on the target point cloud and calculating the cluster centroid to obtain the feature point cloud specifically includes:

[0020] The target point cloud is segmented based on a preset segmentation mechanism, wherein:

[0021] Acquire point cloud data of each object in the target point cloud based on AI segmentation;

[0022] Based on point cloud cluster segmentation, Euclidean clustering is performed using domain search and connected domain analysis in combination with the point cloud data to obtain multiple independent clusters;

[0023] The centroid of each independent cluster is calculated, and multiple centroids are combined into one point cloud to obtain the feature point cloud.

[0024] In this solution, the matching pose is obtained by performing rough feature registration based on the feature point cloud and the pre-trained template point cloud, specifically including:

[0025] Selecting a first reference point based on the feature point cloud;

[0026] Selecting a second reference point based on the template point cloud;

[0027] Performing rough feature registration based on the first reference point and the second reference point to obtain an initial matching result;

[0028] The matching posture is filtered based on the initial matching result to obtain the matching posture, wherein the filtering mechanism includes score filtering, angle constraint and distance constraint.

[0029] In this solution, filtering and precisely registering the matching poses based on a preset algorithm to obtain a registration result specifically includes:

[0030] The matching poses are filtered and precisely registered using a preset closest point iterative algorithm, wherein:

[0031] Performing scene transformation based on the matching posture to obtain a transformation center of gravity and a transformation point cloud, and performing precise registration based on the transformation center of gravity and the center of gravity point to obtain an object point cloud;

[0032] The registration result is obtained by performing precise registration on the transformed point cloud and the object point cloud.

[0033] A second aspect of the present invention further provides a robot point cloud feature registration system, comprising a memory and a processor, wherein the memory comprises a robot point cloud feature registration method program, and when the robot point cloud feature registration method program is executed by the processor, the following steps are implemented:

[0034] Collecting scene point clouds and preprocessing the scene point clouds to obtain target point clouds;

[0035] Performing point cloud clustering segmentation based on the target point cloud and calculating the cluster centroid to obtain a feature point cloud;

[0036] Performing rough feature registration based on the feature point cloud and a preset point cloud template to obtain a matching pose;

[0037] The matching postures are filtered and precisely registered based on a preset algorithm to obtain a registration result.

[0038] In this solution, training to obtain the point cloud template specifically includes:

[0039] Obtaining the template point cloud of the scene where the object being measured is located;

[0040] A normal vector for each point in the point cloud is calculated based on the template point cloud;

[0041] The point cloud template is obtained by performing feature extraction and training of a point pair feature descriptor based on the normal vector and the template point cloud, wherein the point cloud template includes a point pair information feature database.

[0042] In this solution, the acquisition of scene point clouds and preprocessing of the scene point clouds to obtain target point clouds specifically include:

[0043] Collecting the scene point cloud based on a preset 3D camera;

[0044] The scene point cloud is subjected to interference removal preprocessing to obtain the target point cloud, and the processing method includes deleting interference points and deleting noise points, wherein,

[0045] Delete points outside the preset detection area, and delete background points within the preset detection area that do not belong to the object being detected;

[0046] The scene point cloud is filtered to remove noise points, wherein the spatial connection between adjacent objects is disconnected after the noise points are removed.

[0047] In this solution, performing point cloud clustering segmentation based on the target point cloud and calculating the cluster centroid to obtain the feature point cloud specifically includes:

[0048] The target point cloud is segmented based on a preset segmentation mechanism, wherein:

[0049] Acquire point cloud data of each object in the target point cloud based on AI segmentation;

[0050] Based on point cloud cluster segmentation, Euclidean clustering is performed using domain search and connected domain analysis in combination with the point cloud data to obtain multiple independent clusters;

[0051] The centroid of each independent cluster is calculated, and multiple centroids are combined into one point cloud to obtain the feature point cloud.

[0052] In this solution, the matching pose is obtained by performing rough feature registration based on the feature point cloud and the pre-trained template point cloud, specifically including:

[0053] Selecting a first reference point based on the feature point cloud;

[0054] Selecting a second reference point based on the template point cloud;

[0055] Performing rough feature registration based on the first reference point and the second reference point to obtain an initial matching result;

[0056] The matching posture is filtered based on the initial matching result to obtain the matching posture, wherein the filtering mechanism includes score filtering, angle constraint and distance constraint.

[0057] In this solution, filtering and precisely registering the matching poses based on a preset algorithm to obtain a registration result specifically includes:

[0058] The matching poses are filtered and precisely registered using a preset closest point iterative algorithm, wherein:

[0059] Performing scene transformation based on the matching posture to obtain a transformation center of gravity and a transformation point cloud, and performing precise registration based on the transformation center of gravity and the center of gravity point to obtain an object point cloud;

[0060] The registration result is obtained by performing precise registration on the transformed point cloud and the object point cloud.

[0061] The third aspect of the present invention provides a computer-readable storage medium, which includes a robot point cloud feature registration method program for a machine. When the robot point cloud feature registration method program is executed by a processor, it implements the steps of a robot point cloud feature registration method as described in any one of the above items.

[0062] The present invention discloses a robot point cloud feature registration method, system, and readable storage medium. Through an improved point cloud feature registration method, the system reduces reliance on initial position estimation. Furthermore, through an improved discrete feature hash mapping method for point-to-feature pairings, the system improves the uniqueness and speed of feature mapping and enhances robustness. In particular, when faced with a large number of disorderly stacked objects, the present invention uses multiple filtering methods to assist in eliminating erroneous poses, thereby improving coarse registration accuracy and reducing computational complexity. Specific beneficial effects are as follows:

[0063] 1. Improved matching performance for flat or gently curved surfaces. This feature calculates contour points from projected images, and then uses the tangent vectors of these contour points as the normal vectors for the boundary points, improving matching performance for flat or gently curved surfaces.

