Intelligent mobile robot laser repositioning method based on dense point cloud registration

Abstract

The invention belongs to the field of intelligent mobile robot relocation, and particularly relates to an intelligent mobile robot laser relocation method based on dense point cloud registration. The problem that an existing intelligent mobile robot is low in starting positioning precision in an application scene is solved. According to the invention, a ScanContext technology is used to realize point cloud retrieval, and a point cloud registration network (AGHL) based on mixed local feature extraction and anchor point guidance is provided. The network provides a mixed local feature extraction module which is used for integrating local geometric features and local semantic features of the point cloud. Besides, an anchor point guided cross attention mechanism is designed in the network, noise and interference of irrelevant areas are reduced, the accuracy of point cloud registration is improved, and the problem that the intelligent mobile robot is low in start-up positioning precision in an application scene is solved.

Description

Technical Field

[0001] The invention belongs to the field of intelligent mobile robot repositioning, and in particular relates to an intelligent mobile robot laser repositioning method based on dense point cloud registration. Background Art

[0002] With the rapid development of society, intelligent mobile robots play an increasingly critical role in daily life and are widely used in restaurant delivery, exhibition hall interactive services and many other aspects. LiDAR-based global relocalization technology is a key technology to measure the intelligence level of mobile robots, which directly affects the positioning accuracy of the robot after it is powered on and restarted. In this technology, the point cloud registration algorithm plays a core role. It is used to build the geometric relationship between the current position of the robot and the global map, which greatly affects the global relocalization accuracy of the mobile robot. However, when the overlapping area between the two sets of point clouds is small or the viewing angle changes drastically, the existing point cloud registration algorithm often fails, resulting in the failure of global relocalization of the intelligent mobile robot. Therefore, how to improve the existing point cloud registration algorithm and improve the relocalization accuracy of the intelligent mobile robot has become a research hotspot.

[0003] At present, the commonly used point cloud registration algorithms are divided into direct registration methods and matching-based registration methods. Among them, the direct registration method directly estimates the rigid transformation between two point clouds without establishing an explicit correspondence. The traditional iterative closest point (ICP) algorithm and its variants achieve the registration effect by minimizing the distance between two point clouds. However, a large number of experiments have shown that this type of method is sensitive to the initial pose and fails for point cloud pairs with large viewpoint changes and low overlap. The matching-based registration method establishes a matching relationship between two point clouds, and then uses the SVD algorithm to solve the relative pose. CofiNet introduces a two-stage matching framework from coarse to fine. It first establishes a matching relationship between super points and then expands the matching space to the entire point cloud. GeoTransformer improves the robustness of point cloud registration to rotation changes by introducing spatial distance and angle information of point clouds. RoITr constructs a new encoder-decoder architecture and designs two parallel geometric and global context information extraction branches to generate discriminative features, thereby generating accurate point cloud matching results. REGTR abandons the traditional paradigm of generating matching results based on the attention matrix, and instead directly predicts the coordinates of the matching points in the point cloud based on the input query point cloud and the retrieval point cloud. Therefore, the existing intelligent mobile robots have the problem of low startup positioning accuracy in application scenarios. Summary of the invention

[0004] The purpose of the present invention is to solve the problem of low restart positioning accuracy of existing intelligent mobile robots in application scenarios. A laser repositioning method for intelligent mobile robots based on dense point cloud registration is proposed, comprising:

[0005] Step 1: Build a point cloud database for intelligent mobile robot usage scenarios;

[0006] Step 2: Construct a pose database based on the point cloud database obtained in step 1;

[0007] Step 3: Construct a description sub-database based on the point cloud database obtained in step 1;

[0008] Step 4: Collect outdoor data sets; build a point cloud registration AGHL network based on hybrid local feature extraction and anchor guidance.

[0009] The AGHL network is trained according to the outdoor data set to obtain a trained AGHL network;

[0010] Step 5: The intelligent mobile robot is turned on, and the intelligent mobile robot uses the built laser radar to collect the startup query point cloud; the intelligent mobile robot is located at any position in the use scene;

[0011] Use the lidar data environment recognition method to extract the point cloud descriptor of the power-on query point cloud;

[0012] Calculate the CKA similarity between the point cloud descriptor of the startup query point cloud and all descriptors in the descriptor database; select the point cloud corresponding to the descriptor with the highest CKA similarity as the startup retrieval point cloud;

[0013] The CKA similarity is a method for measuring the similarity of neural network representations. It evaluates the representation differences between different layers or different models in a neural network by calculating the similarity between two kernel matrices. The core idea of ​​CKA is to use kernel functions to extract the structural similarity of representations and eliminate the influence of scale changes through normalization, thereby providing a stable similarity measurement method. It is a similarity calculation method well known to people in the field.

[0014] Step 6: Input the power-on query point cloud and the power-on retrieval point cloud into the trained AGHL network to obtain the matching relationship between the power-on query point cloud and the power-on retrieval point cloud;

[0015] Use SVD algorithm to process the matching relationship between the power-on query point cloud and the power-on retrieval point cloud, and obtain the relative pose of the power-on query point cloud and the power-on retrieval point cloud;

[0016] The SVD algorithm is an important matrix decomposition method in linear algebra, and can obtain the relative position of the power-on query point cloud and the power-on retrieval point cloud according to the matching relationship between the power-on query point cloud and the power-on retrieval point cloud; it is a processing process well known to those skilled in the art;

[0017] Step 7: Extract the pose of the power-on retrieval point cloud from the pose database.

[0018] According to the relative pose of the startup query point cloud and the startup retrieval point cloud in step 6 and the pose of the startup retrieval point cloud, the startup pose of the intelligent mobile robot is calculated.

[0019] The specific process of constructing the point cloud database of the intelligent mobile robot usage scenario in step 1 is as follows:

[0020] Use the laser radar carried by the intelligent mobile robot to collect point cloud data of the usage scene, and save the collected point cloud data of the usage scene in the industrial computer of the intelligent mobile robot to form a point cloud database of the usage scene of the intelligent mobile robot;

[0021] The model of the laser radar is Hesai Pandar 32-line laser radar;

[0022] The point cloud data of the usage scene is collected by controlling the intelligent mobile robot to move in the usage scene using a remote controller;

[0023] The point cloud data of the usage scenario includes: X frames of point cloud data; X is a positive integer;

[0024] The specific process of using the laser radar carried by the intelligent mobile robot to collect point cloud data of the use scene is to use a remote control to control the mobile robot to move in the scene, and use the Hesai Pandar 32-line laser radar carried by the robot to collect point cloud data, save it in the industrial computer, and form a point cloud database.

[0025] The specific process of constructing the pose database according to the point cloud database obtained in step 1 in step 2 is as follows:

[0026] Use the laser SLAM algorithm to process each frame of the point cloud database obtained in step 1 to obtain the position and posture of the intelligent mobile robot corresponding to each frame of the point cloud; construct the position and posture of the intelligent mobile robot corresponding to each frame of the point cloud into a position and posture database;

[0027] The posture of the intelligent mobile robot is a six-degree-of-freedom posture.

[0028] The origin of the intelligent mobile robot posture is the intelligent mobile robot posture obtained by processing the first frame point cloud through the laser SLAM algorithm;

[0029] The laser SLAM algorithm used is the DLO algorithm;

[0030] The six degrees of freedom posture includes: movement along the x-axis: the movement of the object in the x-axis direction. Movement along the y-axis: the movement of the object in the y-axis direction. Movement along the z-axis: the movement of the object in the z-axis direction. Rotation around the x-axis: the rotation of the object around the x-axis. Rotation around the y-axis: the rotation of the object around the y-axis. Rotation around the z-axis: the rotation of the object around the z-axis.

