Point cloud registration method and system
By preprocessing, edge feature extraction and clustering of point clouds, combining feature association algorithms and similarity scoring functions, the robustness and accuracy of point cloud registration are achieved, and the problems of poor robustness and low computing efficiency in the existing technology are solved.
Patent Information
- Application Number
- CN202410668277.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-28
- Publication Date
- 2025-03-28
AI Technical Summary
The existing point cloud registration methods are poorly robust and have low computational efficiency, resulting in lower accuracy and efficiency of the precise registration process.
A point cloud registration method is proposed, including obtaining preprocessing of source point clouds and initial point clouds, edge point feature extraction, clustering and construction of triangle descriptors, feature association algorithms and similarity scoring functions for coarse registration and precision registration.
It improves the robustness and accuracy of point cloud registration, reduces the computational complexity, and is suitable for point cloud registration in complex scenarios, especially in the application scenarios of cylindrical filling ports.
Smart Images
Figure CN119722755A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of image processing, and specifically designs a point cloud registration method and a system thereof. Background Art
[0002] With the rapid development of machine vision, processing point cloud data can produce good image effects. Point cloud is a massive set of points that express the spatial distribution and surface characteristics of the target in the same spatial reference system. The points in the point cloud can be expressed by coordinates.
[0003] Point cloud registration is to align point cloud data taken from different perspectives or at different times so that they represent the same scene or object in the same coordinate system. The main goal of registration is to find the best transformation between two or more point clouds to minimize the difference between them. The registration process is usually divided into two stages: coarse registration and fine registration. Coarse registration is a relatively rough registration when the transformation between the two point clouds is completely unknown. Its main purpose is to provide a good initial value for fine registration. Coarse registration provides a good starting point for fine registration so that subsequent iterative precise matching is more effective.
[0004] The existing point cloud registration methods have poor robustness and low computational efficiency, which reduces the accuracy and efficiency of the fine registration process. Therefore, the industry is in urgent need of a new point cloud registration technology to improve the robustness and registration accuracy, so as to facilitate the registration of point clouds in complex scenes. Summary of the invention
[0005] The present invention provides a point cloud registration method and system thereof to solve the technical problem that the existing point cloud registration method has poor robustness and low computational efficiency, which makes the accuracy and efficiency of the precise registration process low. The point cloud registration method provided by the present invention is mainly applicable to cylindrical filling ports, which facilitates accurate positioning and realizes safe and efficient filling tasks.
[0006] To achieve the above object, the present invention provides a point cloud registration method, the method comprising:
[0007] Obtaining a source point cloud and an initial point cloud; wherein the application scenario of the point cloud registration is a cylindrical filling port;
[0008] Preprocessing the initial point cloud to obtain a target point cloud with optimal density;
[0009] Extract edge point features from the target point cloud with the optimal density to obtain an edge feature point set;
[0010] Clustering the edge feature point set to obtain a cluster set;
[0011] Constructing a triangle descriptor of the cluster set according to the stability of the triangle;
[0012] Pairing the triangle descriptors according to a feature association algorithm and a similarity scoring function to obtain a descriptor pair set, and roughly registering the descriptor pair set to obtain a target point cloud after the rough registration;
[0013] According to the target point cloud after rough registration, the least squares method is used to fit the precise registration transformation matrix;
[0014] The precise registration of the source point cloud and the target point cloud is achieved according to the precise registration transformation matrix.
[0015] Furthermore, the initial point cloud is preprocessed to obtain a target point cloud with optimal density, including:
[0016] Obtaining the approximate three-dimensional pose of the cylindrical filling port;
[0017] Segmenting the initial point cloud according to the cylindrical physical model of the filling port and the approximate coordinates to obtain local point cloud data containing only the filling port;
[0018] The local point cloud data is filtered according to an adaptive filtering threshold function to obtain target point cloud data with optimal density.
[0019] Furthermore, obtaining a rough three-dimensional position of the cylindrical filling port includes:
[0020] Get the two-dimensional pixel coordinates of the center point of the cylindrical filling port;
[0021] According to the depth map collected by the depth camera, the pixel value of the two-dimensional pixel coordinates of the center point of the filling port is obtained, and the depth value of the center of the filling port of the tank truck is obtained through the affine function;
[0022] According to the depth value, the approximate three-dimensional position and posture of the cylindrical filling port in the camera coordinate system is obtained using the camera three-dimensional projection model.