[0064] 2. Improved the uniqueness and speed of feature mapping, especially by improving the PPF (Point Pair Feature) discrete feature hash mapping method;

[0065] 3. Flexible point cloud data acquisition methods: Two optional methods are available for obtaining point cloud data of individual objects in the scene point cloud, combined with AI segmentation or traditional point cloud segmentation algorithms, improving flexibility and applicability;

[0066] 4. Improved coarse registration accuracy. Multiple filtering methods are used to help eliminate incorrect poses, thereby improving coarse registration accuracy.

[0067] 5. Reduced computational effort. During the precise registration process, the cluster with the closest centroid distance to the template point cloud after searching and transformation is used as the object point cloud for precise registration. This reduces computational effort while eliminating interference from other objects on the precise registration. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 A flow chart of a robot point cloud feature registration method according to the present invention is shown;

[0069] Figure 2 A block diagram of a robot point cloud feature registration system according to the present invention is shown. DETAILED DESCRIPTION

[0070] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0071] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0072] Figure 1 A flow chart of a robot point cloud feature registration method of the present application is shown.

[0073] like Figure 1 As shown, the present application discloses a robot point cloud feature registration method, comprising the following steps:

[0074] S102, collecting scene point clouds, and preprocessing the scene point clouds to obtain target point clouds;

[0075] S104, performing point cloud clustering segmentation based on the target point cloud and calculating the cluster centroid to obtain a feature point cloud;

[0076] S106, performing feature rough registration based on the feature point cloud and a preset point cloud template to obtain a matching pose;

[0077] S108: Perform filtering and precise registration on the matching postures based on a preset algorithm to obtain a registration result.

[0078] It should be noted that, in this embodiment, a robot point cloud feature registration method of the present application is specifically implemented in two main parts when it comes to point cloud feature registration for 3D disordered grasping: offline template point cloud feature training and online scene point cloud feature matching. Among them, after collecting the online scene point cloud for pre-processing to obtain the target point cloud, clustering and segmentation are performed to obtain the feature point cloud, so that the matching pose can be obtained by coarsely aligning the feature point cloud of the online scene with the point cloud template obtained by the preset offline template point cloud feature training, and then the matching pose can be filtered and finely aligned using the preset ICP algorithm to obtain the registration result. Among them, the ICP (Iterative Closest Point) algorithm corresponds to the nearest point iterative algorithm, which is the most classic data registration algorithm.

[0079] According to an embodiment of the present invention, training to obtain the point cloud template specifically includes:

[0080] Obtaining the template point cloud of the scene where the object being measured is located;

[0081] A normal vector for each point in the point cloud is calculated based on the template point cloud;

[0082] The point cloud template is obtained by performing feature extraction and training of a point pair feature descriptor based on the normal vector and the template point cloud, wherein the point cloud template includes a point pair information feature database.

[0083] It should be noted that, in this embodiment, the template point cloud of the scene in which the object to be measured is located is first obtained, wherein the template point cloud is the basis for subsequent feature matching, and its quality directly affects the accuracy and reliability of the registration. Specifically, the template point cloud can be collected by a 3D camera (such as a structured light camera, a laser scanner, etc.). During the collection process, attention should be paid to maintaining a uniform sampling density and minimizing noise and occlusion to obtain high-quality point cloud data. In actual application, the object to be measured is placed in the center of the field of view, and the 3D camera collects the scene point cloud data containing the object to be measured, and separates the point cloud data belonging to the object to be measured from the scene point cloud as the template point cloud.

[0084] Furthermore, the normal vector of each point in the point cloud is calculated based on the template point cloud. In this embodiment, two modes of point cloud templates are described, one is the direct surface point cloud as a template, and the other is the point cloud boundary contour as a template. The former is suitable for objects with complex shapes, and the latter is suitable for objects with flat surfaces or gentle changes. Specifically, when using the surface point cloud as a template, the main processing process is to remove noise, reduce the point cloud density, and calculate the normal vector of each point in the point cloud.

[0085] Among them, when the point cloud boundary contour is used as a template, the main processing process is to extract the boundary points and calculate the tangent vector of each boundary point as the normal vector of the point. The specific steps are as follows: (1) The template point cloud is projected into a height image according to the height range, and invalid points are filled with black pixels; (2) The height image is used to extract the contour, and morphological operations (such as expansion and erosion) are used to remove small contours. For the same contour that is incorrectly divided into multiple parts, the distance between the parts is calculated. If the distance between the two parts is less than a preset threshold, they are considered to belong to the same contour; (3) For each contour in the image, the three-dimensional point corresponding to the contour is extracted, and the tangent vector of each contour point is calculated: the center point of the bounding box of the contour point pi is calculated. Select a global reference vector (such as the Z axis) n f ;, used to determine the directional consistency of the tangent vector; for each contour point, calculate the neighboring points within the neighborhood range, calculate the center of gravity and main direction of the neighborhood points, and select the eigenvector corresponding to the maximum eigenvalue as the tangent vector of the contour point Correct the direction of the tangent vector and calculate the vector from the center point to the contour point Calculating vectors and the tangent vector The dot product of the two vectors is used to obtain the direction vector perpendicular to the plane where the two vectors are located. If the vector is inconsistent with the reference direction, the direction of the tangent vector is adjusted. The formula is as follows: but Among them, n f is the global reference vector in the Z-axis direction, is the vector from the center point to the contour point, is the tangent vector of the contour point.

