A point cloud registration method, system, storage medium, and device based on a diffusion model.
By using a point cloud registration method based on a diffusion model, we can extract point cloud features and simulate the thermal diffusion process through graph neural networks. This solves the problems of accuracy and computational resources in point cloud registration under conditions of sparse features or occlusion, and achieves more efficient and accurate point cloud registration.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2024-03-21
- Publication Date
- 2026-05-26
AI Technical Summary
In existing technologies, point cloud registration methods are not accurate when features are sparse or occluded, and optimization-based methods consume a lot of computational resources and are sensitive to initial parameters, making them prone to getting trapped in local optima.
A point cloud registration method based on a diffusion model is adopted. Through a point cloud feature extraction module, a feature interaction module, a point cloud generation module, and a transformation matrix prediction module, local features are extracted using a graph neural network, and the registration problem is transformed into a generation problem. The diffusion model is used to simulate the thermal diffusion process to generate aligned point clouds.
It improves the accuracy and robustness of point cloud registration, reduces computational resource consumption, and enhances registration speed and accuracy, effectively resisting the effects of noise and outliers.
Smart Images
Figure CN117994307B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision technology, and specifically to a point cloud registration method, system, storage medium, and device based on a diffusion model. Background Technology
[0002] Point cloud registration refers to the process of aligning two or more point cloud datasets in the same coordinate system. A point cloud is a dataset composed of a large number of discrete 3D points, typically obtained through methods such as laser scanning, photogrammetry, or 3D scanning, representing the geometric shape and positional information of a target surface. The goal of point cloud registration is to find a transformation that aligns point cloud datasets acquired at different positions or angles, facilitating subsequent applications such as 3D reconstruction, target recognition, or environmental modeling. Point cloud registration has significant applications in computer vision, robot perception, and map building.
[0003] Feature-matching methods typically rely on extracting feature points from point clouds and finding matching feature point pairs for registration. However, when features in the point cloud are sparse or occluded, the accuracy of feature matching is affected, leading to unstable registration results. Secondly, optimization-based methods usually transform the point cloud registration problem into an optimization problem, solving for the optimal registration transformation parameters by minimizing the registration error. However, these methods often require significant computational resources and time, are sensitive to the choice of initial parameters, and are prone to getting trapped in local optima. Summary of the Invention
[0004] To address the problems existing in the prior art, this invention proposes a point cloud registration method based on a diffusion model. The method includes a point cloud registration model training stage and a point cloud registration stage; wherein the point cloud registration model includes a point cloud feature extraction module, a feature interaction module, a point cloud generation module, and a transformation matrix prediction module.
[0005] The point cloud registration model training phase includes: acquiring training samples, which include source point cloud A and target point cloud B; using a point cloud feature extraction module to extract features from source point cloud A and target point cloud B respectively, obtaining features Feat1 and Feat2; using a feature interaction module to fuse features Feat1 and Feat2, obtaining fused features Fusion1 and Fusion2; using a point cloud generation module to generate two aligned point clouds Ag and Bg based on fused features Fusion1 and Fusion2; using a transformation matrix prediction module to calculate the correspondence between the original point cloud A and the generated point cloud Ag, and between the original point cloud B and the generated point cloud Bg, obtaining the predicted transformation matrix and translation vector; calculating the model's loss function, adjusting the model parameters, and completing the model training when the loss function converges.
[0006] The point cloud registration stage includes: inputting the source point cloud and target point cloud of the actual object to be measured into the trained point cloud registration model to obtain the transformation matrix R and translation vector p; and using the transformation matrix R and translation vector p to register the source point cloud.
[0007] A point cloud registration system based on a diffusion model, comprising: a point cloud acquisition module, a point cloud feature extraction module, a feature interaction module, a point cloud generation module, and a transformation matrix prediction module;
[0008] The point cloud acquisition module is used to acquire the point cloud to be registered and the target point cloud;
[0009] The point cloud feature extraction module is used to extract the feature map of the point cloud information to be registered and the feature map of the target point cloud;
[0010] The feature interaction module is used to perform feature interaction fusion between the feature map of the point cloud information to be registered and the feature map of the target point cloud to obtain two fused feature maps.
[0011] The point cloud generation module generates two aligned point clouds based on the fused feature maps Fusion1 and Fusion2.
[0012] The transformation matrix prediction module predicts the transformation matrix and translation vector based on the aligned point cloud, and registers the point cloud to be tested based on the transformation matrix and translation vector.
[0013] To achieve the above objectives, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any of the above-described point cloud registration methods based on the diffusion model.
[0014] To achieve the above objectives, the present invention also provides a point cloud registration device based on a diffusion model, comprising a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory and is used to execute the computer program stored in the memory, so that the point cloud registration device based on a diffusion model performs any of the above-described point cloud registration methods based on a diffusion model.
[0015] The beneficial effects of this invention are:
[0016] The point cloud registration method based on a diffusion model of this invention transforms the registration problem into a generation problem. In the point cloud feature extraction module, the pose difference of the point clouds to be registered is first normalized, and then a graph neural network is used to extract the local features of the points. The obtained local features are then processed using Token-mixing MLP and Channel-mixing MLP to ensure that the point cloud features maintain global and local structural information at multiple levels. This structure enables the network to effectively process point cloud data, while taking into account the spatial distribution of the point cloud and the feature details of each point, which is beneficial for more accurate point cloud analysis and processing. The point cloud generation module of this invention uses a diffusion model, treating the points in the point cloud as particles in contact with a hot bath in a thermodynamic system. These particles diffuse from the original distribution to the noise distribution, thereby improving the relationship between the generated point cloud and the original point cloud and improving the accuracy of point cloud registration. This application analyzes the geometric structure of the point cloud by simulating the thermal diffusion process, transforming the registration problem into a generation problem. This can resist the influence of noise and outliers to a certain extent, thus having good robustness, accelerating the speed of point cloud registration, and improving the accuracy of point cloud registration. Attached Figure Description
[0017] Figure 1 This is a flowchart of the point cloud registration process based on the diffusion model of the present invention;
[0018] Figure 2 The flowchart of the point cloud feature extraction module provided by this invention is shown below;
[0019] Figure 3 This is a flowchart of the Token-mixing MLP and Channel-mixing MLP of the present invention;
[0020] Figure 4 The flowchart of the feature interaction module provided by this invention. Detailed Implementation
[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0022] A point cloud registration method based on a diffusion model, such as Figure 1As shown, the method includes: a point cloud registration model training stage and a point cloud registration stage; wherein the point cloud registration model includes a point cloud feature extraction module, a feature interaction module, a point cloud generation module, and a transformation matrix prediction module; the point cloud registration model training stage includes: acquiring training samples, which include source point cloud A and target point cloud B; using the point cloud feature extraction module to extract features from source point cloud A and target point cloud B respectively, to obtain features Feat1 and Feat2; using the feature interaction module to fuse features Feat1 and Feat2 to obtain fused features Fusion1 and Fusion2; using the point cloud generation module to generate two aligned point clouds Ag and Bg based on the fused features Fusion1 and Fusion2; using the transformation matrix prediction module to calculate the correspondence between the original point cloud A and the generated point cloud Ag, and between the original point cloud B and the generated point cloud Bg, to obtain the predicted transformation matrix and translation vector; calculating the model's loss function, adjusting the model parameters, and completing the model training when the loss function converges.
[0023] The point cloud registration stage includes: inputting the source point cloud and target point cloud of the actual object to be measured into the trained point cloud registration model to obtain the transformation matrix R and translation vector p; and using the transformation matrix R and translation vector p to register the source point cloud.
[0024] In this embodiment, the object to be tested can be an aircraft, robot, car, etc.
[0025] Specifically, such as Figure 2 As shown, in the point cloud feature extraction module, the pose difference of the point cloud to be registered is first normalized, and then a graph neural network is used to extract the local features of the points, which includes a graph construction layer, a convolutional layer, and a ReLU activation function. Figure 3 As shown, using Token-mixing MLP and Channel-mixing MLP to obtain point cloud features preserves global and local structural information at multiple levels. This structure enables the network to effectively process point cloud data while considering the spatial distribution of the point cloud and the feature details of each point, which is beneficial for more accurate point cloud analysis and processing.
[0026] The graph neural network used is the T-Net network, a spatial transformation network. The inputs are point clouds A and B. T-Net first extracts global features from the two input point clouds, then uses these connected global features to learn six free parameters (three for rotation and three for translation). These six parameters are then converted into a 4×4 rigid transformation matrix, which is used to transform one of the input point clouds.
[0027] Specifically, the results of inputting point clouds A and B are:
[0028] x1 = ReLU(Conv1d(A))
[0029] x1 = ReLU(Conv1d(x1))
[0030] x3 = ReLU(Conv1d(x2))
[0031] Where ReLU represents the ReLU activation function, Conv1d represents one-dimensional convolution, and A represents all points in the point cloud, with each point having x, y, and z coordinates.
[0032] By repeatedly performing convolution on the point cloud data, followed by a ReLU activation function, and repeating this operation three times, the features of the point cloud are progressively increased in dimensionality. That is:
[0033] x global =Mapxooling(x3)
[0034] Maxpooling is the maximum pooling method used to obtain the global features x of point cloud A. global Similarly, point cloud B can also obtain global features. By concatenating the two global features together, a new x can be obtained. global .
[0035] y=Linear(ReLU(Linear(ReLU(Linear(x global ))))
[0036] Among them, Linear is a fully connected layer. After passing through three fully connected layers, six parameters are obtained: three rotation angles and three translation angles. These six parameters are then converted into a 4×4 rigid transformation matrix to transform one of the input point clouds.
[0037] Obtain the k-NN point set of the i-th point in the point cloud, extract the features of each point and its neighbors, obtain the features of the k-NN neighborhood of the i-th point by copying, expanding and splicing, and obtain the local features of each point by performing two-dimensional convolution and ReLU activation function.
[0038] f z =ReLU(Conv2d(f))
[0039] Where Conv2d represents two-dimensional convolution, and f represents all points, where each point has features from the k-NN domain.
[0040] In this embodiment, the Token-mixing MLP (Multilayer Perceptron) is first performed on the point spatial dimension, which helps capture global information. The network attempts to understand the relationships between points, thereby capturing the global structure of the entire point cloud. Channel-mixing MLP is then performed on the channel dimension of each point to preserve the local information of each point. This step focuses on the feature channels of a point, aiming to understand how the features within each point interact, preserving the local structural information of the point. Each MLP contains two fully connected layers. The final features can preserve global and local structural information at multiple levels. This structure enables the network to efficiently process point cloud data, taking into account the spatial distribution of the point cloud and the feature details of each point, which is beneficial for more accurate point cloud analysis and processing.
[0041] Feature f z After passing through two MLPs, respectively performing calculations on the spatial dimension and the channel dimension of the points, the following results are obtained:
[0042] f g =MLP(layernorm(f z ) T ) T +f z
[0043] f l =MLP(layernorm(f g ))+f g
[0044] Here, Layernorm is the normalization operation, + is skip link, and T is transpose.
[0045] like Figure 4 As shown, the feature interaction module is used to fuse the features of source point cloud A and target point cloud B to obtain features Fusion1 and Fusion2 that have the features of each other. Based on features Fusion1 and Fusion2, which have global and local features, structural information features Fusion1 and Fusion2 of the intersection point cloud are obtained, including:
[0046] Step 1: Perform max pooling on Feature1 and Feature2 to obtain global features G1 and G2:
[0047] G1 = softmax(Feat1)
[0048] G2 = softmax(Feat2)
[0049] Step 2: Calculate the differences between Feature 1 and global feature G1, and between Feature 2 and global feature G2.
[0050] feat_diff_12=feat1-feat2
[0051] feat_diff_21=feat2-feat1
[0052] Step 3: Repeat these feature differences along each batch and feature dimension to concatenate them with features Feat1 and Feat2:
[0053] Fusion1=feat1+repeat(feat_diff_12,N)
[0054] Fusion2=feat2+repeat(feat_diff_21,N)
[0055] Where feat_diff_12 is the difference between Feature1 and global feature G1, feat_diff_21 is the difference between Feature2 and global feature G2, and N is the number of points in the point cloud.
[0056] The point cloud generation module uses the recently popular diffusion model, treating the points in the point cloud as particles in a thermodynamic system in contact with a hot bath, diffusing from the original distribution to a noise distribution. Therefore, point cloud generation is equivalent to learning a reverse diffusion process that transforms the noise distribution into the desired shape distribution. Specifically, the reverse diffusion process of the point cloud is modeled as a Markov chain based on certain latent shape variables.
[0057] Specifically, each point x2 in the point cloud can be considered as a sample independent of a point distribution, whose distribution can be represented as... Here, z is the shape latent code that determines the distribution of points, where the shape latent code is also the fusion features Fusion1 and Fusion2. Physically, over time, these points gradually diffuse into a chaotic set of points. This process is called diffusion, which transforms the originally meaningful point distribution into a noisy distribution. The forward diffusion process is modeled as a Markov chain:
[0058]
[0059] in This is a Markov diffusion kernel, where T is the number of steps required for diffusion. This kernel adds noise at the point in the previous time step and models the distribution of points in the next time step. The diffusion kernel is defined as:
[0060]
[0061] Where β1…β T It is the variance scheduling hyperparameter controlling the diffusion rate of the process. The goal is to generate point clouds with meaningful shapes from the latent representation z-encoded. The generation process is viewed as the reverse of the diffusion process, passing points through a reverse Markov chain to ultimately form the desired shape. Unlike the forward diffusion process, which simply adds noise to the points, the reverse process aims to recover the desired shape from the input noise (learning to remove the added noise, thus recovering the desired point cloud shape distribution), which requires training on the data. The generation of the reverse diffusion process is expressed as:
[0062]
[0063] p θ (x (t-1) |x (t) ,z)=N(x (t-1) |υ θ x (t) ,t,z)β t I)
[0064] Where μ θ The estimated mean is learned by a neural network with parameters θ. z is the latent code encoding the target shape of the point cloud. The initial distribution is set to a standard normal distribution N(0, I). Given a shape latent z, a set of values is passed through a reverse Markov chain from... Points are sampled from the sampled points to obtain a point cloud with the target shape.
[0065] In this embodiment, a point cloud data format conforming to a Gaussian distribution is generated. Using the previously fused features Fusion1 and Fusion2, sampling and denoising are performed according to the principles described above. Through progressive denoising, and after T steps, a point cloud with the target shape is obtained.
[0066] In the above generation method, the transformation matrix prediction module utilizes the proposed loss function to maintain the correspondence between the generated point cloud and the original point cloud. Therefore, these correspondences can be used to estimate the transformation matrix, specifically including:
[0067] Step 1: Given the original point cloud A and the generated point cloud Ag, establish the point-to-point correspondence:
[0068] H=(h1,,,,h N )
[0069] Since the number of point clouds is N, H has N corresponding points.
[0070] Perform m random samplings on the correspondence between all points in the original point cloud and the generated point cloud in H to obtain the pool M = (x1, x2, ..., x...). mThe pool contains k corresponding elements for each element, for example, k=4. The transformation matrix set T for each element in M is calculated using SVD decomposition based on these correspondences. M =(T1,T2...,T) m ).
[0071] Step 2: Calculate the projection error of the transformation matrix and select the matrix with the minimum projection error T. A Similarly, T can also be calculated from the original point cloud B and the generated point cloud Bg. B Its formula is:
[0072] T A =min∑ i∈m ||T i AA g ||2
[0073] Among them, T i For the transformation matrix set T M The i-th transformation matrix in
[0074] Step 3, the correspondence between the original point cloud A and the generated point cloud Ag, and between the original point cloud B and the generated point cloud Bg is as follows:
[0075] [A g ,1]=[A,1]T A
[0076] [B g ,1]=[B,1]T B
[0077] Since Ag and Bg are the generated aligned point clouds, the final transformation matrices of the input point clouds A and B can be calculated as follows:
[0078] T est =(T A ) -1 T B
[0079] During training, the loss function L of the point cloud registration model of this invention is:
[0080] L = l abs +l relative +l cyc +l T
[0081] Among them, l abs Represented as absolute loss, l relative Represented as relative loss, l cyc Represented as cycle consistency loss, l T This is represented as the transformation loss.
[0082] Absolute lossabs The Earth Mover's Distance (EMD) is calculated using the generated point clouds Ag and Bg. The purpose is to align the generated point clouds Ag and Bg. The corresponding formula is:
[0083] l abs =EMD(A g B g )
[0084] EMD(X,Y)=min θ：X→Y ||x-θ(x)||2
[0085] EMD relies on solving an optimization problem to find a one-to-one bijective mapping: θ: X→Y, which appears only in points in the Ag and Bg point clouds, and then to compute the sum of the pairwise distances between x and θ(x) and minimize it.
[0086] l relative Relative loss: This method maintains relative structural similarity between the generated and original point clouds, ensuring that the i-th point in the original point cloud corresponds to the i-th point in the generated point cloud. This guarantees that the generated and original point clouds maintain consistency in both shape and data. The calculation is as follows:
[0087]
[0088] Where MAE represents the mean absolute error, E A E Ag E B E Bg The edge set is calculated as follows:
[0089] E = |P i -P (i+1)％(N-1) |2
[0090] Where N is the number of points in the point cloud, and an edge is the connection between two adjacent points in the data.
[0091] l cyc Cycle consistency loss: The network can arbitrarily transform the input point cloud to match the distribution of the target domain. The cycle consistency loss ensures that the reverse transformation of each point cloud A should bring A back to the original point cloud, calculated as follows:
[0092] l cyc =EMD(F(G(A)),A)+EMD(F(G(B)),B)
[0093] Here, F and G are two functions that share the same encoder parameters.
[0094] l TTransformation loss: The difference between the predicted transformation matrix and the ground truth, minimizing the geometric difference between the registered source and target point clouds, is expressed as follows:
[0095]
[0096] Among them, T est It is the predicted transformation matrix, T gt -1 The transformation loss is obtained by multiplying the two matrices and then subtracting the corresponding identity matrices, which are the ground truth values.
[0097] This invention presents a point cloud registration method based on a diffusion model, transforming the registration problem into a generation problem. The point cloud feature extraction module first normalizes the pose difference of the points to be registered, then uses a graph neural network to extract local features of the points, including a graph construction layer, a convolutional layer, and a ReLU activation function. Token-mixing MLP and Channel-mixing MLP are used on the obtained local features to ensure that the point cloud features maintain global and local structural information at multiple levels. This structure allows the network to effectively process point cloud data, taking into account the spatial distribution of the point cloud and the feature details of each point, which is beneficial for more accurate point cloud analysis and processing. The point cloud generation module uses the recently popular diffusion model, treating the points in the point cloud as particles in contact with a hot bath in a thermodynamic system, diffusing from the original distribution to a noise distribution. Therefore, point cloud generation is equivalent to learning a reverse diffusion process that transforms the noise distribution into the desired shape distribution. Specifically, the reverse diffusion process of the point cloud is modeled as a Markov chain based on certain shape latent variables. While our method can directly generate aligned point clouds and is applicable to 3D matching and retrieval applications, other applications, such as 3D localization and pose estimation, require knowledge of the transformation between the two input point clouds. In the aforementioned generation method, our proposed loss function preserves the correspondence between the generated and original point clouds. Therefore, these correspondences can be used to estimate the transformation matrix. By simulating the thermal diffusion process to analyze the geometric structure of the point cloud, the registration problem is transformed into a generation problem, which can resist the influence of noise and outliers to a certain extent, thus exhibiting good robustness, accelerating the point cloud registration speed, and improving the point cloud registration accuracy.
[0098] A point cloud registration system based on a diffusion model, comprising: a point cloud acquisition module, a point cloud feature extraction module, a feature interaction module, a point cloud generation module, and a transformation matrix prediction module;
[0099] The point cloud acquisition module is used to acquire the point cloud to be registered and the target point cloud;
[0100] The point cloud feature extraction module is used to extract the feature map of the point cloud information to be registered and the feature map of the target point cloud;
[0101] The feature interaction module is used to perform feature interaction fusion between the feature map of the point cloud information to be registered and the feature map of the target point cloud to obtain two fused feature maps.
[0102] The point cloud generation module generates two aligned point clouds based on the fused two feature maps.
[0103] The transformation matrix prediction module predicts the transformation matrix and translation vector based on the aligned point cloud, and registers the point cloud to be tested based on the transformation matrix and translation vector.
[0104] In this embodiment, the specific implementation of the system is the same as that of the method.
[0105] In one embodiment of the present invention, the present invention further includes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described point cloud registration methods based on diffusion models.
[0106] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented using computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.
[0107] A point cloud registration device based on a diffusion model includes a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory and is used to execute the computer program stored in the memory, so that the point cloud registration device based on a diffusion model performs any of the above-described point cloud registration methods based on a diffusion model.
[0108] Specifically, the memory includes various media capable of storing program code, such as ROM, RAM, magnetic disk, USB flash drive, memory card, or optical disk.
[0109] Preferably, the processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.
[0110] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A point cloud registration method based on a diffusion model, characterized in that, include: Point cloud registration model training phase and point cloud registration phase; The point cloud registration model includes a point cloud feature extraction module, a feature interaction module, a point cloud generation module, and a transformation matrix prediction module. The point cloud registration model training phase includes: acquiring training samples, which include source point cloud A and target point cloud B; using a point cloud feature extraction module to extract features from source point cloud A and target point cloud B respectively, obtaining features Feat1 and Feat2; using a feature interaction module to fuse features Feat1 and Feat2, obtaining fused features Fusion1 and Fusion2; using a point cloud generation module to generate two aligned point clouds Ag and Bg based on fused features Fusion1 and Fusion2; using a transformation matrix prediction module to calculate the correspondence between the original point cloud A and the generated point cloud Ag, and between the original point cloud B and the generated point cloud Bg, obtaining the predicted transformation matrix and translation vector; calculating the model's loss function, adjusting the model parameters, and completing the model training when the loss function converges. The point cloud registration stage includes: inputting the source point cloud and target point cloud of the actual object to be measured into the trained point cloud registration model to obtain the transformation matrix R and translation vector p; and using the transformation matrix R and translation vector p to register the source point cloud.
2. The point cloud registration method based on a diffusion model according to claim 1, characterized in that, The point cloud feature extraction module extracts point cloud features by: normalizing the pose difference of the point cloud to be registered; using a graph neural network to extract local features of the point cloud; and using a token-mixing MLP network and a channel-mixing MLP to process the local features to obtain global and local structural information.
3. The point cloud registration method based on a diffusion model according to claim 2, characterized in that, The method of extracting local features of point clouds using graph neural networks includes: using a T-Net network to extract global features from two input point clouds, and learning six free parameters from these global features; converting the six free parameters into a 4×4 rigid transformation matrix and transforming one of the input point clouds; obtaining the k-NN point set of the i-th point in the point cloud based on the transformed features, extracting features from each point and its neighbors, and obtaining the features of the k-NN neighborhood of the i-th point through copying, expanding, and concatenation; and performing two-dimensional convolution and ReLU activation on the features of the k-NN neighborhood of the i-th point to obtain the local features of each point.
4. The point cloud registration method based on a diffusion model according to claim 1, characterized in that, The feature interaction module is used to fuse features Feat1 and Feat2, which includes: performing max pooling on features Feat1 and Feat2 to obtain global features G1 and G2; calculating the difference between Feat1 and global feature G1, and calculating the difference between Feat2 and global feature G2; repeating the feature difference along each batch and feature dimension, and concatenating it with features Feat1 and Feat2 to obtain fused features Fusion1 and Fusion2.
5. The point cloud registration method based on a diffusion model according to claim 1, characterized in that, The formulas for obtaining the fused features Fusion1 and Fusion2 are as follows: Fusion1=Feat1+repeat(feat_diff_12,N) Fusion2=Feat2+repeat(feat_diff_21,N) Where feat_diff_12 is the difference between Feature1 and global feature G1, feat_diff_21 is the difference between Feature2 and global feature G2, and N is the number of points in the point cloud.
6. The point cloud registration method based on a diffusion model according to claim 1, characterized in that, The point cloud generation module uses a diffusion model to diffuse the point cloud from its original distribution to a noise distribution. In other words, the reverse diffusion process of the point cloud is modeled as a Markov chain based on latent variables.
7. The point cloud registration method based on a diffusion model according to claim 1, characterized in that, The transformation matrix prediction module calculates the transformation matrix by performing m random samplings on the correspondence between all points in the original point cloud and the generated point cloud, resulting in a pool M = (x1, x2, ..., x...). m ), where each element contains k corresponding elements; calculate the projection error of the transformation matrix based on the obtained pool, and select the one with the smallest projection error T. A Similarly, the minimum projection error T between the original point cloud B and the generated point cloud Bg is calculated. B According to the minimum projection error T A and T B Calculate the final transformation matrix.
8. A point cloud registration system based on a diffusion model, the system being used to execute the point cloud registration method based on a diffusion model as described in claim 1, characterized in that, The system includes: a point cloud acquisition module, a point cloud feature extraction module, a feature interaction module, a point cloud generation module, and a transformation matrix prediction module; The point cloud acquisition module is used to acquire the point cloud to be registered and the target point cloud; The point cloud feature extraction module is used to extract the feature map of the point cloud information to be registered and the feature map of the target point cloud; The feature interaction module is used to perform feature interaction fusion between the feature map of the point cloud information to be registered and the feature map of the target point cloud to obtain two fused feature maps. The point cloud generation module generates two aligned point clouds based on the fused two feature maps. The transformation matrix prediction module predicts the transformation matrix and translation vector based on the aligned point cloud, and registers the point cloud to be tested based on the transformation matrix and translation vector.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor to implement any one of the point cloud registration methods based on a diffusion model according to claims 1 to 7.
10. A point cloud registration device based on a diffusion model, characterized in that, It includes a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory and is used to execute the computer program stored in the memory, so that the point cloud registration device based on the diffusion model performs any one of the point cloud registration methods based on the diffusion model according to claims 1 to 7.