[0023] Further, edge point feature extraction is performed on the target point cloud data of the optimal density to obtain an edge feature point set; including:
[0024] The five points before and after each point in the target point cloud data with the optimal density are selected as the basis, the curvature of each point in the target point cloud data with the optimal density is calculated, and the edge feature point set is screened out according to the smoothness threshold.
[0025] Furthermore, clustering the edge feature point set to obtain a cluster set; including:
[0026] Construct the functional relationship between the clustering threshold and the optimal voxel filtering parameters,
[0027] Clustering the edge feature point set using a Euclidean clustering algorithm to obtain a cluster set;
[0028] Among them, the functional relationship between the clustering threshold and the optimal voxel filtering parameter is:
[0029] Threshold cluster =5σ filter ;
[0030] Threshold cluster The clustering threshold is expressed as, σ filter is the optimal voxel filtering parameter.
[0031] Further, according to the stability of the triangle, a triangle descriptor of the cluster set is constructed, including:
[0032] Get the coordinates of the center point of each cluster, get the direction vector of each cluster through PCA principal component analysis, project the coordinates of the center point of each cluster onto the direction vector, get the interval between two adjacent points, and get the length value of each cluster;
[0033] Traverse the cluster set, search for the five nearest neighbor clusters of each cluster, traverse all the pairwise non-repeating combination clusters from the nearest neighbor clusters and use the cluster as the vertex to build a stable triangle descriptor, and at the same time remove the cluster from the cluster set.
[0034] Further, the triangle descriptors are paired according to a feature association algorithm and a similarity scoring function to obtain a descriptor pair set, and the descriptor pair set is roughly aligned to obtain a target descriptor pair and a rough transformation pose of the target descriptor pair, including:
[0035] In the first phase, the descriptor set is traversed, and a hash table is constructed with the sum of the side lengths of the descriptors as the key value and the triangle descriptor as the target value;
[0036] Traverse the length clustering set of the target point cloud, query the number of triangle descriptors whose length difference is within the interval in the hash table of the source point cloud, and get the potential descriptor pair set. The number is the first-stage score of each triangle descriptor, and sort the scores from large to small to get the score set and the total descriptor pair set;
[0037] According to the total number of triangle descriptors, the top 20% of the descriptors are retained to enter the second stage for feature association, and the other descriptors are removed from the set to obtain a new set of descriptor pairs;
[0038] Traverse the new set of descriptor pairs, and for a set of paired descriptors, subtract the vertex length information, direction vector information, and three edge length information of the two to obtain the vertex length difference, direction vector difference, and edge length difference;
[0039] For each paired descriptor, determine whether it meets the condition that the three differences are all less than the threshold. If not, remove it. Finally, retain the descriptor pair corresponding to the minimum sum of the three differences in each paired descriptor set;
[0040] Traverse the set of retained descriptor pairs, calculate the pose transformation between each pair of descriptors, perform pose transformation on the remaining descriptor pairs, and calculate whether the distance between the three vertices of the transformed descriptor is within the threshold range. If the threshold requirement is met, the score is increased by 1, and finally a score collection is obtained. The descriptor pair corresponding to the maximum score is the target descriptor pair, and the corresponding pose transformation is the target point cloud after coarse alignment.
[0041] Further, according to the target point cloud after the rough registration, a least squares method is used to fit the fine registration transformation matrix; according to the fine registration transformation matrix, the fine registration of the source point cloud and the target point cloud is realized, including:
[0042] The least square relationship between the target point cloud and the source point cloud is established according to the target point cloud after rough registration;
[0043] Solving the least squares relationship using the Gauss-Newton method to obtain a precise registration transformation matrix;
[0044] The precise registration transformation matrix is multiplied with the source point cloud to achieve precise registration of the source point cloud and the target point cloud.
