Space target three-dimensional reconstruction method based on radar multi-view observation under low signal-to-noise ratio
An adaptive 3D reconstruction network designed using self-supervised constraints and multi-angle observation information solves the accuracy and robustness problems of radar 3D reconstruction under low signal-to-noise ratio and no prior information, achieving high-precision 3D reconstruction results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING INST OF TECH
- Filing Date
- 2026-01-30
- Publication Date
- 2026-05-12
Smart Images
Figure CN122023656A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for three-dimensional reconstruction of spatial targets based on radar multi-view observation under low signal-to-noise ratio conditions, belonging to the field of radar technology. Background Technology
[0002] Radar imaging technology can acquire high-resolution images of targets under all-weather, all-day conditions, and is therefore widely used in remote sensing, target identification, and topographic mapping. Traditional radar imaging is mainly two-dimensional, such as synthetic aperture radar (SAR) and inverse synthetic aperture radar (ISAR) imaging methods, whose main output is range-Doppler two-dimensional images. Compared to two-dimensional imaging, radar three-dimensional imaging can more completely describe the spatial structure information of the target, and has important application value in complex target analysis and detailed modeling.
[0003] However, in practical applications, radar imaging is inevitably affected by factors such as system noise and environmental clutter, resulting in a low signal-to-noise ratio in the imaging results, which severely impacts the accuracy and robustness of radar 3D reconstruction. Existing radar 3D imaging and reconstruction methods typically rely on high-quality observation data or prior target models, such as known target scattering characteristics, sparsity assumptions, or geometric constraints. When the observation environment is complex and the target's morphological characteristics are unknown, these traditional methods struggle to obtain high-precision 3D reconstruction results.
[0004] Furthermore, multi-view radar observation can provide information about targets under different observation geometries, but existing methods often fail to fully utilize the geometric consistency relationship between multi-view observations, and usually require manual annotation or real 3D models as supervision information, which limits their application in real-world scenarios.
[0005] Therefore, there is an urgent need to study a method that can fully utilize multi-view radar observation data to achieve high-precision 3D reconstruction in the absence of prior information on the target and scene. Based on this, this invention proposes a self-supervised radar 3D reconstruction framework that combines J-invariance and geometric consistency. By constructing self-supervised constraints through multi-angle observation information, it effectively suppresses noise and improves the accuracy of target 3D reconstruction without requiring real 3D topography. Summary of the Invention
[0006] To address the failure of traditional radar 3D reconstruction methods when prior information about the target scene is unknown and the observation signal-to-noise ratio is low, a new 3D reconstruction method for spatial targets based on radar multi-view observation under low signal-to-noise ratio conditions is proposed.
[0007] This invention is achieved through the following technical solution: Step 1: Quantitatively analyze the projection relationship between the target's three-dimensional structure and the radar's two-dimensional image, and establish a radar three-dimensional imaging model; Step 2: Through the multi-angle observation process, the mutual independence of radar images acquired from each angle is used to construct a mapping that conforms to J-invariance. Then, a self-supervised 3D reconstruction architecture that utilizes multi-angle observation information is designed to achieve end-to-end mapping from low signal-to-noise ratio 2D images to high-precision 3D structures. Step 3: Design an adaptive 3D reconstruction network based on multi-angle observation information. Constrain the network through J-invariance and geometric structure consistency to improve the accuracy and robustness of 3D reconstruction. Attached Figure Description
[0008] Figure 1 Flowchart of the proposed method; Figure 2 A schematic diagram of CST electromagnetic simulation of a space target; Figure 3 Three-dimensional reconstruction results based on CST electromagnetic simulation data under different methods; Figure 4 Quantitative evaluation results of three-dimensional reconstruction results obtained from CST electromagnetic simulation data under different methods.