[0086] Furthermore, based on the normal vector and the template point cloud, feature extraction and training of the point pair feature descriptor are performed to obtain the point cloud template, wherein the point cloud template includes a point pair information feature database, wherein, after obtaining the template point cloud and calculating the normal vector (or tangent vector) of each point, the template point cloud is subjected to PPF (Point Pitch Correction) PairFeature (point pair feature descriptor) feature extraction and training to provide a reliable reference for subsequent online matching. The specific steps are as follows: (1) Point pair feature (PPF) calculation: For each point pair (p1, p2) in the template point cloud, calculate its point pair feature (PPF); (2) Feature discretization: In order to make the feature values consistent at different scales, the point pair features are discretized; (3) Feature index construction: In order to speed up the online matching process, an index structure hash table is established for the features of the template point cloud - it can quickly find similar features and significantly improve the matching efficiency; (4) Feature library generation: The PPF features of all template point clouds and their corresponding point pair information are stored in a feature library for use in online matching. Among them, the feature library supports efficient query operations to adapt to different application scenarios.

[0087] It should be noted that PPF (Point Pair Feature) is a local feature descriptor that is expressed by analyzing the geometric relationship between a pair of points and is rotation and translation invariant. For any two points p1 and p2 in the point cloud, and are the normal vectors corresponding to p1 and p2 respectively, and their PPF features are calculated as follows:

[0088]

[0089] The PPF feature between two points consists of four parameters, which describe the geometric relationship between the point pairs. Specifically, It is expressed as the straight-line distance between two points, that is, Represented as the normal vector of point p1 With vector The angle between Represented as the normal vector of point p2 With vector The angle between Normal vectors of points p1 and p2 and The angle between them; the above four parameters can be combined into a four-dimensional vector F(p1,p2) to represent the geometric characteristics of the point pair (p1,p2).

[0090] Furthermore, to discretize the eigenvalues and generate a hash table, it is necessary to calculate the index value of the hash table. The specific method of point-to-point feature discretization is as follows:

[0091]

[0092] Among them, it is necessary to set Δ dist and Δ Angle Two step size parameters, Δ dist Represents the distance discrete parameter, with a value range of [0, 10], Δ Angle It is an angle discrete parameter with a value range of [1, 360] degrees. Each element of the eigenvalue F(p1, p2) is divided by the step size and rounded to the integer to obtain the discrete feature four-dimensional vector H(p1, p2) = H(h1, h2, h3, h4). Then, the feature index is constructed to generate the feature library.

[0093] Furthermore, it should be noted that when constructing the PPF feature index, the discrete feature vector H(p1, p2) is used to generate the key value of the hash table, and the point pair information is stored under the corresponding key value in the hash table. When generating the hash key value, the key is to ensure that the hash function has good uniform distribution, consistency and efficiency, while avoiding conflicts. By mapping each component of the four-dimensional vector H(h1, h2, h3, h4) into a fixed-length string, and then concatenating these strings to form the final hash key value, the specific implementation steps are as follows:

[0094] (1) Define the string format of each component: For the first component h1, convert it to a string of length 5, and fill the insufficient part with leading zeros; for the second to fourth components h2, h3, h4, convert them to strings of length 4 respectively, and also fill them with leading zeros to ensure fixed length.

[0095] (2) String concatenation: connect the strings of the above four components in sequence to form a complete hash key value. For example, for the four-dimensional vector (1,2,3,4), its corresponding hash key value is "00001000200030004", which has the following characteristics: fixed length, the string length of each component is fixed, ensuring the consistency of the length of the hash key value, which is convenient for storage and comparison; easy to implement, which can be achieved through simple string formatting operations, with clear logic, easy to understand and maintain; scalability, the string length of each component can be adjusted as needed to adapt to different application scenarios; and high efficiency, due to the use of fixed-length string representation, it can maintain high efficiency in large-scale data processing; and range restriction, in order to ensure the validity of the string representation, the input integer should be within a reasonable range. It should be noted that the maximum value of h1 should not exceed "99999", and the maximum values of h2, h3, and h4 should not exceed "9999".

[0096] According to an embodiment of the present invention, collecting a scene point cloud and preprocessing the scene point cloud to obtain a target point cloud specifically includes:

[0097] Collecting the scene point cloud based on a preset 3D camera;

[0098] The scene point cloud is subjected to interference removal preprocessing to obtain the target point cloud, and the processing method includes deleting interference points and deleting noise points, wherein,

[0099] Delete points outside the preset detection area, and delete background points within the preset detection area that do not belong to the object being detected;

[0100] The scene point cloud is filtered to remove noise points, wherein the spatial connection between adjacent objects is disconnected after the noise points are removed.

[0101] It should be noted that, in this embodiment, scene point cloud data is collected based on a preset 3D camera, and the scene point cloud is preprocessed for interference removal to obtain the target point cloud, wherein points outside the preset detection area are deleted, and background points within the preset detection area that do not belong to the object being measured are deleted, and the preset detection area includes a rectangular detection area, and then the scene point cloud is filtered to remove noise points, wherein the spatial connection between adjacent objects is disconnected after the noise points are removed.

[0102] According to an embodiment of the present invention, performing point cloud clustering segmentation based on the target point cloud and calculating the cluster centroid to obtain a feature point cloud specifically includes:

[0103] The target point cloud is segmented based on a preset segmentation mechanism, wherein:

[0104] Acquire point cloud data of each object in the target point cloud based on AI segmentation;

[0105] Based on point cloud cluster segmentation, Euclidean clustering is performed using domain search and connected domain analysis in combination with the point cloud data to obtain multiple independent clusters;

[0106] The centroid of each independent cluster is calculated, and multiple centroids are combined into one point cloud to obtain the feature point cloud.

