Robust point cloud registration method and system based on invariant and unique features

Abstract

This invention discloses a robust point cloud registration method and system based on invariant and unique features. This system utilizes deep learning techniques to construct a trainable feature extraction and matching model. The proposed feature extraction module helps the network efficiently learn local features and focus on surrounding, distinguishing landmark features, thereby improving feature matching accuracy and point cloud registration precision. This technical solution can achieve high-precision registration in low-overlap indoor and outdoor scenarios, offering high accuracy, strong robustness, and applicability to a variety of indoor and outdoor scenarios.

Description

Technical Field

[0001] The present invention belongs to the field of mobile robots, and in particular relates to a robust point cloud registration method and system based on invariant and unique features. Background Art

[0002] Point cloud registration is an important technical foundation in fields such as SLAM, 3D reconstruction, and augmented reality. It refers to the process of estimating the relative pose relationship between two frames of partially overlapping point cloud scenes or targets. General sensors such as depth cameras and lidar can only obtain point cloud information from a partial perspective or local scene. In order to obtain point cloud information from a complete perspective or global scene, point cloud registration is necessary. In addition, obtaining the relative pose relationship of robots through point cloud registration technology is also crucial for various intelligent unmanned systems to achieve other advanced tasks. Sometimes, image registration can also obtain the relative pose relationship of robots. However, because image registration is extremely sensitive to illumination changes, requires rich color and texture features, and has difficulty estimating depth information, point cloud registration is widely used in scenes such as cultural relic reconstruction, oral modeling, and at night. However, in actual scenarios, point cloud registration mainly faces the following challenges: 1) Noise interference, due to environmental interference or sensor deficiencies, a large amount of noise will be generated; 2) Non-overlapping feature interference, non-overlapping point clouds will interfere with matching and cause mismatches; 3) Repetitive structure interference, repetitive structures will produce similar features, which will in turn cause incorrect feature matching; 4) Occlusion and missing, perspective changes or dynamic objects will cause occlusion, causing some features to disappear, and factors such as exceeding the sensor's working range will cause point cloud missing; 5) Data resolution changes, the inherent close-close and far-sparse characteristics of lidar or the different resolutions of different sensors will cause the impact. Depending on whether feature point pairs are required, point cloud registration is mainly divided into: direct registration method and correspondence-based registration method. Direct registration method generally does not require the extraction of local feature descriptors, but learns global features through the network, and uses a pose estimation network to directly regress and generate transformed poses to complete point cloud registration. This method of regressing relative pose from global features is a rough point cloud registration method. The advantage of this method is that it is robust to changes in point cloud resolution, and the complexity of the algorithm does not increase significantly with the increase in the number of points. However, this type of method is often highly dependent on the high overlap of point cloud data, and is therefore not applicable to the low-overlap scene registration problem that is currently attracting much attention. With the continuous development of point cloud registration technology, researchers have begun to pay attention to the successful registration of point clouds in low-overlap scenes. Correspondence-based registration methods have attracted more and more attention from scholars. This type of method first extracts the correspondence between key points in a pair of point clouds, and then performs point cloud registration through robust pose estimation. GeoTransformer is the most representative model. It extracts local features through a pyramid-structured convolutional feature extractor and enhances local features using a global transformer with geometric position encoding. Subsequently, a coarse-to-fine feature matching framework is adopted to obtain dense point cloud correspondences, and then point cloud registration is completed through a robust pose estimator.This model significantly improves the success rate of point cloud registration in low-overlap scenarios. However, in some scenarios with extremely low overlap or many repetitive structures, the obtained dense point cloud correspondences still contain many erroneous correspondences, which in turn affects the accuracy or success rate of registration.

[0003] In summary, how to design a feature extractor to obtain more correct point correspondences and thus complete point cloud registration more accurately and robustly is a technical challenge that needs to be solved urgently. Therefore, it is necessary to design a new point cloud registration method. Summary of the Invention

[0004] In response to the defects in the existing technology, a robust point cloud registration method and system based on invariant and unique features are provided. This technical solution is based on the point cloud registration method of multi-region MLP and Transformer joint learning, which has the advantages of high accuracy, strong robustness, and is suitable for a variety of indoor and outdoor scenarios.

[0005] The technical solutions provided by the present invention are as follows:

[0006] On the one hand, a robust point cloud registration method based on invariant and unique features, comprising:

[0007] Step 1: Construct the RID-Net model;

[0008] The RID-Net model includes a connected feature extractor and a coarse-to-fine point matcher. The feature extractor includes four AMR modules, a global Transformer module, and four feature forward propagation modules.

