Multi-level point cloud registration method based on overlapped region prediction

Abstract

The invention provides a multi-level point cloud registration method based on overlapped region prediction, and mainly solves the problem that the registration precision is influenced by the noise interference of a point cloud non-overlapped region in the existing point cloud registration technology. The implementation scheme is as follows: 1) acquiring a data set and a label; 2) constructing a point cloud registration model; 3) constructing a loss function; 4) training a point cloud registration model; and 5) obtaining a point cloud registration transformation result. According to the constructed point cloud registration model, the feature extraction efficiency is improved by utilizing the structural features of the pixel grids through the multi-level point cloud feature extraction module based on pixel grid mapping; through an overlapping region prediction module based on a convolutional neural network, predicting an overlapping region of the point cloud on each feature level so as to reduce interference of a non-overlapping region in a subsequent registration process; through an attitude transformation regression module based on frequency domain feature fusion, an attitude transformation result of the point cloud is obtained more accurately from the fused frequency domain features and cross neighborhood features.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, specifically relating to a multi-level point cloud registration method based on overlapping region prediction. Background Technology

[0002] With the rapid development of LiDAR technology, its applications in autonomous driving, industrial automation, and robotics are becoming increasingly widespread. Point cloud registration, as a key technology, aims to estimate the relative rotation and translation relationship between two frames of point clouds acquired at different locations. It is the foundation of 3D mapping and localization tasks, and registration accuracy directly affects the effectiveness of downstream perception, decision-making, and localization tasks. Traditional point cloud registration methods are usually based on feature extraction and matching, dividing the registration task into two stages: feature description and correspondence solving. However, these methods have high requirements for the consistency of feature descriptions and are complex and computationally expensive in solving correspondences, resulting in low efficiency in large-scale point cloud data. End-to-end registration methods based on deep learning utilize the nonlinear fitting capability of neural networks to directly regress the registration results between two point clouds, avoiding complex point-to-point matching and improving registration efficiency. However, while most existing methods perform well on small datasets with complete overlap, their performance is unsatisfactory in partially overlapping LiDAR point cloud data. Noise interference in non-overlapping areas affects registration accuracy. Furthermore, using only a set of global feature descriptions fails to fully utilize the local geometric information of the point cloud, limiting its adaptability to complex scenes. Therefore, existing methods are insufficient to meet the requirements of high accuracy and high robustness in large-scale, partially overlapping lidar point cloud registration tasks. Summary of the Invention

[0003] This invention fully considers the shortcomings of existing methods and aims to provide a multi-level point cloud registration method based on overlapping region prediction. It utilizes a multi-level point cloud feature extraction module based on pixel grid mapping to accelerate the efficiency of point cloud feature extraction by leveraging the structural characteristics of pixel grids. A convolutional neural network-based overlapping region prediction module predicts the overlapping regions of the point cloud at each feature level, thereby reducing interference from non-overlapping regions in subsequent registration processes. Finally, a pose transformation regression module based on frequency domain feature fusion can more accurately capture the pose transformation of the point cloud from the fused frequency domain features and cross-neighborhood features.

[0004] I. Technical Principles

[0005] Most current deep learning-based point cloud registration methods first use feature extractors to extract global features from two point clouds, and then complete the registration by minimizing the reprojection error between the two feature sets. This method is convenient and fast, and does not require establishing a correspondence, thus improving the registration efficiency. However, such methods do not take into account the large number of non-overlapping points often present in laser point cloud registration data. Although they perform well on small datasets with complete overlap, their performance is unsatisfactory in partially overlapping laser point cloud data. Considering the incomplete overlap characteristics of laser point cloud data, to reduce the impact of non-overlapping regions on the registration process and to register LiDAR point cloud data more quickly, this invention designs a multi-level point cloud feature extraction module based on pixel grid mapping. This module utilizes the pixel grid characteristics to perform multi-level sampling of point cloud data, increasing the efficiency of point cloud sampling. An overlapping region prediction module based on convolutional neural networks is proposed to eliminate the influence of outliers in non-overlapping regions on point cloud registration. A posture transformation regression module based on frequency domain feature fusion is designed to improve the accuracy of the model's regression of posture transformation parameters.

[0006] II. Based on the above principles, the present invention is implemented through the following scheme:

[0007] A multi-level point cloud registration method based on overlapping region prediction includes the following steps:

[0008] (1) Obtain the dataset and labels:

[0009] Obtain the KITTI dataset and its corresponding labels for pose estimation and point cloud registration tasks;

[0010] (2) Constructing a point cloud registration model: This model consists of a backbone network, an overlapping region prediction module based on a convolutional neural network, and a pose transformation regression module based on frequency domain feature fusion. The specific construction process includes the following steps:

[0011] (2-a) Constructing the backbone network: The multi-level point cloud feature extraction module based on pixel grid mapping is used as the backbone network. The two input point clouds are processed by the backbone network to obtain four pairs of feature maps with decreasing scale.

