Foundation radar asteroid three-dimensional reconstruction method based on convolutional neural network

Abstract

The invention belongs to the technical field of asteroid shape reconstruction, and particularly relates to a ground-based radar asteroid three-dimensional reconstruction method based on a convolutional neural network, and the method specifically comprises the steps: constructing an asteroid shape model data set: generating asteroid three-dimensional point cloud models of different shapes and different angles, and processing the surfaces of the three-dimensional point cloud models; constructing and training an asteroid three-dimensional reconstruction neural network: setting the asteroid three-dimensional reconstruction neural network to comprise an asteroid image feature extraction network and a multi-stage inversion three-dimensional point cloud network, wherein the asteroid image feature extraction network is an improved VGG16 network; and ground-based radar signal asteroid three-dimensional reconstruction: inputting an asteroid two-dimensional distance Doppler image acquired by a ground-based radar into the trained asteroid three-dimensional reconstruction neural network, and processing an asteroid three-dimensional point cloud model output by the neural network to obtain a final asteroid three-dimensional point cloud model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of asteroid shape reconstruction, and particularly relates to a ground-based radar asteroid three-dimensional reconstruction method based on a convolutional neural network. BACKGROUND

[0002] Asteroids are ancient celestial bodies of the solar system, and are materials left over from the planetary formation process of the solar system, which are crucial to the study of the origin and evolution of the solar system. Asteroids have various shapes, such as gyroscope type, elongated type, double petal type and the like, which reveal a large number of asteroid dynamic evolution mechanisms, and are of great significance to asteroid scientific research. The asteroids can also provide accurate information support for asteroid impact on Earth warning, space resource development and utilization, and have important social and economic value. Ground-based radar is not limited by weather conditions and time, and can realize all-weather and all-time observation.

[0003] The traditional ground-based radar asteroid three-dimensional reconstruction method is based on the parameter inversion method of the range-Doppler image using the SHAPE software. First, a shape model based on spherical harmonic coefficients is established, and an initial ellipsoid model is established according to the size estimated by the time delay Doppler image. Then, an optimization function is established, the original delay Doppler image is compared with the simulation model image, and continuous improvement is carried out. Finally, a suitable optimization algorithm is selected to solve the optimization problem, so as to generate a three-dimensional asteroid model.

[0004] Since most near-Earth asteroids have complex shapes, the traditional reconstruction method is time-consuming and easy to fall into a local optimal solution, so that the three-dimensional shape of the inversion is distorted. The signal-to-noise ratio required for obtaining the range-Doppler image is relatively high, and it is difficult to process such a super-long distance target as an asteroid. SUMMARY

[0005] Therefore, the application provides a ground-based radar asteroid three-dimensional reconstruction method based on deep learning, which can realize three-dimensional reconstruction of a super-long distance asteroid.

[0006] The technical scheme of the application is as follows:

[0007] In the first aspect, the application provides a ground-based radar asteroid three-dimensional reconstruction method based on deep learning, and the specific process is as follows:

[0008] Asteroid shape model data set construction: generate three-dimensional point cloud models of asteroids with different shapes and angles, and uniformly process the surface of the three-dimensional point cloud models; mark unobservable points in the three-dimensional point cloud models, and perform two-dimensional imaging simulation on the three-dimensional point cloud models, wherein the unobservable points do not participate in the simulation, and generate two-dimensional range-Doppler images; the three-dimensional point cloud model is taken as a label, and the Doppler image is used to construct a data set together;

[0009] The asteroid three-dimensional reconstruction neural network construction and training includes an asteroid image feature extraction network and a multi-stage inversion three-dimensional point cloud network, wherein the asteroid image feature extraction network is an improved VGG16 network, and the improvement manner is to reduce the convolution layer and the pooling layer, add a Dropout layer and delete a classifier; the asteroid three-dimensional reconstruction neural network is trained by using the data set;

[0010] The ground-based radar signal asteroid three-dimensional reconstruction is performed by inputting the asteroid two-dimensional range Doppler image collected by the ground-based radar into the trained asteroid three-dimensional reconstruction neural network, and processing the asteroid three-dimensional point cloud model output by the neural network to obtain the final asteroid three-dimensional point cloud model.