[0107] It should be noted that, in this embodiment, point cloud segmentation to extract point cloud data of a single object specifically includes AI segmentation and point cloud clustering segmentation. The core of AI segmentation is to use a deep learning model to classify each pixel in the image and mark it as background or belonging to a specific object. The main steps include: data labeling, data enhancement, model training, reasoning and post-processing, and object area extraction. Specifically, when performing AI segmentation, the pixel information of the area where each object is located is identified, and then the point cloud data of each object in the target point cloud is obtained based on AI segmentation in the scene point cloud.

[0108] Furthermore, when performing point cloud clustering segmentation, by setting an appropriate distance threshold and using neighborhood search and connected region analysis, Euclidean clustering can separate relatively separated objects from the point cloud and generate multiple independent clusters; secondly, the centroid of each independent cluster is calculated and recorded as It is used to accurately align and quickly locate scene object point cloud data, reduce the amount of calculation, and eliminate interference; finally, multiple cluster centroid points are merged into one point cloud data for subsequent feature matching, that is, the feature point cloud is obtained based on the merging of multiple centroid points into one point cloud.

[0109] According to an embodiment of the present invention, performing coarse feature registration based on the feature point cloud and the pre-trained template point cloud to obtain a matching pose specifically includes:

[0110] Selecting a first reference point based on the feature point cloud;

[0111] Selecting a second reference point based on the template point cloud;

[0112] Performing rough feature registration based on the first reference point and the second reference point to obtain an initial matching result;

[0113] The matching posture is filtered based on the initial matching result to obtain the matching posture, wherein the filtering mechanism includes score filtering, angle constraint and distance constraint.

[0114] It should be noted that, in this embodiment, the feature point cloud of the real-time scene and the pre-trained template point cloud are feature matched. First, a first reference point is selected based on the feature point cloud, and a second reference point is selected based on the template point cloud, so that a rough feature alignment can be performed based on the first reference point and the second reference point to obtain an initial matching result, and then matching posture filtering is performed based on the initial matching result to obtain the matching posture.

[0115] Specifically, the steps to obtain the matching pose are as follows:

[0116] Step 1: Feature point cloud (denoted as S n ) uniformly select a portion of points as the first reference point for feature matching, generally select "5%-20%" points as reference points, denoted as S r ={s r ∈S n |r=0,1,2…,N r}, where N r is the quantity value of the first reference point;

[0117] Step 2: Uniformly select a portion of points s from the template point cloud i ∈S n As the second reference point, denoted as S i ={s i ∈S n |i=0,1,2…,N s}, where N s is the quantity value of the second reference point;

[0118] Step 3: For S r Each reference point s in r , perform the following operations:

[0119] Reference point r and S i Each point in the form a point pair s r s i , according to the point cloud feature calculation method described in the offline modeling part, calculate the point pair s r s i PPF characteristics F(s r ,s i ) and characteristic discrete vector H(s r ,s i ), using the discrete eigenvector H(s r ,s i ) into the key value of the hash table;

[0120] Search and s in the model feature library r s i The point pair combination with the same feature hash key value (mj ,m k ), listed as s r s i A set of corresponding points of the feature. If there are multiple sets of point pairs with the same key value in the model, a fixed number of point pairs are taken to obtain multiple sets of corresponding points. A pose can be calculated between each set of corresponding points to represent the transformation relationship between the template point cloud and the scene point cloud (also called matching pose). All transformation relationships are mapped into the pose space;

[0121] Reference point r and S i Each point in the form a point pair s r s i , a total of N s Each point pair searches for corresponding points in the template and maps them into the pose space. The pose with the most counts in the pose space is taken as the reference point s r The final candidate pose of N r Group;

[0122] Step 4: According to Step 3, the final candidate pose of each reference point is obtained. Similar poses may represent the matching results of the same object and template. The pose clustering method is used to merge similar poses and finally obtain matching results that are suspected to belong to multiple objects.

[0123] Step 5: Perform matching pose filtering. The filtering mechanism includes score filtering, angle constraint and distance constraint. Specifically, score filtering corresponds to: according to Step 4, obtain the matching poses of multiple objects, transform the template point cloud according to the matching results to coincide with the object point cloud, calculate the matching score, and delete the matching poses with lower scores in descending order; angle constraint corresponds to: if the template point cloud is flipped during the transformation, delete the pose; distance constraint corresponds to: sort the poses by matching score, and then transform the template point cloud center of gravity according to the matching pose to obtain multiple sets of transformed template point cloud center of gravity points, traverse the transformed template point cloud center of gravity points, delete the matching poses corresponding to other center of gravity points closer to the current center of gravity point, until the traversal is completed, and retain the output of the remaining matching poses.

[0124] Step 6: Output the filtered coarse registration result, i.e. the matching pose.

[0125] According to an embodiment of the present invention, filtering and precisely registering the matching poses based on a preset algorithm to obtain a registration result specifically includes:

[0126] The matching poses are filtered and precisely registered using a preset closest point iterative algorithm, wherein:

[0127] Performing scene transformation based on the matching posture to obtain a transformation center of gravity and a transformation point cloud, and performing precise registration based on the transformation center of gravity and the center of gravity point to obtain an object point cloud;

[0128] The registration result is obtained by performing precise registration on the transformed point cloud and the object point cloud.

[0129] It should be noted that, in this embodiment, in order to further improve the registration accuracy, the ICP algorithm is used to optimize the coarse registration result, and finally obtain a more accurate matching result. The coarse registration scores multiple coarse registration matching poses, and for each pose, the fine registration matching result is calculated. The specific steps are as follows:

[0130] S1: Based on the matching pose, the scene is transformed to obtain a transformation center of gravity and a transformation point cloud. Specifically, the template point cloud is transformed according to the coarse registration result until it roughly coincides with an object in the scene, and the center of gravity of the transformed template point cloud is calculated;

[0131] S2: Based on the transformation center of gravity and the center of gravity point, the object point cloud is obtained by precise registration. Specifically, The cluster with the closest centroid distance to the template point cloud after searching and transformation is used as the object point cloud for precise registration;

[0132] S3: Finally, the registration result is obtained based on the precise registration of the transformed point cloud and the object point cloud. Specifically, the transformed point cloud and the object point cloud are precisely registered, which can reduce the amount of calculation while eliminating the influence of other objects on the precise registration effect, and output the precise registration result.