[0012] The multi-level point cloud feature extraction module based on pixel mesh mapping is constructed as follows:

[0013] ① LiDAR point cloud pixel grid mapping: Assuming there exists an H×W pixel grid, where H equals the number of LiDAR beams and W is the number of points in one revolution of a single beam, the coordinates (u,v) of each LiDAR point in the pixel grid are calculated using the following formula based on the 3D coordinates (x,y,z) of each LiDAR point in the point cloud:

[0014]

[0015] Where β is the vertical resolution of the lidar, i.e. the angle between two adjacent rays, arctan(·) is the arctangent function, and arcsin() is the arcsine function;

[0016] ② Point cloud feature extraction based on pixel grid: Once a one-to-one mapping between point cloud data and pixel grid positions is obtained, the point cloud data can be downsampled in multiple layers based on pixel positions by setting different sampling intervals in the H and W directions; after sampling, the features of deeper points are obtained from the features of the K shallower points surrounding the pixel; for any sampling point in the Lth layer... Its features are derived from the feature set of its k nearest neighbors in layer L-1. and coordinates The extraction process involves three convolutional layers for each feature. Each layer contains a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function. Then, max pooling is performed on these k features along dimension k to obtain the sampling points. eigenvalues ​​f i L This process can be represented as:

[0017]

[0018] in The feature is the concatenation of the k nearest neighbors of the (L-1)th layer. Maxpool(·) is the max pooling operation, and cat() is the concatenation operation. By performing the above operation on all sampled points of this layer, the feature extraction of the point cloud of this layer can be completed. The pixel grid-based downsampling method does not require point cloud traversal of the K nearest neighbor operation. It can achieve efficient downsampling by utilizing the structured information of pixels. In addition, by using the neighborhood feature information of the (L-1)th layer to extract the feature of the Lth layer point cloud, the Lth layer features have a larger receptive field. With this progressive feature extraction, the receptive field of the deeper features is larger, the feature information is more comprehensive and richer, and it is more conducive to pose calculation.

[0019] (2-b) Construct an overlapping region prediction module based on a convolutional neural network: This module first samples the features of the source point cloud P and the target point cloud Q at the same sampling level. Max pooling is used to extract the global features g of the two point clouds in this layer. P g Q Global features contain overall information about the point cloud and can be used as reference features for comparing the similarity of two point clouds; then g P with f Q splicing, g Q with f PThe points are concatenated and then fed into a weight-sharing convolutional neural network. The network will obtain a mask M for the overlapping region based on the similarity between the global features of the reference point cloud and its own features. P and M Q This process can be expressed as the following formula:

[0020]

[0021] Maxpool(·) is the max pooling operation, ConvBlock(·) is the convolution operation module with weight sharing mentioned above, and FC(·) is the fully connected layer.

[0022] (2-c) Construct a pose transformation regression module based on frequency domain feature fusion: This module consists of three parts: a feature enhancement module based on frequency domain characteristics, a feature fusion module based on cross-neighborhood, and a pose regression module;

[0023] ①The feature enhancement module based on frequency domain characteristics is constructed as follows:

[0024] First, use mask M. P and M Q Multiplying the corresponding features of the two point clouds yields the masked features. Next, the discrete frequency domain cosine eigentransform matrix D is used to transform mf P and mf Q A one-dimensional discrete frequency domain transform is performed to obtain df P and df Q Then, the frequency domain attention weights for each feature are obtained through multi-layer convolution and sigmoid. w P and w Q respectively with feature df P and df Q Weighting yields the frequency domain feature f1. P and F1 Q Since the discrete frequency domain transformation causes some loss of feature information, an additional original feature branch is added to compensate for this loss. In this branch, mf... P and mf Q The feature f2 is obtained by performing convolution weighting operations. P and f2 Q Then, the features from the two branches are added together to obtain the frequency domain weighted features f corresponding to the two point clouds. P and f Q Finally, a cross-attention mechanism is used to model the long-distance dependency between the two point cloud features, fully fusing the frequency domain characteristics of the two point clouds. This process can be expressed as the following formula:

[0025] mf P =MP ·f P ,mf Q =M Q ·f Q

[0026] df P =D×mf P ,df Q =D×mf Q

[0027]

[0028] f P ,f Q =CrossAttn(f3) P f3 Q )

[0029] Where D is an N×N matrix, and the elements in D are calculated as follows:

[0030]

[0031] Where i is the row index and j is the column index;

[0032] ②The feature fusion module based on cross-neighborhood is constructed as follows:

[0033] For any sampling point in the source point cloud P of layer L... Its in f P The corresponding feature is use The pixel grid coordinates are used to find the k nearest neighbors in the grid of the point cloud Q, and the k nearest neighbors are obtained in f. Q The corresponding feature set N Q and After concatenation, a feature set f1 is obtained; then, a neural network is used to extract the weight of each feature, and the k features are weighted and summed according to their weights to obtain the final feature set f1. Cross-point cloud fusion features Then use The pixel grid coordinates are used to find the k nearest neighbors in the grid of the point cloud P, and the k nearest neighbors are obtained in f. P The corresponding feature set N P and After splicing, the fusion features are obtained in the same way. Finally and After splicing, we get Enhanced features

[0034] ③ The attitude regression module is constructed as follows:

[0035] Enhanced features for the Lth layer point cloud It is convolved with the L+1th layer feature u after upsampling L+1 One by one, the intermediate feature c1 is obtained through two convolutional modules. c1 is convolved twice and the feature weight wc1 is obtained through Softmax. Finally, the quaternion representation q of the pose transformation and the translation representation t are regressed from wc1 using a fully connected layer.

[0036] (3) Construct the loss function:

[0037] Construct the following hybrid loss function L:

[0038]

[0039] Where L M For the overlapping loss term, L q L is the rotational loss term. t For the translation loss term, l represents the number of layers, and β, γ, and λ are adjustment factors used to adjust the weight of each loss term;

[0040] The specific construction process includes the following steps:

[0041] (3-a) Construct the overlap loss term L M :

[0042] L M =∑-αM gt log(M)-(1-α)(1-M gt log(1-M)

[0043] Where M is the mask predicted by the model, M gt α represents the true value of the mask, and is the weighting parameter.

[0044] (3-b) Construct the rotational loss term L q :

[0045] L q =||q gt -q||

[0046] Where q gt q represents the true rotation quaternion obtained from the labeled data, and q represents the rotation quaternion predicted by the network model.

[0047] (3-c) Construct the translation loss term L t :

[0048] L t =||t gt -t||

[0049] Where t gt Let t be the true translation vector obtained from the labeled data, and t be the translation vector predicted by the network model.

[0050] (4) Training the point cloud registration model:

[0051] The point cloud registration model constructed in step (2) is trained using the dataset obtained in step (1); the error between the prediction result output by the model and the label is calculated using the loss function L constructed in step (3); the Adam algorithm is used to update the model parameters during the training process until the loss function converges, and the trained point cloud registration model is obtained.

[0052] (5) Obtain the point cloud registration transformation results:

[0053] After pixel-level mapping, the two test point clouds are input into the trained point cloud registration model, and the model output is the final point cloud registration transformation result.

[0054] In step (2-a), the initial pixel grid parameters H and W after point cloud projection are preferably 64 and 1792, respectively;

[0055] In step (3), α, β, γ and λ are preferably 0.7, 1, 0.9 and 0.99, respectively;

[0056] Compared with the prior art, the present invention has the following advantages:

[0057] (1) The multi-level point cloud feature extraction module based on pixel grid mapping constructed in this invention can quickly perform multi-level neighborhood sampling of point clouds using the structure of pixel grid, and make full use of the information lost during sampling when extracting features.

[0058] (2) The point cloud overlapping region prediction module based on convolutional neural network constructed in this invention can predict part of the overlapping region of two point clouds based on the global and local features of the two point clouds, thereby eliminating point cloud data and features of non-overlapping regions during registration and increasing registration accuracy.

[0059] (3) The attitude transformation regression module based on frequency domain feature fusion constructed in this invention can fully integrate the frequency domain features and neighborhood features of two point clouds and accurately capture the motion characteristics of the point cloud. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the overall architecture of the multi-level point cloud registration method based on overlapping region prediction according to an embodiment of the present invention.

[0061] Figure 2 A structural diagram of the pixel-grid-based point cloud feature extraction module according to an embodiment of the present invention;

[0062] Figure 3 A structural diagram of the overlapping region prediction module based on a convolutional neural network according to an embodiment of the present invention;

[0063] Figure 4 A structural diagram of the feature enhancement module based on frequency domain characteristics according to an embodiment of the present invention;

[0064] Figure 5 Structure diagram of the feature fusion module based on cross-neighborhood in an embodiment of the present invention;

[0065] Figure 6 A structural diagram of the attitude regression module according to an embodiment of the present invention;

[0066] Figure 7 A comparison of the registration success rates of the point cloud registration method of this invention with other registration methods in the KITTI dataset under different RTE and RRE error thresholds;

[0067] Figure 8(A) Visualization of the registration effect of the point cloud registration method of the present invention on the KITTI dataset. Figure 1 ;

[0068] Figure 8(B) Visualization of the registration effect of the point cloud registration method of the present invention on the KITTI dataset. Figure 2 . Detailed Implementation

[0069] To make the above-mentioned objectives, features, and advantages of the present invention clearer and easier to understand, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be noted that the accompanying drawings of the present invention are all in a simplified form and use non-precise proportions, and are only used to facilitate and clearly assist in illustrating the implementation of the present invention; the "several" mentioned in the present invention are not limited to the specific number shown in the examples in the accompanying drawings; the orientations or positional relationships indicated by terms such as "front," "middle," "rear," "left," "right," "up," "down," "top," "bottom," and "center" mentioned in the present invention are all based on the orientations or positional relationships shown in the accompanying drawings of the present invention, and do not indicate or imply that the device or component referred to must have a specific orientation, nor should they be construed as limitations on the present invention.