[0031] The DLO algorithm is an open source algorithm published in 2022 and is well known to those skilled in the art. DLO: Direct LiDAR Odometry - Fast Localization Using Dense Point Clouds - Jointly Open Sourced by UCLA and Caltech; The constructed scene point cloud map is as follows Figure 3 In this process, the present invention regards the position of the mobile robot corresponding to the first frame point cloud as the origin

[0032] In step 3, a descriptor database is constructed based on the point cloud database obtained in step 1; the specific process is as follows:

[0033] The laser radar data environment recognition method is used to extract each frame of point cloud data in the point cloud database obtained in step 1, and the point cloud descriptor of each frame of point cloud data is obtained, which is expressed as D train ; The point cloud descriptors of each frame of point cloud data constitute a descriptor database;

[0034] Among them, the point cloud descriptor of the k-th frame point cloud data is represented by d k , D train ={d1, d2, ....d k ...d X}; The laser radar data environment recognition method is the ScanContext method;

[0035] Scan Context technology is an open source environment recognition and matching method based on LiDAR data. It is a geometric feature-based environment perception method known to those skilled in the art. Scan Context combines LiDAR data processing, environment feature extraction, and low-dimensional representation technologies to achieve efficient environment matching and positioning, especially in complex three-dimensional environments. The dimension of each point cloud descriptor extracted by Scan Context technology is 1200.

[0036] The beneficial effects of the present invention are:

[0037] 1. The present invention optimizes the point cloud registration algorithm in the global relocalization technology based on lidar, and proposes a point cloud registration network (AGHL) based on hybrid local feature extraction and anchor guidance, which improves the power-on positioning accuracy of the intelligent mobile robot and provides technical support for its long-term autonomous operation.

[0038] 2. In view of the problem that the current point cloud registration technology usually regards point clouds as one-dimensional sequences and uses Transformer to realize global context information interaction, which leads to the neglect of the local geometric and semantic correlation between points and adjacent points, the AGHL network proposed in this invention uses a hybrid local feature extraction module. This module extracts the local geometric features and semantic features of the point cloud at the same time, thereby providing more descriptive point cloud features for the point cloud registration task.

[0039] 3. To address the problem that the global context information interaction of Transformer inevitably introduces noise and interference from irrelevant areas, the AGHL network proposed in the present invention uses an anchor-guided cross-attention mechanism to focus the network's attention on the anchor points of relevant areas, thereby ensuring reliable information interaction by reducing noise and interference from irrelevant areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a structural schematic diagram of an intelligent mobile robot laser relocalization technology based on dense point cloud registration proposed in the present invention;

[0041] Figure 2 Schematic diagram of the intelligent mobile robot experimental platform built for the present invention;

[0042] Figure 3 This is a schematic diagram of a scene point cloud map constructed using the DLO laser SLAM algorithm in the present invention;

[0043] Figure 4 Schematic diagram of the point cloud registration network (AGHL) based on hybrid local feature extraction and anchor guidance of the present invention;

[0044] Figure 5 Schematic diagram of a hybrid local feature extraction module used in AGHL of the present invention;

[0045] Figure 6 Schematic diagram of the anchor-guided cross-attention mechanism in AGHL of the present invention;

[0046] Figure 7 A schematic diagram of a mobile robot retrieval point cloud trajectory, a mobile robot query point cloud trajectory, and a mobile robot prediction trajectory of the present invention;

[0047] Figure 8 This is a schematic diagram of the input point cloud and AGHL registration effect of the present invention. DETAILED DESCRIPTION

[0048] Specific implementation method 1: Combination Figure 1 The invention is described, comprising:

[0049] Step 1: Build a point cloud database for intelligent mobile robot usage scenarios;

[0050] Step 2: Construct a pose database based on the point cloud database obtained in step 1;

[0051] Step 3: Construct a description sub-database based on the point cloud database obtained in step 1;

[0052] Step 4: Collect outdoor data sets; build a point cloud registration AGHL network based on hybrid local feature extraction and anchor guidance.

[0053] The AGHL network is trained according to the outdoor data set to obtain a trained AGHL network;

[0054] Step 5: The intelligent mobile robot is turned on, and the intelligent mobile robot uses the built laser radar to collect the startup query point cloud; the intelligent mobile robot is located at any position in the use scene;

[0055] Use the lidar data environment recognition method to extract the point cloud descriptor of the power-on query point cloud;

[0056] Calculate the CKA similarity between the point cloud descriptor of the startup query point cloud and all descriptors in the descriptor database; select the point cloud corresponding to the descriptor with the highest CKA similarity as the startup retrieval point cloud;

[0057] The CKA similarity is a method for measuring the similarity of neural network representations. It evaluates the representation differences between different layers or different models in a neural network by calculating the similarity between two kernel matrices. The core idea of ​​CKA is to use kernel functions to extract the structural similarity of representations and eliminate the influence of scale changes through normalization, thereby providing a stable similarity measurement method. It is a similarity calculation method well known to people in the field.

[0058] Step 6: Input the power-on query point cloud and the power-on retrieval point cloud into the trained AGHL network to obtain the matching relationship between the power-on query point cloud and the power-on retrieval point cloud;

[0059] Use SVD algorithm to process the matching relationship between the power-on query point cloud and the power-on retrieval point cloud, and obtain the relative pose of the power-on query point cloud and the power-on retrieval point cloud;

[0060] The SVD algorithm is an important matrix decomposition method in linear algebra, and can obtain the relative position of the power-on query point cloud and the power-on retrieval point cloud according to the matching relationship between the power-on query point cloud and the power-on retrieval point cloud; it is a processing process well known to those skilled in the art;

[0061] Step 7: Extract the pose of the power-on retrieval point cloud from the pose database.

[0062] According to the relative pose of the power-on query point cloud and the power-on retrieval point cloud in step 6 and the pose of the power-on retrieval point cloud, the power-on pose of the intelligent mobile robot is calculated;

[0063] The present invention constructs a laser relocalization technology for intelligent mobile robots based on dense point cloud registration. The present invention uses ScanContext technology to realize point cloud retrieval, and proposes a point cloud registration network (AGHL) based on hybrid local feature extraction and anchor guidance. The network proposes a hybrid local feature extraction module for integrating the local geometric features and local semantic features of the point cloud. In addition, the network designs an anchor-guided cross-attention mechanism to reduce noise and interference in irrelevant areas, thereby improving the accuracy of point cloud registration. Finally, the present invention built a set of intelligent mobile robot experimental platform and tested it in a real scene with an area of ​​18m×31m. The experimental results show that the present invention achieves a positioning error of 5.137cm.

[0064] Specific implementation method 2: The difference between this implementation method and specific implementation method 1 is that:

[0065] The specific process of constructing the point cloud database of the intelligent mobile robot usage scenario in step 1 is as follows:

[0066] Use the laser radar carried by the intelligent mobile robot to collect point cloud data of the usage scene, and save the collected point cloud data of the usage scene in the industrial computer of the intelligent mobile robot to form a point cloud database of the usage scene of the intelligent mobile robot;

[0067] The model of the laser radar is Hesai Pandar 32-line laser radar;

[0068] The point cloud data of the usage scene is collected by controlling the intelligent mobile robot to move in the usage scene using a remote controller;

[0069] The point cloud data of the usage scenario includes: X frames of point cloud data; X is a positive integer;

[0070] The specific process of using the laser radar carried by the intelligent mobile robot to collect point cloud data of the use scene is to use a remote control to control the mobile robot to move in the scene, and use the Hesai Pandar 32-line laser radar carried by the robot to collect point cloud data, save it in the industrial computer, and form a point cloud database.

[0071] The specific process of constructing the pose database according to the point cloud database obtained in step 1 in step 2 is as follows:

[0072] Use the laser SLAM algorithm to process each frame of the point cloud database obtained in step 1 to obtain the position and posture of the intelligent mobile robot corresponding to each frame of the point cloud; construct the position and posture of the intelligent mobile robot corresponding to each frame of the point cloud into a position and posture database;

[0073] The posture of the intelligent mobile robot is a six-degree-of-freedom posture.

[0074] The origin of the intelligent mobile robot posture is the intelligent mobile robot posture obtained by processing the first frame point cloud through the laser SLAM algorithm;

[0075] The laser SLAM algorithm used is the DLO algorithm;

[0076] The six degrees of freedom posture includes: movement along the x-axis: the movement of the object in the x-axis direction. Movement along the y-axis: the movement of the object in the y-axis direction. Movement along the z-axis: the movement of the object in the z-axis direction. Rotation around the x-axis: the rotation of the object around the x-axis. Rotation around the y-axis: the rotation of the object around the y-axis. Rotation around the z-axis: the rotation of the object around the z-axis.