[0009] The AMR module includes an Abstract layer, an MLP module enhanced by local rotation-invariant position coding, and a ring attention module connected in sequence;

[0010] Step 2: Model training;

[0011] Multiple pairs of point cloud data and the corresponding relative pose truth values ​​of a scene are obtained as training data, where the relative pose truth values ​​are converted into super-point matching relationships and dense point matching relationships. Each pair of point cloud data is input into the constructed RID-Net model in turn, and the super-point matching relationships and dense point matching relationships of each pair of point clouds predicted by the RID-Net model are used to obtain the trained RID-Net model parameters when the matching loss converges.

[0012] Step 3: Model registration;

[0013] The collected point cloud data is input into the trained RID-Net model to predict the super-point matching relationship and dense point matching relationship between the two frames of point clouds. Based on the obtained matching relationship, the relative pose of the two frames of point clouds is solved by a robust pose estimator. The source point cloud data is then converted to the target point cloud coordinate system through the relative pose to complete the alignment.

[0014] The robust pose estimator relies entirely on traditional geometric knowledge to solve the robust registration pose based on the predicted matching relationship without the need for training;

[0015] Furthermore, the specific working process of the MLP module based on local rotation-invariant position coding enhancement is as follows:

[0016] Step A1: Obtain the local neighborhood point coordinates and aggregation features of each point;

[0017] The initial feature of each point is set to 1, and the aggregated feature of each point is obtained by aggregating the initial features through the Abstract layer;

[0018] Step A2: Calculate the reference axis perpendicular to the normal line of each point;

[0019]

[0020] in, represents the i-th point processed on the first layer AMR module, K l represents the number of neighborhood points used by the AMR module in the first layer, k l Indicates the kth layer of the lth layer l Neighborhood points, Indicated by point point to vector, express kth l Neighborhood points, express The neighborhood point set of is the calculated perpendicular to the point Normal vector, r is set to The distance to the farthest neighbor point, v i is the final weighted perpendicular to The reference axis of the normal; each point The normal of

[0021] Step A3: Based on the reference axis and normal, construct a high-dimensional rotation-invariant position encoding of each point and its neighboring points;

[0022]

[0023] Among them, HRIPE() represents the high-dimensional rotation-invariant position encoding operation, and rpe() represents the rotation-invariant position representation; Indicates a point and High-dimensional rotation-invariant position encoding between Indicates that the input feature dimension is 5, and the feature dimension of the output of the lth AMR module is C l The linear layer operation, Norm represents the normalization operation, Indicates that the input feature dimension is C l , the feature dimension of the output of the lth AMR module is C l Linear layer operation;

[0024] Step A4: Linearly transform the high-dimensional rotation-invariant position encoding of all neighboring points of each point and the first 1 / 4 of the channels after adding the downsampled and aggregated sampling point features of each point through the Abstract layer;

[0025] Step A5: The result of the linear transformation in step A4 is aggregated using a maximum pooling layer, and then two linear layers are used to learn deeper features.

[0026] Step A6: Perform a residual connection between the features obtained in step A5 and the features obtained by aggregation in the Abstract layer to obtain the updated features to alleviate the gradient disappearance.

[0027] Furthermore, the working process of the ring attention module is as follows:

[0028] Step B1: The input of the ring attention module is the set of target points to be processed The set of annular neighborhood points around the target point is recorded as and the corresponding feature set

[0029] in, Indicates distance point In addition to the recent K l The nearest point outside Point index, feature set According to this index Indexed from the features obtained from the MLP module based on the local rotation-invariant position encoding enhancement;

[0030] Step B2: According to the high-dimensional rotation-invariant position encoding operation, Relative to The high-dimensional rotation-invariant position encoding of Using five learnable matrices W G ,W E ,W Q ,WK and W V To project the input, high-dimensional rotation-invariant position encoding By W G ,W E It is further encoded into geometric code G and position code E, and then the new features of the target point in the lth layer are obtained by weighting the geometric code and the value

[0031]

[0032] Among them, G and E are the learned geometric and positional encodings respectively, and the learned context features can be projected into q, K and V:

[0033]

[0034] in, is the feature of the i-th point in the l-th layer, F l It is the point feature updated by the MLP module based on the local rotation-invariant position encoding enhancement, and c0 is the feature dimension size of G and V;

[0035] Step B3: Through linear layers and residual connections and After adding and updating, we get the final feature F of the target point in the lth layer. l .

[0036] The ring attention module senses the ring area around the target point and focuses on the surrounding landmark structures, making locally similar features unique;

[0037] Furthermore, the four AMR modules of the feature extractor are connected in sequence, wherein the downsampling coefficients in each AMR module are [1, 0.25, 0.25, 0.25] in sequence; the feature dimensions of each AMR module are [64, 128, 256, 256] in sequence; the number of neighborhood points is set to [8, 16, 16, 16] in sequence, the last AMR module is connected to the global Transformer module, each AMR module is residually connected to a feature forward propagation module, and the four feature forward propagation modules are connected in reverse sequence, the input end of the first feature forward propagation module and the input end of the coarse-to-fine point matcher are both connected to the output end of the global Transformer module, and the output end of the last feature forward propagation module and the global Transformer module are connected to the input end of the coarse-to-fine point matcher, thereby obtaining the predicted super-point matching relationship and dense point matching relationship through the coarse-to-fine point matcher.