[0009] Beneficial effects 1. This invention achieves high-precision 3D reconstruction of radar targets under conditions where the target's 3D structure and prior scene information are unknown. By quantitatively establishing the projection relationship between the target's 3D structure and the radar's 2D image, and combining multi-angle observation information to construct a radar 3D imaging model, it provides a model foundation for 3D reconstruction under unknown targets and complex scenes; 2. This invention significantly improves the accuracy of radar 3D reconstruction under low signal-to-noise ratio (SNR) observation conditions. It constructs a self-supervised architecture conforming to J-invariance by utilizing the independence between multi-angle radar observation images. Without requiring real 3D labels or high-quality reference data, it achieves effective noise suppression, overcoming the problem of traditional 3D reconstruction methods easily failing under low SNR conditions. 3. This invention fully utilizes the consistency of the target's geometric structure under different observation perspectives, thereby improving the robustness of the 3D reconstruction results. By introducing geometric consistency constraints into the self-supervised 3D reconstruction network, the reconstruction results under different observation angles maintain consistency in spatial structure, effectively reducing reconstruction errors caused by changes in observation angle, noise interference, or incomplete data, thus improving the robustness of the 3D imaging results. Detailed Implementation
[0010] The implementation of the method of the present invention will be described below with reference to the accompanying drawings and embodiments. The flowchart of the method proposed in this invention is shown below. Figure 1 As shown.
[0011] Step 1: Quantitatively analyze the projection relationship between the target's three-dimensional structure and the radar's two-dimensional image, and establish a radar three-dimensional imaging model.
[0012] First, according to the radar imaging mechanism, a two-dimensional radar image of a target is a projection of its three-dimensional structure onto the imaging plane, which can be mathematically modeled as follows: (1) in, kr Represents the distance projection vector. kd Represents the Doppler projection vector. r Represents the target distance coordinates. d Indicates the Doppler coordinates of the target. q This represents the target's three-dimensional coordinates. It should be noted that all methods for 3D reconstruction based on multi-view 2D image sequences revolve around this projection relationship.
[0013] Step 2: Through multi-angle observation, the independent nature of radar images acquired from each angle is used to construct a mapping that conforms to J-invariance. Then, a self-supervised 3D reconstruction architecture that utilizes multi-angle observation information is designed to achieve end-to-end mapping from low signal-to-noise ratio 2D images to high-precision 3D structures.
[0014] First, multi-view radar observation images need to be acquired and preprocessed. Specifically, observations are performed along the height direction to acquire... M Multi-view observation echoes of several targets, with the azimuth synthetic aperture length corresponding to each echo being... M Next, the entire synthetic aperture is divided into... N There are 10 sub-apertures, and the length of each sub-aperture is 1000. L / N Therefore, the total amount that can be obtained is... Nall = M × N A noisy radar echo. Furthermore, Nall The noisy radar echoes are divided into Q Groups and each group contains K A noisy radar echo. Then, using radar imaging algorithms, a series of noisy radar images can be obtained, i.e. .
[0015] Introducing a radar image grouper It can Divided into and its complement Since the acquisition process of each radar image in multi-view observation is independent and the radar imaging operators do not overlap, a system that conforms to this principle can be constructed. J Mapping of invariance g ,Right now .
[0016] Therefore, the optimal mapping g* must minimize the following expression: (2) in, , Through mathematical derivation, we can obtain: (3) It should be noted that, for The unbiased estimate, a property that allows noise to be suppressed through statistical averaging, thus approximating a clean image. Therefore, based on the above theoretical analysis, equation (3) can be equivalent to a supervised process and a fixed noise floor. This shows that the architecture proposed in this invention can solve the problem of poor image quality under low signal-to-noise ratio observation conditions by utilizing multi-view observation data. In addition, as can be seen from the aforementioned analysis of the geometric projection relationship of three-dimensional imaging, multi-view information can also be used for three-dimensional reconstruction. Therefore, based on multi-view information, a three-dimensional reconstruction architecture based on a self-supervised architecture can be designed.
[0017] Step 3: Design an adaptive 3D reconstruction network based on multi-angle observation information. Constrain the network through J-invariance and geometric structure consistency to improve the accuracy and robustness of 3D reconstruction.