[0133] Figure 2 A block diagram of a robot point cloud feature registration system according to the present invention is shown.

[0134] like Figure 2 As shown, the present invention discloses a robot point cloud feature registration system, including a memory and a processor. The memory includes a robot point cloud feature registration method program. When the robot point cloud feature registration method program is executed by the processor, the following steps are implemented:

[0135] Collecting scene point clouds and preprocessing the scene point clouds to obtain target point clouds;

[0136] Performing point cloud clustering segmentation based on the target point cloud and calculating the cluster centroid to obtain a feature point cloud;

[0137] Performing rough feature registration based on the feature point cloud and a preset point cloud template to obtain a matching pose;

[0138] The matching postures are filtered and precisely registered based on a preset algorithm to obtain a registration result.

[0139] It should be noted that, in this embodiment, a robot point cloud feature registration method of the present application is specifically implemented in two main parts when it comes to point cloud feature registration for 3D disordered grasping: offline template point cloud feature training and online scene point cloud feature matching. Among them, after collecting the online scene point cloud for pre-processing to obtain the target point cloud, clustering and segmentation are performed to obtain the feature point cloud, so that the matching pose can be obtained by coarsely aligning the feature point cloud of the online scene with the point cloud template obtained by the preset offline template point cloud feature training, and then the matching pose can be filtered and finely aligned using the preset ICP algorithm to obtain the registration result. Among them, the ICP (Iterative Closest Point) algorithm corresponds to the nearest point iterative algorithm, which is the most classic data registration algorithm.

[0140] According to an embodiment of the present invention, training to obtain the point cloud template specifically includes:

[0141] Obtaining the template point cloud of the scene where the object being measured is located;

[0142] A normal vector for each point in the point cloud is calculated based on the template point cloud;

[0143] The point cloud template is obtained by performing feature extraction and training of a point pair feature descriptor based on the normal vector and the template point cloud, wherein the point cloud template includes a point pair information feature database.

[0144] It should be noted that, in this embodiment, the template point cloud of the scene in which the object to be measured is located is first obtained, wherein the template point cloud is the basis for subsequent feature matching, and its quality directly affects the accuracy and reliability of the registration. Specifically, the template point cloud can be collected by a 3D camera (such as a structured light camera, a laser scanner, etc.). During the collection process, attention should be paid to maintaining a uniform sampling density and minimizing noise and occlusion to obtain high-quality point cloud data. In actual application, the object to be measured is placed in the center of the field of view, and the 3D camera collects the scene point cloud data containing the object to be measured, and separates the point cloud data belonging to the object to be measured from the scene point cloud as the template point cloud.

[0145] Furthermore, the normal vector of each point in the point cloud is calculated based on the template point cloud. In this embodiment, two modes of point cloud templates are described, one is the direct surface point cloud as a template, and the other is the point cloud boundary contour as a template. The former is suitable for objects with complex shapes, and the latter is suitable for objects with flat surfaces or gentle changes. Specifically, when using the surface point cloud as a template, the main processing process is to remove noise, reduce the point cloud density, and calculate the normal vector of each point in the point cloud.

[0146] Among them, when the point cloud boundary contour is used as a template, the main processing process is to extract the boundary points and calculate the tangent vector of each boundary point as the normal vector of the point. The specific steps are as follows: (1) The template point cloud is projected into a height image according to the height range, and invalid points are filled with black pixels; (2) The height image is used to extract the contour, and morphological operations (such as expansion and erosion) are used to remove small contours. For the same contour that is incorrectly divided into multiple parts, the distance between the parts is calculated. If the distance between the two parts is less than a preset threshold, they are considered to belong to the same contour; (3) For each contour in the image, the three-dimensional point corresponding to the contour is extracted, and the tangent vector of each contour point is calculated: the center point of the bounding box of the contour point pi is calculated. Select a global reference vector (such as the Z axis) n f ;, used to determine the directional consistency of the tangent vector; for each contour point, calculate the neighboring points within the neighborhood range, calculate the center of gravity and main direction of the neighborhood points, and select the eigenvector corresponding to the maximum eigenvalue as the tangent vector of the contour point Correct the direction of the tangent vector and calculate the vector from the center point to the contour point Calculating vectors and the tangent vector The dot product of the two vectors is used to obtain the direction vector perpendicular to the plane where the two vectors are located. If the vector is inconsistent with the reference direction, the direction of the tangent vector is adjusted. The formula is as follows: but Among them, n f is the global reference vector in the Z-axis direction, is the vector from the center point to the contour point, is the tangent vector of the contour point.

[0147] Furthermore, based on the normal vector and the template point cloud, feature extraction and training of the point pair feature descriptor are performed to obtain the point cloud template, wherein the point cloud template includes a point pair information feature database, wherein, after obtaining the template point cloud and calculating the normal vector (or tangent vector) of each point, the template point cloud is subjected to PPF (Point Pitch Correction) PairFeature (point pair feature descriptor) feature extraction and training to provide a reliable reference for subsequent online matching. The specific steps are as follows: (1) Point pair feature (PPF) calculation: For each point pair (p1, p2) in the template point cloud, calculate its point pair feature (PPF); (2) Feature discretization: In order to make the feature values consistent at different scales, the point pair features are discretized; (3) Feature index construction: In order to speed up the online matching process, an index structure hash table is established for the features of the template point cloud - it can quickly find similar features and significantly improve the matching efficiency; (4) Feature library generation: The PPF features of all template point clouds and their corresponding point pair information are stored in a feature library for use in online matching. Among them, the feature library supports efficient query operations to adapt to different application scenarios.