[0070] This embodiment:

[0071] Figure 1 The diagram shows the overall architecture of the multi-level point cloud registration method based on overlapping region prediction according to an embodiment of the present invention. The specific steps are as follows:

[0072] Step 1, Obtain the dataset and labels:

[0073] Obtain the KITTI dataset and its corresponding labels for pose estimation and point cloud registration tasks;

[0074] Step 2, Constructing the point cloud registration model: This model consists of a backbone network, an overlapping region prediction module based on convolutional neural networks, and a pose transformation regression module based on frequency domain feature fusion. The specific construction process includes the following steps:

[0075] (2-a) Constructing the backbone network: The multi-level point cloud feature extraction module based on pixel grid mapping is used as the backbone network. The two input point clouds are processed by the backbone network to obtain four pairs of feature maps with decreasing scale.

[0076] The multi-level point cloud feature extraction module based on pixel mesh mapping is constructed as follows:

[0077] ① LiDAR point cloud pixel grid mapping: Assuming there exists an H×W pixel grid, where H equals the number of LiDAR beams and W is the number of points in one revolution of a single beam, the coordinates (u,v) of each LiDAR point in the pixel grid are calculated using the following formula based on the 3D coordinates (x,y,z) of each LiDAR point in the point cloud:

[0078]

[0079] Where β is the vertical resolution of the lidar, i.e. the angle between two adjacent rays, arctan() is the arctangent function, and arcsin() is the arcsine function;

[0080] ② Point cloud feature extraction based on pixel grid: Once a one-to-one mapping between point cloud data and pixel grid positions is obtained, multiple layers of downsampling can be performed on the point cloud data based on pixel positions by setting different sampling intervals in the H and W directions. After sampling, the features of deeper points are obtained from the features of the K shallower points surrounding the pixel. The feature extraction method is as follows: Figure 2 As shown; for any sampling point in the Lth layer Its features are derived from the feature set of its k nearest neighbors in layer L-1. and coordinates The extraction process involves three convolutional layers for each feature. Each layer contains a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function. Then, max pooling is performed on these k features along dimension k to obtain the sampling points. eigenvalues ​​f i L This process can be represented as:

[0081]

[0082] in The feature is the concatenation of the k nearest neighbors of the (L-1)th layer. Maxpool(·) is the max pooling operation, and cat() is the concatenation operation. By performing the above operation on all sampled points of this layer, the feature extraction of the point cloud of this layer can be completed. The pixel grid-based downsampling method does not require point cloud traversal of the K nearest neighbor operation. It can achieve efficient downsampling by utilizing the structured information of pixels. In addition, by using the neighborhood feature information of the (L-1)th layer to extract the feature of the Lth layer point cloud, the Lth layer features have a larger receptive field. With this progressive feature extraction, the receptive field of the deeper features is larger, the feature information is more comprehensive and richer, and it is more conducive to pose calculation.

[0083] (2-b) Construct an overlapping region prediction module based on a convolutional neural network: Figure 3 The diagram shown is a structural diagram of the overlapping region prediction module based on a convolutional neural network according to an embodiment of the present invention. This module first samples the features of the source point cloud P and the target point cloud Q at the same sampling level. Max pooling is used to extract the global features g of the two point clouds in this layer. P g Q Global features contain overall information about the point cloud and can be used as reference features for comparing the similarity of two point clouds; then g P with f Q splicing, g Q with f P The points are concatenated and then fed into a weight-sharing convolutional neural network. The network will obtain a mask M for the overlapping region based on the similarity between the global features of the reference point cloud and its own features. P and M Q This process can be expressed as the following formula:

[0084]

[0085] Maxpool(·) is the max pooling operation, ConvBlock(·) is the convolution operation module with weight sharing mentioned above, and FC(·) is the fully connected layer.