[0038] Furthermore, the matching loss is calculated by the overlap-aware circle loss and the negative log-likelihood loss. The matching loss is obtained by adding the overlap-aware circle loss and the negative log-likelihood loss in a one-to-one ratio.

[0039] The overlap-aware circle loss is calculated using the predicted value of the super-point matching relationship and the super-point matching relationship derived from the pose truth as input;

[0040] The negative log-likelihood loss is calculated by taking the dense point matching relationship prediction value and the pose truth value as input.

[0041] In the second aspect, a robust point cloud registration system based on invariant and unique features is provided, comprising:

[0042] RID-Net model construction unit: The RID-Net model includes a connected feature extractor and a coarse-to-fine point matcher. The feature extractor includes four AMR modules, a global Transformer module, and four feature forward propagation modules.

[0043] The AMR module includes an Abstract layer, an MLP module enhanced by local rotation-invariant position coding, and a ring attention module connected in sequence;

[0044] Model training unit: By obtaining multiple pairs of point cloud data and the corresponding relative pose truth values ​​of a certain scene as training data, where the relative pose truth values ​​are converted into super-point matching relationships and dense point matching relationships; each pair of point cloud data is input into the constructed RID-Net model in turn, and the super-point matching relationship and dense point matching relationship of each pair of point clouds predicted by the RID-Net model are used to obtain the trained RID-Net model parameters when the matching loss converges (in practice, 40 rounds for indoor data and 80 rounds for outdoor data);

[0045] Model registration unit: The collected point cloud data is input into the trained RID-Net model, and the super-point matching relationship and dense point matching relationship corresponding to the two frames of point clouds are predicted. Based on the obtained matching relationship, the relative pose of the two frames of point clouds is solved by a robust pose estimator, and then the source point cloud data is converted to the target point cloud coordinate system through the relative pose to complete the registration.

[0046] Furthermore, the specific working process of the MLP module based on local rotation-invariant position coding enhancement is as follows:

[0047] Step A1: Obtain the local neighborhood point coordinates and aggregation features of each point;

[0048] The initial feature of each point is set to 1, and the aggregated feature of each point is obtained by aggregating the initial features through the Abstract layer;

[0049] Step A2: Calculate the reference axis perpendicular to the normal line of each point;

[0050]

[0051] in, represents the i-th point processed on the first layer AMR module, K l represents the number of neighborhood points used by the AMR module in the first layer, k l Indicates the kth layer of the lth layer l Neighborhood points, Indicated by point point to vector, express kth l Neighborhood points, express The neighborhood point set of is the calculated perpendicular to the point Normal vector, r is set to The distance to the farthest neighbor point, v i is the final weighted perpendicular to The reference axis of the normal; each point The normal of

[0052] Step A3: Based on the reference axis and normal, construct a high-dimensional rotation-invariant position encoding of each point and its neighboring points;

[0053]

[0054] Among them, HRIPE() represents the high-dimensional rotation-invariant position encoding operation, and rpe() represents the rotation-invariant position representation; Indicates a point and High-dimensional rotation-invariant position encoding between Indicates that the input feature dimension is 5, and the feature dimension of the output of the lth AMR module is C l The linear layer operation, Norm represents the normalization operation, Indicates that the input feature dimension is C l , the feature dimension of the output of the lth AMR module is C l Linear layer operation;

[0055] Step A4: Linearly transform the high-dimensional rotation-invariant position encoding of all neighboring points of each point and the first 1 / 4 of the channels after adding the downsampled and aggregated sampling point features of each point through the Abstract layer;

[0056] Step A5: The result of the linear transformation in step A4 is aggregated using a maximum pooling layer, and then two linear layers are used to learn deeper features.

[0057] Step A6: Perform a residual connection between the features obtained in step A5 and the features obtained by aggregation in the Abstract layer to obtain the updated features to alleviate the gradient disappearance.