[0077] The DLO algorithm is an open source algorithm published in 2022 and is well known to those skilled in the art. DLO: Direct LiDAR Odometry - Fast Localization Using Dense Point Clouds - Jointly Open Sourced by UCLA and Caltech; The constructed scene point cloud map is as follows Figure 3 In this process, the present invention regards the position of the mobile robot corresponding to the first frame point cloud as the origin

[0078] In step 3, a descriptor database is constructed based on the point cloud database obtained in step 1; the specific process is as follows:

[0079] The laser radar data environment recognition method is used to extract each frame of point cloud data in the point cloud database obtained in step 1, and the point cloud descriptor of each frame of point cloud data is obtained, which is expressed as D train ; The point cloud descriptors of each frame of point cloud data constitute a descriptor database;

[0080] Among them, the point cloud descriptor of the k-th frame point cloud data is represented by d k , D train ={d1, d2, ....d k ...d X}; The laser radar data environment recognition method is the ScanContext method;

[0081] Scan Context technology is an open source environment recognition and matching method based on LiDAR data. It is a method of environment perception based on geometric features that is well known to those skilled in the art. Scan Context combines technologies such as LiDAR data processing, environment feature extraction, and low-dimensional representation, mainly to achieve efficient environment matching and positioning, especially in complex three-dimensional environments; the dimension of each point cloud descriptor extracted by Scan Context technology is 1200 dimensions; other steps and parameters are the same as those in Specific Implementation Method 1.

[0082] Specific implementation method three: This implementation method is different from the specific implementation method one in that:

[0083] The outdoor dataset in step 4 is the KITTI outdoor dataset;

[0084] The KITTI dataset contains a large amount of outdoor point cloud data, including multiple groups of point clouds in the same scene, and has good applicability as a training set.

[0085] The AGHL network includes: a backbone network, a point cloud feature enhancement network and a two-stage matching network;

[0086] The backbone network is a KPConv-FPN network;

[0087] The AGHL network is trained according to the outdoor data set to obtain a trained AGHL network; the specific process is:

[0088] Step 41: Select two sets of point clouds of the same scene in the KITTI outdoor dataset as the input of the AGHL network and set the number of training times; the two sets of point clouds of the same scene are represented as: query point cloud Retrieving point clouds

[0089] Step 42: Query point cloud Input the backbone network for processing and get The query point cloud downsampled points and query features; each sampling point in the query point cloud corresponds to a query feature; is a positive integer;

[0090] Will The query point cloud sampling points form the query point cloud, which is expressed as Represents dimension;

[0091] Will query features form a query feature vector, expressed as Represents a dimensional variable;

[0092] Retrieve point cloud Input the backbone network for processing and get The downsampled points of the retrieved point cloud and retrieval features; each sampling point in the retrieval point cloud corresponds to a retrieval feature; is a positive integer;

[0093] Will The sampling points of the retrieval point cloud constitute the retrieval sampling point cloud, which is expressed as

[0094] Will The retrieval feature vector composed of retrieval features is expressed as

[0095] Because there is The coordinate dimension of each down-sampled point in the query point cloud is 3, so the dimension of the composed vector is The query feature, the retrieval point cloud sampling points, and the retrieval feature are the same;

[0096] Step 43: Assign all points in the query point cloud to the nearest query point cloud downsampling point according to the distance; Query local point cloud block G x as well as The query local feature F gx ;

[0097] Among them, the query local point cloud block composed of the sampling points of the i-th query point cloud is expressed as G i x The query local feature is expressed as n represents the number of points in the query local point cloud block; n is a positive integer;

[0098] All points in the retrieval point cloud are assigned to the nearest down-sampling point in the retrieval point cloud according to the distance; Retrieve local point cloud blocks G y as well as The local features F are retrieved gy ;

[0099] Among them, the retrieval local point cloud block composed of the sampling points under the jth retrieval point cloud is expressed as G j y The local feature of retrieval is expressed as m represents the number of points in the local point cloud block to be retrieved; m is a positive integer;

[0100] For example, if the distances of a certain point from sampling point 1, sampling point 2, and sampling point 3 of the query point cloud are 10, 20, and 15 respectively, then this point is assigned to sampling point 1 of the query point cloud;

[0101] Step 44: Subsample the query point cloud obtained in step 42 Retrieve the downsampled point cloud Query feature vector and retrieve feature vector Input the point cloud feature enhancement module to obtain the final query point cloud feature F att x And the final retrieved point cloud feature F att y ;

[0102] Step 45: The final query point cloud feature F obtained in step 44 att x And the final retrieved point cloud feature F att y ;、 Query local point cloud block G x , The query local feature F gx ; Retrieve local point cloud blocks G y as well as The local features F are retrieved gy Input the two-stage matching module, and output the dense point cloud matching relationship between the query point cloud and the retrieval point cloud;

[0103] Step 46: Perform iterative training according to the input and output of the AGHL network. When the number of training times reaches the set number of training times, stop the training and obtain the trained AGHL network.

[0104] The other steps and parameters are the same as those in the first to second embodiments.

[0105] Specific implementation method 4: This implementation method is different from specific implementation methods 1 to 4 in that:

[0106] The point cloud feature enhancement module in step 44 includes: a hybrid local feature extraction module, a self-attention mechanism module, and an anchor-guided cross-attention mechanism module;

[0107] The query downsampled point cloud obtained in step 42 is Retrieve the downsampled point cloud Query feature vector and retrieve feature vector Input the point cloud feature enhancement module to obtain the final query point cloud feature F att xAnd the final retrieved point cloud feature F att y ; The specific process is:

[0108] Step 441: Query the sampled point cloud And the query feature vector Input the mixed local feature extraction module to obtain the query fusion feature; the query fusion feature is expressed as The downsampled point cloud will be retrieved And retrieve the feature vector Input the mixed local feature extraction module to obtain the retrieval fusion feature; the retrieval fusion feature is expressed as

[0109] Step 442: Fusion query features Input the self-attention mechanism module to obtain enhanced query features; the enhanced query features are represented as F sa x ; Retrieve fusion features Input the self-attention mechanism module to obtain enhanced retrieval query features; the enhanced retrieval query features are represented as F sa y ;

[0110] The self-attention mechanism module is a standard Transformer network;

[0111] Step 443: Subsample the query point cloud obtained in step 42 Retrieve the downsampled point cloud Enhanced query feature F sa x and enhanced retrieval query features F sa y At the same time, the anchor-guided cross-attention mechanism module is input to obtain feature-enhanced query point cloud features and feature-enhanced retrieval point cloud features;

[0112] Step 444: The query point cloud features obtained by feature enhancement are used as the query feature vector of the downsampled point cloud for the next iteration; the retrieval point cloud features obtained by feature enhancement are used as the retrieval feature vector of the downsampled point cloud for the next iteration; repeat steps 441 to 444 for L times, and the enhanced query point cloud features obtained in the last iteration are represented as F att x ; L is a positive integer; the enhanced retrieval point cloud feature obtained in the last iteration is represented by F att y ; Other steps and parameters are the same as those in Specific Implementation Methods 1 to 3.

[0113] Specific implementation method 5: This implementation method is different from specific implementation methods 1 to 4 in that:

[0114] In step 441, the downsampled point cloud is queried. And the query feature vector Input the mixed local feature extraction module to obtain the query fusion feature; the query fusion feature is expressed as

[0115] The downsampled point cloud will be retrieved And retrieve the feature vector Input the mixed local feature extraction module to obtain the retrieval fusion feature; the retrieval fusion feature is expressed as The specific process is:

[0116] Step 4411: Use KNN algorithm to process the query downsampled point cloud get The K closest points to the query point cloud sampling points; The K nearest points of the query point cloud sampling points constitute the query closest point cloud, which is expressed as In the query feature vector Find the query closest point feature vector corresponding to the query closest point cloud, expressed as

[0117] Use KNN algorithm to process the sampled point clouds separately get The K nearest points of the retrieval point cloud sampling points; The K nearest points of the retrieval point cloud sampling points constitute the retrieval nearest point cloud, which is expressed as In retrieving feature vectors Find the retrieval nearest point feature vector corresponding to the query nearest point cloud, expressed as

[0118] Specifically, K points with the smallest distance are selected as the nearest points, and the process of retrieving the point cloud is the same as that of querying the point cloud.

[0119] Each query point cloud point corresponds to a query feature. If K query nearest points are found, then the corresponding query features are also found. The vector composed of the corresponding query features is used as the query nearest point feature vector. The point cloud retrieval process is the same as the query point cloud process.

[0120] Step 4412: Query the nearest point cloud And query the sampled point cloud Constructing the query geometry vector Retrieve the nearest point cloud and retrieve the downsampled point cloud Constructing the retrieval geometry vector According to the query nearest point feature vector and retrieve feature vector Constructing query semantic vector According to the nearest point feature vector and retrieve feature vector Constructing the retrieval geometry vector The formula is:

[0121]

[0122] In the formula, [·,·,·] represents the concatenation of features along the channel dimension, and Π() represents the concatenation of dimensions from Convert to || ||2 means calculating the Euclidean distance, represents the sampling point under the i-th query, represents the jth retrieval sampling point; represents the query feature corresponding to the sampling point under the i-th query, represents the retrieval feature corresponding to the sampling point under the jth retrieval;

[0123] The first item in the splice feature Represents the absolute coordinates of the K nearest points.

[0124] Item 2 Represents the relative coordinates of the K nearest points.

[0125] Item 3 represents the Euclidean distance.

[0126] Step 4413: According to the query geometry vector E x Get the query geometric features, which are expressed as According to the search geometry vector E y Get the retrieval geometric features, the query geometric features are expressed as According to the query semantic vector S x Get the query semantic features, which are expressed as According to the retrieval semantic vector S y Get the query semantic features, and the retrieval semantic features are expressed as The formula is:

[0127] F G x =LN(MLP(E x ))

[0128] F G y =LN(MLP(E y ))

[0129] F S x =LN(MLP(S x ))

[0130] F S y =LN(MLP(S y ))

[0131] Wherein, LN() represents the normalization layer processing, MLP() represents the MLP neural network layer processing; the MLP neural network layer processing includes an input layer, a hidden layer, and an output layer, which are neural network layers well known to those skilled in the art;

[0132] Step 4414: Based on the query geometric feature F G x Get the query geometry weighting coefficient, which is expressed as W G x ; According to the retrieval geometric feature F G y Get the retrieval geometry weighting coefficient, which is expressed as W G x ; According to the query semantic feature F S x Get the query semantic weighting coefficient, which is expressed as W S x ; According to the retrieval semantic feature F S y Get the retrieval semantic weighting coefficient, which is expressed as W S x ; expressed as:

[0133] W G x =softmax(MLP(LN(MLP(F G x ))))

[0134] W G y =softmax(MLP(LN(MLP(F G y ))))

[0135] W S x =softmax(MLP(LN(MLP(F S x ))))

[0136] W S y =softmax(MLP(LN(MLP(F S y ))))

[0137] In the formula, softmax() represents the softmax function processing;

[0138] Step 4415: Based on the query geometric feature F G x and query semantic features F S x Get the query connection feature; the query connection feature is represented by F m x The geometric features F will be retrieved G y and retrieve semantic features F S y Get the retrieval connection feature; the retrieval connection feature is represented by F m y ; expressed as:

[0139] F m x =LN(MLP([F G x ,F S x ]))

[0140] F m y =LN(MLP([F G y ,F S y ]))

[0141] Where [·,·] represents the concatenation of features along the channel dimension;

[0142] Step 4416: Connect feature F based on query m x and the query geometry weight coefficient is denoted as W G x The query geometry weighted fusion feature is obtained, and the query geometry weighted fusion feature is expressed as F fg x ; Connect feature F according to query m x and the query semantic weight coefficient is denoted as W S x The query semantic weighted fusion feature is obtained, and the query semantic weighted fusion feature is expressed as F fs x ; According to the retrieval connection feature F m y and the retrieval geometric weight coefficient is denoted as W G y The retrieval geometry weighted fusion feature is obtained, and the retrieval geometry weighted fusion feature is expressed as F fg y; According to the retrieval connection feature F m y and the retrieval semantic weight coefficient is expressed as W S y The retrieval semantic weighted fusion feature is obtained, and the retrieval semantic weighted fusion feature is expressed as F fs y ; expressed as:

[0143] F fg x =F m x ⊙W G x +F G x

[0144] F fs x =F m x ⊙W S x +F S x

[0145] F fg y =F m y ⊙W G y +F G y

[0146] F fs y =F m y ⊙W S y +F S y

[0147] In the formula, ⊙ represents element-by-element multiplication processing;

[0148] Step 4417: Weighted fusion features F according to query geometry fg x and query semantic weighted fusion feature F fs x Get the query fusion feature; the query fusion feature is expressed as According to the retrieval geometry weighted fusion feature F fg y The weighted fusion feature of the retrieval semantics is represented as F fs y Get the retrieval fusion feature, which is expressed as The other steps and parameters are the same as those in Specific Embodiments 1 to 4.

[0149] Specific implementation method 6: This implementation method is different from specific implementation methods 1 to 5 in that:

[0150] In step 4417, the weighted fusion feature F is obtained according to the query geometry. fg x and query semantic weighted fusion feature F fs x Get query fusion features; weight fusion features F according to retrieval geometry fg y The weighted fusion feature of the retrieval semantics is represented as F fs y Get the retrieval fusion features. The specific process is as follows:

[0151] Step 44171: Weighted fusion of query geometry features F fg x Sum in the second dimension to get the query intermediate feature F fg1 x ; Weight the query semantics and fuse the features F fs x Sum in the second dimension to get the query intermediate feature F fs2 x ; Retrieve geometric weighted fusion features F fg y Sum in the second dimension to get the retrieval intermediate feature F fg1 y ; The retrieval semantic weighted fusion feature F fs y Sum in the second dimension to get the retrieval intermediate feature F fs2 y ;

[0152] Step 44172: Query the intermediate feature F fg1 x and query intermediate features F fs1 x Get the query fusion feature F f x ; According to the retrieval intermediate feature F fg1 y and retrieve the intermediate feature F fs2 y Get the retrieval fusion feature F f y ; expressed as:

[0153]

[0154] ; The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5.

[0155] Specific implementation method 7: The difference between this implementation method and specific implementation methods 1 to 6 is that in step 443, the query downsampled point cloud obtained in step 42 is Retrieve the downsampled point cloud Enhanced query feature F sa x and enhanced retrieval query features F sa y At the same time, the anchor-guided cross-attention mechanism module is input to obtain the feature-enhanced query point cloud features and the feature-enhanced retrieval point cloud features; the specific process is:

[0156] Step 4431: Use the KNN algorithm to process the query downsampled point cloud obtained in step 42 Get the K0 closest points of each query point cloud sampling point; the K0 closest points of each query point cloud sampling point form the query point cloud sampling point closest point cloud, expressed as

[0157] Query point cloud sampling points Each query point cloud sampling point has K0 nearest points, so there are a total of points, each point has 3-dimensional coordinates, so the nearest point cloud dimension of the query point cloud sampling point is

[0158] According to the enhanced query feature F sa x Calculate the inlier similarity subvector of each query sampling point; combine the inlier similarity subvectors of each query sampling point into a query inlier similarity vector, expressed as

[0159] Step 4432: Enhance the query feature F sa x Perform enhanced search query feature F sa y Inner product processing is performed to obtain the soft allocation matrix A, which is expressed as According to the soft assignment matrix A, the response point cloud and the external point similarity vector are constructed; the response point cloud is expressed as The outlier similarity vector is expressed as

[0160] The inner product calculation method between the vectors is a calculation method known to those skilled in the art.

[0161] Step 4433: Based on the inlier similarity vector S ra and the external point similarity vector S er Construct a significance score; the significance score is expressed as The formula is:

[0162] S s =S ra ☉S er

[0163] According to the significance score S s In response to the point cloud P ag y Select K1 initial anchor points of each query sampling point, and form an initial anchor point cloud with K1 initial anchor points of each query sampling point; the initial anchor point cloud is represented as Use KNN algorithm to process the response point cloud P ag y , get the K1 closest points of each initial anchor point; get the K1 closest points of each initial anchor point to form an extended anchor point cloud, the extended anchor point cloud is expressed as

[0164] Step 4434: Expand the anchor point cloud P an y And query the sampled point cloud Construct a point cloud mask, which is represented by

[0165] Step 4435: Based on the enhanced query feature F sa x , Enhanced retrieval query features F sa y and point cloud mask M xy Generate intermediate feature Info; use FNN network to process intermediate feature Info to obtain feature-enhanced query point cloud features and feature-enhanced retrieval point cloud features;

[0166] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.