[0011] Optionally, in the ground-based radar asteroid three-dimensional reconstruction process, the specific process of processing the asteroid three-dimensional point cloud model is as follows:

[0012] Firstly, the minimum distance of each point in the point cloud model to other points in the point cloud is calculated, and then several points with the maximum minimum distance are excluded;

[0013] Secondly, the voxel size of upsampling is set, the coordinate range of the point cloud is obtained, and then a uniform grid is generated in the coordinate range of the point cloud for interpolation;

[0014] Finally, the interpolated grid is converted into a point cloud format to obtain the final three-dimensional point cloud model.

[0015] Optionally, the process of obtaining the unobservable point is as follows: firstly, the N neighbor points of a point are calculated by using the KdTree algorithm, the decentralized covariance matrix is calculated by using the neighbor points, the eigenvalue and eigenvector are obtained by performing eigenvalue decomposition on the covariance matrix, and the normalized eigenvector is the normal vector of the point; then, the included angle between each point vector and the radar line-of-sight observation vector is calculated, and the point is an observable point when the included angle is less than 90 degrees, and is an unobservable point when the included angle is greater than or equal to 90 degrees.

[0016] Optionally, the process of generating the two-dimensional range Doppler image is as follows: the asteroid three-dimensional model is imported into matlab for radar two-dimensional imaging simulation, and then the radar echo receiving signal is processed by distance compression and translation correction to obtain two-dimensional range Doppler images at different angles.

[0017] Optionally, the present application further includes data set expansion, specifically: for two-dimensional range Doppler images at different angles, data expansion is performed by horizontal flipping and vertical flipping.

[0018] Optionally, the present application adds Gaussian noise and uniform noise to the generated two-dimensional range Doppler image to simulate the real environment.

[0019] Optionally, the asteroid three-dimensional point cloud model is up-sampled, then down-sampled to obtain a point cloud model of 1000 points, and the point cloud is taken as a label point cloud.

[0020] Optionally, the convolution layer and the pooling layer of the asteroid image feature extraction network are 10 layers, and the network output is 2048 points; the multi-level inversion three-dimensional point cloud network comprises three layers, the first layer reduces the number of points of the point cloud from 2048 to 1365 points, the third layer reduces the number of points of the point cloud from 1365 to 1000 points, and then comparison is made with the label to achieve the purpose of back propagation; wherein the second layer of the multi-level inversion three-dimensional point cloud network selects part of the points to be 0 during training.

[0021] Optionally, the loss function during network training comprises a mean square error loss L MSE , a chamfer distance loss L CD and a point cloud difference loss L PCD .

[0022] L=λ1L MSE +λ2L CD +λ3L PCD

[0023] Wherein, λ1, λ2, λ3 are the set weights.

[0024] In the second aspect, the application provides a ground-based radar asteroid three-dimensional reconstruction device based on deep learning, comprising:

[0025] The asteroid three-dimensional reconstruction neural network comprises an asteroid image feature extraction network and a multi-level inversion three-dimensional point cloud network, and is trained by using the above process, wherein the asteroid image feature extraction network is an improved VGG16 network, and the improvement manner is to reduce the convolution layer and the pooling layer, add a Dropout layer and delete a classifier; the asteroid image feature extraction network is used for feature point extraction on an input ground-based radar two-dimensional range Doppler image; and the multi-level inversion three-dimensional point cloud network is used for inverting the extracted feature points into an asteroid three-dimensional point cloud model.

[0026] The ground-based radar asteroid three-dimensional reconstruction module is used for processing the asteroid three-dimensional point cloud model to obtain a final asteroid three-dimensional point cloud model.

[0027] Advantages

[0028] The application uses deep learning to propose an asteroid three-dimensional reconstruction method.