[0058] Furthermore, the working process of the ring attention module is as follows:

[0059] Step B1: The input of the ring attention module is the set of target points to be processed The set of annular neighborhood points around the target point is recorded as and the corresponding feature set in, Indicates distance point In addition to the recent K l The nearest point outside Point index, feature set According to this index Indexed from the features obtained from the MLP module based on the local rotation-invariant position encoding enhancement;

[0060] Step B2: According to the high-dimensional rotation-invariant position encoding operation, Relative to The high-dimensional rotation-invariant position encoding of Using five learnable matrices W G ,W E ,W Q ,W K and W V To project the input, high-dimensional rotation-invariant position encoding By W G ,W E It is further encoded into geometric code G and position code E, and then the new features of the target point in the lth layer are obtained by weighting the geometric code and the value

[0061]

[0062] Among them, G and E are the learned geometric and positional encodings respectively, and the learned context features can be projected into q, K and V:

[0063]

[0064] in, is the feature of the i-th point in the l-th layer, F lIt is the point feature updated by the MLP module based on the local rotation-invariant position encoding enhancement, and c0 is the feature dimension size of G and V;

[0065] Step B3: Through linear layers and residual connections and After adding and updating, we get the final feature F of the target point in the first layer. l .

[0066] According to a third aspect, a computer device includes at least:

[0067] one or more processors;

[0068] a memory storing one or more computer programs;

[0069] The processor calls the computer program to implement:

[0070] The above steps of a robust point cloud registration method based on invariant and unique features.

[0071] In a fourth aspect, a computer-readable storage medium stores a computer program, wherein the computer program is invoked by a processor to implement:

[0072] The above steps of a robust point cloud registration method based on invariant and unique features.

[0073] Beneficial effects

[0074] This invention provides a robust point cloud registration method and system based on invariant and unique features. It utilizes deep learning techniques to construct a trainable feature extraction and matching model. The proposed feature extraction module helps the network efficiently learn local features and focus on surrounding, distinguishing landmark features, thereby improving feature matching accuracy and point cloud registration precision. This technical solution can achieve high-precision registration in low-overlap indoor and outdoor scenarios, offering high accuracy, strong robustness, and applicability to a variety of indoor and outdoor scenarios.

[0075] Compared with the existing point cloud registration method based on deep learning, the technical solution of the present invention is advanced in the following aspects:

[0076] 1. The technical solution of this invention proposes an MLP feature encoder enhanced with rotation-invariant position encoding. By mapping the rotation-invariant relative position into the feature space and fusing it with the point features, the network's ability to extract local features is effectively enhanced, thereby improving the accuracy of point cloud registration.

[0077] 2. The technical solution of this invention proposes a feature enhancement strategy that fuses MLP local features with an attention mechanism. First, the MLP module enhanced by rotation-invariant position encoding efficiently extracts local features from the 0-k nearest neighbors around the feature point. Then, the attention mechanism extracts and fuses the discriminative landmark features within the k-2k nearest neighbors of the feature point. This enables the network to accurately extract local features while paying more attention to the surrounding discriminative structural features, thereby significantly improving the point matching inlier rate.

[0078] 3. Based on the efficient local feature extraction capability and surrounding feature enhancement strategy of the technical solution of the present invention, the test results of the technical solution of the present invention on indoor and outdoor datasets show that in indoor scenarios, the technical solution of the present invention exceeds the currently advanced Geotransformer and RolTr in terms of matching inlier rate and registration recall rate. In outdoor autonomous driving scenarios, the technical solution of the present invention still has advantages in registration accuracy and success rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 It is the overall flow chart of the present invention;

[0080] Figure 2 This is the specific network architecture of the network model RID-Net proposed in this invention. DETAILED DESCRIPTION

[0081] The present invention is further described in detail below with reference to the accompanying drawings and embodiments:

[0082] Example 1

[0083] like Figure 1 As shown, a robust point cloud registration method based on invariant and unique features includes:

[0084] Step 1: Construct the RID-Net model;

[0085] The RID-Net model includes a connected feature extractor and a coarse-to-fine point matcher. The feature extractor includes four AMR modules, a global Transformer module, and four feature forward propagation modules.

[0086] The AMR module includes an Abstract layer, an MLP module enhanced by local rotation-invariant position coding, and a ring attention module connected in sequence;

[0087] The specific working process of the MLP module based on local rotation-invariant position coding enhancement is as follows:

[0088] Step A1: Obtain the local neighborhood point coordinates and aggregation features of each point;

[0089] The initial feature of each point is set to 1, and the aggregated feature of each point is obtained by aggregating the initial features through the Abstract layer;

[0090] Step A2: Calculate the reference axis perpendicular to the normal line of each point;

[0091]

[0092] in, represents the i-th point processed on the first layer AMR module, K l represents the number of neighborhood points used by the AMR module in the first layer, k l Indicates the kth layer of the lth layer l Neighborhood points, Indicated by point point to vector, express kth l Neighborhood points, express The neighborhood point set of is the calculated perpendicular to the point Normal vector, r is set to The distance to the farthest neighbor point, v i is the final weighted perpendicular to The reference axis of the normal; each point The normal of Step A3: Based on the reference axis and normal, construct a high-dimensional rotation-invariant position encoding of each point and its neighboring points;

[0093]

[0094] Among them, HRIPE() represents the high-dimensional rotation-invariant position encoding operation, and rpe() represents the rotation-invariant position representation; Indicates a point and High-dimensional rotation-invariant position encoding between Indicates that the input feature dimension is 5, and the feature dimension of the output of the lth AMR module is C l The linear layer operation, Norm represents the normalization operation, Indicates that the input feature dimension is C l , the feature dimension of the output of the lth AMR module is C l Linear layer operation;

[0095] Step A4: Linearly transform the high-dimensional rotation-invariant position encoding of all neighboring points of each point and the first 1 / 4 of the channels after adding the downsampled and aggregated sampling point features of each point through the Abstract layer;

[0096] Step A5: The result of the linear transformation in step A4 is aggregated using a maximum pooling layer, and then two linear layers are used to learn deeper features.

[0097] Step A6: Perform a residual connection between the features obtained in step A5 and the features obtained by aggregation in the Abstract layer to obtain the updated features to alleviate the gradient disappearance.

[0098] The working process of the ring attention module is as follows:

[0099] Step B1: The input of the ring attention module is the set of target points to be processed The set of annular neighborhood points around the target point is recorded as and the corresponding feature set in, Indicates distance point In addition to the recent K l The nearest point outside Point index, feature set According to this index Indexed from the features obtained from the MLP module based on the local rotation-invariant position encoding enhancement;

[0100] Step B2: According to the high-dimensional rotation-invariant position encoding operation, Relative to The high-dimensional rotation-invariant position encoding of Using five learnable matrices W G ,W E ,W Q ,W K and W V To project the input, high-dimensional rotation-invariant position encoding By W G ,W E It is further encoded into geometric code G and position code E, and then the new features of the target point in the lth layer are obtained by weighting the geometric code and the value

[0101]

[0102] Among them, G and E are the learned geometric and positional encodings respectively, and the learned context features can be projected into q, K and V:

[0103]

[0104] in, is the feature of the i-th point in the l-th layer, F lIt is the point feature updated by the MLP module based on the local rotation-invariant position encoding enhancement, and c0 is the feature dimension size of G and V;

[0105] Step B3: Through linear layers and residual connections and After adding and updating, we get the final feature F of the target point in the lth layer. l The ring attention module senses the ring area around the target point to focus on the surrounding landmark structures, making the locally similar features unique;

[0106] The four AMR modules of the feature extractor are connected in sequence, wherein the downsampling coefficients in each AMR module are [1, 0.25, 0.25, 0.25] in sequence; the feature dimensions of each AMR module are [64, 128, 256, 256] in sequence; the number of neighborhood points is set to [8, 16, 16, 16] in sequence, the last AMR module is connected to the global Transformer module, each AMR module is residually connected to a feature forward propagation module, and the four feature forward propagation modules are connected in reverse sequence, the input end of the first feature forward propagation module and the input end of the coarse-to-fine point matcher are both connected to the output end of the global Transformer module, and the output end of the last feature forward propagation module and the global Transformer module are connected to the input end of the coarse-to-fine point matcher, thereby obtaining the predicted super-point matching relationship and dense point matching relationship through the coarse-to-fine point matcher.

[0107] Step 2: Model training;

[0108] Multiple pairs of point cloud data and the corresponding relative pose truth values ​​of a certain scene are obtained as training data, where the relative pose truth values ​​are converted into super-point matching relationships and dense point matching relationships. Each pair of point cloud data is input into the constructed RID-Net model in turn, and the super-point matching relationship and dense point matching relationship of each pair of point clouds predicted by the RID-Net model are used to match the convergence of the matching loss. In actual training, 40 rounds of iterative matching are performed on indoor data and 80 rounds of iterative matching are performed on outdoor data to obtain the trained RID-Net model parameters.

[0109] The matching loss is calculated by overlapping circle loss and negative log-likelihood loss. The matching loss is obtained by adding the overlapping circle loss and the negative log-likelihood loss in a one-to-one ratio.

[0110] The overlap-aware circle loss is calculated using the predicted value of the super-point matching relationship and the super-point matching relationship derived from the pose truth as input;

[0111] The negative log-likelihood loss is calculated by taking the dense point matching relationship prediction value and the pose truth value as input.

[0112] The Adam optimizer is used for training, and the weight decay factor is set to 10. -6 , the batch size is set to 4, for the 3DMatch dataset, the number of training rounds is set to 40 rounds, and the initial learning rate is 10 -4 The learning rate is decayed by 0.05 per round. For the KITTI dataset, the number of training rounds is set to 80 and the initial learning rate is 10 -4 The training result of the last round is taken as the optimal model.

[0113] Step 3: Model registration;

[0114] The collected point cloud data is input into the trained RID-Net model to predict the super-point matching relationship and dense point matching relationship between the two frames of point clouds. Based on the obtained matching relationship, the relative pose of the two frames of point clouds is solved by a robust pose estimator. The source point cloud data is then converted to the target point cloud coordinate system through the relative pose to complete the alignment.

[0115] The robust pose estimator relies entirely on traditional geometric knowledge to solve the robust registration pose based on the predicted matching relationship without the need for training;

[0116] The experimental results of the present invention are shown in Table 1. Three public datasets containing indoor and outdoor scene types are used to test the present invention. The error between the relative pose prediction values ​​of all the point cloud pairs to be registered in a certain scene and the true values ​​provided by the dataset is used to evaluate the experimental results.

[0117] Table 1 shows the comparative experimental results based on the 3DMatch dataset and the 3DLoMatch dataset, which are derived from the same indoor dataset. This dataset collects data from 62 scenes, 54 of which are used for training and 8 for evaluation. 3DLoMatch, for example, is more challenging because its point cloud pairs overlap less than 30%. As shown in the table, the proposed method achieves a higher registration success rate (RR) and matching inlier rate (IR) in all indoor scenes, while also achieving a high feature match recall rate (FMR).

[0118] Table 1 Performance comparison of the proposed algorithm and existing algorithms on indoor datasets

[0119]

[0120]

[0121] Table 2 shows the effects of the proposed method when sampling different numbers of matching points. As shown in the table, the proposed method is still stable and has excellent effects when there are only a small number of corresponding points.

[0122] Table 2 Performance comparison of the proposed method and existing algorithms on indoor datasets with different numbers of matching points

[0123]

[0124]

[0125] Table 3 shows the results of comparative experiments based on the KITTI dataset, currently the largest international dataset for evaluating computer vision algorithms in autonomous driving scenarios. The point cloud data in this dataset is collected by a vehicle-mounted Velodyne 64-line 3D lidar, and the ground truth data is provided by a GPS navigation system. The dataset includes real road scenes collected in urban, rural, and highway environments. As shown in the table, the proposed method still has a significant advantage on outdoor scene data.

[0126] Table 3 KITTI dataset algorithm effect

[0127]

[0128] According to the above experimental analysis, the algorithm of the present invention is superior to existing algorithms in terms of registration accuracy and robustness, and is also applicable to various indoor and outdoor scenes.

[0129] Example 3

[0130] Corresponding to the aforementioned method of this embodiment, this example further provides a computer device, comprising at least:

[0131] one or more processors;

[0132] a memory storing one or more computer programs;

[0133] The processor calls the computer program to implement:

[0134] The above steps of a robust point cloud registration method based on invariant and unique features.

[0135] The specific implementation process of each step can refer to the detailed process of the aforementioned method, and the present invention will not elaborate on this in detail.

[0136] Example 4

[0137] Corresponding to the aforementioned method of this embodiment, this example further provides a computer storage medium storing a computer program, which is called by a processor to execute:

[0138] The above steps of a robust point cloud registration method based on invariant and unique features.

[0139] The specific implementation process of each step can refer to the detailed process of the aforementioned method, and the present invention will not elaborate on this in detail.

[0140] Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned readable storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0141] It should be understood that in the embodiments of the present invention, the processor referred to may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The memory may include a read-only memory and a random access memory, and provides instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0142] It should be emphasized that the examples described in the present invention are illustrative rather than restrictive. Therefore, the present invention is not limited to the examples described in the specific embodiments. Any other embodiments derived by those skilled in the art based on the technical solution of the present invention that do not depart from the purpose and scope of the present invention, whether modified or replaced, also fall within the scope of protection of the present invention.