[0018] Based on the self-supervised 3D reconstruction architecture described above, we designed a network that fully utilizes multi-view information, namely, an adaptive 3D imaging network based on a self-supervised architecture.
[0019] First, define a radar image grouper, namely Its function is to divide the radar image sequence into an input part and a label part. Specifically, F The grouping strategy can be expressed as The input section contains K - One radar image is used, and the label part contains one radar image. Secondly, to achieve 3D reconstruction, the network input should also include projection vectors of the distance and Doppler dimensions. Therefore, the network input consists of a noisy image and its corresponding projection vector, while the network label part is a noisy image (complementary to the noisy image in the input part). Finally, to improve the effectiveness and robustness of the designed self-supervised adaptive 3D imaging network, a loss function based on multi-view observation information was designed. f . f It consists of two parts, namely based on J Invariant loss function for image denoising Loss functions that improve the robustness of 3D reconstruction based on the consistency of spatial target geometry .
[0020] First, construct the conformance. J Invariant loss function The purpose of network training is to find optimal parameters. This minimizes the right-hand side of the following equation, i.e. (4) in, F C indicates F The complement of the network. Theoretical analysis shows that the training process based on the self-supervised architecture described above is equivalent to a supervised process with a fixed noise floor. Therefore, the network output... It can be represented as: (5) in, This indicates the denoising result of the radar image. Nfix This represents a fixed noise floor. This noise floor can be estimated and suppressed by introducing global and local convolutional modules into the designed network.
[0021] Secondly, a loss function that conforms to the consistency of multi-view geometric structure is constructed. Given that space targets possess rigid body characteristics, we assume... v Based on the shared consistent 3D geometric structure among the 3D reconstruction results, a 3D geometric consistency loss function can be introduced. This is done to constrain and enhance the geometric consistency of 3D reconstruction results from different perspectives, and its specific form is as follows: (6) in, S This is the result of 3D reconstruction. Furthermore, it can be... and This combination enables adaptive and robust low signal-to-noise ratio, high-precision 3D reconstruction, as shown in the following equation: (7) In summary, the proposed method organically combines multi-view observation with a self-supervised architecture. Through the design of a 3D reconstruction network based on J-invariance constraints and geometric consistency constraints, it can achieve high-precision 3D reconstruction of space targets with low signal-to-noise ratio.
[0022] To verify the algorithm's performance, a high-precision 3D reconstruction experiment of space targets was carried out based on an electromagnetic simulation dataset, and qualitative and quantitative analyses were conducted.
[0023] Example 1 This experiment was used to analyze the 3D reconstruction accuracy of the algorithm. The experimental parameters are shown in Table 1. The experiment used CST electromagnetic simulation software to conduct electromagnetic simulations on a space target and performed 3D reconstruction processing under low signal-to-noise ratio conditions. A schematic diagram of the CAD model simulation of the space target is shown below. Figure 2 As shown.
[0024] The proposed method is compared with other radar 3D reconstruction methods, such as factorization-based 3D reconstruction, ISEA (Integrated Sound Energy Evolution), and ISAR-NeRF. The comparison of 3D reconstruction results is shown in the figure. Figure 3 As shown. Due to the low signal-to-noise ratio and the complex scattering of space targets, the 3D reconstruction method based on factorization is difficult to obtain the complete target structure; therefore, it is not used for qualitative and quantitative evaluation. Figure 3 It can be seen that the proposed method can effectively achieve 3D reconstruction while preserving structural information such as the main body and solar panels of the space target under low signal-to-noise ratio conditions. In contrast, other methods have poor 3D reconstruction results and poor ability to restore local details under low signal-to-noise ratio conditions.