[0045] The present invention also provides a point cloud registration system, the system comprising:
[0046] A data acquisition module, used to acquire source point cloud and initial point cloud;
[0047] A data processing module, used for preprocessing the initial point cloud to obtain a target point cloud with optimal density;
[0048] The first registration module is used to extract edge point features from the target point cloud with the optimal density to obtain an edge feature point set; cluster the edge feature point set to obtain a cluster set; construct a triangle descriptor of the cluster set according to the stability of the triangle; pair the triangle descriptors according to a feature association algorithm and a similarity scoring function to obtain a descriptor pair set, and roughly register the descriptor pair set to obtain a target point cloud after the rough registration;
[0049] The second registration module is used to obtain a precise registration transformation matrix by least squares fitting according to the target point cloud after the rough registration; and realize precise registration of the source point cloud and the target point cloud according to the precise registration transformation matrix.
[0050] Furthermore, the data processing module includes:
[0051] A segmentation processing module, used to obtain the approximate three-dimensional position and posture of the cylindrical filling port, and segment the initial point cloud according to the cylindrical physical model of the filling port and the approximate coordinates to obtain local point cloud data containing only the filling port;
[0052] The filtering processing module is used to filter the local point cloud data according to an adaptive filtering threshold function to obtain target point cloud data with optimal density.
[0053] The present invention has the following beneficial effects:
[0054] (1) The present invention provides a point cloud registration method, which is mainly applicable to cylindrical filling ports. First, the approximate three-dimensional position and posture of the cylindrical filling port is obtained, and then the initial point cloud is segmented according to the cylindrical physical model of the filling port and the approximate coordinates to obtain local point cloud data containing only the filling port, so as to reduce the interference caused by redundant point clouds; at the same time, the local point cloud data is filtered according to an adaptive filtering threshold function to obtain target point cloud data with optimal density, thereby avoiding the adjustment of filtering parameters and achieving optimal processing of the target point cloud, thereby speeding up the 3D point cloud processing speed while ensuring accuracy.
[0055] (2) The present invention provides a point cloud registration method. First, edge point features are extracted from target point cloud data with optimal density based on point cloud smoothness. Five points before and after each point in the target point cloud data with optimal density are selected as a basis. The curvature of each point in the target point cloud data with optimal density is calculated. A set of edge feature points is screened out according to a smoothness threshold. Edge points with greater information are used to perform point cloud registration tasks. Then, a functional relationship between a clustering threshold and an optimal voxel filtering parameter is constructed, and the set of edge feature points is clustered using a Euclidean clustering algorithm to obtain a cluster set. A triangle descriptor of the cluster set is constructed with the aid of the stability concept of a triangle, and the descriptor has spatial and temporal invariance. At the same time, the pairing of triangle descriptors according to a two-stage feature association algorithm and a similarity scoring function can not only accelerate the pairing process of the descriptors, but also improve the success rate of pairing. The least squares method is further used to achieve precise registration, thereby improving the accuracy of the final registration.
[0056] (3) The present invention provides a point cloud registration method, which is suitable for cylindrical filling ports and can be widely used in the field of automatic filling technology, facilitating accurate positioning and achieving safe and efficient filling tasks. At the same time, the method of the present invention can make the filling port positioning fast, accurate, and robust, and can greatly improve the safety of the filling process.
[0057] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:
[0059] Figure 1 A flow chart of a point cloud registration method provided by the present invention;
[0060] Figure 2 A schematic structural diagram of an embodiment of a point cloud registration system provided by the present invention. DETAILED DESCRIPTION
[0061] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0062] like Figure 1 As shown, the present invention provides a point cloud registration method, comprising the following steps:
[0063] Step 101, obtaining a source point cloud and an initial point cloud; wherein the application scenario of the point cloud registration is a cylindrical filling port.
[0064] Specifically, a depth camera or a laser radar may be used to obtain image data of a target area in real time, wherein the target area includes the cylindrical filling port and the environment information around it.
[0065] Step 102: preprocess the initial point cloud to obtain a target point cloud with optimal density. Specifically, it includes:
[0066] Obtaining the approximate three-dimensional pose of the cylindrical filling port;
[0067] Segmenting the initial point cloud according to the cylindrical physical model of the filling port and the approximate coordinates to obtain local point cloud data containing only the filling port;
[0068] The local point cloud data is filtered according to an adaptive filtering threshold function to obtain target point cloud data with optimal density.