Claims

1. A robust point cloud registration method based on invariant and unique features, characterized in that include: Step 1: Construct the RID-Net model; The RID-Net model includes a connected feature extractor and a coarse-to-fine point matcher. The feature extractor includes four AMR modules, a global Transformer module, and four feature forward propagation modules. The four AMR modules are connected in sequence, and the last AMR module is connected to the global Transformer module. Each AMR module is residually connected to a feature forward propagation module. The four feature forward propagation modules are connected in reverse order. The input of the first feature forward propagation module and the input of the coarse-to-fine point matcher are connected to the output of the global Transformer module. The output of the last feature forward propagation module and the global Transformer module are connected to the input of the coarse-to-fine point matcher. The AMR module includes an Abstract layer, an MLP module enhanced by local rotation-invariant position coding, and a ring attention module connected in sequence; Step 2: Model training; Multiple pairs of point cloud data and the corresponding relative pose truth values ​​of a scene are obtained as training data, where the relative pose truth values ​​are converted into super-point matching relationships and dense point matching relationships. Each pair of point cloud data is input into the constructed RID-Net model in turn, and the super-point matching relationships and dense point matching relationships of each pair of point clouds predicted by the RID-Net model are used to obtain the trained RID-Net model parameters when the matching loss converges. Step 3: Model registration; The collected point cloud data is input into the trained RID-Net model to predict the super-point matching relationship and dense point matching relationship between the two frames of point clouds. Based on the obtained matching relationship, the relative pose of the two frames of point clouds is solved by a robust pose estimator. The source point cloud data is then converted to the target point cloud coordinate system through the relative pose to complete the registration; The specific working process of the MLP module based on local rotation-invariant position coding enhancement is as follows: Step A1: Obtain the local neighborhood point coordinates and aggregation features of each point; The initial feature of each point is set to 1, and the aggregated feature of each point is obtained by aggregating the initial features through the Abstract layer; Step A2: Calculate the reference axis perpendicular to the normal line of each point; in, represents the i-th point processed on the l-th layer AMR module, K l represents the number of neighborhood points used by the AMR module in the first layer, k l Indicates the kth layer of the lth layer l Neighborhood points, Indicated by point point to vector, express kth l Neighborhood points, express The neighborhood point set of is the calculated perpendicular to the point Normal vector, r is set to The distance to the farthest neighbor point, is the final weighted l-th layer perpendicular to The reference axis of the normal; each point The normal of Step A3: Based on the reference axis and normal, construct a high-dimensional rotation-invariant position encoding of each point and its neighboring points; Among them, HRIPE() represents the high-dimensional rotation-invariant position encoding operation, and rpe() represents the rotation-invariant position representation; Indicates a point and High-dimensional rotation-invariant position encoding between Indicates that the input feature dimension is 5, and the feature dimension of the output of the lth AMR module is C l The linear layer operation, Norm represents the normalization operation, Indicates that the input feature dimension is C l , the feature dimension of the output of the lth AMR module is C l Linear layer operation; Step A4: Linearly transform the high-dimensional rotation-invariant position encoding of all neighboring points of each point and the first 1 / 4 of the channels after adding the downsampled and aggregated sampling point features of each point through the Abstract layer; Step A5: The result of the linear transformation in step A4 is aggregated using a maximum pooling layer, and then two linear layers are used to learn deeper features. Step A6: Perform a residual connection between the features obtained in step A5 and the features obtained by aggregation in the Abstract layer to obtain the updated features to alleviate the gradient disappearance.

2. The method according to claim 1, characterized in that The working process of the ring attention module is as follows: Step B1: The input of the ring attention module is the set of target points to be processed The set of annular neighborhood points around the target point is recorded as and the corresponding feature set in, Indicates distance point In addition to the recent K l The nearest point outside Point index, feature set According to this index Indexed from the features obtained from the MLP module based on the local rotation-invariant position encoding enhancement; Step B2: According to the high-dimensional rotation-invariant position encoding operation, Relative to The high-dimensional rotation-invariant position encoding is denoted as Using five learnable matrices W G , W E , W Q , W K and W V To project the input, high-dimensional rotation-invariant position encoding By W G , W E It is further encoded into geometric code G and position code E, and then the new features of the target point in the lth layer are obtained by weighting the geometric code and the value Among them, G and E are the learned geometric and positional encodings respectively, and the learned context features can be projected into q, K and V: in, is the feature of the i-th point in the l-th layer, F l It is the point feature updated by the MLP module based on the local rotation-invariant position encoding enhancement, and c0 is the feature dimension size of G and V; Step B3: Through linear layers and residual connections and After adding and updating, we get the final feature F of the target point in the lth layer. l .

3. The method according to claim 1, characterized in that The four AMR modules of the feature extractor are connected in sequence, wherein the downsampling coefficients in each AMR module are [1, 0.25, 0.25, 0.25] in sequence; the feature dimensions of each AMR module are [64, 128, 256, 256] in sequence; the number of neighborhood points is set to [8, 16, 16, 16] in sequence, the last AMR module is connected to the global Transformer module, each AMR module is residually connected to a feature forward propagation module, and the four feature forward propagation modules are connected in reverse sequence, the input end of the first feature forward propagation module and the input end of the coarse-to-fine point matcher are both connected to the output end of the global Transformer module, and the output end of the last feature forward propagation module and the global Transformer module are connected to the input end of the coarse-to-fine point matcher, thereby obtaining the predicted super-point matching relationship and dense point matching relationship through the coarse-to-fine point matcher.

4. The method according to claim 1, wherein The matching loss is calculated by overlapping circle loss and negative log-likelihood loss. The matching loss is obtained by adding the overlapping circle loss and the negative log-likelihood loss in a one-to-one ratio. The overlap-aware circle loss is calculated using the predicted value of the super-point matching relationship and the super-point matching relationship derived from the pose truth as input; The negative log-likelihood loss is calculated by taking the dense point matching relationship prediction value and the pose truth value as input.

