CN117351052A - Point cloud fine registration method based on feature consistency and spatial consistency - Google Patents

Abstract

The invention provides a point cloud precise registration method based on feature consistency and spatial consistency, and belongs to the field of computer vision technology. The method includes: processing point pairs after coarse registration of point clouds to obtain feature consistency and spatial consistency; using confidence to select seed points based on feature consistency and spatial consistency; defining k nearest neighbors. In the feature space, with the seed point as the center, its nearest neighbor points are found in the k-nearest neighbor feature space to form a cluster. The feature consistency and spatial consistency are combined with non-local attention to obtain the point cloud feature update formula to obtain the points in the cluster. Updated point cloud features containing long-distance information; generate a probability matching matrix through the updated point cloud features. Based on the probability matching matrix, the rotation vector and translation matrix are calculated through singular value decomposition to achieve point cloud registration. By adopting the present invention, the accuracy and speed of point cloud registration can be improved.

Description

A point cloud precision registration method based on feature consistency and spatial consistency

Technical field

The invention relates to the field of computer vision technology, and in particular refers to a point cloud precision registration method based on feature consistency and spatial consistency.

Background technique

Active point cloud Target point cloud/> Among them, the number of points of the source point cloud and the target point cloud is different, that is, N _x ≠ N _y . Point cloud registration is to find the coincident points of the source point cloud and the target point cloud, that is, the interior points, and calculate the rotation matrix and translation vector between the interior points. In recent years, point cloud precision registration methods can be divided into traditional methods and deep learning-based methods.

Traditional methods are represented by ICP and RANSAC, and ICP variant algorithms and RANSAC variant algorithms continue to appear. However, the common problems of traditional methods are: they can only obtain local optimal solutions and have low computational efficiency.

With the advancement of deep learning technology in recent years, more and more scholars are focusing on deep learning methods. In 2019, Wang proposed the DCP algorithm to achieve fast point cloud registration through the DGCNN network and attention mechanism, but it could not solve the problem of outlier elimination in point cloud registration. In 2021, Bai proposed the PointDSC method, which integrates spatial consistency into feature extraction and can achieve outlier elimination. However, it ignores the geometric features between point clouds, resulting in low accuracy of point cloud registration.

Contents of the invention

Embodiments of the present invention provide a point cloud precise registration method based on feature consistency and spatial consistency, which can improve the accuracy and speed of point cloud registration. The technical solutions are as follows:

On the one hand, a point cloud precision registration method based on feature consistency and spatial consistency is provided. The method is applied to electronic devices. The method includes:

Process the point pairs after coarse registration of the point cloud to obtain feature consistency and spatial consistency;

Based on feature consistency and spatial consistency, confidence is used to select seed points;

Define the k-nearest neighbor feature space, take the seed point as the center, find its nearest neighbor points in the k-nearest neighbor feature space, form a cluster, fuse feature consistency and spatial consistency with non-local attention, and obtain the point cloud feature update formula to obtain The updated point cloud features containing long-distance information of the points in the cluster;

A probability matching matrix is generated through the updated point cloud features. Based on the probability matching matrix, the rotation vector and translation matrix are calculated through singular value decomposition to achieve point cloud registration.

Further, processing the point pairs after coarse registration of the point cloud to obtain feature consistency and spatial consistency includes:

The normalized value of the feature difference of a point pair is defined as α. The α value reflects the feature consistency of the point pair. The smaller the α value, the more consistent the features are. The specific formula is as follows:

Δf _n =||g(x _n )-g(y _n )||

α _n ∈α, n=1, 2,...,N

Among them _, g(· _{) represents} the dynamic _{graph convolutional} neural network _; The nth initial point pair, the initial point pair is obtained by coarse registration of the point cloud, N represents the total number of point pairs, that is: Represents a 6-dimensional real number space,/> represents a 3-dimensional real number space; Δf _n represents the feature difference of the n-th point pair; α _n represents the feature consistency of the n-th point pair; ||·|| represents the Euclidean distance; [·] ₊ is a non-negative operation , indicating that α _n value is greater than or equal to 0;

Define the point-pair spatial distance difference between each two sets of point pairs as β. The β value reflects the spatial consistency. The smaller the β value, the closer the spatial position characteristics between the point pair and the point pair match. Suppose the i-th group of point pairs The Euclidean distance between is d _i , then the Euclidean distance difference d _ij between the i-th group of point pairs and the j-th group of point pairs is:

d _ij =||d _i -d _j ||

The spatial consistency β _ij between the i-th group of point pairs and the j-th group of point pairs is:

β _ij ∈ β, i=1, 2,...,N, j==1, 2,...,N.