[0069] The approximate three-dimensional position of the cylindrical filling port can be directly obtained by using a laser radar, or by using a depth camera. The depth camera is used to obtain the position, specifically including:
[0070] Get the two-dimensional pixel coordinates of the center point of the cylindrical filling port;
[0071] According to the depth map collected by the depth camera, the pixel value of the two-dimensional pixel coordinates of the center point of the filling port is obtained, and the depth value of the center of the filling port of the tank truck is obtained through the affine function;
[0072] According to the depth value, the approximate three-dimensional position and posture of the cylindrical filling port in the camera coordinate system is obtained using the camera three-dimensional projection model.
[0073] Specifically, the depth camera is used to obtain the image data of the filling port in real time, including the surrounding environment information and potential risk factors of the filling port. First, the acquired image data is preprocessed and feature extracted. Among them, the acquired image is preprocessed, including noise removal, missing value filling, and image noise reduction. Then, the image data of different resolutions are combined and processed by multi-scale information fusion technology to obtain a training data set; then the training data set is input into the yolov5 model based on temporal information learning and attention mechanism for training to obtain the yolo visual detection model; then the environmental image of the filling port is input into the yolo visual detection model for target detection and feature extraction, and the two-dimensional pixel coordinates of the center point of the filling port are output; finally, the two-dimensional pixel coordinates of the center point are converted into the approximate three-dimensional pose of the filling port.
[0074] Among them, temporal information learning and attention mechanism are added to the yolov5 model training process to learn and pay attention to the visual features of different time steps, so as to improve the recognition speed and cope with factors such as light changes and interference from debris, so as to improve the robustness and adaptability of the filling port detection system; the specific implementation process is as follows:
[0075] 1) In the training data, the temporal information of the image data is extracted by means of timestamps or sequence numbers, so as to introduce temporal learning in the model training process.
[0076] 2) Design a suitable time series feature representation method to combine time series information with image features to form an input data format suitable for model learning.
[0077] 3) Based on the time series characteristics, a time series model structure is designed, and LSTM is used as the model structure for time series data to achieve learning and representation of visual features at different time steps.
[0078] 4) Introducing the attention mechanism into the timing model enables the model to dynamically focus on and learn important information at different time steps, so as to improve the adaptability of the filling port detection system to factors such as light changes and interference from debris.
[0079] 5) Model training and optimization: Based on the designed timing model and attention mechanism, the training data is trained, and the YOLOv5 model parameters and structure are continuously optimized to achieve accurate recognition and positioning of the filling port.
[0080] 6) Performance evaluation and adjustment: Perform performance evaluation on the trained YOLO visual inspection model, adjust and optimize the model based on the evaluation results to ensure the robustness and adaptability of the filling port inspection system under different environmental conditions.
[0081] Specifically, the pixel coordinates of the center of the filling port output from the yolo visual inspection model are center (u0,v0), get the Coordinate of the center of the filling port according to the depth map collected by the camera center The pixel value at the tanker is obtained by applying an affine function to obtain the depth value z at the center of the tanker filling port. center , using the camera 3D projection model to obtain the approximate 3D pose T in the camera coordinate system center (x center ,y center ,z center ), the relevant formula is as follows:
[0082]
[0083] Among them, c x ,c y ,f x ,f y is the camera internal parameter.
[0084] According to the cylindrical physical model of the filling port, the physical parameters of the filling port {radius: r; cylindrical axis vector: (a center ,b center ,c center ); Cylinder height: H center} and the filling port contour model model. The specific implementation process is:
[0085] 1) Select a number of points on the cylindrical surface and fit the plane formed by these points A , and then get the unit normal vector n of the plane A , which is the normal vector of the point. Perform this operation for each point on the cylinder to obtain the unit normal vector of each point. Treat the unit normal vector of each point as a point P0, and fit these points again to obtain the plane normal vector n0, which is the initial value of the cylinder axis vector (a0, b0, c0);
[0086] 2) After obtaining the axis, perform coordinate transformation on the cylinder, transform the cylinder axis vector (a0, b0, c0) into a vector parallel to the Z axis of the camera coordinate system, so that the (x, y) coordinates of the points on the cylindrical surface form a plane circle. Then fit these points to obtain the center Center0, i.e. (x1, y1, z1) and radius r.
