Remote sensing sample annotation method based on 3D point cloud

Abstract

The present invention discloses a remote sensing sample labeling method based on three-dimensional point clouds, which relates to the field of remote sensing image classification. The method includes: collecting three-dimensional point cloud data of a remote sensing sample set to obtain a first point cloud set; obtaining a point cloud library, the point cloud library forming a public point cloud space, and using a screening method to screen the point cloud library to obtain a second point cloud set containing label information; aligning the first point cloud set and the second point cloud set in the public point cloud space; constructing and training a segmentation model, and using the segmentation model to obtain label information of the first point cloud set; reprojecting the first point cloud set onto the remote sensing sample set to obtain label information of the remote sensing sample set. The present invention applies three-dimensional point cloud technology to remote sensing sample labeling, making full use of the information of the previous three-dimensional point cloud, and migrating it according to the feature invariance of the three-dimensional point cloud to obtain label information of the remote sensing sample set to be labeled. The method saves a lot of manpower and material resources, and the obtained labeling results are relatively accurate and have low noise.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing image classification, and in particular to a remote sensing sample labeling method based on three-dimensional point cloud. Background Art

[0002] Currently, remote sensing technology is widely used in geoscience applications such as forest resource planning, crop yield estimation, environmental assessment, and disaster monitoring. To better utilize remote sensing images, it is necessary to first convert them into various thematic information (such as forest cover maps of different forest types, thematic maps of disaster-stricken areas, etc.). Remote sensing image classification technology is an important technical means to convert remote sensing images from data into information. Based on whether training samples are required, remote sensing image classification methods are divided into supervised classification and unsupervised classification. The most commonly used method in remote sensing image classification is supervised classification. Supervised classification methods require manual determination of the classification system and selection of training samples for each category. Supervised classification methods can ensure relatively good classification accuracy. However, labeling remote sensing samples is a time-consuming and labor-intensive process.

[0003] Point cloud is a commonly used 3D data format that preserves the original geometric information of objects in three-dimensional space. With the rapid development of imaging technology in recent years, 3D information has attracted increasing attention. Currently, any image captured by a camera can be used to collect 3D point cloud data, including satellite remote sensing images, digital cameras, photos taken with mobile phones, or tablets. Existing 3D point cloud libraries are enormous, and the point clouds in these libraries contain rich prior knowledge and labeling information. However, these libraries are currently underutilized, resulting in a significant waste of point cloud information. Summary of the Invention

[0004] Based on the above-mentioned problems of information waste in the annotation of remote sensing samples and three-dimensional point clouds in the existing technology, the present invention provides a method for annotating remote sensing samples by combining three-dimensional point clouds with transfer learning. This method can make full use of the prior knowledge of the point cloud library. By utilizing three-dimensional point cloud data, remote sensing samples can be annotated more quickly and better annotation results can be obtained.

[0005] In order to solve the above technical problems, the present invention provides a remote sensing sample annotation method based on three-dimensional point cloud, which includes:

[0006] S1 obtains a set of remote sensing sample sets, collects three-dimensional point cloud data on the remote sensing sample sets, and obtains a first point cloud set;

[0007] S2 acquires a point cloud library, forms a public point cloud space through the point cloud library, and filters the point cloud library using a preset filtering method to obtain a second point cloud set, wherein the second point cloud set includes label information;

[0008] S3 aligning the first point cloud set and the second point cloud set in the common point cloud space;

[0009] S4: constructing and training a segmentation model, and using the segmentation model to segment the first point cloud set to obtain label information of the first point cloud set;

[0010] S5 reprojects the first point cloud set onto the remote sensing sample set, and obtains label information of the remote sensing sample set based on the correspondence between the point cloud in the first point cloud set and the pixel points in the remote sensing sample set.

[0011] In a specific embodiment of the present invention, the preset screening method is a perceptual hash algorithm.

[0012] In a specific embodiment of the present invention, the preset screening method is a generated query network screening method.

[0013] In one embodiment of the present invention, step S2 includes:

[0014] S21 acquires a point cloud library, and forms a public point cloud space through the point cloud library;

[0015] S22 builds a generative query network, including a representation network and a generation network;

[0016] S23 inputs the first point cloud set and the point cloud library into a generative query network for training and representation learning, thereby obtaining a scene representation of the first point cloud set and a scene representation of the point cloud library, wherein during training, the point cloud library is completed according to the generative network and the representation network;

[0017] S24 compares the similarity between the scene representation of the point cloud library and the scene representation of the first point cloud set, and classifies the point clouds in the point cloud library that reach a similarity threshold into the second point cloud set, wherein the second point cloud set includes label information.

[0018] In one embodiment of the present invention, step S3 includes:

[0019] In the common point cloud space, coarsely registering the first point cloud set and the second point cloud set using a first registration method;

[0020] The first point cloud set and the second point cloud set after the rough registration are finely registered according to the second registration method to achieve alignment of the first point cloud set and the second point cloud set.

[0021] In one embodiment of the present invention, step S4 includes:

[0022] In the public point cloud space, the constructed segmentation model is trained using the second point cloud set and the label information of the second point cloud set as training data to obtain a trained segmentation model;

[0023] The first point cloud set is input into the trained segmentation model for segmentation to obtain label information of the first point cloud set.

[0024] In one embodiment of the present invention, step S4 includes:

[0025] Mapping the first point cloud set and the second point cloud set to a two-dimensional space;

[0026] In the two-dimensional space, the constructed segmentation model is trained using the second point cloud set and the label information of the second point cloud set as training data to obtain a trained segmentation model;

[0027] The first point cloud set is segmented using the trained segmentation model, and the segmented first point cloud set is mapped back to the public point cloud space to obtain label information of the first point cloud set.

[0028] The beneficial effects of the present invention are as follows: the present invention provides a remote sensing sample labeling method based on three-dimensional point cloud, firstly, three-dimensional point cloud data of the remote sensing sample set is collected to obtain a first point cloud set, then a point cloud library is obtained, the point cloud library constitutes a public point cloud space, the point cloud library is screened by a screening method to obtain a second point cloud set, wherein the second point cloud set contains label information, in the public point cloud space, the first point cloud set and the second point cloud set are aligned, and then a segmentation model is constructed and trained, the segmentation model is used to segment the first point cloud set to obtain label information of the first point cloud set, the first point cloud set is reprojected onto the remote sensing sample set, and the label information of the remote sensing sample set is obtained according to the correspondence between the point cloud in the first point cloud set and the pixel points in the remote sensing sample set. The present invention applies 3D point cloud technology to remote sensing sample labeling, making full use of the information of the previous 3D point cloud, and aligning the point cloud set, that is, migrating it according to the feature invariance of the 3D point cloud. Based on the one-to-one correspondence between the pixels of the 3D point cloud and the remote sensing image, the label information of the remote sensing sample set to be labeled can be obtained by converting the label content contained in the 3D point cloud through the correspondence. This method saves a lot of manpower and material resources, and the obtained labeling results are relatively accurate and have low noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0030] Figure 1 This is a flowchart of a remote sensing sample labeling method based on three-dimensional point cloud according to an embodiment of the present invention;

[0031] Figure 2 Schematic diagram of the PointNet network architecture according to an embodiment of the present invention. DETAILED DESCRIPTION

[0032] It should be understood that the following will be combined with the accompanying drawings to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, and not all of the embodiments. It should be noted that as long as no conflict arises, the various embodiments and various features of each embodiment of the present invention can be combined with each other, and the resulting technical solutions are all within the scope of protection of the present invention.

[0033] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0034] See also Figure 1 , Figure 1 The present invention provides a method for labeling remote sensing samples based on three-dimensional point clouds, which includes:

[0035] S1 obtains a set of remote sensing sample sets, collects three-dimensional point cloud data for the remote sensing sample sets, and obtains a first point cloud set.

[0036] S2 acquires a point cloud library, forms a public point cloud space through the point cloud library, and filters the point cloud library using a preset filtering method to obtain a second point cloud set, wherein the second point cloud set includes label information.

[0037] S3 aligns the first point cloud set and the second point cloud set in the common point cloud space.

[0038] S4 constructs and trains a segmentation model, and uses the segmentation model to segment the first point cloud set to obtain label information of the first point cloud set.

[0039] S5 reprojects the first point cloud set onto the remote sensing sample set, and obtains label information of the remote sensing sample set based on the correspondence between the point cloud in the first point cloud set and the pixel points in the remote sensing sample set.

[0040] The technical ideas of the present invention are as follows: 1) using 3D modeling technology to collect 3D point cloud data of the obtained remote sensing sample set, the remote sensing sample set contains multiple remote sensing sample images belonging to the same scene, and the 3D point cloud obtained after the collection is a first point cloud set; 2) directly obtaining a point cloud library from a database, or collecting a large number of geographically tagged images, performing 3D modeling on the geographically tagged images, obtaining multiple 3D point cloud images, and then forming a point cloud library, the point cloud library forming a public point cloud space, and then screening the point cloud library based on the scene to obtain a 3D point cloud that matches the remote sensing sample set as a second point cloud set, since the images corresponding to the point cloud library have been marked, the second point cloud set contains label information; 3) mapping the first point cloud set to the public point cloud space, and aligning the first point cloud set and the second point cloud set;

[0041] 4) In the public point cloud space, the second point cloud set is migrated, the prior knowledge of the second point cloud set is migrated to the first point cloud set, the first point cloud set is segmented (i.e., classified), and the label information of the first point cloud set is obtained; 5) Since there is a specific relationship between each 3D point cloud in the first point cloud set and the pixel points in the remote sensing sample set, the first point cloud set containing the label information is reprojected from the public point cloud space to the remote sensing sample set. The pixel labels can be obtained through the labels of the 3D point cloud, and then the label information of the remote sensing sample set can be obtained.

[0042] In a specific implementation, step S1 can be implemented by obtaining multiple remote sensing sample images to be labeled within a scene, dividing the remote sensing sample images into a certain number of groups, and obtaining multiple remote sensing sample sets. These multiple remote sensing sample sets are then used as a group of remote sensing sample sets. Then, three-dimensional point cloud data is collected from the remote sensing sample sets to obtain a first point cloud set. In this embodiment, each remote sensing sample set includes five remote sensing sample images.

[0043] The method of collecting the three-dimensional point cloud in the above step S1 may include the following steps:

[0044] (1) Generate seed points through a facet-based multi-view stereo method

[0045] Input the relevant parameters of the remote sensing sample set, such as image window parameters, pixel parameters, and surface parameters. Based on the internal and external orientation elements of the image in the remote sensing sample set, the image projection matrix is ​​calculated to generate the initial seed point cloud. The seed point cloud includes the 3D coordinates (X, Y, Z) of the seed point, the normal vector (a, b, c) of the surface element, and the color (R, G, B) of the seed point.

[0046] (2) Point cloud diffusion based on object aspect

[0047] The point cloud is diffused according to the number of pixels of the diffusion bin. If the diffusion bin includes m×m pixels, the number of point clouds can be diffused at most by m×m times.

[0048] First, the seed cloud point is projected onto each remote sensing sample image in the remote sensing sample set to obtain the corresponding projection point coordinates (x i ,y i )(i=1,2,...,n), where n is the number of remote sensing sample images in the remote sensing sample set. In this embodiment, n=5. Then, an image with the smallest distance between the projection point and the center point of the remote sensing sample image is selected from the remote sensing sample set as a reference image. A window with m×m pixels is set on the reference image. Each pixel in the window is projected onto an object element to obtain m×m object points, which are added to the diffused point element set to obtain a first cloud point set.

[0049] Of course, in this embodiment, the acquisition method for the 3D point cloud can also be a multi-view dense matching method, such as a dense reconstruction method based on PatchMatch. The specific process is as follows:

[0050] (1) Calculate the depth map of remote sensing image samples

[0051] The PatchMatch algorithm is used to calculate the depth map. First, sparse 3D points are obtained by SFM. The depth and normal values ​​corresponding to other pixels on the remote sensing image sample are initialized through meshing and interpolation. Then, the depth and normal values ​​between pixels are propagated and randomly searched. The depth and normal values ​​of the tangent plane associated with the pixels are further optimized through multiple iterations to obtain the depth information corresponding to each remote sensing image sample.

[0052] (2) Depth map fusion

[0053] After completing the above process, a series of depth maps for the remote sensing image samples are obtained, and all of these depth maps are then combined to represent a single scene. A point cloud generation method based on visibility-based multi-view depth map fusion is used to perform a weighted average of the depths of pairs of identical points in depth maps from different angles. The fused depth value is then calculated, and pixel depths that deviate by more than a certain threshold are considered outliers. This depth information yields a 3D point cloud for the remote sensing image samples, the first cloud point set.

[0054] In the above step S2, the public point cloud space is composed of a point cloud library, which may include many modeled three-dimensional point clouds. The point cloud library can be directly screened from the point cloud library according to the scene requirements of the remote sensing sample set to obtain a second point cloud set, which contains label information.

[0055] It should be noted that the preset screening method in the above step S2 can be a perceptual hash method. Perceptual hashing is derived from cryptographic hashing technology and can compress original information into fixed-bit numbers, thereby greatly reducing the amount of data.

[0056] The retrieval process based on perceptual hashing mainly includes: preprocessing, hash feature extraction, feature matching, and query result return. Preprocessing involves data reduction of the point cloud library, eliminating redundancy, and performing transformations to facilitate retrieval; hash feature extraction refers to converting the preprocessed point cloud library into a fixed-length hash code; hash feature matching refers to performing similarity comparisons on the obtained hash codes. Since most hash functions produce binary bit strings, hash code comparison methods in cryptographic hashing, such as Hamming distance, can be used; query result return refers to returning similar point clouds to the queryer. The first point cloud set must first be converted into a hash feature code as a hash feature index library. By comparing the hash feature code of the query point cloud library with the hash feature index library, the system returns point clouds within the perceptual threshold range to obtain the second point cloud set.

[0057] In addition, the point cloud library can also be composed of 3D point clouds generated from acquired geotagged images. Specifically, multiple geotagged images from multiple perspectives are first acquired, and 3D point cloud data is collected from the geotagged images to obtain a 3D point cloud set, which is used as the point cloud library. Considering that in practical applications, the point cloud library may have missing points and need to be supplemented, a query network can be used to supplement and filter the point cloud library to obtain a second point cloud set. The process is as follows:

[0058] S21 acquires a point cloud library, and forms a public point cloud space through the point cloud library.

[0059] S22 builds a generative query network, including a representation network and a generation network.

[0060] S23 inputs the first point cloud set and the point cloud library into a generation query network for training and representation learning to obtain a scene representation of the first point cloud set and a scene representation of the point cloud library, wherein during training, the point cloud library is completed according to the generation network and the representation network.

[0061] The representation network observes the first point cloud set and the point cloud library, observing and training the spatial position, color, and normal vectors of the point clouds, and generates a representation vector describing the underlying scene. During training, the generative network learns the characteristics, relationships, and patterns of the point clouds. Once the representation network generates the representation vector, it can predict point clouds from missing perspectives and complete the 3D point cloud set.

[0062] S24 compares the similarity between the scene representation of the point cloud library and the scene representation of the first point cloud set, and classifies the point clouds in the point cloud library that reach a similarity threshold into the second point cloud set, wherein the second point cloud set includes label information.

[0063] Specifically, the similarity between scene representations can be calculated using the cosine similarity of vectors.

[0064] In this embodiment, the similarity threshold may be set to 0.5.

[0065] In the above steps, the point cloud library is obtained by geo-tagged images, which can make full use of a large number of geo-tagged images. These images contain very rich labels and content information, and are complemented according to the generative query network, which can improve the utilization value of geo-tagged images. At the same time, the generative query network can also learn the scene representation of the point cloud. According to the similarity between the scene representation of the point cloud library and the scene representation of the first point cloud set, the point cloud with higher similarity to the scene of the remote sensing image set is selected as the second point cloud set. When the scene is more relevant, the label of the remote sensing image set obtained by subsequent migration will be more accurate and more valuable.

[0066] In this embodiment, the above step S3 may specifically include: aligning the first point cloud set and the second point cloud set in the common point cloud space, including coarse registration and fine registration.

[0067] The first registration method may be the LORAX method, and the specific process is as follows:

[0068] A collection of small part of point cloud is defined as super point SP, and SP is used as the basic unit in the whole registration process.

[0069] (1) First, randomly select a point P that does not belong to any SP, use P as the center, and use R sphere The points inside the circle are the points contained in the new SP, R sphere The calculation is as follows:

[0070]

[0071] Among them, 0.64 means that in the random spherical blocks, non-repeated spheres will account for 64%; V loval is the volume of the sphere containing the point cloud to be processed; m refers to the number of point cloud features used for matching.

[0072] (2) A normalized coordinate system is established for each SP. The origin of the coordinate system is the center of mass of the SP. The directions of the three axes are obtained by performing SVD decomposition on the covariance matrix of the SP. The z-axis is set as the third eigenvector. The SP is divided into discrete arc slices and statistically analyzed as an epipolar histogram, with the x-axis as the direction with the highest value.

[0073] (3) After converting each SP into a coordinate system and comparing them, the x and y axes of all points are projected onto an image grid and mapped into a depth map.

[0074] (4) Construct a deep neural network, including four fully connected hidden layers, input SP into the deep neural network, and obtain a SAF matrix.

[0075] (5) By detecting the Euclidean distance between the SAF matrices, each SP from the first point cloud set is paired with the SP from the second point cloud set to perform rough registration.

[0076] The second registration method may be an ICP algorithm, and the process is as follows:

[0077] Find the nearest corresponding point between the two point clouds, calculate an RT matrix from the corresponding points that minimizes the distance between the corresponding points after transformation according to the rotation and translation matrix, and perform the transformation. Repeat the above steps until the number of times the above steps are repeated reaches the preset number, then stop the iteration to accurately align the first point cloud set with the second point cloud set.

[0078] In this embodiment, step S4 may specifically include: after aligning the first point cloud set and the second point cloud set, constructing and training a segmentation model, and using the segmentation model to segment the first point cloud set to obtain label information of the first point cloud set.

[0079] The present invention can segment the first point cloud set in two ways. The first is to construct and train a segmentation model in the common point cloud space and directly segment the first point cloud set using the segmentation model. The second is to first map the first point cloud set and the second point cloud set to a two-dimensional space. Then, in the two-dimensional space, the second point cloud set and its label information are used as training data to train a segmentation model. The segmentation model is then used to segment the first point cloud set. The segmented first point cloud set is then mapped back to the common point cloud space to obtain the label information of the first point cloud set.

[0080] For the first segmentation method, the segmentation model can be a PointNet network segmentation model.

[0081] The process of training the PointNet network segmentation model using the second point cloud set and its label information as training data is as follows:

[0082] PointNet is the first deep neural network to directly process unordered point cloud data. The PointNet network segmentation model has three core modules, including a transformation network, a maximum pooling layer that is a symmetric function that aggregates all voxel information, and a multilayer perceptron network. Point clouds have three core properties: (1) point clouds are unordered, that is, a network that consumes N 3D point cloud sets needs to maintain the original order of the N arranged input point cloud data; (2) holistic expressiveness, that is, point clouds are not isolated, and each point cloud and its neighbors form a meaningful point cloud subset; (3) invariance to transformations, that is, deep learning represents point cloud sets that should be invariant to certain transformations. Therefore, it is necessary to design a symmetric function that is independent of the order of the scanned points in the voxels.

[0083] The PointNet network is represented as:

[0084]

[0085] in, is the input unordered point cloud in the jth voxel; 1024 is the number of input points per voxel; f is a continuous set function that maps a set of points to a vector; γ represents a multilayer perceptron network, and h is a combination of a univariate function and a max pooling function. The value of f remains unchanged regardless of the input order of the point clouds.

[0086] like Figure 2 As shown, Figure 2 The PointNet architecture consists of two transformation matrix prediction networks (T-Nets), three multi-layer perceptrons (MLPs), and a max pooling layer. The network takes 1024 points in a voxel as input, applies input and feature transformations, and aggregates point features through a max pooling layer. The output is the predicted probability of the point cloud dataset's classification.

[0087] The first point cloud set is segmented using the trained PointNet network segmentation model. Based on the information learned from the network architecture, the first point cloud set is classified into categories to obtain its label information. As can be seen from the above steps of generating a point cloud, there is a specific intrinsic relationship between point clouds and pixels. Based on the label information of the first point cloud set, the first point cloud set is reprojected onto the remote sensing sample set to obtain the label information of the remote sensing sample set, thus completing the annotation of the remote sensing sample set.

[0088] For the second segmentation method, the segmentation model can be an image semantic segmentation network model, such as FCN, U-Net, or other segmentation networks. The first and second point cloud sets are first projected into a two-dimensional space. Within this two-dimensional space, the image semantic segmentation network model is trained using the second point cloud set and its label information. The first point cloud set is then segmented using the trained image semantic segmentation network model to obtain its label information.

[0089] The beneficial effects of the present invention are as follows: the present invention provides a remote sensing sample labeling method based on three-dimensional point cloud, firstly, three-dimensional point cloud data of the remote sensing sample set is collected to obtain a first point cloud set, then a point cloud library is obtained, the point cloud library constitutes a public point cloud space, the point cloud library is screened by a screening method to obtain a second point cloud set, wherein the second point cloud set contains label information, in the public point cloud space, the first point cloud set and the second point cloud set are aligned, and then a segmentation model is constructed and trained, the segmentation model is used to segment the first point cloud set to obtain label information of the first point cloud set, the first point cloud set is reprojected onto the remote sensing sample set, and the label information of the remote sensing sample set is obtained according to the correspondence between the point cloud in the first point cloud set and the pixel points in the remote sensing sample set. The present invention applies 3D point cloud technology to remote sensing sample labeling, making full use of the information of the previous 3D point cloud, and aligning the point cloud set, that is, migrating it according to the feature invariance of the 3D point cloud. Based on the one-to-one correspondence between the pixels of the 3D point cloud and the remote sensing image, the label information of the remote sensing sample set to be labeled can be obtained by converting the label content contained in the 3D point cloud through the correspondence. This method saves a lot of manpower and material resources, and the obtained labeling results are relatively accurate and have low noise.

[0090] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A remote sensing sample annotation method based on three-dimensional point cloud, characterized in that: The method includes: S1 obtains a set of remote sensing sample sets, collects three-dimensional point cloud data for the remote sensing sample sets, and obtains an initial first point cloud set; S2 obtains a point cloud library, forms a common point cloud space through the point cloud library, and filters the point cloud library using a preset filtering method to obtain a second point cloud set, wherein the second point cloud set includes label information; S3 aligns the initial first point cloud set and the second point cloud set in the common point cloud space to obtain an aligned first point cloud set; S4 constructs and trains a segmentation model, and uses the segmentation model to segment the aligned first point cloud set to obtain label information of the aligned first point cloud set; S5 reprojects the aligned first point cloud set onto the remote sensing sample set, and obtains label information of the remote sensing sample set based on a correspondence between point clouds in the aligned first point cloud set and pixel points in the remote sensing sample set; Wherein, step S2 includes: S21 acquires a point cloud library, and forms a public point cloud space through the point cloud library; S22 builds a generative query network, including a representation network and a generation network; S23 inputs the initial first point cloud set and the point cloud library into a generative query network for training and representation learning, thereby obtaining a scene representation of the initial first point cloud set and a scene representation of the point cloud library, wherein during training, the point cloud library is completed according to the generative network and the representation network; S24 compares the similarity between the scene representation of the point cloud library and the scene representation of the initial first point cloud set, and classifies the point clouds in the point cloud library that meet the similarity threshold into the second point cloud set, wherein the second point cloud set includes label information; Step S3 includes: In the common point cloud space, coarsely registering the initial first point cloud set and the second point cloud set using a first registration method; The first point cloud set and the second point cloud set after the rough registration are finely registered according to the second registration method to achieve alignment between the initial first point cloud set and the second point cloud set, thereby obtaining an aligned first point cloud.

2. The remote sensing sample annotation method based on three-dimensional point cloud according to claim 1, characterized in that: The preset screening method is the perceptual hash algorithm.

3. The remote sensing sample annotation method based on three-dimensional point cloud according to claim 1, characterized in that: The preset screening method is a generation query network screening method.

4. The remote sensing sample annotation method based on three-dimensional point cloud according to claim 3, characterized in that: Step S4 includes: In the public point cloud space, the constructed segmentation model is trained using the second point cloud set and the label information of the second point cloud set as training data to obtain a trained segmentation model; The aligned first point cloud set is input into the trained segmentation model for segmentation to obtain label information of the first point cloud set.

5. The remote sensing sample annotation method based on three-dimensional point cloud according to claim 3, characterized in that: Step S4 includes: Mapping the aligned first point cloud set and the second point cloud set to a two-dimensional space; In the two-dimensional space, the constructed segmentation model is trained using the second point cloud set and the label information of the second point cloud set as training data to obtain a trained segmentation model; The trained segmentation model is used to segment the aligned first point cloud set, and the segmented aligned first point cloud set is mapped back to the common point cloud space to obtain a label of the aligned first point cloud set.

Images