Further, on the basis of feature consistency and spatial consistency, using confidence to select seed points includes:

Define seed confidence C as:

C＝α+β

Among them, the α value reflects the feature consistency of the point pair, and the β value reflects the spatial consistency;

The point pairs are sorted from large to small according to the confidence C value, and the point pairs with the top p% ranking of confidence are selected as seed points, where p is a constant.

Further, the k-nearest neighbor feature space is defined, with the seed point as the center, and its nearest neighbor points are found in the k-nearest neighbor feature space to form a cluster. The feature consistency and spatial consistency are combined with non-local attention to obtain point cloud features. The update formula to obtain the updated point cloud features containing long-distance information for the points in the cluster includes:

Define the k nearest neighbor feature space, with the seed point as the center, find its nearest neighbor points in the k nearest neighbor feature space, and record the index of the seed point and its nearest neighbor points as index, where index is a two-dimensional matrix with a size of [n _seed , n _k ], n _seed is the number of seed points, and n _k is the number of neighboring points centered on the seed point; the seed point and its neighboring points form a cluster. If the point is within the cluster, it is considered an interior point. If the point is within If it is outside the cluster, it is considered to be an outlier;

By fusing feature consistency and spatial consistency with non-local attention, the point cloud feature update formula in the k-nearest neighbor feature space is obtained:

Among them, f on the right side of the equal sign is the point pair feature obtained after one layer of convolution of the initial point pair coordinates obtained by rough registration; f on the left side of the equal sign is the updated feature, and MLP(·) is multi-layer perception machine model, softmax(·) is the activation function, χ represents feature similarity, Represents vector cross product, h(·) is a linear mapping function, α′ expands α, that is:

Among them, N _c represents the feature dimension in the feature space;

Through the obtained point cloud feature update formula, the updated point cloud features containing long-distance information of the points in the cluster are obtained.

Further, the probability matching matrix is generated by the updated point cloud features. Based on the probability matching matrix, the rotation vector and the translation matrix are calculated through singular value decomposition. The realization of point cloud registration includes:

Perform L2 normalization and non-negative operations on the updated point cloud features, and calculate their main eigenvalues to obtain a probability matching matrix;

Based on the probability matching matrix, the rotation vector and translation matrix are calculated through singular value decomposition to achieve point cloud registration.

Further, the updated features are subjected to L2 normalization and non-negative operations, and their main eigenvalues are calculated to obtain a probability matching matrix including:

Through the updated feature information, the probability matching matrix e is generated:

e＝L{m}

Among them, m is normalized by the updated point cloud feature f output from the k-nearest neighbor feature space. L{·} represents the calculation of the main feature vector using the power iteration method. [·] ₊ is a non-negative operation, indicating that the m value is non-negative.

Further, the rotation vector R and translation matrix t calculated through singular value decomposition are expressed as:

In one aspect, an electronic device is provided. The electronic device includes a processor and a memory. At least one instruction is stored in the memory. The at least one instruction is loaded and executed by the processor to achieve the above feature-based consistency. and spatially consistent point cloud precision registration method.

On the one hand, a computer-readable storage medium is provided. At least one instruction is stored in the storage medium. The at least one instruction is loaded and executed by a processor to implement the above-mentioned point cloud precision based on feature consistency and spatial consistency. Registration method.

The beneficial effects brought by the technical solutions provided by the embodiments of the present invention include at least:

(1) In terms of point cloud registration, in order to solve the problem that traditional methods easily fall into local optimal solutions, embodiments of the present invention propose a point cloud precise registration method based on feature consistency and spatial consistency. This method can quickly , accurately give the rotation vector and translation matrix, and obtain the global optimal solution.

(2) In point cloud registration, when the number of source point clouds and target point clouds is different and the number of low overlap rates is low, it is difficult for previous algorithms to deal with the problem of outliers. The embodiment of the present invention eliminates outliers through feature consistency and spatial consistency. outlier points, and filter out matching interior points through a confidence-based seed selection mechanism; in this way, by processing point pairs after coarse point cloud registration, low-level points can be completed with higher accuracy and faster speed. Point cloud registration under overlap ratio.

(3) In view of the problem that previous algorithms ignore topological information between point clouds in point cloud registration, the embodiment of the present invention obtains geometric information between point clouds by introducing a dynamic graph convolutional neural network; by fusing non-local attention, obtains Point cloud features containing long-distance information greatly improve the accuracy of point cloud registration.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without exerting creative efforts.