[0029] First, a three-dimensional point cloud model of asteroids of different shapes is generated for the problem of various shapes of near-earth asteroids, and the surface of the three-dimensional point cloud model is uniformly processed to avoid the problem of unclear profile of imaging simulation results caused by uneven distribution of points of the three-dimensional point cloud model of asteroids, and considering the occlusion relationship in the real environment when radar two-dimensional imaging simulation of asteroids is considered, a part of points need to be selected not to participate in the radar two-dimensional imaging simulation, the occlusion algorithm based on the normal vector is used to screen the unobservable points, then two-dimensional imaging simulation of different angles is carried out for each typical asteroid, and multi-angle range-doppler images of asteroids of different shapes are obtained, which provides an effective data set for the three-dimensional reconstruction method of asteroids based on deep learning.

[0030] Second, the three-dimensional reconstruction method of asteroids based on convolutional neural network is proposed to solve the problem that the traditional reconstruction method is time-consuming and easy to fall into local optimal solution, and the three-dimensional shape of the inversion is distorted. In the method, the theoretical knowledge of deep learning is combined, and considering that the original two-dimensional convolutional layer can only extract the features of one picture and cannot consider the picture features of multiple angles of an asteroid, the number of channels corresponding to the angle is used when the neural network is used to extract the multi-angle range-doppler image features of asteroids of different shapes. The traditional VGG16 network is used for classification problem, but the problem to be solved by the present application is regression problem, so the full connection neural network is used to regress the specific point cloud coordinate parameters while the feature extraction part of the VGG16 network is applied, and the chamfer distance is selected instead of the cross-entropy loss function of the original network. At the same time, the number of layers of the feature extraction part of the VGG16 network is too large, which can easily lead to model overfitting, so the Dropout layer is added on the basis of the VGG16 network to reduce the number of convolutional layers. After the network output, the reconstructed feature point cloud is obtained, and because the spatial resolution of the feature point cloud is too low, the feature point cloud is up-sampled to obtain the three-dimensional reconstruction result. Through comparison and analysis with the traditional method, the method proposed by the present application improves the problem of the traditional method and has faster reconstruction speed. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creating laborious work.

[0032] Figure 1 The three-dimensional model of the asteroid in the form of point cloud;

[0033] Figure 2 The comparison diagram of the uniformized asteroid point cloud model, (a) is the original point cloud, and (b) is the point cloud after uniformization;

[0034] Figure 3 Figure (a) is the original VGG16 network, and figure (b) is the asteroid image feature extraction network improved by the application;

[0035] Figure 4 Figure is a network structure diagram of the application;

[0036] Figure 5 Figure is a point cloud model diagram without up-sampling processing;

[0037] Figure 6 Figure is a point cloud model diagram with up-sampling processing;

[0038] Figure 7 Figure is a three-dimensional reconstruction model of asteroid 1958TL1;

[0039] Figure 8 Figure is a period measurement error at different shapes and initial phases. DETAILED DESCRIPTION

[0040] The embodiments of the application will be described in detail below with reference to the accompanying drawings.

[0041] It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict; and all other embodiments obtained by those skilled in the art based on the embodiments in the present disclosure without creative labor are within the scope of protection of the present disclosure.

[0042] It should be noted that various aspects of the embodiments described below are within the scope of the appended claims. As will be apparent, the aspects described herein can be implemented in various ways, and that any particular structure is used is for illustration only. One of ordinary skill in the art, having the benefit of the present disclosure, will appreciate that one aspect described herein can be implemented independent of any other aspect, and that two or more aspects can be combined in any way. For example, an apparatus can be implemented and / or a method can be practiced using any number of the aspects set forth herein. In addition, such an apparatus can be implemented and / or such a method can be practiced using other structure and / or functionality in addition to or other than one or more of the aspects set forth herein.

[0043] The application embodiment is a ground-based radar asteroid three-dimensional reconstruction method based on a convolutional neural network, which specifically includes three parts of an asteroid shape model data set construction, an asteroid three-dimensional reconstruction neural network construction and training, and a ground-based radar signal asteroid three-dimensional reconstruction.

[0044] The asteroid shape model dataset is constructed: a three-dimensional point cloud model of an asteroid with different shapes and different angles is generated, and the surface of the three-dimensional point cloud model is uniformly processed; the unobservable points are marked in the three-dimensional point cloud model, the three-dimensional point cloud model is imaged in two dimensions, the unobservable points do not participate in simulation, and a two-dimensional range Doppler image is generated; the three-dimensional point cloud model is taken as a label, and the Doppler image is used to construct a dataset together;

[0045] The asteroid three-dimensional reconstruction neural network is constructed and trained: the asteroid three-dimensional reconstruction neural network is set, including an asteroid image feature extraction network and a multi-level inversion three-dimensional point cloud network, wherein the asteroid image feature extraction network is an improved VGG16 network, and the improvement manner is: reducing the convolution layer and the pooling layer, adding the Dropout layer and deleting the classifier; the data set is used to train the asteroid three-dimensional reconstruction neural network;

[0046] The ground-based radar signal asteroid three-dimensional reconstruction: the two-dimensional range Doppler image of the asteroid collected by the ground-based radar is input into the trained asteroid three-dimensional reconstruction neural network, and the three-dimensional point cloud model of the asteroid output by the neural network is processed to obtain the final three-dimensional point cloud model of the asteroid.

[0047] The above three processes are described in detail as follows:

[0048] Step one, constructing an asteroid shape model dataset

[0049] The present application improves the problems caused by the traditional SHAPE method by using the method of deep learning, and first needs to construct a data set for neural network training.

[0050] (1) Generation of three-dimensional point cloud models of asteroids with different shapes: in view of the problem of the variety of shapes of near-Earth asteroids, the main shape characteristics of asteroids are analyzed, and three-dimensional point cloud models of asteroids of irregular type (such as 1958TL1), ellipsoid type (such as Vesta), and gyroscope type (such as Bennu) are generated.

[0051] Generation of three-dimensional point cloud models of asteroids with different angles: in order to improve the north-south ambiguity problem, the data set of the neural network should be multiple distance Doppler images of asteroids at different angles. In order to obtain distance Doppler images at different angles, three-dimensional point cloud models of asteroids at different angles need to be generated. Different angle three-dimensional point cloud models of asteroids are generated by multiplying the three-dimensional point cloud model of the asteroid by the rotation matrix around the z-axis (formula (1)).

[0052]

[0053] Wherein, θ is the rotation angle.

[0054] (2) Point cloud model surface homogenization: The uneven surface of the asteroid point cloud model can cause the profile of the radar two-dimensional imaging simulation result to be unclear. In order to improve this problem, the surface of the asteroid three-dimensional point cloud model needs to be homogenized. The present application proposes an interpolation algorithm based on screening Poisson surface reconstruction to make the point distribution of the asteroid three-dimensional point cloud model uniform. Figure 2 is a comparison diagram of the original point cloud model and the homogenized point cloud model.

[0055] (3) Single asteroid data matrix generation: Considering the occlusion relationship of objects in the real physical environment when radar imaging, a part of the points need to be selected not to participate in the radar two-dimensional imaging simulation. The present application uses an occlusion algorithm based on the normal vector. The algorithm first uses the KdTree algorithm to calculate 10 neighbor points (including the point, a total of 11 points) of a certain point, uses the neighbor points to calculate the decentralized covariance matrix, performs eigenvalue decomposition on the covariance matrix to obtain the eigenvector and eigenvalue, and the normalized eigenvector is the normal vector of the point. Then the angle between each point vector and the radar line-of-sight observation vector is calculated. The angle less than 90 degrees is an observable point, and the angle greater than or equal to 90 degrees is an unobservable point.

[0056] Then the matlab is used to perform multi-angle radar two-dimensional echo imaging simulation on the real three-dimensional point cloud model of the asteroid. The specific configuration parameters are shown in Table 1. The xyz point cloud format of the asteroid three-dimensional model similar to Figure 1 is imported into matlab for radar two-dimensional imaging simulation, and then the radar echo receiving signal is processed by distance compression, translation correction, etc. Finally, the two-dimensional range-Doppler images at different angles are obtained.

[0057] Table 1 Simulation experiment parameters

[0058]

[0059]

[0060] (4) Data set expansion: The present application uses the flip operation in the geometric transformation method to expand the data by horizontally flipping and vertically flipping the original range-Doppler image. This operation does not change the characteristics of the data, but only changes the viewing angle, expands the data set, and makes the model more easily learn the features in the image. Compared with other operations, the flip method is relatively simple and efficient. At the same time, considering that the ground-based radar has different echo signal-to-noise ratios for asteroid detection, in order to ensure the quality and reliability of the data set, the present application selects to add different degrees of Gaussian noise and uniform noise to simulate the real environment, and expands the data set on the basis of the original image in this way. Finally, 80% of the data set is selected as the training set, and 20% is selected as the test set.

[0061] Meanwhile, in order to prevent the uneven asteroid real point cloud model from causing adverse effects on neural network training, the application first performs upsampling on the asteroid three-dimensional point cloud model, and then performs downsampling to obtain a point cloud model of 1000 points, and the point cloud is taken as a label point cloud.

[0062] Step two: construction of the asteroid three-dimensional reconstruction neural network

[0063] The asteroid three-dimensional reconstruction neural network established by the application is based on Python, and the deep learning framework used is Pytorch.

[0064] Table 2: experimental environment configuration

[0065]

[0066] (1) Asteroid image feature extraction network: the application improves the VGG16 network to extract the range-Doppler image features of the asteroid. The network is used to obtain the point target position of the asteroid, and a comparison diagram of the network and the VGG16 network is as shown in Figure 3 Unlike the normal VGG16 network, the network used in the embodiment only retains the original convolutional layer and the pooling layer and is reduced to 10 layers, and a Dropout layer is added to prevent network overfitting, and the original network classifier is deleted.

[0067] The final effect of the asteroid three-dimensional reconstruction is related to the size of the extracted features. Therefore, as many points as possible should be retained when extracting the features. The actual three-dimensional point cloud model has only 574 points, but the output of the network is set to 2048 points.

[0068] (2) Multi-level inversion three-dimensional point cloud network: the multi-level inversion three-dimensional point cloud network of the asteroid is based on a fully connected neural network, and the features obtained by the asteroid image feature extraction network are inverted, and the weight of each point is updated step by step, and finally the number of points (1000 points) equal to the label point cloud is obtained.

[0069] Figure 4 The network structure diagram of the application is shown. The input of the multi-level inversion three-dimensional point cloud network is 2048 points obtained by the asteroid image feature extraction network. The first layer of processing changes the number of point clouds from 2048 points to 1365 points; the second layer randomly selects a part of the points to be 0 in the training process to prevent model overfitting; the third layer changes the number of point clouds from 1365 points to 1000 points, and then compares with the label to achieve the purpose of back propagation.

[0070] (3) Loss function selection: In the process of restoring the accurate three-dimensional model of asteroids round by round, compared with traditional three-dimensional reconstruction methods and convolutional neural networks, the invention uses mean square error loss, chamfer distance loss and point cloud difference loss to improve the overall reconstruction effect.

[0071] The purpose of the mean square error loss (formula (2)) is to measure the difference between the model predicted value and the true value, and adjust the model parameters as an optimization target, so that the predicted point cloud coordinate value is as close to the true value as possible.

[0072]

[0073] Chamfer distance (formula (3)) is an index for measuring the similarity between two point clouds, and the smaller the chamfer distance, the higher the similarity between the two point clouds.

[0074]

[0075] Point cloud difference loss (formula (4)) is to measure the predicted and true point clouds, thereby enhancing the network's ability to learn fine details.

[0076]

[0077] The total loss function (formula (5)) is a weighted combination of the following items, and the weights are selected to balance their respective contributions.

[0078] L = λ1L MSE + λ2L CD + λ3L PCD (5)

[0079] Step three, ground-based radar asteroid three-dimensional reconstruction: the point cloud model obtained through the above steps is unevenly distributed, has low spatial resolution, and cannot determine its reconstruction accuracy. Upsampling can restore spatial resolution, increase point cloud density, improve reconstruction accuracy and preserve detailed information. Therefore, the point cloud model obtained in step two will be upsampled. Figure 5

[0080] First, calculate the minimum distance of each point in the point cloud to other points in the point cloud, and then exclude the points with the largest minimum distance; set the voxel size of upsampling, obtain the coordinate range of the point cloud, then generate a uniform grid within the point cloud range, and perform interpolation; finally, convert the interpolated grid to point cloud format to obtain the upsampled point cloud model similar to Figure 6 .