[0086] (2-c) Construct a pose transformation regression module based on frequency domain feature fusion: This module consists of three parts: a feature enhancement module based on frequency domain characteristics, a feature fusion module based on cross-neighborhood, and a pose regression module;

[0087] ①The feature enhancement module based on frequency domain characteristics is constructed as follows:

[0088] The feature enhancement module structure based on frequency domain characteristics is as follows: Figure 4 As shown; first, use mask M P and M QMultiplying the corresponding features of the two point clouds yields the masked features. Next, the discrete frequency domain cosine eigentransform matrix D is used to transform mf P and mf Q A one-dimensional discrete frequency domain transform is performed to obtain df P and df Q Then, the frequency domain attention weights for each feature are obtained through multi-layer convolution and sigmoid. w P and w Q respectively with feature df P and df Q Weighting yields the frequency domain feature f1. P and F1 Q Since the discrete frequency domain transformation causes some loss of feature information, an additional original feature branch is added to compensate for this loss. In this branch, mf... P and mf Q The feature f2 is obtained by performing convolution weighting operations. P and f2 Q Then, the features from the two branches are added together to obtain the frequency domain weighted features f corresponding to the two point clouds. P and f Q Finally, a cross-attention mechanism is used to model the long-distance dependency between the two point cloud features, fully fusing the frequency domain characteristics of the two point clouds. This process can be expressed as the following formula:

[0089] mf P =M P ·f P ,mf Q =M Q ·f Q

[0090] df P =D×mf P ,df Q =D×mf Q

[0091]

[0092] f P ,f Q =CrossAttn(f3) P f3 Q )

[0093] Where D is an N×N matrix, and the elements in D are calculated as follows:

[0094]

[0095] Where i is the row index and j is the column index;

[0096] ②The feature fusion module based on cross-neighborhood is constructed as follows:

[0097] The feature fusion module structure based on cross-neighborhood is as follows: Figure 5 As shown; for any sampling point in the source point cloud P of layer L. Its in f P The corresponding feature is use The pixel grid coordinates are used to find the k nearest neighbors in the grid of the point cloud Q, and the k nearest neighbors are obtained in f. Q The corresponding feature set N Q and After concatenation, a feature set f1 is obtained; then, a neural network is used to extract the weight of each feature, and the k features are weighted and summed according to their weights to obtain the final feature set f1. Cross-point cloud fusion features Then use The pixel grid coordinates are used to find the k nearest neighbors in the grid of the point cloud P, and the k nearest neighbors are obtained in f. P The corresponding feature set N P and After splicing, the fusion features are obtained in the same way. Finally and After splicing, we get Enhanced features

[0098] ③ The attitude regression module is constructed as follows:

[0099] The structure of the attitude regression module is as follows: Figure 6 As shown; for the enhanced features of the Lth layer point cloud It is convolved with the L+1th layer feature u after upsampling L+1 One by one, the intermediate feature c1 is obtained through two convolutional modules. c1 is convolved twice and the feature weight wc1 is obtained through Softmax. Finally, the quaternion representation q of the pose transformation and the translation representation t are regressed from wc1 using a fully connected layer.

[0100] Step 3, construct the loss function:

[0101] Construct the following hybrid loss function L:

[0102]

[0103] Where L M For the overlapping loss term, Lq L is the rotational loss term. t For the translation loss term, l represents the number of layers, and β, γ, and λ are adjustment factors used to adjust the weight of each loss term;

[0104] The specific construction process includes the following steps:

[0105] (3-a) Construct the overlap loss term L M :

[0106] L M =∑-αM gt log(M)-(1-α)(1-M gt log(1-M)

[0107] Where M is the mask predicted by the model, M gt α represents the true value of the mask, and is the weighting parameter.

[0108] (3-b) Construct the rotational loss term L q :

[0109] L q =||q gt -q||

[0110] Where q gt q represents the true rotation quaternion obtained from the labeled data, and q represents the rotation quaternion predicted by the network model.

[0111] (3-c) Construct the translation loss term L t :

[0112] L t =||t gt -t||

[0113] Where t gt Let t be the true translation vector obtained from the labeled data, and t be the translation vector predicted by the network model.

[0114] Step 4, train the point cloud registration model:

[0115] The point cloud registration model constructed in step (2) is trained using the dataset obtained in step (1); the error between the prediction result output by the model and the label is calculated using the loss function L constructed in step (3); the Adam algorithm is used to update the model parameters during the training process until the loss function converges, and the trained point cloud registration model is obtained.

[0116] Step 5, obtain the point cloud registration transformation results:

[0117] After pixel-level mapping, the two test point clouds are input into the trained point cloud registration model, and the model output is the final point cloud registration transformation result.

[0118] Example 1: Ablation Test

[0119] Table 1 presents the ablation experiment results for each module in the model proposed in this invention. Through ablation experiments, the importance of each module in the model and its impact on registration performance can be compared and studied. The main analysis focuses on the impact of the convolutional neural network-based overlapping region prediction module (denoted as OP) and the frequency domain-based feature enhancement module (denoted as D) on the overall registration performance. For the ablation experiments, sequences 08, 09, and 10 from the KITTI dataset were used as the benchmark dataset, with RTE, RRE, and Recall as evaluation metrics. The threshold for RTE was set to 2m, and the threshold for RRE was set to 5°.