Figure 1 is a schematic flow chart of a point cloud precise registration method based on feature consistency and spatial consistency provided by an embodiment of the present invention;

Figure 2 is a schematic structural diagram of a point cloud registration network provided by an embodiment of the present invention;

Figure 3(a) is a schematic diagram of feature consistency provided by an embodiment of the present invention;

Figure 3(b) is a schematic diagram of spatial consistency provided by an embodiment of the present invention;

Figure 4 is a schematic structural diagram of a feature consistency network provided by an embodiment of the present invention;

Figure 5 is a schematic diagram of the k-nearest neighbor feature space structure provided by the embodiment of the present invention;

Figure 6 is a schematic diagram of the registration results on the 3DMatch public data set provided by the embodiment of the present invention;

FIG. 7 is a schematic structural diagram of an electronic device provided by an embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

As shown in Figures 1 and 2, embodiments of the present invention provide a point cloud precise registration method based on feature consistency and spatial consistency. This method is used for point cloud coarse registration. Pairs usually contain wrong matching points and outliers, and are easy to fall into local optimal solutions; through this method, outliers can be eliminated, registration accuracy improved, and the global optimal solution obtained. The method can be implemented by an electronic device, which can be a terminal or a server. The method includes:

S101, process the point pairs after coarse registration of the point cloud to obtain feature consistency and spatial consistency; specifically, the following steps may be included:

A1, define the normalized value of the feature difference of a point pair as α. The α value reflects the feature consistency of the point pair. The smaller the α value, the more consistent the features are. The specific formula is as follows:

Δf _n =||g(x _n )-g(y _n )||

α _n ∈α, n=1, 2,...,N

Among them, g(·) represents the Dynamic Graph Convolutional Neural Network (DGCNN); x _n and y _n are decomposed from c _n , and x _n , y _n and c _n respectively represent the nth source point cloud , the nth target point cloud and the nth initial point pair, N represents the total number of point pairs, that is: Represents a 6-dimensional real number space,/> Represents a 3-dimensional real number space, and the initial point pair is obtained by rough registration of the point cloud; Δf _n represents the feature difference of the nth point pair; α _n represents the feature consistency of the nth point pair; ||·|| represents Euclidean Rider distance; [·] ₊ is a non-negative operation, indicating that the α _n value is greater than or equal to 0;

A2, define the point-pair spatial distance difference between each two sets of point pairs as β. The β value reflects the spatial consistency. The smaller the β value, the closer the spatial position characteristics between the point pair and the point pair match. Suppose the i-th group The Euclidean distance between point pairs is d _i , then the Euclidean distance difference d _ij between the i-th group of point pairs and the j-th group of point pairs is:

d _ij =||d _i -d _j ||

The spatial consistency β _ij between the i-th group of point pairs and the j-th group of point pairs is:

β _ij ∈ β, i=1, 2,...,N, j==1, 2,...,N.

As shown in Figure 3, Figure 3(a) is a schematic diagram of feature consistency. c1 and c2 are correct point pairs, c3 is an incorrect point pair, and Δf ₁ , Δf ₂ and Δf ₃ represent the feature differences between point pairs: | Δf ₁ -Δf ₂ |＜|Δf ₂ -Δf ₃ |; Figure 3(b) is a schematic diagram of spatial consistency, c1 and c2 are correct point pairs, c3 is an incorrect point pair, d1, d2, d3 and d4 represent points and The Euclidean distance between points is |d1-d2|<|d3-d4|.

Figure 4 is a schematic diagram of the DGCNN network in feature consistency calculation. Specifically, the network inputs the source point cloud x _n and the target point cloud y _n , performs edge convolution operations on the source point cloud and the target point cloud respectively, and passes through four layers of edges. Convolution extraction, the number of output channels of the four layers of convolution is [16, 16, 32, 64], the size of the convolution kernel of each layer is 1×1, the convolution output of each layer is added, and passed through the multi-layer perceptron and Maximum pooling layer to obtain the final feature value. Difference the source point cloud feature value and the target point cloud feature value to obtain the feature difference value Δf _n . After normalizing Δf _n , the feature consistency α is obtained.