[0081] ​The embodiment analyzes main shape features of asteroids, performs two-dimensional imaging simulation at different angles for each typical asteroid, obtains multi-angle range-Doppler images of asteroids with different shapes, and provides an effective data set for a three-dimensional reconstruction method of asteroids based on a convolutional neural network; a multi-angle range-Doppler image feature of an asteroid with different shapes is extracted by using a neural network, a chamfer distance is selected as one of loss functions, a three-dimensional point cloud network is output after three-dimensional reconstruction through multi-level inversion, and a three-dimensional reconstruction result is obtained by upsampling the feature point cloud.

[0082] Simulation experiment

[0083] To verify the feasibility of the method, a simulation experiment of three-dimensional reconstruction of asteroids is performed.

[0084] The experiment sets take the asteroid 1958TL1 as an example, reconstruct a three-dimensional model of the asteroid, and obtain a reconstruction accuracy of 93.06%. Figure 7

[0085] The multi-angle range-Doppler images of the reconstructed model are compared with the real range-Doppler images, and the similarities of the images are analyzed. Figure 8 It can be concluded that the range-Doppler images at different angles are similar, which indicates that the method proposed in the application can better reconstruct the three-dimensional model of the asteroid.

[0086] In view of the problem of various shapes of asteroids, in order to verify the effectiveness of the algorithm proposed in the application, representative asteroids with different shapes are selected, three-dimensional reconstruction is performed on them, the reconstruction accuracy of the asteroids with different shapes is analyzed, and finally a comparison table of reconstruction accuracy of asteroids with different shapes in table 3 is obtained.

[0087] Table 3 Comparison table of reconstruction accuracy of asteroids with different shapes

[0088]

[0089] In view of the problem of long time consumption of the traditional method, the method of deep learning is used for improvement. In order to verify the effect, the method is compared with the time required by the traditional method to obtain table 4.

[0090] Table 4 Comparison of three-dimensional reconstruction time of asteroids by different methods

[0091]

[0092] ​The experimental results are analyzed and summarized, and the method for three-dimensional reconstruction of asteroids based on deep learning of the ground-based radar proposed in the application is evaluated from the aspects of reconstruction accuracy and reconstruction range, and it is concluded that the method proposed in the application has the advantages of high reconstruction accuracy, fast reconstruction speed and low requirement for equipment. Meanwhile, the analysis of Table 3 can conclude that the three-dimensional reconstruction effect of the method proposed in the application on some special-shaped asteroids is poor. In order to improve this problem, more distance Doppler images of asteroids of other shapes can be added to the data set.

[0093] To sum up, the above is only a preferred embodiment of the application, and is not used to limit the protection scope of the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. A method for 3D reconstruction of asteroids using ground-based radar based on deep learning, characterized in that: The specific process is: Asteroid shape model dataset construction: Generate 3D point cloud models of asteroids of different shapes and angles, and homogenize the surface of the 3D point cloud models; mark unobservable points in the 3D point cloud models, perform 2D imaging simulation on the 3D point cloud models, where unobservable points are not included in the simulation, and generate 2D range Doppler images; use the 3D point cloud models as labels and construct a dataset together with the Doppler images; Construction and training of asteroid 3D reconstruction neural network: Setting up an asteroid 3D reconstruction neural network including an asteroid image feature extraction network and a multi-stage inversion 3D point cloud network, wherein the asteroid image feature extraction network is an improved VGG16 network improved by reducing convolutional layers and pooling layers, adding a dropout layer, and deleting the classifier; using the dataset to train the asteroid 3D reconstruction neural network; 3D reconstruction of asteroids using ground-based radar signals: The 2D range Doppler images of asteroids collected by ground-based radar are input into the trained 3D reconstruction neural network for asteroids, and the 3D point cloud model output by the neural network is processed to obtain the final 3D point cloud model of asteroids.

2. The method for ground-based radar asteroid 3D reconstruction based on deep learning according to claim 1, characterized in that: During the ground-based radar asteroid 3D reconstruction process, the specific process of processing the asteroid 3D point cloud model is as follows: First, calculate the minimum distance between each point in the point cloud model and other points in the point cloud, and then exclude the points with the largest minimum distance; Secondly, set the upsampled voxel size, obtain the coordinate range of the point cloud, and then generate a uniform grid within the coordinate range of the point cloud for interpolation; Finally, the interpolated mesh is converted into point cloud format to obtain the final three-dimensional point cloud model.

3. The method for ground-based radar asteroid 3D reconstruction based on deep learning according to claim 2, characterized in that: The process of obtaining unobservable points is as follows: first, the KdTree algorithm is used to calculate the N neighboring points of a certain point, the neighboring points are used to calculate the decentralized covariance matrix, the covariance matrix is ​​subjected to eigenvalue decomposition to obtain the eigenvectors and eigenvalues, and the normalized eigenvector is the normal vector of the point. Then, the angle between each point vector and the radar line of sight observation vector is calculated. If the angle is less than 90 degrees, it is an observable point, and if it is greater than or equal to 90 degrees, it is an unobservable point.

4. The method for ground-based radar asteroid 3D reconstruction based on deep learning according to claim 2, characterized in that: The two-dimensional range Doppler image generation process is as follows: the three-dimensional asteroid model is imported into MATLAB to perform radar two-dimensional imaging simulation, and then the radar echo reception signal is range compressed and corrected to obtain two-dimensional range Doppler images at different angles.

5. The method for ground-based radar asteroid 3D reconstruction based on deep learning according to claim 2, characterized in that: It also includes data set expansion, specifically: for two-dimensional range Doppler images at different angles, data expansion is performed by horizontal flipping and vertical flipping.

6. The method for ground-based radar asteroid 3D reconstruction based on deep learning according to claim 2, characterized in that: Gaussian noise and uniform noise are added to the generated two-dimensional range Doppler image to simulate the real environment.

7. The method for ground-based radar asteroid 3D reconstruction based on deep learning according to claim 2, characterized in that: The asteroid 3D point cloud model is upsampled and then downsampled to obtain a point cloud model of 1000 points, which is used as the label point cloud.

8. The method for ground-based radar asteroid 3D reconstruction based on deep learning according to claim 7, characterized in that: The asteroid image feature extraction network has 10 convolutional and pooling layers, and the network output is 2048 points. The multi-level inversion 3D point cloud network contains three layers. The first layer reduces the number of point clouds from 2048 to 1365 points, and the third layer reduces the number of point clouds from 1365 to 1000 points, which are then compared with the labels to achieve the purpose of backpropagation. When training the multi-level inversion 3D point cloud network, some points in the second layer are set to 0.

9. The method for ground-based radar asteroid 3D reconstruction based on deep learning according to claim 8, characterized in that: The loss function during network training includes the mean square error loss L MSE , chamfer distance loss L CD And point cloud difference loss L PCD ; L=λ1L MSE +λ2L CD +λ3L PCD Among them, λ1, λ2, and λ3 are the set weights.

10. A ground-based radar asteroid 3D reconstruction device based on deep learning, characterized in that: include: An asteroid three-dimensional reconstruction neural network includes an asteroid image feature extraction network and a multi-level inversion three-dimensional point cloud network, which are trained using the above process, wherein the asteroid image feature extraction network is an improved VGG16 network improved by reducing convolutional layers and pooling layers, adding a Dropout layer, and deleting a classifier; the asteroid image feature extraction network is used to extract feature points from an input ground-based radar two-dimensional range Doppler image; and the multi-level inversion three-dimensional point cloud network is used to invert the extracted feature points into an asteroid three-dimensional point cloud model. The ground-based radar asteroid 3D reconstruction module is used to process the asteroid 3D point cloud model to obtain the final asteroid 3D point cloud model.

Images