[0087] 3) The distance d from the cylindrical surface point to the cylindrical axis is always the radius. Therefore, according to the above initial value, the error equation f can be established to solve the filling port fitting model. The specific formula is as follows:
[0088]
[0089] Then, in the pose T center The initial point cloud is segmented by fitting the model model as the contour to obtain a local point cloud Pointcloud containing only the methanol filling port local ; According to the depth value z center Design an adaptive filter threshold function to calculate the optimal voxel filter parameter σ filter , the relevant calculation formula is as follows:
[0090]
[0091] Among them, aim is the ideal accuracy.
[0092] Finally, according to σ filter For local point cloud Pointcloud local Perform voxel downsampling to obtain the target point cloud with optimal density. target .
[0093] Step 103: extract edge point features from the target point cloud with the optimal density to obtain an edge feature point set; cluster the edge feature point set to obtain a cluster set; and construct a triangle descriptor of the cluster set based on the stability of the triangle.
[0094] Specifically: First, select the target point cloud Pointcloud target The 5 points before and after each point in the target point cloud are used as the basis to calculate the target point cloud. target The curv of each point in is calculated according to the smoothness threshold Threshold smooth Filter out the edge feature point set Edge{Point1,Point2,...,Point n}, the specific calculation formula is as follows:
[0095]
[0096] Among them, S represents the set of 5 points before and after.
[0097] Then, construct the clustering threshold Threshold cluster and the optimal voxel filter parameter σ filter The functional relationship between them is used to obtain the cluster set Clusters {cluster1(Point a ,Point b ,...), cluster2(Point a ,Point b ,...),...,cluster m (Point a ,Point b ,...)}, the specific clustering threshold calculation formula is as follows:
[0098] Threshold cluster =5σ filter ;
[0099] Then, for the above cluster sets, calculate the coordinates of their respective center points (x center_cluster ,y center_cluster ), and obtain its main direction through PCA principal component analysis That is, the direction vector of each cluster. Project the points in each cluster point cloud onto the main direction and calculate the distance between the two farthest points to get the length value of each cluster. clusters {length1,length2,…,length m}, the specific calculation formula is as follows:
[0100]
[0101] Among them, Num represents the number of points in the cluster.
[0102] Finally, traverse the cluster set Clusters and search for each cluster cluster i The 5 nearest neighbor clusters near {cluster near_1 ,…,cluster near_5}, to traverse all pairwise non-repeating combinations of clusters and clusters from the nearest neighbor clusters i Construct a stable triangle descriptor Tri_desc for the vertex i {tri_desc1,tri_desc2,…,tri_desc l}, and Clustersi Remove the set, each triangle descriptor tri_desc i Contains the length information of three vertices {length_desc1, length_desc2, length_desc3} and direction vector information And three side length information {length_side1, length_side2, length_side3}.
[0103] Step 104: Pair the triangle descriptors according to the feature association algorithm and the similarity scoring function to obtain a set of descriptor pairs, and roughly align the set of descriptor pairs to obtain a target point cloud after rough alignment. Specifically:
[0104] 1) The first stage traverses the descriptor set Tri_desc of the source point cloud i , to describe the sub-tri_desc i The sum of the side lengths sum_side i Key value, triangle descriptor tri_desc i Build a hash table Hash for the target value desc {(sum_side1:tri_desc1),(sum_side2:tri_desc2),…,(sum_side l :tri_desc l )};
[0105] 2) Traverse the length clustering set Sum_side_group of the target point cloud target , for the triangle descriptor tri_desc i_target , in the hash table Hash of the source point cloud desc_source The query length difference is in the interval (sum_side i_target -Threshold hash ,Sum_side i_target +Threshold hash ) The number of triangle descriptors within sum i , and get tri_desc i_target Potential descriptor pair set Target i {(tri_desc i_target ,tri_desc a_source ),(tri_desc i_target ,tri_desc b_source ),…}, quantity sum i That is, the first stage grade of each triangle descriptor i, and sort the scores from large to small to get the score set Grade{grade max ,…,grade min} and the total descriptor pair set Match sum {Match1,Match2,…,Match l};
[0106] 3) According to the total number of triangle descriptors, the top 20% of the descriptors are retained to enter the second stage for feature association, and the other descriptors are removed from the set to obtain Match sum {Match1,Match2,...,Match k}, where k = 0.2*l;
[0107] 4) Traverse Match sum , for a pair of descriptors (tri_desc i_target ,tri_desc a_source ), the length information of the two vertices {length_desc1, length_desc2, length_desc3}, the direction vector information And the three side length information {length_side1, length_side2, length_side3} are subtracted to get the vertex length difference Δ desc , direction vector difference Δ n And the side length difference Δ side ;
[0108] 5) For each paired descriptor (tri_desc i_target ,tri_desc a_source ) to determine whether the three differences are all less than the threshold {Threshold desc Thereshold n ,Threshold side} condition, if it is not met, it will be removed, and finally each pairing descriptor set Match will be retained i The sum of the three differences Δ sum The descriptor pair corresponding to the minimum value (tri_desc min_target ,tri_desc min_source ), Δ sum The calculation formula is as follows:
[0109] Δ sum =Δ desc +Δ n +Δ side ;
[0110] 6) Traverse the set of retained descriptor pairs and calculate the pose transformation T between each pair of descriptors ij (R ij ,t ij ), for the remaining descriptors Tri_desc remain Conduct T ij Transformation, calculate whether the distance between the three vertices of the transformed descriptor is within the threshold Threshold dis If the threshold requirement is met, the vote is increased by 1, and the final score collection vote is obtained. sum {vote1,vote2,...,vote h}, the descriptor pair corresponding to the maximum score vote is the target descriptor pair (tri_desc tar_target ,tri_desc tar_source ), and the corresponding pose transformation T ij This is the target point cloud after rough registration.
[0111] Step 105: According to the target point cloud after the rough registration, the least square method is used to fit the fine registration transformation matrix; and the fine registration of the source point cloud and the target point cloud is achieved according to the fine registration transformation matrix. Specifically:
[0112] First, create the target point cloud Pointcloud after rough registration t ` arget With the source point cloud Pointcloud source The least squares relationship between them is expressed as follows:
[0113]
[0114] Then, the Gauss-Newton method is used to solve the least squares function to obtain the precise registration transformation matrix T'(R', t'), and precise registration of the source point cloud and the target point cloud is achieved according to the precise registration transformation matrix.
[0115] Among them, the conversion relationship between the source point cloud and the target point cloud is:
[0116] T target =T`T source
[0117] Among them, T target is the target point cloud, T source is the source point cloud.
[0118] The present invention provides a point cloud registration method, and its application scenario is mainly suitable for cylindrical filling ports. First, the approximate three-dimensional posture of the cylindrical filling port is obtained, and then the initial point cloud is segmented according to the cylindrical physical model of the filling port and the approximate coordinates to obtain local point cloud data containing only the filling port, so as to reduce the interference caused by redundant point clouds; at the same time, the local point cloud data is filtered according to an adaptive filtering threshold function to obtain target point cloud data with optimal density, thereby avoiding the adjustment of filtering parameters, achieving optimal processing of the target point cloud, and accelerating the 3D point cloud processing speed while ensuring accuracy.
[0119] The present invention provides a point cloud registration method. Firstly, edge point features are extracted from target point cloud data with optimal density based on point cloud smoothness, 5 points before and after each point in the target point cloud data with optimal density are selected as a basis, the curvature of each point in the target point cloud data with optimal density is calculated, and a set of edge feature points is screened out according to a smoothness threshold, and edge points with greater information are used to perform point cloud registration tasks; then, a functional relationship between a clustering threshold and an optimal voxel filtering parameter is constructed, and a Euclidean clustering algorithm is used to cluster the edge feature point set to obtain a cluster set; and with the aid of the stability concept of a triangle, a triangle descriptor of the cluster set is constructed, and the descriptor has spatial and temporal invariance; at the same time, according to a two-stage feature association algorithm and a similarity scoring function, the triangle descriptors are paired, which not only accelerates the pairing process of the descriptors, but also improves the success rate of pairing, and further adopts a least squares method to achieve precise registration, thereby improving the accuracy of the final registration.