S102, based on feature consistency and spatial consistency, use confidence to select seed points; specifically, the following steps may be included:

B1, define the seed confidence C as:

C＝α+β

Among them, the α value reflects the feature consistency of the point pair, and the β value reflects the spatial consistency. The greater the confidence C value of the point pair, the more the point pair satisfies the feature consistency and spatial consistency, and it is correct. The greater the likelihood of matching points;

B2, sort the point pairs according to the confidence C value from large to small, and select the point pairs with the top p% (for example, 10%) confidence level as the seed points, where p is a constant.

S103, define the k-nearest neighbor feature space, take the seed point as the center, find its nearest neighbor points in the k-nearest neighbor feature space, form a cluster, fuse feature consistency and spatial consistency with non-local attention, and obtain the point cloud feature update formula. To obtain updated point cloud features containing long-distance information for points in the cluster; the specific steps may include the following:

C1, define the k nearest neighbor feature space, take the seed point as the center, find its nearest neighbor points in the k nearest neighbor feature space, record the index of the seed point and its neighbor points as index, index is a two-dimensional matrix, its size is [n _seed , n _k ], where n _seed is the number of seed points, and n _k is the number of neighboring points centered on the seed point; the seed point and its neighboring points form a cluster. If a point is within the cluster, it is considered an interior point. If If a point is outside the cluster, it is considered an outlier;

C2, fuse feature consistency and spatial consistency with non-local attention, and obtain the point cloud feature update formula in k-nearest neighbor feature space as:

Among them, f on the right side of the equal sign is the point cloud position feature (abbreviation: point cloud feature) obtained after one layer of convolution of the initial point pair coordinates obtained by rough registration; f on the left side of the equal sign is the updated feature. MLP(·) is the multi-layer perceptron model, softmax(·) is the activation function, χ represents feature similarity, Represents vector cross product, h(·) is a linear mapping function, α′ expands α, that is:

Among them, N _c represents the feature dimension in the feature space;

In order to better understand the point cloud feature update formula described in this embodiment, it is necessary to explain the classic non-local attention model. The classic non-local attention model can be expressed as:

Non-local attention models have the ability to relate contextual information.

The feature update formula in this embodiment integrates feature consistency and spatial consistency with non-local attention to obtain long-distance features combined with contextual information in the feature space, and fuses them with the original features to complete the feature update. .

As shown in Figure 5, the k-nearest neighbor feature space network inputs point cloud features f, f undergoes a layer of convolution to obtain features φ and/> After transposition, we get θ,/> φ and θ perform matrix multiplication to obtain the feature similarity χ; after fusing χ with the spatial consistency β, the whole is used as a weight and added to the feature consistency α′, and the updated feature f is obtained through the multi-layer perceptron MLP. The new Features combine feature information and spatial information and increase the receptive field.

C3. Through the obtained point cloud feature update formula, the updated point cloud features containing long-distance information of the points in the cluster are obtained. Each point pair has a corresponding point cloud feature. The point cloud features of the point pair are sorted according to the index, that is, the points in the cluster are filtered out. The number of points in the cluster is n _seed × n _k ; for point pairs not in the index, Its point cloud features will be discarded to complete outlier elimination.

S104. Generate a probability matching matrix through the updated point cloud features. Based on the probability matching matrix, obtain the rotation vector and translation matrix through singular value decomposition calculation to achieve point cloud registration; specifically, the following steps may be included:

D1, perform L2 normalization and non-negative operations on the updated point cloud features, and calculate its main eigenvalues to obtain a probability matching matrix;

In this embodiment, the probability matching matrix e is generated through the updated feature information:

e＝L{m}

Among them, m is normalized by the updated point cloud feature f output from the k-nearest neighbor feature space. L{·} represents the calculation of the main feature vector using the power iteration method. [·] ₊ is a non-negative operation, indicating that the m value is non-negative.

D2, based on the probability matching matrix, calculates the rotation vector and translation matrix through singular value decomposition to achieve point cloud registration.

In this embodiment, the rotation vector R and translation matrix t calculated through singular value decomposition are expressed as:

In order to verify the effect of the point cloud precise registration method based on feature consistency and spatial consistency provided in this embodiment, this embodiment uses the following evaluation indicators:

(1) Recall rate RR: Recall rate is one of the important indicators to measure the performance of the registration algorithm. It is used to evaluate the algorithm's ability to identify correctly matched point pairs. The recall rate is defined as the ratio of the number of correctly matched point pairs to the total number of correctly matched point pairs. The recall rate ranges from 0 to 1. The value The larger the value, the higher the robustness and accuracy of the algorithm, and the better it can identify the correct matching point pairs.