[0148] It should be noted that PPF (Point Pair Feature) is a local feature descriptor that is expressed by analyzing the geometric relationship between a pair of points and is rotation and translation invariant. For any two points p1 and p2 in the point cloud, and are the normal vectors corresponding to p1 and p2 respectively, and their PPF features are calculated as follows:

[0149]

[0150] The PPF feature between two points consists of four parameters, which describe the geometric relationship between the point pairs. Specifically, It is expressed as the straight-line distance between two points, that is, Represented as the normal vector of point p1 With vector The angle between Represented as the normal vector of point p2 With vector The angle between Normal vectors of points p1 and p2 and The angle between them; the above four parameters can be combined into a four-dimensional vector F(p1,p2) to represent the geometric characteristics of the point pair (p1,p2).

[0151] Furthermore, to discretize the eigenvalues and generate a hash table, it is necessary to calculate the index value of the hash table. The specific method of point-to-point feature discretization is as follows:

[0152]

[0153] Among them, it is necessary to set Δ dist and Δ Angle Two step size parameters, Δ dist Represents the distance discrete parameter, with a value range of [0, 10], Δ Angle It is an angle discrete parameter with a value range of [1, 360] degrees. Each element of the eigenvalue F(p1, p2) is divided by the step size and rounded to the integer to obtain the discrete feature four-dimensional vector H(p1, p2) = H(h1, h2, h3, h4). Then, the feature index is constructed to generate the feature library.

[0154] Furthermore, it should be noted that when constructing the PPF feature index, the discrete feature vector H(p1, p2) is used to generate the key value of the hash table, and the point pair information is stored under the corresponding key value in the hash table. When generating the hash key value, the key is to ensure that the hash function has good uniform distribution, consistency and efficiency, while avoiding conflicts. By mapping each component of the four-dimensional vector H(h1, h2, h3, h4) into a fixed-length string, and then concatenating these strings to form the final hash key value, the specific implementation steps are as follows:

[0155] (1) Define the string format of each component: For the first component h1, convert it to a string of length 5, and fill the insufficient part with leading zeros; for the second to fourth components h2, h3, h4, convert them to strings of length 4 respectively, and also fill them with leading zeros to ensure fixed length.

[0156] (2) String concatenation: connect the strings of the above four components in sequence to form a complete hash key value. For example, for the four-dimensional vector (1,2,3,4), its corresponding hash key value is "00001000200030004", which has the following characteristics: fixed length, the string length of each component is fixed, ensuring the consistency of the length of the hash key value, which is convenient for storage and comparison; easy to implement, which can be achieved through simple string formatting operations, with clear logic, easy to understand and maintain; scalability, the string length of each component can be adjusted as needed to adapt to different application scenarios; and high efficiency, due to the use of fixed-length string representation, it can maintain high efficiency in large-scale data processing; and range restriction, in order to ensure the validity of the string representation, the input integer should be within a reasonable range. It should be noted that the maximum value of h1 should not exceed "99999", and the maximum values of h2, h3, and h4 should not exceed "9999".

[0157] According to an embodiment of the present invention, collecting a scene point cloud and preprocessing the scene point cloud to obtain a target point cloud specifically includes:

[0158] Collecting the scene point cloud based on a preset 3D camera;

[0159] The scene point cloud is subjected to interference removal preprocessing to obtain the target point cloud, and the processing method includes deleting interference points and deleting noise points, wherein,

[0160] Delete points outside the preset detection area, and delete background points within the preset detection area that do not belong to the object being detected;

[0161] The scene point cloud is filtered to remove noise points, wherein the spatial connection between adjacent objects is disconnected after the noise points are removed.

[0162] It should be noted that, in this embodiment, scene point cloud data is collected based on a preset 3D camera, and the scene point cloud is preprocessed for interference removal to obtain the target point cloud, wherein points outside the preset detection area are deleted, and background points within the preset detection area that do not belong to the object being measured are deleted, and the preset detection area includes a rectangular detection area, and then the scene point cloud is filtered to remove noise points, wherein the spatial connection between adjacent objects is disconnected after the noise points are removed.

[0163] According to an embodiment of the present invention, performing point cloud clustering segmentation based on the target point cloud and calculating the cluster centroid to obtain a feature point cloud specifically includes:

[0164] The target point cloud is segmented based on a preset segmentation mechanism, wherein:

[0165] Acquire point cloud data of each object in the target point cloud based on AI segmentation;

[0166] Based on point cloud cluster segmentation, Euclidean clustering is performed using domain search and connected domain analysis in combination with the point cloud data to obtain multiple independent clusters;

[0167] The centroid of each independent cluster is calculated, and multiple centroids are combined into one point cloud to obtain the feature point cloud.

[0168] It should be noted that, in this embodiment, point cloud segmentation to extract point cloud data of a single object specifically includes AI segmentation and point cloud clustering segmentation. The core of AI segmentation is to use a deep learning model to classify each pixel in the image and mark it as background or belonging to a specific object. The main steps include: data labeling, data enhancement, model training, reasoning and post-processing, and object area extraction. Specifically, when performing AI segmentation, the pixel information of the area where each object is located is identified, and then the point cloud data of each object in the target point cloud is obtained based on AI segmentation in the scene point cloud.

[0169] Furthermore, when performing point cloud clustering segmentation, by setting an appropriate distance threshold and using neighborhood search and connected region analysis, Euclidean clustering can separate relatively separated objects from the point cloud and generate multiple independent clusters; secondly, the centroid of each independent cluster is calculated and recorded as It is used to accurately align and quickly locate scene object point cloud data, reduce the amount of calculation, and eliminate interference; finally, multiple cluster centroid points are merged into one point cloud data for subsequent feature matching, that is, the feature point cloud is obtained based on the merging of multiple centroid points into one point cloud.