[0120] Based on the same inventive concept, the present invention also proposes a point cloud registration system, which can be implemented by the above point cloud registration method, so the repeated parts will not be repeated. Figure 2 In one embodiment, the point cloud registration system may include a data acquisition module 100 , a data processing module 200 , a first registration module 300 and a second registration module 400 .
[0121] The data acquisition module 100 is used to acquire the source point cloud and the initial point cloud.
[0122] The data processing module 200 is used to pre-process the initial point cloud to obtain a target point cloud with optimal density.
[0123] The first registration module 300 is used to extract edge point features of the target point cloud with optimal density to obtain an edge feature point set; cluster the edge feature point set to obtain a cluster set; construct a triangle descriptor of the cluster set according to the stability of the triangle; pair the triangle descriptors according to a feature association algorithm and a similarity scoring function to obtain a descriptor pair set, and coarsely register the descriptor pair set to obtain a target point cloud after coarse registration.
[0124] The second registration module 400 is used to obtain a precise registration transformation matrix by least square fitting according to the target point cloud after the rough registration; and to realize precise registration of the source point cloud and the target point cloud according to the precise registration transformation matrix.
[0125] Furthermore, the data processing module includes:
[0126] A segmentation processing module, used to obtain the approximate three-dimensional position and posture of the cylindrical filling port, and segment the initial point cloud according to the cylindrical physical model of the filling port and the approximate coordinates to obtain local point cloud data containing only the filling port;
[0127] The filtering processing module is used to filter the local point cloud data according to an adaptive filtering threshold function to obtain target point cloud data with optimal density.
[0128] The present invention provides a point cloud registration method and system, which is suitable for cylindrical filling ports and can be widely used in the field of automatic filling technology, facilitating accurate positioning and achieving safe and efficient filling tasks. At the same time, the method of the present invention can make the filling port positioning fast, accurate, and robust, and can greatly improve the safety of the filling process.
[0129] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.
Claims
1. A point cloud registration method, characterized in that: The method comprises: Obtaining a source point cloud and an initial point cloud; wherein the application scenario of the point cloud registration is a cylindrical filling port; Preprocessing the initial point cloud to obtain a target point cloud with optimal density; Extract edge point features from the target point cloud with the optimal density to obtain an edge feature point set; Clustering the edge feature point set to obtain a cluster set; Constructing a triangle descriptor of the cluster set according to the stability of the triangle; Pairing the triangle descriptors according to a feature association algorithm and a similarity scoring function to obtain a descriptor pair set, and roughly registering the descriptor pair set to obtain a target point cloud after the rough registration; According to the target point cloud after rough registration, the least squares method is used to fit the precise registration transformation matrix; The precise registration of the source point cloud and the target point cloud is achieved according to the precise registration transformation matrix.
2. A point cloud registration method according to claim 1, characterized in that: Preprocessing the initial point cloud to obtain a target point cloud with optimal density includes: Obtaining the approximate three-dimensional pose of the cylindrical filling port; Segmenting the initial point cloud according to the cylindrical physical model of the filling port and the approximate coordinates to obtain local point cloud data containing only the filling port; The local point cloud data is filtered according to an adaptive filtering threshold function to obtain target point cloud data with optimal density.
3. A point cloud registration method according to claim 2, characterized in that: Get the approximate 3D pose of the cylindrical filling port; including: Get the two-dimensional pixel coordinates of the center point of the cylindrical filling port; According to the depth map collected by the depth camera, the pixel value of the two-dimensional pixel coordinates of the center point of the filling port is obtained, and the depth value of the center of the filling port of the tank truck is obtained through the affine function; According to the depth value, the approximate three-dimensional position and posture of the cylindrical filling port in the camera coordinate system is obtained using the camera three-dimensional projection model.
4. A point cloud registration method according to claim 1, characterized in that: Extracting edge point features from the target point cloud data of the optimal density to obtain an edge feature point set; comprising: The five points before and after each point in the target point cloud data with the optimal density are selected as the basis, the curvature of each point in the target point cloud data with the optimal density is calculated, and the edge feature point set is screened out according to the smoothness threshold.
5. A point cloud registration method according to claim 1, characterized in that: Clustering the edge feature point set to obtain a cluster set; comprising: Construct the functional relationship between the clustering threshold and the optimal voxel filtering parameters, The edge feature point set is clustered using a Euclidean clustering algorithm to obtain a cluster set.