(2) Rotation error RE: Rotation error (Rotation Error) is used to measure the accuracy of the rotation matrix in the point cloud transformation matrix, and is defined as follows:

in, is the predicted rotation matrix value, R ^* is the rotation matrix true value, arccos is the inverse cosine function, and Tr is the trace operation.

(3) Translation error TE: Translation error (Translation Error) is used to measure the accuracy of the translation vector in the point cloud transformation matrix, and is defined as follows:

in, is the predicted translation vector value, and t ^* is the true value of the translation vector.

In order to verify the performance of the point cloud precise registration method based on feature consistency and spatial consistency provided by the embodiment of the present invention, in this embodiment, the public data set 3DMatch is used to compare with the classic algorithms in recent years; the recall rate is used RR, rotation error RE and translation error TE are used as evaluation criteria, and the point cloud registration results are shown in Table 1. As shown in Figure 6, the source point cloud and the target point cloud do not completely overlap; according to Table 1 and Figure 6, it can be seen that the point cloud registration results provided by the embodiment of the present invention are better.

Table 1 Point cloud registration results

In summary, the point cloud precise registration method based on feature consistency and spatial consistency described in the embodiment of the present invention has at least the following beneficial effects:

(1) In terms of point cloud registration, in order to solve the problem that traditional methods easily fall into local optimal solutions, embodiments of the present invention propose a point cloud precise registration method based on feature consistency and spatial consistency. This method can quickly , accurately give the rotation vector and translation matrix, and obtain the global optimal solution.

(2) In point cloud registration, when the number of source point clouds and target point clouds is different and the number of low overlap rates is low, it is difficult for previous algorithms to deal with the problem of outliers. The embodiment of the present invention eliminates outliers through feature consistency and spatial consistency. outlier points, and filter out matching interior points through a confidence-based seed selection mechanism; in this way, by processing point pairs after coarse point cloud registration, low-level points can be completed with higher accuracy and faster speed. Point cloud registration under overlap ratio.

(3) In view of the problem that previous algorithms ignore the topological information between point clouds in point cloud registration, the embodiment of the present invention obtains the geometric information between point clouds by introducing a dynamic graph convolutional neural network (wherein, the nth point cloud in the source point cloud The geometric information of points is g(x _n ), and the geometric information of the nth point in the target point cloud is g(y _n )). By fusing non-local attention, point cloud features containing long-distance information are obtained, which greatly improves The accuracy of point cloud registration is improved.

FIG. 7 is a schematic structural diagram of an electronic device 600 provided by an embodiment of the present invention. The electronic device 600 may vary greatly due to different configurations or performance, and may include one or more processors (central processing units, CPUs) 601 and one or more memories 602, wherein at least one instruction is stored in the memory 602, and the at least one instruction is loaded and executed by the processor 601 to implement the above point cloud based on feature consistency and spatial consistency. Precise registration method.

In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including instructions. The instructions can be executed by a processor in a terminal to complete the above-mentioned point cloud precise registration based on feature consistency and spatial consistency. method. For example, the computer-readable storage medium may be ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, optical data storage device, etc.

Those of ordinary skill in the art can understand that all or part of the steps to implement the above embodiments can be completed by hardware, or can be completed by instructing relevant hardware through a program. The program can be stored in a computer-readable storage medium. The above-mentioned The storage media mentioned can be read-only memory, magnetic disks or optical disks, etc.

The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection of the present invention. within the range.

Claims (7)