5. A robust point cloud registration system based on invariant and unique features, characterized by: include: RID-Net model construction unit; the RID-Net model includes a connected feature extractor and a coarse-to-fine point matcher, wherein the feature extractor includes four AMR modules, a global Transformer module, and four feature forward propagation modules; The four AMR modules are connected in sequence, and the last AMR module is connected to the global Transformer module. Each AMR module is residually connected to a feature forward propagation module. The four feature forward propagation modules are connected in reverse order. The input of the first feature forward propagation module and the input of the coarse-to-fine point matcher are connected to the output of the global Transformer module. The output of the last feature forward propagation module and the global Transformer module are connected to the input of the coarse-to-fine point matcher. The AMR module includes an Abstract layer, an MLP module enhanced by local rotation-invariant position coding, and a ring attention module connected in sequence; Model training unit: By obtaining multiple pairs of point cloud data and the corresponding relative pose truth values ​​of a certain scene as training data, where the relative pose truth values ​​are converted into super-point matching relationships and dense point matching relationships; each pair of point cloud data is input into the constructed RID-Net model in turn, and the super-point matching relationship and dense point matching relationship of each pair of point clouds predicted by the RID-Net model are used to obtain the trained RID-Net model parameters when the matching loss converges; Model registration unit: The collected point cloud data is input into the trained RID-Net model, and the super-point matching relationship and dense point matching relationship corresponding to the two frames of point clouds are predicted. Based on the obtained matching relationship, the relative pose of the two frames of point clouds is solved by a robust pose estimator. The source point cloud data is then converted to the target point cloud coordinate system through the relative pose to complete the registration; The specific working process of the MLP module based on local rotation-invariant position coding enhancement is as follows: Step A1: Obtain the local neighborhood point coordinates and aggregation features of each point; The initial feature of each point is set to 1, and the aggregated feature of each point is obtained by aggregating the initial features through the Abstract layer; Step A2: Calculate the reference axis perpendicular to the normal line of each point; in, represents the i-th point processed on the l-th layer AMR module, K l represents the number of neighborhood points used by the AMR module in the first layer, k l Indicates the kth layer of the lth layer l Neighborhood points, Indicated by point point to vector, express kth l Neighborhood points, express The neighborhood point set of is the calculated perpendicular to the point Normal vector, r is set to The distance to the farthest neighbor point, is the final weighted perpendicular to The reference axis of the normal; each point The normal of Step A3: Based on the reference axis and normal, construct a high-dimensional rotation-invariant position encoding of each point and its neighboring points; Among them, HRIPE() represents the high-dimensional rotation-invariant position encoding operation, and rpe() represents the rotation-invariant position representation; Indicates a point and High-dimensional rotation-invariant position encoding between Indicates that the input feature dimension is 5, and the feature dimension of the output of the lth AMR module is C l The linear layer operation, Norm represents the normalization operation, Indicates that the input feature dimension is C l , the feature dimension of the output of the lth AMR module is C l Linear layer operation; Step A4: Linearly transform the high-dimensional rotation-invariant position encoding of all neighboring points of each point and the first 1 / 4 of the channels after adding the downsampled and aggregated sampling point features of each point through the Abstract layer; Step A5: The result of the linear transformation in step A4 is aggregated using a maximum pooling layer, and then two linear layers are used to learn deeper features. Step A6: Perform a residual connection between the features obtained in step A5 and the features obtained by aggregation in the Abstract layer to obtain the updated features to alleviate the gradient disappearance.

6. The system according to claim 5, characterized in that The working process of the ring attention module is as follows: Step B1: The input of the ring attention module is the set of target points to be processed The set of annular neighborhood points around the target point is recorded as and the corresponding feature set in, Indicates distance point In addition to the recent K l The nearest point outside Point index, feature set According to this index Indexed from the features obtained from the MLP module based on the local rotation-invariant position encoding enhancement; Step B2: According to the high-dimensional rotation-invariant position encoding operation, Relative to The high-dimensional rotation-invariant position encoding is denoted as Using five learnable matrices W G , W E , W Q , W K and W V To project the input, high-dimensional rotation-invariant position encoding By W G , W E It is further encoded into geometric code G and position code E, and then the new target point features of the lth layer are obtained by weighting the geometric code and the value Value. Among them, G and E are the learned geometric and positional encodings respectively, and the learned context features can be projected into q, K and V: in, is the feature of the i-th point in the l-th layer, F l It is the point feature updated by the MLP module based on the local rotation-invariant position encoding enhancement, and c0 is the feature dimension size of G and V; Step B3: Through linear layers and residual connections and After adding and updating, we get the final feature F of the target point in the lth layer. l .

7. A computer device, characterized in that: At least: one or more processors; a memory storing one or more computer programs; The processor calls the computer program to implement: The steps of the method according to any one of claims 1 to 4.

8. A computer-readable storage medium, characterized in that: A computer program is stored, which is called by a processor to implement: The steps of the method according to any one of claims 1 to 4.

Images