[0170] According to an embodiment of the present invention, performing coarse feature registration based on the feature point cloud and the pre-trained template point cloud to obtain a matching pose specifically includes:

[0171] Selecting a first reference point based on the feature point cloud;

[0172] Selecting a second reference point based on the template point cloud;

[0173] Performing rough feature registration based on the first reference point and the second reference point to obtain an initial matching result;

[0174] The matching posture is filtered based on the initial matching result to obtain the matching posture, wherein the filtering mechanism includes score filtering, angle constraint and distance constraint.

[0175] It should be noted that, in this embodiment, the feature point cloud of the real-time scene and the pre-trained template point cloud are feature matched. First, a first reference point is selected based on the feature point cloud, and a second reference point is selected based on the template point cloud, so that a rough feature alignment can be performed based on the first reference point and the second reference point to obtain an initial matching result, and then matching posture filtering is performed based on the initial matching result to obtain the matching posture.

[0176] Specifically, the steps to obtain the matching pose are as follows:

[0177] Step 1: Feature point cloud (denoted as S n) uniformly select a portion of points as the first reference point for feature matching, generally select "5%-20%" points as reference points, denoted as S r ={s r ∈S n |r=0,1,2…,N r}, where N r is the quantity value of the first reference point;

[0178] Step 2: Uniformly select a portion of points s from the template point cloud i ∈S n As the second reference point, denoted as S i ={s i ∈S n |i=0,1,2…,N s}, where N s is the quantity value of the second reference point;

[0179] Step 3: For S r Each reference point s in r , perform the following operations:

[0180] Reference point r and S i Each point in the form a point pair s r s i , according to the point cloud feature calculation method described in the offline modeling part, calculate the point pair s r s i PPF characteristics F(s r ,s i ) and characteristic discrete vector H(s r ,s i ), using the discrete eigenvector H(s r ,s i ) into the key value of the hash table;

[0181] Search and s in the model feature library r s i The point pair combination with the same feature hash key value (m j ,m k ), listed as s r s i A set of corresponding points of the feature. If there are multiple sets of point pairs with the same key value in the model, a fixed number of point pairs are taken to obtain multiple sets of corresponding points. A pose can be calculated between each set of corresponding points to represent the transformation relationship between the template point cloud and the scene point cloud (also called matching pose). All transformation relationships are mapped into the pose space;

[0182] Reference point r and S i Each point in the form a point pair sr s i , a total of N s Each point pair searches for corresponding points in the template and maps them into the pose space. The pose with the most counts in the pose space is taken as the reference point s r The final candidate pose of N r Group;

[0183] Step 4: According to Step 3, the final candidate pose of each reference point is obtained. Similar poses may represent the matching results of the same object and template. The pose clustering method is used to merge similar poses and finally obtain matching results that are suspected to belong to multiple objects.

[0184] Step 5: Perform matching pose filtering. The filtering mechanism includes score filtering, angle constraint and distance constraint. Specifically, score filtering corresponds to: according to Step 4, obtain the matching poses of multiple objects, transform the template point cloud according to the matching results to coincide with the object point cloud, calculate the matching score, and delete the matching poses with lower scores in descending order; angle constraint corresponds to: if the template point cloud is flipped during the transformation, delete the pose; distance constraint corresponds to: sort the poses by matching score, and then transform the template point cloud center of gravity according to the matching pose to obtain multiple sets of transformed template point cloud center of gravity points, traverse the transformed template point cloud center of gravity points, delete the matching poses corresponding to other center of gravity points closer to the current center of gravity point, until the traversal is completed, and retain the output of the remaining matching poses.

[0185] Step 6: Output the filtered coarse registration result, i.e. the matching pose.

[0186] According to an embodiment of the present invention, filtering and precisely registering the matching poses based on a preset algorithm to obtain a registration result specifically includes:

[0187] The matching poses are filtered and precisely registered using a preset closest point iterative algorithm, wherein:

[0188] Performing scene transformation based on the matching posture to obtain a transformation center of gravity and a transformation point cloud, and performing precise registration based on the transformation center of gravity and the center of gravity point to obtain an object point cloud;

[0189] The registration result is obtained by performing precise registration on the transformed point cloud and the object point cloud.

[0190] It should be noted that, in this embodiment, in order to further improve the registration accuracy, the ICP algorithm is used to optimize the coarse registration result, and finally obtain a more accurate matching result. The coarse registration scores multiple coarse registration matching poses, and for each pose, the fine registration matching result is calculated. The specific steps are as follows:

[0191] S1: Based on the matching pose, the scene is transformed to obtain a transformation center of gravity and a transformation point cloud. Specifically, the template point cloud is transformed according to the coarse registration result until it roughly coincides with an object in the scene, and the center of gravity of the transformed template point cloud is calculated;

[0192] S2: Based on the transformation center of gravity and the center of gravity point, the object point cloud is obtained by precise registration. Specifically, The cluster with the closest centroid distance to the template point cloud after searching and transformation is used as the object point cloud for precise registration;

[0193] S3: Finally, the registration result is obtained based on the precise registration of the transformed point cloud and the object point cloud. Specifically, the transformed point cloud and the object point cloud are precisely registered, which can reduce the amount of calculation while eliminating the influence of other objects on the precise registration effect, and output the precise registration result.

[0194] The third aspect of the present invention provides a computer-readable storage medium, which includes a robot point cloud feature registration method program. When the robot point cloud feature registration method program is executed by a processor, it implements the steps of a robot point cloud feature registration method as described in any one of the above items.