1. A point cloud precision registration method based on feature consistency and spatial consistency, which is characterized by: Process the point pairs after coarse registration of the point cloud to obtain feature consistency and spatial consistency; Based on feature consistency and spatial consistency, confidence is used to select seed points; Define the k-nearest neighbor feature space, take the seed point as the center, find its nearest neighbor points in the k-nearest neighbor feature space, form a cluster, fuse feature consistency and spatial consistency with non-local attention, and obtain the point cloud feature update formula to obtain The updated point cloud features containing long-distance information of the points in the cluster; A probability matching matrix is generated through the updated point cloud features. Based on the probability matching matrix, the rotation vector and translation matrix are calculated through singular value decomposition to achieve point cloud registration. 2. The point cloud precise registration method based on feature consistency and spatial consistency according to claim 1, characterized in that, the point pairs after the point cloud coarse registration are processed to obtain feature consistency and spatial consistency. Consistency includes: The normalized value of the feature difference of a point pair is defined as α. The α value reflects the feature consistency of the point pair. The smaller the α value, the more consistent the features are. The specific formula is as follows: Δf _n =‖g(x _n )-g(y _n )‖ Among them, g(·) represents the dynamic graph convolutional neural network; x _n and y _n are decomposed by c _n , x _n , y _n and c _n represent the nth source point cloud, nth target point cloud and The nth initial point pair, the initial point pair is obtained by coarse registration of the point cloud, N represents the total number of point pairs, that is: Represents a 6-dimensional real number space,/> represents a 3-dimensional real number space; Δf _n represents the feature difference of the n-th point pair; α _n represents the feature consistency of the n-th point pair; ‖·‖ represents the Euclidean distance; [·] ₊ is a non-negative operation, indicating α _n value is greater than or equal to 0; Define the point-pair spatial distance difference between each two sets of point pairs as β. The β value reflects the spatial consistency. The smaller the β value, the closer the spatial position characteristics between the point pair and the point pair match. Suppose the i-th group of point pairs The Euclidean distance between is d _i , then the Euclidean distance difference d _ij between the i-th group of point pairs and the j-th group of point pairs is: d _ij =||d _i -d _j || The spatial consistency β _ij between the i-th group of point pairs and the j-th group of point pairs is: 3. The point cloud precision registration method based on feature consistency and spatial consistency according to claim 1, characterized in that, on the basis of feature consistency and spatial consistency, confidence is used to select seed points. include: Define seed confidence C as: C＝α+β Among them, the α value reflects the feature consistency of the point pair, and the β value reflects the spatial consistency; The point pairs are sorted from large to small according to the confidence C value, and the point pairs with the top p% ranking of confidence are selected as seed points, where p is a constant. 4. The point cloud precise registration method based on feature consistency and spatial consistency according to claim 1, characterized in that the k-nearest neighbor feature space is defined, with the seed point as the center, and the k-nearest neighbor feature space is searched. Its neighboring points form a cluster, and the feature consistency and spatial consistency are combined with non-local attention to obtain the point cloud feature update formula to obtain the point cloud features containing long-distance information after updating the points in the cluster, including: Define the k nearest neighbor feature space, with the seed point as the center, find its nearest neighbor points in the k nearest neighbor feature space, and record the index of the seed point and its nearest neighbor points as index, where index is a two-dimensional matrix with a size of [n _seed , n _k ], n _seed is the number of seed points, n _k is the number of neighboring points centered on the seed point; the seed point and its neighboring points form a cluster. If the point is within the cluster, it is considered to be an interior point. If the point is within If it is outside the cluster, it is considered to be an outlier; By fusing feature consistency and spatial consistency with non-local attention, the point cloud feature update formula in the k-nearest neighbor feature space is obtained: Among them, f on the right side of the equal sign is the point pair feature obtained after one layer of convolution of the initial point pair coordinates obtained by rough registration; f on the left side of the equal sign is the updated feature, and MLP(·) is multi-layer perception machine model, softmax(·) is the activation function, χ represents feature similarity, Represents vector cross product, h(·) is a linear mapping function, α′ expands α, that is: Among them, N _c represents the feature dimension in the feature space; Through the obtained point cloud feature update formula, the updated point cloud features containing long-distance information of the points in the cluster are obtained. 5. The point cloud precise registration method based on feature consistency and spatial consistency according to claim 1, characterized in that the probability matching matrix is generated through the updated point cloud features, based on the probability matching matrix, through singular Value decomposition calculation is used to obtain the rotation vector and translation matrix. The implementation of point cloud registration includes: Perform L2 normalization and non-negative operations on the updated point cloud features, and calculate their main eigenvalues to obtain a probability matching matrix; Based on the probability matching matrix, the rotation vector and translation matrix are calculated through singular value decomposition to achieve point cloud registration. 6. The point cloud precision registration method based on feature consistency and spatial consistency according to claim 5, characterized in that the updated features are subjected to L2 normalization and non-negative operations, and their main points are calculated. Eigenvalues, the probability matching matrix obtained includes: Through the updated feature information, the probability matching matrix e is generated: e＝L{m} Among them, m is normalized by the updated point cloud feature f output from the k-nearest neighbor feature space. L{·} represents the calculation of the main feature vector using the power iteration method. [·] ₊ is a non-negative operation, indicating that the m value is non-negative. 7. The point cloud precise registration method based on feature consistency and spatial consistency according to claim 6, characterized in that the rotation vector R and the translation matrix t calculated by singular value decomposition are expressed as: