A self-supervised registration method for multi-source remote sensing point clouds
By extracting global features from multi-source remote sensing point clouds using a self-supervised mask autoencoder network, the problem of large errors in multi-source point cloud registration is solved, achieving high-precision registration under unsupervised conditions, especially the effective fusion of lidar point clouds collected by UAVs and point clouds reconstructed from images.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2024-11-11
- Publication Date
- 2026-06-26
Smart Images
Figure CN119515937B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of remote sensing point cloud registration, specifically relating to a self-supervised registration method for multi-source remote sensing point clouds. Background Technology
[0002] With the continuous development of 3D detection technology, there are now various methods for acquiring remote sensing point clouds. Mainstream technologies include actively acquiring point clouds using lidar or generating point clouds using photogrammetry. Different methods of acquiring point clouds have their own advantages and limitations. To integrate and utilize the advantages of multiple source point clouds, necessary registration processing is first required. Existing point cloud registration methods are mostly designed for point clouds acquired from the same sensor, without considering the differences in point cloud characteristics caused by different acquisition principles. Furthermore, deep learning-based methods require pre-training with supervisory information for registration ground truth, which is difficult to obtain in actual multi-source remote sensing point clouds. Therefore, existing technologies rely on supervisory information, leading to large registration errors when supervisory information is lacking. Summary of the Invention
[0003] The purpose of this invention is to address the problem of large registration errors in existing technologies that rely on supervisory information, leading to a lack of such information. A self-supervised registration method for multi-source remote sensing point clouds is proposed, comprising:
[0004] Step 1: Acquire multi-source remote sensing point clouds;
[0005] Step 2: Construct a masked autoencoder point cloud reconstruction network. Train the constructed masked autoencoder point cloud reconstruction network based on multi-source remote sensing point clouds to obtain a trained masked autoencoder point cloud reconstruction network.
[0006] Step 3: Obtain the multi-source remote sensing point cloud and template point cloud to be registered; reconstruct the network based on the multi-source remote sensing point cloud to be registered, the template point cloud and the masked autoencoder point cloud trained in Step 2 to obtain the transformation relationship of the multi-source remote sensing point cloud to be registered.
[0007] Step 4: Transform the multi-source remote sensing point cloud to be registered according to the transformation relationship of the multi-source remote sensing point cloud to be registered, and obtain the registered multi-source remote sensing point cloud.
[0008] The beneficial effects of this invention are as follows:
[0009] 1. This invention addresses the characteristic differences between 3D reconstructed point clouds and lidar point clouds by designing a self-supervised registration framework for multi-source point clouds based on mask autoencoder features. The framework uses a mask autoencoder and self-supervised learning method to robustly extract global features from the distribution differences of multi-source point clouds and uses feature alignment to register the multi-source point clouds, thereby improving the registration accuracy of multi-source point clouds.
[0010] 2. This invention enables the registration of multi-source remote sensing point clouds with significant characteristic differences even without known supervisory information, thus improving the registration accuracy of multi-source point clouds. To verify the performance of the proposed algorithm, experiments were conducted on lidar point clouds acquired by UAVs and image-based 3D reconstructed point clouds. The experimental results validated the effectiveness of the proposed method. Attached Figure Description
[0011] Figure 1 This is a schematic diagram of the multi-source point cloud self-supervised registration method of the present invention;
[0012] Figure 2 This is a schematic diagram of the registration experiment of the multi-source point cloud self-supervised registration method of the present invention. Detailed Implementation
[0013] Specific implementation method one: Combining Figure 1 This invention describes a self-supervised registration method for multi-source remote sensing point clouds, comprising:
[0014] Step 1: Acquire multi-source remote sensing point clouds;
[0015] Step 2: Construct a masked autoencoder point cloud reconstruction network. Train the constructed masked autoencoder point cloud reconstruction network based on multi-source remote sensing point clouds to obtain a trained masked autoencoder point cloud reconstruction network.
[0016] Step 3: Obtain the multi-source remote sensing point cloud and template point cloud to be registered; reconstruct the network based on the multi-source remote sensing point cloud to be registered, the template point cloud and the masked autoencoder point cloud trained in Step 2 to obtain the transformation relationship of the multi-source remote sensing point cloud to be registered.
[0017] Step 4: Transform the multi-source remote sensing point cloud to be registered according to the transformation relationship of the multi-source remote sensing point cloud to be registered, and obtain the registered multi-source remote sensing point cloud.
[0018] This invention addresses the characteristic differences between 3D reconstructed point clouds and lidar point clouds by designing a self-supervised registration framework for multi-source point clouds based on mask autoencoder features. The framework uses a mask autoencoder and self-supervised learning method to robustly extract global features from the distribution differences of multi-source point clouds, and uses feature alignment to register the multi-source point clouds, thereby improving the registration accuracy of multi-source point clouds.
[0019] This invention enables the registration of multi-source remote sensing point clouds with significant characteristic differences even without known supervisory information, thus improving the registration accuracy of multi-source point clouds. To verify the performance of the proposed algorithm, experiments were conducted on lidar point clouds acquired by UAVs and image-based 3D reconstructed point clouds. The experimental results validate the effectiveness of the proposed method.
[0020] Specific Implementation Method Two: The difference between this implementation method and Specific Implementation Method One is that...
[0021] The multi-source remote sensing point cloud in step 1 includes: lidar point cloud and 3D reconstructed point cloud;
[0022] The lidar used in this invention is the VELODYNE VLP16 LITE, and the camera used for image acquisition is the Micasenserededge-MX DUAL. The acquisition location is located in the Harbin Institute of Technology Science Park.
[0023] The lidar and camera used in this invention can be selected according to the actual usage.
[0024] The other steps and parameters are the same as in Specific Implementation Method 1.
[0025] Specific Implementation Method Three: The difference between this implementation method and Specific Implementation Method One is that...
[0026] The mask autoencoder point cloud reconstruction network in step 2 includes: a random mask module, an encoder network, and a decoder network;
[0027] The process involves training a masked autoencoder point cloud reconstruction network based on multi-source remote sensing point clouds to obtain a trained masked autoencoder point cloud reconstruction network. The specific steps are as follows:
[0028] Step 2.1: Use the multi-source remote sensing point cloud as the input point cloud P, input it into the random masking module, and obtain the masked point cloud P';
[0029] Step 2.2: Input the masked point cloud P' into the encoder network to obtain the global feature F';
[0030] Step 2.3: Input the global feature F' into the decoder network to obtain the reconstructed point cloud PC;
[0031] Step 2.4: Calculate the loss function between the reconstructed point cloud PC and the input point cloud P. Train the masked autoencoder point cloud reconstruction network based on the loss function. Training is complete when the loss function is minimized, resulting in the trained masked autoencoder point cloud reconstruction network.
[0032] This step employs a self-supervised training method, rather than relying on known transformation relationships for pre-training to extract local features from the point cloud.
[0033] The other steps and parameters are the same as in one of the specific implementation methods one or two.
[0034] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One through Four in that...
[0035] In step 2.1, the multi-source remote sensing point cloud is used as the input point cloud P, which is then input into the random masking module to obtain the masked point cloud P'. The specific process is as follows:
[0036] Step 2.1.1: Set the number of sampling points to N, where N is a positive integer, and randomly select one point from the input point cloud P as the initial sampling point P1;
[0037] Step 2.1.2: Calculate the distance between the unselected points in the input point cloud P and the initial sampling point P1, and select the point with the largest distance as the sampling point P2;
[0038] Step 2.1.3: Calculate the minimum distance from the unselected points in the input point cloud P to the existing set of sampled points, and select the point with the largest minimum distance as the sampled point;
[0039] Step 2.1.4: Repeat step 2.1.3 until N sampling points are obtained, then stop execution and proceed to step 2.1.5; where the nth sampling point is denoted as Pn;
[0040] Step 2.1.5: Combine the N sampling points to form the farthest point sampling point cloud PS; determine the mask point cloud PK based on the sampling point cloud PS;
[0041] Step 2.1.6: Based on the masked point cloud PK obtained in Step 2.1.5, perform masking processing on the input point cloud P to obtain the masked point cloud P'. The specific process is as follows:
[0042] The above sampling points in the input point cloud P are masked and marked as invisible points (the remaining points in P are marked as visible points, i.e., P').
[0043] The other steps and parameters are the same as those in one of the specific implementation methods one to three.
[0044] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that the specific process of calculating the minimum distance from the unselected points in the input point cloud P to the existing set of sampled points in step 2.1.3 is as follows:
[0045] Calculate the distance from the unselected point to the nearest sampled point in the existing sampled point set as the minimum distance from the unselected point to the existing sampled point set in the input point cloud P;
[0046] Assuming the existing set of sampling points includes: initial sampling point P1 and sampling point P2; then the distance from the o-th unselected point to the initial sampling point P1 is o1, and the distance from the o-th point to the sampling point P2 is o2. If o1 is greater than o2, then the minimum distance from the o-th point to the set of sampling points is o2.
[0047] When calculating the minimum distance from unselected points in the input point cloud P to each point in the existing sampled point set, the minimum value of the previously calculated sampling distance is retained as the distance from that point to the sampled points to avoid duplicate calculations.
[0048] The specific process of determining the mask point cloud PK based on the sampled point cloud PS in step 2.1.5 is as follows:
[0049] In the input point cloud P, find the K nearest neighbors of each point in the sample point cloud PS to form the mask point cloud PK;
[0050] For example, if there are N points in PS, find the K nearest neighbors for each point in the input point cloud P, for a total of NK points. However, duplicates are allowed when finding the K nearest neighbors for different points, so the final total number of points is less than or equal to NK. The other steps and parameters are the same as in one of the specific implementation methods one to four.
[0051] Specific Implementation Method Six: The difference between this implementation method and Specific Implementation Methods One to Five is that...
[0052] The encoder network structure in step 2.2 includes, in sequence, an MLP layer and a max pooling layer;
[0053] The MLP layer sequentially includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, a fourth hidden layer, and an output layer; the first hidden layer, the second hidden layer, the third hidden layer, and the fourth hidden layer are all convolutional layers;
[0054] The kernel size of the first hidden layer is 1×3; the kernel size of the second, third, and fourth hidden layers is 1×1.
[0055] In step 2.2, the masked point cloud P' is input into the encoder network to obtain the global feature F'; the specific process is as follows:
[0056] The masked point cloud P' serves as the input to the encoder network, with a dimension of N'×3. The encoder network consists of four MLP layers and one max-pooling layer. The MLP convolutional kernels have 64, 64, 128, and 1024 channels, respectively. After passing through the four MLP layers, the masked point cloud P' outputs N'×1024-dimensional intermediate features. The max-pooling layer maximizes the intermediate features extracted by the fourth MLP layer, ultimately yielding 1024 global features F'.
[0057] Other steps and parameters are the same as in any of the specific implementation methods one to five.
[0058] Specific Implementation Method Seven: The difference between this implementation method and Specific Implementation Methods One through Six is that...
[0059] The decoder network structure in step 2.3 includes, in sequence, a first fully connected layer, a second fully connected layer, a third fully connected layer, and a fourth fully connected layer;
[0060] The activation functions for the first, second, third, and fourth fully connected layers are all LeakyReLU;
[0061] In step 2.3, the global feature F' is input into the decoder network to obtain the reconstructed point cloud PC; the specific process is as follows:
[0062] The global feature F' is used as the input to the decoder, which then passes through four fully connected layers. The final layer predicts an output with a dimension of N×3. This output serves as the reconstructed point cloud PC.
[0063] The other steps and parameters are the same as those in any of the specific implementation methods one to six.
[0064] Specific Implementation Method Eight: The difference between this implementation method and Specific Implementation Methods One to Seven is that...
[0065] The loss function calculation formula for the reconstructed point cloud PC and the input point cloud P in step 2.4 is expressed as follows:
[0066]
[0067] Where x represents a point in the input point cloud P, y represents a point in the reconstructed point cloud PC; Pnum and PCnum represent the number of points in the input point cloud P and the number of points in the reconstructed point cloud PC. express.
[0068] Based on loss function minimization, a masked autoencoder point cloud reconstruction network is trained in a self-supervised manner using the backpropagation algorithm to optimize network weights. The network weights that minimize the loss function are used as the trained masked autoencoder point cloud reconstruction network. Benefiting from the masked reconstruction training method, the trained network has the ability to extract robust features from the overall point cloud and can improve the registration ability of multi-source point clouds with local missing features.
[0069] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.
[0070] Specific Implementation Method Nine: The difference between this implementation method and Specific Implementation Methods One through Eight is that...
[0071] In step 3, the multi-source remote sensing point cloud to be registered and the template point cloud are obtained; based on the multi-source remote sensing point cloud to be registered, the template point cloud, and the masked autoencoder point cloud trained in step 2, the network is reconstructed to obtain the transformation relationship of the multi-source remote sensing point cloud to be registered; the specific process is as follows:
[0072] Step 3.1: Obtain the multi-source remote sensing point cloud S to be registered and the template point cloud T;
[0073] The multi-source point cloud refers to point cloud data obtained in the same scene using two or more methods. One of the point cloud data obtained by one method can be designated as the template point cloud T, and the other as the point cloud to be registered S.
[0074] Step 3.2: Set the mask rate of the random mask module in the trained mask autoencoder point cloud reconstruction network to 0;
[0075] Step 3.3: Input the multi-source remote sensing point cloud S to be registered into the random mask module and encoder network in the trained mask autoencoder point cloud reconstruction network in sequence to obtain the intermediate feature G(S);
[0076] The template point cloud T is sequentially input into the random mask module and encoder network in the trained mask autoencoder point cloud reconstruction network to obtain the intermediate feature G(T);
[0077] Step 3.4: Construct a registration matrix based on intermediate features G(S) and G(T); the registration matrix is expressed by the formula:
[0078] G(S)=G(γ -1 ·T) (2)
[0079] γ represents the transformation relationship of the multi-source remote sensing point cloud to be registered;
[0080] Step 3.5: Linearize the registration matrix to obtain the linearized registration matrix, which is expressed as follows:
[0081]
[0082] φ represents the transformation increment; ξ represents the transformation increment.
[0083] Step 3.6: Calculate the Jacobian matrix of the registration matrix after linearization, expressed by the formula:
[0084]
[0085] J denotes the analytic derivative of the Jacobian matrix;
[0086] The analytic derivative of a Jacobian matrix refers to the derivative of each element in the Jacobian matrix, and is usually expressed as a partial derivative.
[0087] Step 3.7: Apply the difference gradient approximation to each column of the Jacobian matrix to obtain the approximate Jacobian matrix. The i-th column J of the approximate Jacobian matrix is... i Expressed as a formula:
[0088]
[0089] t i X is represented as an infinitesimally changing variable of γ, therefore, in practical applications, ti is set to a very small fixed value to obtain better results. i The result of the exponential mapping of the transformation matrix; exp() represents the exponential mapping; X i The result is an exponential mapping of the transformation matrix, which is obtained by solving the Jacobian matrix, a process known to those skilled in the art.
[0090] Step 3.8: Calculate the transformation increment ξ based on the approximate Jacobian matrix J′, which can be expressed by the formula:
[0091] ξ=J + [G(S)-G(T)] (6)
[0092] J + This represents the plus sign generalized inverse matrix of the approximate Jacobian matrix J′;
[0093] The process of solving for the plus generalized inverse matrix of the approximate Jacobian matrix J′ is a computational process known to those skilled in the art.
[0094] Step 3.9: Set the threshold and maximum number of iterations; when the transformation increment ξ obtained in step 3.8 is less than the limit threshold, or reaches the maximum number of iterations, multiply the transformation increments obtained in each round of iteration to obtain the multiplied transformation increment, and then perform exponential mapping on the multiplied transformation increment to obtain the transformation relationship γ;
[0095] When the transformation increment ξ obtained in step 3.8 is not less than the limited threshold and has not reached the maximum number of iterations, the multi-source remote sensing point cloud S to be registered is transformed into point cloud S' according to the transformation increment ξ; expressed by the formula:
[0096] S'=ξ·S (7)
[0097] The point cloud S' is input into the encoder network in the trained mask autoencoder point cloud reconstruction network to obtain the intermediate feature G(S'), and then the process is returned to S3.4 for iteration.
[0098] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.
[0099] This step optimizes the solution based on the gradient direction of the feature difference between the point cloud to be registered and the template point cloud, combining optimization algorithms with deep learning algorithms. This improves the accuracy of multi-source point cloud registration.
[0100] Specific Implementation Method Ten: The difference between this implementation method and Specific Implementation Methods One through Nine is that...
[0101] In step 4, the multi-source remote sensing point cloud to be registered is transformed according to the transformation relationship of the multi-source remote sensing point cloud to be registered, so as to obtain the registered multi-source remote sensing point cloud. The specific process is as follows:
[0102] Based on the transformation relationship γ of the multi-source remote sensing point cloud to be registered, a rigid transformation is performed on the point cloud S to be registered to obtain the registered point cloud S. out This can be expressed as a formula:
[0103] S out =γ·S (8)
[0104] The rigid transformation refers to a combination of rotation and translation transformations, a definition well-known to those skilled in the art.
[0105] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.
[0106] Based on the description, a registration experiment was conducted on a LiDAR point cloud dataset collected by an UAV and an image-based 3D reconstructed point cloud dataset using the multi-source point cloud registration method designed in this invention. The registration results are as follows: Figure 2 As shown:
[0107] The lidar used was a VELODYNE VLP16 LITE, and the image acquisition camera was a Micasenserededge-MXDUAL. The acquisition location was located in the Harbin Institute of Technology Science Park. The root mean square error of the nearest neighbor point was used as the evaluation index for registration. The registration accuracy of this method reached 0.35m, which is higher than that of existing methods.
[0108] The above description is merely of preferred embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention, and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.
Claims
1. A self-supervised registration method for multi-source remote sensing point clouds, characterized in that, Includes the following steps: Step 1: Acquire multi-source remote sensing point clouds; including: lidar point clouds and 3D reconstructed point clouds; Step 2: Construct a masked autoencoder point cloud reconstruction network, including: a random mask module, an encoder network, and a decoder network; the encoder network structure includes, in sequence: an MLP layer and a max pooling layer; The MLP layer sequentially includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, a fourth hidden layer, and an output layer; the first hidden layer, the second hidden layer, the third hidden layer, and the fourth hidden layer are all convolutional layers; The kernel size of the first hidden layer is 1×3; the kernel size of the second, third and fourth hidden layers is 1×1. The decoder network structure includes, in sequence, a first fully connected layer, a second fully connected layer, a third fully connected layer, and a fourth fully connected layer; The activation functions for the first, second, third, and fourth fully connected layers are all LeakyReLU; The masked autoencoder point cloud reconstruction network is trained based on multi-source remote sensing point clouds to obtain a trained masked autoencoder point cloud reconstruction network; the specific process is as follows: Step 2.1: Use the multi-source remote sensing point cloud as the input point cloud P, input it into the random masking module, and obtain the masked point cloud P'; Step 2.2: Input the masked point cloud P' into the encoder network to obtain the global feature F'; Step 2.3: Input the global feature F' into the decoder network to obtain the reconstructed point cloud PC; Step 2.4: Calculate the loss function between the reconstructed point cloud PC and the input point cloud P. Train the masked autoencoder point cloud reconstruction network according to the loss function. The training is completed when the loss function is minimized, and the trained masked autoencoder point cloud reconstruction network is obtained. Step 3: Obtain the multi-source remote sensing point cloud and template point cloud to be registered; reconstruct the network based on the multi-source remote sensing point cloud to be registered, the template point cloud, and the masked autoencoder point cloud trained in Step 2 to obtain the transformation relationship of the multi-source remote sensing point cloud to be registered; the specific process is as follows: Step 3.1: Obtain the multi-source remote sensing point cloud S to be registered and the template point cloud T; Step 3.2: Set the mask rate of the random mask module in the trained mask autoencoder point cloud reconstruction network to 0; Step 3.3: Input the multi-source remote sensing point cloud S to be registered into the random mask module and encoder network in the trained mask autoencoder point cloud reconstruction network in sequence to obtain the intermediate feature G(S); Then, the template point cloud T is sequentially input into the random mask module and encoder network in the trained mask autoencoder point cloud reconstruction network to obtain the intermediate feature G(T); Step 3.4: Construct the registration matrix based on the intermediate features G(S) and G(T); The registration matrix is expressed by the formula: (2) in, Transformation relationship of multi-source remote sensing point clouds to be registered; Step 3.5: Linearize the registration matrix to obtain the linearized registration matrix. The linearized registration matrix is expressed by the following formula: (3) in Indicates the change increment; Step 3.6: Calculate the Jacobian matrix of the registration matrix after linearization. , The Jacobian matrix of the linearized registration matrix is expressed by the formula: (4) Step 3.7: Apply the difference gradient approximation to each column of the Jacobian matrix to obtain the approximate Jacobian matrix. , The approximate Jacobian matrix The i-th column is represented as , The formula is expressed as: (5) Represented as The changing variables, For intermediate variables; exp() represents exponential mapping; Step 3.8: Based on the approximate Jacobian matrix Calculations are performed to obtain the transformation increment. For, expressed by the formula: (6) in Represents the approximate Jacobian matrix The plus sign generalized inverse matrix; Step 3.9: Set the threshold and maximum number of iterations; when the transformation increment obtained in step 3.8... When the value is less than the threshold or the maximum number of iterations is reached, the transformation increments obtained from each iteration are multiplied to obtain the multiplied transformation increment. Then, the multiplied transformation increment is subjected to exponential mapping to obtain the transformation relationship of the multi-source remote sensing point cloud to be registered. ; When the transformation increment obtained in step 3.8 When the number of iterations is not less than the specified threshold and has not reached the maximum number of iterations, According to the transformation increment Transform the multi-source remote sensing point cloud S to be registered into point cloud S'; expressed by the formula: (7) The point cloud S' is input into the encoder network in the trained mask autoencoder point cloud reconstruction network to obtain the intermediate feature G(S'), and then the process is returned to S3.4 for iteration. Step 4: Transform the multi-source remote sensing point cloud to be registered according to the transformation relationship of the multi-source remote sensing point cloud to be registered, and obtain the registered multi-source remote sensing point cloud.
2. The self-supervised registration method for multi-source remote sensing point clouds according to claim 1, characterized in that, In step 2.1, the multi-source remote sensing point cloud is used as the input point cloud P, which is then input into the random masking module to obtain the masked point cloud P'. The specific process is as follows: Step 2.1.1: Set the number of sampling points to N, where N is a positive integer, and randomly select one point from the input point cloud P as the initial sampling point P1; Step 2.1.2: Calculate the distance between the unselected points in the input point cloud P and the initial sampling point P1, and select the point with the largest distance as the sampling point P2; Step 2.1.3: Calculate the minimum distance from the unselected points in the input point cloud P to the existing set of sampled points, and select the point with the largest minimum distance as the sampled point; Step 2.1.4: Repeat step 2.1.3 until N sampling points are obtained, then stop execution and proceed to step 2.1.5; where the nth sampling point is denoted as Pn; Step 2.1.5: Combine the N sampling points to form the farthest sampling point cloud PS; Determine the mask point cloud PK based on the sampled point cloud PS; Step 2.1.6: Perform masking processing on the input point cloud P based on the masked point cloud PK obtained in step 2.1.5 to obtain the masked point cloud P'.
3. The self-supervised registration method for multi-source remote sensing point clouds according to claim 2, characterized in that, The specific process of determining the mask point cloud PK based on the sampled point cloud PS in step 2.1.5 is as follows: In the input point cloud P, find the K nearest neighbors of each point in the sample point cloud PS to be used as the mask point cloud PK.
4. The self-supervised registration method for multi-source remote sensing point clouds according to claim 3, characterized in that, The loss function calculation formula for the reconstructed point cloud PC and the input point cloud P in step 2.4 is expressed as follows: (1) Where x represents a point in the input point cloud P, and y represents a point in the reconstructed point cloud PC; This indicates the number of points in the input point cloud P. Indicates the number of points in the reconstructed point cloud PC; This represents the square of the L2 norm.
5. The self-supervised registration method for multi-source remote sensing point clouds according to claim 4, characterized in that, In step 3, the multi-source remote sensing point cloud to be registered and the template point cloud are obtained; the network is reconstructed based on the multi-source remote sensing point cloud to be registered, the template point cloud and the masked autoencoder point cloud trained in step 2 to obtain the transformation relationship of the multi-source remote sensing point cloud to be registered.
6. The self-supervised registration method for multi-source remote sensing point clouds according to claim 5, characterized in that, In step 4, the multi-source remote sensing point cloud to be registered is transformed according to the transformation relationship of the multi-source remote sensing point cloud to be registered, so as to obtain the registered multi-source remote sensing point cloud. The specific process is as follows: Based on the transformation relationship of the multi-source remote sensing point clouds to be registered A rigid transformation is performed on the point cloud S to be registered to obtain the registered point cloud. This can be expressed as a formula: （8）。