[0195] The present invention discloses a robot point cloud feature registration method, system and readable storage medium. Through an improved point cloud feature registration method, the dependence on initial position estimation is reduced. Through an improved discrete feature hash mapping method of point-to-feature, the uniqueness and speed of feature mapping are improved, and the robustness is enhanced. Especially when faced with a large number of disorderly stacked objects, the present invention uses multiple filtering methods to assist in eliminating erroneous postures, thereby improving the coarse registration accuracy and reducing the amount of calculation.

[0196] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0197] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0198] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0199] Those skilled in the art will appreciate that all or part of the steps of the above-mentioned method embodiments may be implemented by hardware associated with program instructions, and the aforementioned program may be stored in a computer-readable storage medium. When the program is executed, the program executes the steps of the above-mentioned method embodiments. The aforementioned storage medium includes various media that can store program codes, such as mobile storage devices, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0200] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

Claims

1. A robot point cloud feature registration method, characterized in that: The following steps are involved: Collecting scene point clouds and preprocessing the scene point clouds to obtain target point clouds; Performing point cloud clustering segmentation based on the target point cloud and calculating the cluster centroid to obtain a feature point cloud; Performing rough feature registration based on the feature point cloud and a preset point cloud template to obtain a matching pose; The matching postures are filtered and precisely registered based on a preset algorithm to obtain a registration result.

2. A robot point cloud feature registration method according to claim 1, characterized in that: The training to obtain the point cloud template specifically includes: Obtaining the template point cloud of the scene where the object being measured is located; A normal vector for each point in the point cloud is calculated based on the template point cloud; The point cloud template is obtained by performing feature extraction and training of a point pair feature descriptor based on the normal vector and the template point cloud, wherein the point cloud template includes a point pair information feature database.

3. A robot point cloud feature registration method according to claim 2, characterized in that: The collecting of scene point clouds and preprocessing of the scene point clouds to obtain target point clouds specifically include: Collecting the scene point cloud based on a preset 3D camera; The scene point cloud is subjected to interference removal preprocessing to obtain the target point cloud, and the processing method includes deleting interference points and deleting noise points, wherein, Delete points outside the preset detection area, and delete background points within the preset detection area that do not belong to the object being detected; The scene point cloud is filtered to remove noise points, wherein the spatial connection between adjacent objects is disconnected after the noise points are removed.

4. A robot point cloud feature registration method according to claim 3, characterized in that: The point cloud is clustered and segmented based on the target point cloud and the cluster centroid is calculated to obtain a feature point cloud, specifically comprising: Performing point cloud segmentation on the target point cloud based on a preset segmentation mechanism, wherein point cloud data of each object in the target point cloud is obtained based on AI segmentation; Based on point cloud cluster segmentation, Euclidean clustering is performed using domain search and connected domain analysis in combination with the point cloud data to obtain multiple independent clusters; The centroid of each independent cluster is calculated, and multiple centroids are combined into one point cloud to obtain the feature point cloud.

5. A robot point cloud feature registration method according to claim 4, characterized in that: The step of performing rough feature registration based on the feature point cloud and the pre-trained template point cloud to obtain a matching pose specifically includes: Selecting a first reference point based on the feature point cloud; Selecting a second reference point based on the template point cloud; Performing rough feature registration based on the first reference point and the second reference point to obtain an initial matching result; The matching posture is filtered based on the initial matching result to obtain the matching posture, wherein the filtering mechanism includes score filtering, angle constraint and distance constraint.

6. A robot point cloud feature registration method according to claim 5, characterized in that: The filtering and precise registration of the matching postures based on a preset algorithm to obtain a registration result specifically includes: The matching poses are filtered and precisely registered using a preset closest point iterative algorithm, wherein: Performing scene transformation based on the matching posture to obtain a transformation center of gravity and a transformation point cloud, and performing precise registration based on the transformation center of gravity and the center of gravity point to obtain an object point cloud; The registration result is obtained by performing precise registration on the transformed point cloud and the object point cloud.

7. A robot point cloud feature registration system, characterized in that: The system comprises a memory and a processor, wherein the memory comprises a robot point cloud feature registration method program, and when the robot point cloud feature registration method program is executed by the processor, the following steps are implemented: Collecting scene point clouds and preprocessing the scene point clouds to obtain target point clouds; Performing point cloud clustering segmentation based on the target point cloud and calculating the cluster centroid to obtain a feature point cloud; Performing rough feature registration based on the feature point cloud and a preset point cloud template to obtain a matching pose; The matching postures are filtered and precisely registered based on a preset algorithm to obtain a registration result.

8. The robot point cloud feature registration system according to claim 7, characterized in that: The training to obtain the point cloud template specifically includes: Obtaining the template point cloud of the scene where the object being measured is located; A normal vector for each point in the point cloud is calculated based on the template point cloud; The point cloud template is obtained by performing feature extraction and training of a point pair feature descriptor based on the normal vector and the template point cloud, wherein the point cloud template includes a point pair information feature database.

9. The robot point cloud feature registration system according to claim 8, characterized in that: The collecting of scene point clouds and preprocessing of the scene point clouds to obtain target point clouds specifically include: Collecting the scene point cloud based on a preset 3D camera; The scene point cloud is subjected to interference removal preprocessing to obtain the target point cloud, and the processing method includes deleting interference points and deleting noise points, wherein, Delete points outside the preset detection area, and delete background points within the preset detection area that do not belong to the object being detected; The scene point cloud is filtered to remove noise points, wherein the spatial connection between adjacent objects is disconnected after the noise points are removed.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a robot point cloud feature registration method program. When the robot point cloud feature registration method program is executed by a processor, the steps of the robot point cloud feature registration method according to any one of claims 1 to 6 are implemented.

Images