[0167] Specific implementation eight: This implementation differs from specific implementations one to seven in that:

[0168] In step 4431, according to the enhanced query feature F sa x The specific process of calculating the inlier similarity subvector of the sampling point under the pth query is:

[0169] Step 44311: Based on the enhanced query feature F sa x Find the enhanced query features corresponding to the sampling point under the p-th query and the enhanced query features corresponding to the K0 nearest points of the sampling point under the p-th query,

[0170] Step 44312: Calculate the inner product between the enhanced query feature corresponding to the sampling point under the p-th query and the enhanced query feature corresponding to the K0 nearest points under the sampling point under the p-th query, obtain the inner product value of the K0 query sampling points, and use the inner product value of the K0 query sampling points to form the inner point similarity subvector of the sampling point under the p-th query

[0171] The vector inner product calculation method is a calculation method well known to those skilled in the art;

[0172] The specific process of constructing the response point cloud and the external point similarity according to the soft allocation matrix A in step 4432 is as follows:

[0173] The element A(i′,j′) in the i′th row and j′th column of the soft assignment matrix A is used as the query downsampled point cloud The i′th query sampling point and the retrieval sampling point cloud The matching probability of the sampling point under the j′th retrieval;

[0174] Find P ag x The maximum matching probability of each point in P ag x The maximum matching probability of each point in the composes the external point similarity S er ,

[0175] Find P ag x Find the maximum matching probability for each point in the sampled point cloud under retrieval The corresponding point in P ag x The response point of each point in P ag x The response points of each point in form a response point cloud P ag y ;

[0176] Because P ag x Each point in the query is a sampled point cloud So P ag x The matching probability of each point in is the value of the corresponding element value in the soft assignment matrix A; find the maximum value of each row of the soft assignment matrix A, which is P ag x The maximum matching probability of each point in the retrieval point cloud. All the points in constitute the response point cloud;

[0177] In step 4434, according to the expanded anchor point cloud P an yAnd query the sampled point cloud Construct a point cloud mask, which is represented by The specific process is:

[0178] The point cloud mask M xy Corresponding to the expanded anchor point cloud P an y And query the sampled point cloud The value of the element is 1; the values ​​of other elements are 0; the point cloud mask M is obtained xy ;

[0179] For example, for query downsampled point cloud The i-th query downsampled point in , assuming that the query downsampled point cloud The index of the anchor point in is {j k |k=1,2,...,K1K2}, the present invention converts M xy (i,j k ) is set to 1

[0180] In step 4435, according to the enhanced query feature F sa x , Enhanced retrieval query features F sa y and point cloud mask M xy Generate intermediate feature Info; expressed as:

[0181]

[0182] Where Q represents the MLP processing F sa x The query tensor generated after, K represents the MLP processing F sa y The key tensor generated after, V represents the MLP processing F sa y The value tensor generated later, K T represents the transpose of K, represents the characteristic dimension; the other steps and parameters are the same as those in any one of the specific implementation modes one to seven.

[0183] Specific implementation method 9: This implementation method is different from specific implementation methods 1 to 8 in that:

[0184] In step 45, the final query point cloud feature F obtained in step 44 is att x , finally retrieve point cloud feature F att y , Query local point cloud block G x , The query local feature F gx, Retrieve local point cloud blocks G y as well as The local features F are retrieved gy Input the two-stage matching module, and output the dense point cloud matching relationship between the query point cloud and the retrieval point cloud; the specific process is:

[0185] Step 451: Based on the final query point cloud feature F att x And the final retrieved point cloud feature F att y Perform downsampling point matching to get downsampling point matching results; the specific process is:

[0186] Step 4511: Based on the final query point cloud feature F att x And the final retrieved point cloud feature F att y Calculate the Gaussian correlation matrix; the Gaussian correlation matrix is ​​expressed as Among them, the element G(i″,j″) in the i″th row and j″th column of the Gaussian correlation matrix G is expressed by the formula:

[0187]

[0188] In the formula, norm() represents normalization processing, represents the final query point cloud feature of the i″th query sampling point, represents the final retrieval point cloud features of the sampling point under the j″th retrieval, It means calculating the second norm and then squaring it, exp() means the exponential function,

[0189] Step 4512: Get the downsampling point matching result H according to the Gaussian correlation matrix G c , expressed as:

[0190]

[0191] In the formula, Indicates the matching relationship between the sampling points in the query point cloud in the i″th row and the sampling points in the query point cloud in the j″th column. topk() indicates selecting the largest top K su Elements; K su is a positive integer;

[0192] Step 452: Match the result H according to the downsampling point c , Query local point cloud block G x , The query local feature F gx ; Retrieve local point cloud blocks Gy as well as The local features F are retrieved gy Perform original point matching to obtain the dense point cloud matching relationship H between the query point cloud and the retrieval point cloud f ; Dense point cloud matching relationship H between query point cloud and retrieval point cloud f In the i″th query local point cloud block and the j″th retrieved local point cloud block The dense point cloud matching relationship between

[0193] Among them, according to the downsampling point matching results, Query local point cloud block G x , The query local feature F gx ; Retrieve local point cloud blocks G y as well as The local features F are retrieved gy Perform original point matching to obtain query local point cloud blocks and retrieve local point cloud blocks Dense point cloud matching relationship between The specific process is:

[0194] Step 4521: Match the results based on the downsampling points Determine the corresponding query local point cloud block Query local features Retrieve local point cloud blocks and retrieve local features According to the i″th query local feature and the j″th retrieved local feature Constructing the cost matrix The formula is:

[0195]

[0196] In the formula, express The transpose of , τ represents the temperature factor, τ = 0.1;

[0197] Step 4522: Use SinkHorn algorithm to process the cost matrix Get the soft assignment matrix

[0198] The SinkHorn algorithm is an iterative algorithm for solving the optimal transmission problem. It is an algorithm well known to those skilled in the art;

[0199] Step 4523: Delete the soft allocation matrix The last row and last column of , we get the soft assignment matrix

[0200] Step 4524: According to the soft allocation matrix Query local point cloud blocks and retrieve local point cloud blocks Perform matching and obtain matching results; the specific process is:

[0201] like The Points and The Points The corresponding soft assignment matrix Elements When the matching conditions are met, and As a result of matching The matching conditions are:

[0202] element The value of is No. One of the three largest element values ​​in the row;

[0203] The value of is No. One of the three largest element values ​​in the column;

[0204] The value of is greater than a predefined threshold, the predefined threshold is 0.05;

[0205] Step 4525: Combine all matching results into query local point cloud blocks and retrieval local point cloud blocks Dense point cloud matching relationship between The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0206] Specific implementation method 10: This implementation method is different from specific implementation methods 1 to 9 in that:

[0207] The laser radar data environment recognition method in step 5 is the ScanContext method;

[0208] In step 7, the power-on posture of the intelligent mobile robot is calculated based on the relative posture of the power-on query point cloud and the power-on retrieval point cloud in step 6 and the posture of the power-on retrieval point cloud, and the formula is expressed as:

[0209] T X =T XY T Y

[0210] Where, T X represents the startup posture of the intelligent mobile robot; represents the startup retrieval point cloud posture T XY , represents the relative pose T of the startup query point cloud and the startup retrieval point cloud Y ;

[0211] The other steps and parameters are the same as those in Specific Implementations 1 to 9.

[0212] Combining the simulation analysis of specific implementation methods one to ten

[0213] Aiming at an intelligent mobile robot laser relocalization technology based on dense point cloud registration, the present invention proposes a point cloud registration network AGHL based on hybrid local feature extraction and anchor guidance. To verify its effectiveness, the present invention first compares it with other point cloud registration methods in the public dataset KITTI, and the comparison results are shown in Table 1. The experimental results show that compared with other methods, AGHL has achieved excellent performance in terms of inlier rate (IR), relative translation error (RTE), relative rotation error (RRE), and recall rate (RR). Specifically, compared with Predator, CofiNet and GeoTransformer, AGHL reduces RTE from 6.8cm, 8.2cm, 6.8cm to 6.2cm, and reduces RRE from 0.27°, 0.41°, 0.24° to 0.22°. In addition, AGHL achieves the highest recall rate of 99.8%.

[0214] Table 1 Comparison results of AGHL and other methods in KITTI dataset

[0215]

[0216] For the AGHL network, the present invention uses 4 NVIDIA TeslaA40 GPUs for training, and the training takes about 12 hours. During the training process, the training epoch is set to 160, the batch is set to 1, and the Adam optimizer is used. The initial learning rate is 1×10 -4 , and then every 4 epochs it becomes 95% of the original.

[0217] During the data set collection process, NoMachine software was used to remotely connect to the industrial computer and the remote control was used to operate the intelligent mobile robot. The real scene area is 18m×31m. The motion trajectory of the intelligent mobile robot is as follows: Figure 7 As shown in the red and green tracks in Figure 2. During this process, a total of 3277 frames of point clouds were collected, of which 1800 frames (red tracks) were used to build the scene database, and the remaining 1477 frames (green tracks) were used for testing.

[0218] The motion trajectory of the intelligent mobile robot predicted by the intelligent mobile robot laser relocalization technology based on dense point cloud registration proposed in this invention is as follows: Figure 7 As shown in the blue track in . Despite the presence of some outliers, the proposed technique always shows highly accurate pose estimation. In addition, Figure 8 The point cloud registration effect shown in further highlights that AGHL can effectively process point clouds obtained from unknown environments.

[0219] The above only describes the preferred embodiments of the present invention. It should be understood that the present invention is not limited to the above specific implementation methods. Although the present invention has been disclosed as above in the preferred embodiments, it is not used to limit the present invention. Any technician familiar with the profession can make some changes or modify the technical contents disclosed above into equivalent embodiments with equivalent changes without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement made to the above embodiments without departing from the content of the technical solution of the present invention, based on the technical essence of the present invention, within the spirit and principles of the present invention, still fall within the protection scope of the technical solution of the present invention.

Claims

1. A laser relocalization method for an intelligent mobile robot based on dense point cloud registration, characterized in that: The following steps are involved: Step 1: Build a point cloud database for intelligent mobile robot usage scenarios; Step 2: Construct a pose database based on the point cloud database obtained in step 1; Step 3: Construct a description sub-database based on the point cloud database obtained in step 1; Step 4: Collect outdoor data sets; build AGHL network, The AGHL network is trained according to the outdoor data set to obtain a trained AGHL network; Step 5: The intelligent mobile robot is turned on, and the intelligent mobile robot uses the built laser radar to collect the startup query point cloud; the intelligent mobile robot is located at any position in the use scene; Use the LiDAR data environment recognition method to extract the point cloud descriptor of the power-on query point cloud; Calculate the CKA similarity between the point cloud descriptor of the startup query point cloud and all descriptors in the descriptor database; Select the point cloud corresponding to the descriptor with the highest CKA similarity as the startup retrieval point cloud; Step 6: Input the power-on query point cloud and the power-on retrieval point cloud into the trained AGHL network to obtain the matching relationship between the power-on query point cloud and the power-on retrieval point cloud; Use SVD algorithm to process the matching relationship between the power-on query point cloud and the power-on retrieval point cloud, and obtain the relative pose of the power-on query point cloud and the power-on retrieval point cloud; Step 7: Extract the pose of the power-on retrieval point cloud from the pose database. According to the relative pose of the power-on query point cloud and the power-on retrieval point cloud in step six and the pose of the power-on retrieval point cloud, the power-on pose of the intelligent mobile robot is calculated.

2. According to claim 1, a laser relocation method for an intelligent mobile robot based on dense point cloud registration is characterized in that: The specific process of constructing the point cloud database of the intelligent mobile robot usage scenario in step 1 is as follows: Use the laser radar carried by the intelligent mobile robot to collect point cloud data of the usage scene, and save the collected point cloud data of the usage scene in the industrial computer of the intelligent mobile robot to form a point cloud database of the usage scene of the intelligent mobile robot; The model of the laser radar is Hesai Pandar 32-line laser radar; The point cloud data of the usage scene is collected by controlling the intelligent mobile robot to move in the usage scene using a remote controller; The point cloud data of the usage scenario includes: X frames of point cloud data; X is a positive integer; The specific process of constructing the pose database according to the point cloud database obtained in step 1 in step 2 is as follows: Use the laser SLAM algorithm to process each frame of the point cloud database obtained in step 1 to obtain the position and posture of the intelligent mobile robot corresponding to each frame of the point cloud; construct the position and posture of the intelligent mobile robot corresponding to each frame of the point cloud into a position and posture database; The posture of the intelligent mobile robot is a six-degree-of-freedom posture. The origin of the intelligent mobile robot posture is the intelligent mobile robot posture obtained by processing the first frame point cloud with the laser SLAM algorithm; the laser SLAM algorithm used is the DLO algorithm; in step three, a descriptor database is constructed based on the point cloud database obtained in step one; the specific process is: The laser radar data environment recognition method is used to extract each frame of point cloud data in the point cloud database obtained in step 1, and the point cloud descriptor of each frame of point cloud data is obtained, which is expressed as D train ; The point cloud descriptors of each frame of point cloud data constitute a descriptor database; Among them, the point cloud descriptor of the k-th frame point cloud data is represented by d k , D train ={d1, d2, ....d k ...d X }; The laser radar data environment recognition method is the ScanContext method.

3. The intelligent mobile robot laser relocation method based on dense point cloud registration according to claim 2 is characterized in that: The outdoor data set in step 4 is the KITTI outdoor data set; the AGHL network includes: a backbone network, a point cloud feature enhancement network and a two-stage matching network; the backbone network is a KPConv-FPN network; the AGHL network is trained according to the outdoor data set to obtain a trained AGHL network; the specific process is: Step 41: Select two sets of point clouds of the same scene in the KITTI outdoor dataset as the input of the AGHL network and set the number of training times; the two sets of point clouds of the same scene are represented as: query point cloud Retrieving point clouds Step 42: Query point cloud Input the backbone network for processing and get The query point cloud downsampled points and query features; each sampling point in the query point cloud corresponds to a query feature; is a positive integer; The query point cloud sampling points form the query point cloud, which is expressed as Represents dimension; query features form a query feature vector, expressed as Represents a dimension variable; the point cloud will be retrieved Input the backbone network for processing and get The downsampled points of the retrieved point cloud and retrieval features; each sampling point in the retrieval point cloud corresponds to a retrieval feature; is a positive integer; The sampling points of the retrieval point cloud constitute the retrieval sampling point cloud, which is expressed as Will The retrieval feature vector composed of retrieval features is expressed as Step 43: Assign all points in the query point cloud to the nearest query point cloud downsampling point according to the distance; Query local point cloud block G x as well as The query local features F gx ; Among them, the query local point cloud block composed of the sampling points under the i-th query point cloud is expressed as G i x The query local feature is expressed as n represents the number of points in the query local point cloud block; n is a positive integer; All points in the retrieval point cloud are assigned to the nearest down-sampling point in the retrieval point cloud according to the distance; Retrieve local point cloud blocks G y as well as The local features F are retrieved gy ; Among them, the retrieval local point cloud block composed of the sampling points under the jth retrieval point cloud is expressed as G j y The local feature of retrieval is expressed as m represents the number of points in the local point cloud block to be retrieved; m is a positive integer; Step 44: Subsample the query point cloud obtained in step 42 Retrieve the downsampled point cloud Query feature vector and retrieve feature vector Input the point cloud feature enhancement module to obtain the final query point cloud feature F att x And the final retrieved point cloud feature F att y ; Step 45: The final query point cloud feature F obtained in step 44 att x And the final retrieved point cloud feature F att y ;、 Query local point cloud block G x , The query local features F gx ; Retrieve local point cloud blocks G y And M retrieval local features F gy Input the two-stage matching module, and output the dense point cloud matching relationship between the query point cloud and the retrieval point cloud; Step 46: Perform iterative training based on the input and output of the AGHL network. When the number of training times reaches the set number of training times, stop training to obtain a trained AGHL network.