[0120] Table 1 Ablation experiments of each module in this invention

[0121]

[0122]

[0123] Based on the basic model, ablation comparison experiments were conducted to verify the performance gain of the proposed convolutional neural network-based overlapping region prediction module. After adding the OP module to the basic model to reduce the influence of noise in non-overlapping point cloud regions, the model showed a significant improvement in registration accuracy across three test sequences. Specifically, in the comparative test results of the 08 sequence, after adding the OP module, the RTE registration error decreased from 0.4691 to 0.2581, an accuracy improvement of approximately 45%; the RRE registration error decreased from 0.7813 degrees to 0.5279 degrees, an accuracy improvement of 32.4%; and the registration success rate increased from 87.39% to 93%, an improvement of 5.61 percentage points. In the comparative test results of the 09 sequence, after adding the OP module, the RTE registration error decreased from 0.3325 to 0.3189, with an accuracy improvement of approximately 4.09%; the RRE registration error decreased from 0.5406 degrees to 0.4924 degrees, with an accuracy improvement of 4.9%; and the registration success rate increased from 80.63% to 90%, an improvement of 0.37 percentage points. In the comparative test results of the 10 sequence, after adding the OP module, the RTE registration error decreased from 0.3471 to 0.3069, with an accuracy improvement of approximately 11.6%; the RRE registration error decreased from 0.719 degrees to 0.6548 degrees, with an accuracy improvement of 1.4%; and the registration success rate increased from 90.76% to 92.19%, an improvement of 1.43 percentage points.

[0124] The ablation learning experiment results for the OP module in Table 2 show that the OP module effectively handles noise in non-overlapping regions of point cloud data, reducing the impact of non-overlapping regions on the point cloud registration process. This increases the translation accuracy, rotation accuracy, and registration success rate of large-scale laser point cloud registration, resulting in a positive gain for improving the performance of the point cloud model.

[0125] Based on the base model with the OP module added, ablation comparison experiments were conducted to verify the impact of the frequency domain characteristic-based feature enhancement module proposed in this invention on the overall performance of the model. After adding the D module to increase the application of point cloud frequency characteristics, the model showed a significant improvement in registration accuracy on the three test sequences. Specifically, in the comparative test results of the 09 sequence, after adding the D module, the RTE registration error decreased from 0.3189 to 0.2964, with an accuracy improvement of about 4.09%; the RRE registration error increased from 0.4924 degrees to 0.5030 degrees, with an accuracy decrease of 2.1%; and the registration success rate increased from 90% to 90.7%, an improvement of 0.7 percentage points. In the comparative test results of 10 sequences, after adding the D module, the RTE registration error decreased from 0.3069 to 0.2827, with an accuracy improvement of about 7.9%; the RRE registration error decreased from 0.6548 degrees to 0.6053 degrees, with an accuracy improvement of 2.1%; and the registration success rate decreased from 92.19% to 90.18%, a decrease of 2.18 percentage points.

[0126] Example 2: Comparative Experiment

[0127] To demonstrate the effectiveness of the model designed in this invention, this embodiment uses five metrics—RTE(m), RRE(deg), Recall, Time(ms) / Points, and NT(s)—to compare the method of this invention with nine other test methods on the KITTI dataset. The evaluation results are shown in Table 2, where the arrow "↑" to the right of the metric indicates that a larger metric value is better, and "↓" indicates that a lower metric value is better.

[0128] Table 2 shows the comparative experiment of the KITTI test dataset.

[0129]

[0130] Compared to non-learning methods such as ICP and RANSAC, this invention, while slightly inferior in RTE and RRE accuracy, achieves significant advantages in registration success rate and efficiency. In practical applications, registration efficiency and success rate are often paramount. Compared to corresponding point matching methods in learning-based approaches, this invention maintains both registration accuracy and efficiency, demonstrating faster point cloud data processing speeds when handling large amounts of point cloud data simultaneously. While it lags behind in registration success rate (approximately 6% lower than DGR and FastReg), this invention improves RTE by 14.1% and 3.6% respectively, and RTE by 1.8% and 32.6% respectively, while also significantly enhancing registration efficiency. Compared to end-to-end learning-based methods, this invention increases point cloud registration accuracy by utilizing point cloud feature information from various levels. Compared to FMR, the RTE registration accuracy of this invention is improved by 59.3%, the RRE registration accuracy by 64.4%, and the registration success rate by 1.6%. Meanwhile, OMTRNet's standardized registration efficiency is 36.9% higher than FMR. This demonstrates the advantage of a multi-level registration structure in fully utilizing point cloud information, resulting in lower registration errors.