[0025] Furthermore, we quantitatively compared the 3D reconstruction accuracy of the proposed method with other methods by evaluating the Integrity Ratio and Chamfer Distance (CD). The results are shown in Table 2. Table 2 shows that for the four different observation scenarios of 8, 16, 24, and 32 viewpoints: compared to ISEA, the proposed method improved the IR by 0.19, 0.21, 0.31, and 0.27, respectively, and reduced the CD by 2.08, 1.77, 1.79, and 1.82, respectively; compared to ISAR-NeRF, the proposed method improved the IR by 0.24, 0.26, 0.23, and 0.21, respectively, and reduced the CD by 1.26, 0.75, 0.54, and 0.43, respectively.
[0026] Furthermore, to fully verify the 3D reconstruction performance of the proposed method, the IR and CD of the proposed method and other methods were quantitatively evaluated at different signal-to-noise ratios. The evaluation results are as follows: Figure 4 As shown. By Figure 4 It can be seen that the proposed method has significantly better 3D reconstruction performance than other methods when the signal-to-noise ratio is extremely low (especially when the image signal-to-noise ratio is below 6dB).
[0027] Table 1 Parameters of Simulation Experiment 1
[0028] Table 2 Quantitative evaluation results based on CST simulation data
[0029] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for three-dimensional reconstruction of spatial targets based on radar multi-view observation under low signal-to-noise ratio, characterized in that... The method includes the following steps: Step 1: Quantitatively analyze the projection relationship between the target's three-dimensional structure and the radar's two-dimensional image, and establish a radar three-dimensional imaging model; Step 2: Through the multi-angle observation process, the mutual independence of radar images acquired from each angle is used to construct a mapping that conforms to J-invariance. Then, a self-supervised 3D reconstruction architecture that utilizes multi-angle observation information is designed to achieve end-to-end mapping from low signal-to-noise ratio 2D images to high-precision 3D structures. Step 3: Design an adaptive 3D reconstruction network based on multi-angle observation information. Constrain the network through J-invariance and geometric structure consistency to improve the accuracy and robustness of 3D reconstruction.
2. The method for three-dimensional reconstruction of spatial targets based on radar multi-view observation under low signal-to-noise ratio as described in claim 1, characterized in that, In step one, the two-dimensional radar image of the target is a projection of its three-dimensional structure onto the imaging plane, which can be mathematically modeled as follows: (1)； in, kr Represents the distance projection vector. kd Represents the Doppler projection vector. r Represents the target distance coordinates. d Indicates the Doppler coordinates of the target. q Represents the three-dimensional coordinates of the target.
3. The method for three-dimensional reconstruction of spatial targets based on radar multi-view observation under low signal-to-noise ratio as described in claim 1, characterized in that, In step two, a self-supervised 3D reconstruction architecture utilizing multi-angle observation information was designed, making the training process based on the self-supervised architecture equivalent to a supervised training process: Input: Noisy image, distance dimension, and Doppler projection vector; Output: 3D reconstruction results.
4. The method for three-dimensional reconstruction of spatial targets based on radar multi-view observation under low signal-to-noise ratio as described in claim 1, characterized in that, In step two, a matching J Mapping of invariance g ,Right now ; The optimal mapping g* must minimize the following expression: (2)； in, , Through mathematical derivation, we can obtain: (3)； for Unbiased estimation, For a clean image.
5. The method for three-dimensional reconstruction of spatial targets based on radar multi-view observation under low signal-to-noise ratio as described in claim 1, characterized in that, In step three, construct the conformity J Invariant loss function Optimal parameters Represented as: (4)； in, F C indicates F The complement; Network output for: (5)； in, This indicates the denoising result of the radar image. N fix This indicates a fixed noise floor.
6. The method for three-dimensional reconstruction of spatial targets based on radar multi-view observation under low signal-to-noise ratio as described in claim 1, characterized in that, In step three, assume v To ensure consistency in 3D geometry among the reconstructed 3D results, a 3D geometric consistency loss function is introduced. This is done to constrain and enhance the geometric consistency of 3D reconstruction results from different perspectives, and its specific form is as follows: (6)； in, S This is the result of three-dimensional reconstruction; furthermore, it can be... and This combination enables adaptive and robust low signal-to-noise ratio, high-precision 3D reconstruction, as shown in the following equation: 。