4. The method for laser relocalization of an intelligent mobile robot based on dense point cloud registration according to claim 3, characterized in that: The point cloud feature enhancement module in step 44 includes: a hybrid local feature extraction module, a self-attention mechanism module, and an anchor-guided cross-attention mechanism module; The query downsampled point cloud obtained in step 42 is Retrieve the downsampled point cloud Query feature vector and retrieve feature vector Input the point cloud feature enhancement module to obtain the final query point cloud feature F att x And the final retrieved point cloud feature F att y ; The specific process is: Step 441: Query the sampled point cloud And the query feature vector Input the hybrid local feature extraction module to obtain the query fusion feature; the query fusion feature is represented by F f x ; The downsampled point cloud will be retrieved And retrieve the feature vector Input the mixed local feature extraction module to obtain the retrieval fusion feature; the retrieval fusion feature is represented by F f y ; Step 442: Fusion query feature F f x Input the self-attention mechanism module to obtain enhanced query features; the enhanced query features are represented as F sa x ; Retrieve fusion feature F f y Input the self-attention mechanism module to obtain enhanced retrieval query features; the enhanced retrieval query features are represented as F sa y ; The self-attention mechanism module is a standard Transformer network; Step 443: Subsample the query point cloud obtained in step 42 Retrieve the downsampled point cloud Enhanced query feature F sa x and enhanced retrieval query features F sa y At the same time, the anchor-guided cross-attention mechanism module is input to obtain feature-enhanced query point cloud features and feature-enhanced retrieval point cloud features; Step 444: The query point cloud features obtained by feature enhancement are used as the query feature vector of the downsampled point cloud for the next iteration; the retrieval point cloud features obtained by feature enhancement are used as the retrieval feature vector of the downsampled point cloud for the next iteration; repeat steps 441 to 444 for L times, and the enhanced query point cloud features obtained in the last iteration are represented as F att x ; L is a positive integer; the enhanced retrieval point cloud feature obtained in the last iteration is represented by F att y .

5. The method for laser relocalization of an intelligent mobile robot based on dense point cloud registration according to claim 4, characterized in that: In step 441, the downsampled point cloud is queried. And the query feature vector Input the hybrid local feature extraction module to obtain the query fusion feature; the query fusion feature is represented by F f x ; The downsampled point cloud will be retrieved And retrieve the feature vector Input the mixed local feature extraction module to obtain the retrieval fusion feature; the retrieval fusion feature is represented by F f y ; The specific process is: Step 4411: Use KNN algorithm to process the query downsampled point cloud get The K closest points to the query point cloud sampling points; The K nearest points of the query point cloud sampling points constitute the query closest point cloud, which is expressed as In the query feature vector Find the query closest point feature vector corresponding to the query closest point cloud, expressed as Use KNN algorithm to process the sampled point clouds separately get The K nearest points of the retrieval point cloud sampling points; The K nearest points of the retrieval point cloud sampling points constitute the retrieval nearest point cloud, which is expressed as In retrieving feature vectors Find the retrieval nearest point feature vector corresponding to the query nearest point cloud, expressed as Step 4412: Query the nearest point cloud And query the sampled point cloud Constructing the query geometry vector Retrieve the nearest point cloud and retrieve the downsampled point cloud Constructing the retrieval geometry vector According to the query nearest point feature vector and retrieve feature vector Constructing query semantic vector According to the nearest point feature vector and retrieve feature vector Constructing the retrieval geometry vector The formula is: Where [·,·,·] represents the concatenation of features along the channel dimension, and Π( ) represents the concatenation of dimensions from Convert to || ||2 means calculating the Euclidean distance, represents the sampling point under the i-th query, represents the jth retrieval sampling point; represents the query feature corresponding to the sampling point under the i-th query, represents the retrieval feature corresponding to the sampling point under the jth retrieval; Step 4413: According to the query geometry vector E x Get the query geometric features, which are expressed as According to the search geometry vector E y Get the retrieval geometric features, the query geometric features are expressed as According to the query semantic vector S x Get the query semantic features, which are expressed as According to the retrieval semantic vector S y Get the query semantic features, and the retrieval semantic features are expressed as The formula is: F G x =LN(MLP(E x )) F G y =LN(MLP(E y )) F S x =LN(MLP(S x )) F S y =LN(MLP(S y )) In the formula, LN() represents the normalization layer processing, MLP() represents the MLP neural network layer processing; Step 4414: Based on the query geometric feature F G x Get the query geometry weighting coefficient, which is expressed as W G x ; According to the retrieval geometric feature F G y Get the retrieval geometry weighting coefficient, which is expressed as W G x ; According to the query semantic feature F S x Get the query semantic weighting coefficient, which is expressed as W S x ; According to the retrieval semantic feature F S y Get the retrieval semantic weighting coefficient, which is expressed as W S x ; expressed as: W G x =softmax(MLP(LN(MLP(F G x )))) W G y =softmax(MLP(LN(MLP(F G y )))) W S x =softmax(MLP(LN(MLP(F S x )))) W S y =softmax(MLP(LN(MLP(F S y )))) In the formula, softmax() represents the softmax function processing; Step 4415: Based on the query geometric feature F G x and query semantic features F S x Get the query connection feature; the query connection feature is represented by F m x ; The geometric feature F will be retrieved G y and retrieve semantic features F S y Get the retrieval connection feature; the retrieval connection feature is represented by F m y ; expressed as: F m x =LN(MLP([F G x ,F S x ])) F m y =LN(MLP([F G y ,F S y ])) Where [·,·] represents the concatenation of features along the channel dimension; Step 4416: Connect feature F based on query m x and the query geometry weight coefficient is denoted as W G x The query geometry weighted fusion feature is obtained, and the query geometry weighted fusion feature is expressed as F fg x ; Connect feature F according to query m x and the query semantic weight coefficient is denoted as W S x The query semantic weighted fusion feature is obtained, and the query semantic weighted fusion feature is expressed as F fs x ; According to the retrieval connection feature F m y and the retrieval geometric weight coefficient is denoted as W G y The retrieval geometry weighted fusion feature is obtained, and the retrieval geometry weighted fusion feature is expressed as F fg y ; According to the retrieval connection feature F m y and the retrieval semantic weight coefficient is expressed as W S y The retrieval semantic weighted fusion feature is obtained, and the retrieval semantic weighted fusion feature is expressed as F fs y ; expressed as: F fg x =F m x ⊙W G x +F G x F fs x =F m x ⊙W S x +F S x F fg y =F m y ⊙W G y +F G y F fs y =F m y ⊙W S y +F S y In the formula, ⊙ represents element-by-element multiplication processing; Step 4417: Weighted fusion features F according to query geometry fg x and query semantic weighted fusion feature F fs x Obtain query fusion features; The query fusion feature is expressed as According to the retrieval geometry weighted fusion feature F fg y The weighted fusion feature of the retrieval semantics is represented as F fs y Get the retrieval fusion feature, which is expressed as 6. The method for laser relocalization of an intelligent mobile robot based on dense point cloud registration according to claim 5, characterized in that: In step 4417, the weighted fusion feature F is obtained according to the query geometry. fg x and query semantic weighted fusion feature F fs x Get query fusion features; weight fusion features F according to retrieval geometry fg y The weighted fusion feature of the retrieval semantics is represented as F fs y Get the retrieval fusion features. The specific process is as follows: Step 44171: Weighted fusion of query geometry features F fg x Sum in the second dimension to get the query intermediate feature F fg1 x ; Weight the query semantics and fuse the features F fs x Sum in the second dimension to get the query intermediate feature F fs2 x ; Retrieve geometric weighted fusion features F fg y Sum in the second dimension to get the retrieval intermediate feature F fg1 y ; The retrieval semantic weighted fusion feature F fs y Sum in the second dimension to get the retrieval intermediate feature F fs2 y ; Step 44172: Query the intermediate feature F fg1 x and query intermediate features F fs1 x Get the query fusion feature F f x ; According to the retrieval intermediate feature F fg1 y and retrieve the intermediate feature F fs2 y Get the retrieval fusion feature F f y ; expressed as:

7. The method for laser relocalization of an intelligent mobile robot based on dense point cloud registration according to claim 6, characterized in that: In step 443, the query downsampled point cloud obtained in step 42 is Retrieve the downsampled point cloud Enhanced query feature F sa x and enhanced retrieval query features F sa y At the same time, the anchor-guided cross-attention mechanism module is input to obtain feature-enhanced query point cloud features and feature-enhanced retrieval point cloud features; The specific process is: Step 4431: Use the KNN algorithm to process the query downsampled point cloud obtained in step 42 Get the K0 closest points of each query point cloud sampling point; the K0 closest points of each query point cloud sampling point form the query point cloud sampling point closest point cloud, expressed as According to the enhanced query feature F sa x Calculate the inlier similarity subvector of each query sampling point; combine the inlier similarity subvectors of each query sampling point into the query inlier similarity vector, expressed as Step 4432: Enhance the query feature F sa x Perform enhanced search query feature F sa y Inner product processing is performed to obtain the soft allocation matrix A, which is expressed as According to the soft assignment matrix A, the response point cloud and the external point similarity vector are constructed; the response point cloud is expressed as The outlier similarity vector is expressed as Step 4433: Based on the inlier similarity vector S ra and the external point similarity vector S er Construct a significance score; the significance score is expressed as The formula is: According to the significance score S s In response to the point cloud P ag y Select K1 initial anchor points of each query sampling point, and form an initial anchor point cloud with the K1 initial anchor points of each query sampling point; The initial anchor point cloud is represented as Use KNN algorithm to process the response point cloud P ag y , get the K1 closest points of each initial anchor point; get the K1 closest points of each initial anchor point to form an extended anchor point cloud, the extended anchor point cloud is expressed as Step 4434: Expand the anchor point cloud P an y And query the sampled point cloud Construct a point cloud mask, which is represented by Step 4435: Based on the enhanced query feature F sa x , Enhanced retrieval query features F sa y and point cloud mask M xy Generate intermediate feature Info; use the FNN network to process the intermediate feature Info to obtain feature-enhanced query point cloud features and feature-enhanced retrieval point cloud features.

8. The intelligent mobile robot laser relocation method based on dense point cloud registration according to claim 7, characterized in that: In step 4431, according to the enhanced query feature F sa x The specific process of calculating the inlier similarity subvector of the sampling point under the pth query is: Step 44311: Based on the enhanced query feature F sa x Find the enhanced query features corresponding to the sampling point under the p-th query and the enhanced query features corresponding to the K0 nearest points of the sampling point under the p-th query, Step 44312: Calculate the inner product between the enhanced query feature corresponding to the sampling point under the p-th query and the enhanced query feature corresponding to the K0 nearest points under the sampling point under the p-th query, obtain the inner product value of the K0 query sampling points, and use the inner product value of the K0 query sampling points to form the inner point similarity subvector of the sampling point under the p-th query The specific process of constructing the response point cloud and the external point similarity according to the soft allocation matrix A in step 4432 is as follows: The element A(i′,j′) in the i′th row and j′th column of the soft assignment matrix A is used as the query downsampled point cloud The i′th query sampling point and the retrieval sampling point cloud The matching probability of the sampling point under the j′th retrieval; Find P ag x The maximum matching probability of each point in P ag x The maximum matching probability of each point in the composes the external point similarity S er , Find P ag x Find the maximum matching probability for each point in the sampled point cloud under retrieval The corresponding point in P ag x The response point of each point in P ag x The response points of each point in form a response point cloud P ag y ; In step 4434, according to the expanded anchor point cloud P an y And query the sampled point cloud Construct a point cloud mask, which is represented by The specific process is: The point cloud mask M xy Corresponding to the expanded anchor point cloud P an y And query the sampled point cloud The value of the element is 1; the values ​​of other elements are 0; the point cloud mask M is obtained xy ; In step 4435, according to the enhanced query feature F sa x , Enhanced retrieval query features F sa y and point cloud mask M xy Generate intermediate feature Info; The formula is: Where Q represents the MLP processing F sa x The query tensor generated after, K represents the MLP processing F sa y The key tensor generated after, V represents the MLP processing F sa y The value tensor generated later, K T represents the transpose of K, Represents the feature dimension.

9. The intelligent mobile robot laser relocation method based on dense point cloud registration according to claim 8, characterized in that: In step 45, the final query point cloud feature F obtained in step 44 is att x , finally retrieve point cloud feature F att y , Query local point cloud block G x , The query local features F gx , Retrieve local point cloud blocks G y as well as The local features F are retrieved gy Input the two-stage matching module, and output the dense point cloud matching relationship between the query point cloud and the retrieval point cloud; the specific process is: Step 451: Based on the final query point cloud feature F att x And the final retrieved point cloud feature F att y Perform downsampling point matching to obtain downsampling point matching results; The specific process is: Step 4511: Based on the final query point cloud feature F att x And the final retrieved point cloud feature F att y Calculate the Gaussian correlation matrix; the Gaussian correlation matrix is ​​expressed as Among them, the element G(i″,j″) in the i″th row and j″th column of the Gaussian correlation matrix G is expressed by the formula: In the formula, norm() represents normalization processing, represents the final query point cloud feature of the i″th query sampling point, represents the final retrieval point cloud features of the sampling point under the j″th retrieval, It means calculating the second norm and then squaring it, exp() means the exponential function, Step 4512: Get the downsampling point matching result H according to the Gaussian correlation matrix G c , expressed as: In the formula, Indicates the matching relationship between the sampling points in the query point cloud in the i″th row and the sampling points in the query point cloud in the j″th column. topk() indicates selecting the largest top K su Elements; K su is a positive integer; Step 452: Match the result H according to the downsampling point c , Query local point cloud block G x , The query local features F gx ; Retrieve local point cloud blocks G y as well as The local features F are retrieved gy Perform original point matching to obtain the dense point cloud matching relationship H between the query point cloud and the retrieval point cloud f ; Dense point cloud matching relationship H between query point cloud and retrieval point cloud f In the i″th query local point cloud block and the j″th retrieved local point cloud block The dense point cloud matching relationship between Among them, according to the downsampling point matching results, Query local point cloud block G x , The query local features F gx ; Retrieve local point cloud blocks G y as well as The local features F are retrieved gy Perform original point matching to obtain query local point cloud blocks and retrieve local point cloud blocks Dense point cloud matching relationship between The specific process is: Step 4521: Match the results based on the downsampling points Determine the corresponding query local point cloud block Query local features Retrieve local point cloud blocks and retrieve local features According to the i″th query local feature and the j″th retrieved local feature Constructing the cost matrix The formula is: In the formula, express The transpose of , τ represents the temperature factor, τ = 0.1; Step 4522: Use SinkHorn algorithm to process the cost matrix Get the soft assignment matrix Step 4523: Delete the soft allocation matrix The last row and last column of , we get the soft assignment matrix Step 4524: According to the soft allocation matrix Query local point cloud blocks and retrieve local point cloud blocks Perform matching and obtain matching results; the specific process is: like The Points and The Points The corresponding soft assignment matrix Elements When the matching conditions are met, and As a result of matching The matching conditions are: element The value of is No. One of the three largest element values ​​in the row; The value of is No. One of the three largest element values ​​in the column; The value of is greater than a predefined threshold, the predefined threshold is 0.05; Step 4525: Combine all matching results into query local point cloud blocks and retrieval local point cloud blocks Dense point cloud matching relationship between 10. The intelligent mobile robot laser relocation method based on dense point cloud registration according to claim 9, characterized in that: The laser radar data environment recognition method in step 5 is the ScanContext method; In step 7, the power-on posture of the intelligent mobile robot is calculated based on the relative posture of the power-on query point cloud and the power-on retrieval point cloud in step 6 and the posture of the power-on retrieval point cloud, and the formula is expressed as: T X =T XY T Y Where, T X represents the startup posture of the intelligent mobile robot; represents the startup retrieval point cloud posture T Y , represents the relative pose T of the startup query point cloud and the startup retrieval point cloud XY .

Images