6. A point cloud registration method according to claim 1, characterized in that: According to the stability of the triangle, a triangle descriptor of the cluster set is constructed, including: Get the coordinates of the center point of each cluster, get the direction vector of each cluster through PCA principal component analysis, project the coordinates of the center point of each cluster onto the direction vector, get the interval between two adjacent points, and get the length value of each cluster; Traverse the cluster set, search for the five nearest neighbor clusters of each cluster, traverse all the pairwise non-repeating combination clusters from the nearest neighbor clusters and use the cluster as the vertex to build a stable triangle descriptor, and at the same time remove the cluster from the cluster set.
7. A point cloud registration method according to claim 1, characterized in that: The triangle descriptors are paired according to a feature association algorithm and a similarity scoring function to obtain a descriptor pair set, and the descriptor pair set is roughly aligned to obtain a target descriptor pair and a rough transformation pose of the target descriptor pair, including: In the first phase, the descriptor set is traversed, and a hash table is constructed with the sum of the side lengths of the descriptors as the key value and the triangle descriptor as the target value; Traverse the length clustering set of the target point cloud, query the number of triangle descriptors whose length difference is within the interval in the hash table of the source point cloud, and get the potential descriptor pair set. The number is the first-stage score of each triangle descriptor, and sort the scores from large to small to get the score set and the total descriptor pair set; According to the total number of triangle descriptors, the top 20% of the descriptors are retained to enter the second stage for feature association, and the other descriptors are removed from the set to obtain a new set of descriptor pairs; Traverse the new set of descriptor pairs, and for a set of paired descriptors, subtract the vertex length information, direction vector information, and three edge length information of the two to obtain the vertex length difference, direction vector difference, and edge length difference; For each paired descriptor, determine whether it meets the condition that the three differences are all less than the threshold. If not, remove it. Finally, retain the descriptor pair corresponding to the minimum sum of the three differences in each paired descriptor set; Traverse the set of retained descriptor pairs, calculate the pose transformation between each pair of descriptors, perform pose transformation on the remaining descriptor pairs, and calculate whether the distance between the three vertices of the transformed descriptor is within the threshold range. If the threshold requirement is met, the score is increased by 1, and finally a score collection is obtained. The descriptor pair corresponding to the maximum score is the target descriptor pair, and the corresponding pose transformation is the target point cloud after coarse alignment.
8. A point cloud registration method according to claim 1, characterized in that: According to the target point cloud after the rough registration, a least square method is used to fit the fine registration transformation matrix; according to the fine registration transformation matrix, the fine registration of the source point cloud and the target point cloud is realized, including: According to the target point cloud after rough registration, the target point cloud is transformed in position and pose, and the least squares relationship between the transformed target point cloud and the source point cloud is established; Solving the least squares relationship using the Gauss-Newton method to obtain a precise registration transformation matrix; The precise registration transformation matrix is multiplied with the source point cloud to achieve precise registration of the source point cloud and the target point cloud.
9. A point cloud registration system, characterized in that: The system comprises: A data acquisition module, used to acquire source point cloud and initial point cloud; A data processing module, used for preprocessing the initial point cloud to obtain a target point cloud with optimal density; The first registration module is used to extract edge point features from the target point cloud with the optimal density to obtain an edge feature point set; cluster the edge feature point set to obtain a cluster set; construct a triangle descriptor of the cluster set according to the stability of the triangle; pair the triangle descriptors according to a feature association algorithm and a similarity scoring function to obtain a descriptor pair set, and roughly register the descriptor pair set to obtain a target point cloud after the rough registration; The second registration module is used to obtain a precise registration transformation matrix by least squares fitting according to the target point cloud after the rough registration; and realize precise registration of the source point cloud and the target point cloud according to the precise registration transformation matrix.
10. A point cloud registration system according to claim 9, characterized in that: The data processing module comprises: A segmentation processing module, used to obtain the approximate three-dimensional position and posture of the cylindrical filling port, and segment the initial point cloud according to the cylindrical physical model of the filling port and the approximate coordinates to obtain local point cloud data containing only the filling port; The filtering processing module is used to filter the local point cloud data according to an adaptive filtering threshold function to obtain target point cloud data with optimal density.