[0131] To demonstrate the performance of the network model of this invention, the registration success rates of the proposed method and the comparison method were statistically analyzed on the KITTI dataset under different RRE and RTE thresholds. The statistical results are as follows: Figure 7 As shown in the figures, this invention achieves significant advantages over deep learning registration methods such as FMR, IDAM, and DCP. While it has disadvantages compared to methods like Teaser and DCP, the results are still very similar. It also demonstrates an overwhelming advantage over non-learning methods such as ICP and RANSAC. To visually demonstrate the registration effect, this invention selected a typical scene from the KITTI dataset and plotted point clouds before and after registration, as shown in Figures 8(A) and 8(B). In the figures, the gray point clouds represent the source and target point clouds, and the black lines represent the corresponding keypoint pairs after registration. The visualization results clearly show that despite challenges such as road occlusion and target sparsity, this method can still accurately align the corresponding structural regions, demonstrating good robustness and registration accuracy.

[0132] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Therefore, any changes made in accordance with the structure and principles of the present invention should be covered within the protection scope of the present invention.

Claims

1. A multi-level point cloud registration method based on overlapping region prediction, characterized in that... Includes the following steps: (1) Obtain the dataset and labels: Obtain the KITTI dataset and its corresponding labels for pose estimation and point cloud registration tasks; (2) Constructing a point cloud registration model: This model consists of a backbone network, an overlapping region prediction module based on a convolutional neural network, and a pose transformation regression module based on frequency domain feature fusion. The specific construction process includes the following steps: (2-a) Constructing the backbone network: The multi-level point cloud feature extraction module based on pixel grid mapping is used as the backbone network. The two input point clouds are processed by the backbone network to obtain four pairs of feature maps with decreasing scale. The multi-level point cloud feature extraction module based on pixel mesh mapping is constructed as follows: ① LiDAR point cloud pixel grid mapping: Assuming there exists an H×W pixel grid, where H equals the number of LiDAR beams and W is the number of points in one revolution of a single beam, the coordinates (u,v) of each LiDAR point in the pixel grid are calculated using the following formula based on the 3D coordinates (x,y,z) of each LiDAR point in the point cloud: Where β is the vertical resolution of the lidar, i.e. the angle between two adjacent rays, arctan() is the arctangent function, and arcsin() is the arcsine function; ② Point cloud feature extraction based on pixel grid: Once a one-to-one mapping between point cloud data and pixel grid positions is obtained, the point cloud data can be downsampled in multiple layers based on pixel positions by setting different sampling intervals in the H and W directions; after sampling, the features of deeper points are obtained from the features of the K shallower points surrounding the pixel; for any sampling point in the Lth layer... Its features are derived from the feature set of its k nearest neighbors in layer L-1. and coordinates Extracting from the set: First, feature extraction is performed on each feature using a three-layer convolutional module; then, max pooling is performed on these k features along dimension k to obtain the sampling points. eigenvalues ​​f i L This process can be represented as: in The feature is the concatenation of the k nearest neighbors of the (L-1)th layer. Maxpool(·) is the max pooling operation, and cat() is the concatenation operation. By performing the above operation on all sampled points of this layer, the feature extraction of the point cloud of this layer can be completed. The pixel grid-based downsampling method does not require point cloud traversal of the K nearest neighbor operation. It can achieve efficient downsampling by utilizing the structured information of pixels. In addition, by using the neighborhood feature information of the (L-1)th layer to extract the feature of the Lth layer point cloud, the Lth layer features have a larger receptive field. With this progressive feature extraction, the receptive field of the deeper features is larger, the feature information is more comprehensive and richer, and it is more conducive to pose calculation. (2-b) Construct an overlapping region prediction module based on a convolutional neural network: This module first samples the features of the source point cloud P and the target point cloud Q at the same sampling level. Max pooling is used to extract the global features g of the two point clouds in this layer. P g Q Global features contain overall information about the point cloud and can be used as reference features for comparing the similarity of two point clouds; then g P with f Q splicing, g Q with f P The points are concatenated and then fed into a weight-sharing convolutional neural network. The network will obtain a mask M for the overlapping region based on the similarity between the global features of the reference point cloud and its own features. P and M Q This process can be expressed as the following formula: Maxpool(·) is the max pooling operation, ConvBlock(·) is the convolution operation module with weight sharing mentioned above, and FC(·) is the fully connected layer. (2-c) Construct a pose transformation regression module based on frequency domain feature fusion: This module consists of three parts: a feature enhancement module based on frequency domain characteristics, a feature fusion module based on cross-neighborhood, and a pose regression module; ①The feature enhancement module based on frequency domain characteristics is constructed as follows: First, use mask M. P and M Q Multiplying the corresponding features of the two point clouds yields the masked features. Next, the discrete frequency domain cosine eigentransform matrix D is used to transform mf P and mf Q A one-dimensional discrete frequency domain transform is performed to obtain df P and df Q Then, the frequency domain attention weights for each feature are obtained through multi-layer convolution and sigmoid. w P and w Q respectively with feature df P and df Q Weighting yields the frequency domain feature f1. P and F1 Q Since the discrete frequency domain transformation causes some loss of feature information, an additional original feature branch is added to compensate for this loss. In this branch, mf... P and mf Q The feature f2 is obtained by performing convolution weighting operations. P and f2 Q Then, the features from the two branches are added together to obtain the frequency domain weighted features f corresponding to the two point clouds. P and f Q Finally, a cross-attention mechanism is used to model the long-distance dependency between the two point cloud features, fully fusing the frequency domain characteristics of the two point clouds. This process can be expressed as the following formula: mf P *M P ·f P ,mf Q *M Q ·f Q df P =D×mf P ,df Q =D×mf Q f P ,f Q =CrossAttn(f3 P f3 Q ) Where D is an N×N matrix, and the elements in D are calculated as follows: Where i is the row index and j is the column index; ②The feature fusion module based on cross-neighborhood is constructed as follows: For any sampling point in the source point cloud P of layer L... Its in f P The corresponding feature is use The pixel grid coordinates are used to find the k nearest neighbors in the grid of the point cloud Q, and the k nearest neighbors are obtained in f. Q The corresponding feature set N Q and After concatenation, a feature set f1 is obtained; then, a neural network is used to extract the weight of each feature, and the k features are weighted and summed according to their weights to obtain the final feature set f1. Cross-point cloud fusion features Then use The pixel grid coordinates are used to find the k nearest neighbors in the grid of the point cloud P, and the k nearest neighbors are obtained in f. P The corresponding feature set N P and After splicing, the fusion features are obtained in the same way. Finally and After splicing, we get Enhanced features ③ The attitude regression module is constructed as follows: Enhanced features for the Lth layer point cloud It is convolved with the L+1th layer feature u after upsampling L+1 One by one, the intermediate feature c1 is obtained through two convolutional modules. c1 is convolved twice and the feature weight wc1 is obtained through Softmax. Finally, the quaternion representation q of the pose transformation and the translation representation t are regressed from wc1 using a fully connected layer. (3) Construct the loss function: Construct the following hybrid loss function L: Where L M For the overlapping loss term, L q L is the rotational loss term. t For the translation loss term, l represents the number of layers, and β, γ, and λ are adjustment factors used to adjust the weight of each loss term; The specific construction process includes the following steps: (3-a) Construct the overlap loss term L M : L M =∑-αM gt log(M)-(1-α)(1-M gt )log(1-M) Where M is the mask predicted by the model, M gt α represents the true value of the mask, and is the weighting parameter. (3-b) Construct the rotational loss term L q : L q =||q gt -q|| Where q gt q represents the true rotation quaternion obtained from the labeled data, and q represents the rotation quaternion predicted by the network model. (3-c) Construct the translation loss term L t : L t =||t gt -t|| Where t gt Let t be the true translation vector obtained from the labeled data, and t be the translation vector predicted by the network model. (4) Training the point cloud registration model: The point cloud registration model constructed in step (2) is trained using the dataset obtained in step (1); the error between the prediction result output by the model and the label is calculated using the loss function L constructed in step (3); the Adam algorithm is used to update the model parameters during the training process until the loss function converges, and the trained point cloud registration model is obtained. (5) Obtain the point cloud registration transformation results: After pixel-level mapping, the two test point clouds are input into the trained point cloud registration model, and the model output is the final point cloud registration transformation result.

2. The multi-level point cloud registration method based on overlapping region prediction according to claim 1, characterized in that, In step (2-a), the backbone network obtains four pairs of feature maps with successively decreasing scales, which are then input in pairs to the subsequent four overlapping region prediction modules based on convolutional neural networks.

3. The multi-level point cloud registration method based on overlapping region prediction according to claim 1, characterized in that, The convolutional module in step (2-a) includes a convolutional layer, a batch normalization layer, and a ReLU activation function.

4. The multi-level point cloud registration method based on overlapping region prediction according to claim 1, characterized in that, The weight-sharing convolutional neural network in step (2-b) adopts a three-layer one-dimensional convolutional structure, with each convolutional layer followed by batch normalization and ReLU activation function.

5. The multi-level point cloud registration method based on overlapping region prediction according to claim 1, characterized in that, The convolutional module in step (2-c) includes a convolutional layer, a batch normalization layer, and a ReLU activation function.

6. The multi-level point cloud registration method based on overlapping region prediction according to claim 1, characterized in that, The original feature branch and the frequency domain feature branch contained in the frequency domain feature enhancement module are fused using a residual connection method.

7. The multi-level point cloud registration method based on overlapping region prediction according to claim 1, characterized in that, In step (2-c), the attitude regression module outputs through two independent fully connected layers when regressing the quaternion representation and the displacement vector, thereby achieving decoupled modeling of the rotation and translation components.

8. The multi-level point cloud registration method based on overlapping region prediction according to claim 1, characterized in that, In step (3), α, β, γ and λ are all positive numbers and satisfy α∈[0.1,0.9], β∈[0.1,1.0], γ∈[0.1,1.0], λ∈[0.1,1.0].

Images