Space target three-dimensional reconstruction method and device fused with attitude estimation

By constructing attitude transformation coordinates and position encoding strategies, and combining them with implicit neural networks to optimize attitude parameters, the problems of applicability and high computational cost in traditional methods are solved, and high-accuracy attitude estimation and 3D reconstruction of non-cooperative spatial targets are achieved.

CN118982633APending Publication Date: 2024-11-19SUN YAT SEN UNIV
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202411012951.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-26
Publication Date
2024-11-19

AI Technical Summary

Technical Problem

In the existing technology, traditional attitude estimation methods have low applicability to non-cooperative spatial targets, high computational cost, and rely on a pre-built target feature database, making it impossible to estimate the attitude of targets not included in the database, resulting in low accuracy.

Method used

By acquiring inverse synthetic aperture radar images, an attitude transformation coordinate and position encoding strategy is constructed. The attitude parameters are then optimized using an implicit neural network to perform 3D reconstruction of spatial targets, reducing computational costs and improving accuracy.

Benefits of technology

It achieves high-precision attitude estimation and 3D reconstruction for all space targets, has wide applicability, and reduces computational costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118982633A_ABST
    Figure CN118982633A_ABST
Patent Text Reader

Abstract

The invention discloses a space target three-dimensional reconstruction method and device fusing attitude estimation. The method comprises the following steps: acquiring an inverse synthetic aperture radar image of a space target; constructing attitude transformation coordinates according to the inverse synthetic aperture radar image and the scene value; constructing a position coding strategy according to the position coding frequency and the adjustable coding parameters; constructing a rendering pixel value according to the attitude transformation coordinate, the position coding strategy and the sight angle direction; according to the rendering pixel value and the image pixel value, target attitude parameters are calculated, and the target attitude parameters comprise target rotation matrix parameters or target offset matrix parameters; and according to the target attitude parameters, carrying out space target three-dimensional reconstruction to obtain a surface mesh model. According to the method, the three-dimensional reconstruction of the space target is realized by combining attitude estimation, the accuracy and the applicability are improved, and the calculation cost is reduced. The method can be widely applied to the technical field of space target recognition.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of space target recognition, and in particular to a method and device for three-dimensional reconstruction of space targets integrating posture estimation. Background Art

[0002] With the development of aerospace technology, the number of space targets in orbit is increasing. In order to improve space safety, high-resolution three-dimensional reconstruction is needed, and three-dimensional reconstruction technology requires high-precision attitude information of space targets. Traditional attitude estimation methods use attitude sensors to obtain attitude data, but they are only applicable to cooperative space targets equipped with corresponding sensors, not to non-cooperative targets, and have low applicability. In addition, when matching attitude data through a database, a large amount of computing resources is consumed, the computing cost is high, and it relies on a pre-built target feature database. It is difficult to estimate the attitude of targets that have not yet been included in the target feature database, and the accuracy is low.

[0003] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the invention

[0004] The embodiments of the present invention provide a method and device for three-dimensional reconstruction of a space target by integrating posture estimation, which effectively improves the accuracy and applicability and reduces the calculation cost.

[0005] On the one hand, an embodiment of the present invention provides a method for three-dimensional reconstruction of a space target by integrating posture estimation, comprising the following steps:

[0006] Acquire inverse synthetic aperture radar images of space targets;

[0007] constructing attitude transformation coordinates according to the inverse synthetic aperture radar image and the scene value;

[0008] Construct a position encoding strategy based on the position encoding frequency and adjustable encoding parameters;

[0009] Constructing a rendering pixel value according to the posture transformation coordinates, the position encoding strategy and the sight angle direction;

[0010] Calculating target posture parameters according to the rendered pixel values ​​and the image pixel values, wherein the target posture parameters include target rotation matrix parameters or target offset matrix parameters;

[0011] According to the target posture parameters, the space target is three-dimensionally reconstructed to obtain a surface mesh model.

[0012] In some embodiments, constructing attitude transformation coordinates according to the inverse synthetic aperture radar image and the scene value includes:

[0013] Initialize the rotation matrix and offset matrix;

[0014] constructing a normal vector according to a range axis projection vector of the inverse synthetic aperture radar image and a Doppler axis projection vector of the inverse synthetic aperture radar image;

[0015] Constructing the three-dimensional coordinates of the starting point according to the range axis projection vector, the Doppler axis projection vector and the point coordinates in the inverse synthetic aperture radar image;

[0016] Constructing the spatial point coordinates in the normal vector according to the normal vector, the three-dimensional coordinates of the starting point and the scene value;

[0017] The posture transformation coordinates are constructed according to the range axis projection vector, the Doppler axis projection vector, the three-dimensional coordinates of the starting point, the rotation matrix and the offset matrix.

[0018] In some embodiments, constructing a position coding strategy according to the position coding frequency and the adjustable coding parameter includes:

[0019] Divide the position coding frequency to obtain multiple coding frequencies;

[0020] Constructing a position coding weighting function according to the coding frequency and the adjustable coding parameter;

[0021] The position encoding strategy is constructed according to a plurality of the position encoding weighting functions.

[0022] In some embodiments, constructing a rendering pixel value according to the posture transformation coordinates, the position encoding strategy and the sight angle direction includes:

[0023] According to the posture transformation coordinates and the sight angle direction, encoding processing is performed through the position encoding strategy to obtain high and low frequency information;

[0024] Inputting the high and low frequency information into an implicit neural network to obtain particle density and backscattering coefficient;

[0025] Constructing a normal line according to the normal vector and the three-dimensional coordinates of the starting point;

[0026] The rendering pixel value is constructed according to the particle density, the backscattering coefficient and the normal line.

[0027] In some embodiments, the calculating the target posture parameter according to the rendering pixel value and the image pixel value includes:

[0028] Constructing a mean square error loss function according to the rendering pixel value, the image pixel value and a set of sample coordinate points;

[0029] Constructing a high dynamic scattering coefficient loss function according to the rendering pixel value, the image pixel value, the sign function and the number of samples involved in loss calculation;

[0030] Constructing a joint optimization loss function according to the mean square error loss function, the high dynamic scattering coefficient loss function and hyperparameters;

[0031] Constructing an objective function according to the joint optimization loss function, rotation matrix parameters, offset matrix parameters and implicit neural network parameters;

[0032] The rotation matrix and the offset matrix are updated by the objective function to obtain the target posture parameters.

[0033] In some embodiments, performing three-dimensional reconstruction of the space target according to the target posture parameters to obtain a surface mesh model includes:

[0034] Initialize 3D voxels;

[0035] extracting a mask image from the inverse synthetic aperture radar image according to the target posture parameter;

[0036] Back-projecting the plurality of mask images into an initial three-dimensional coordinate space to obtain a target three-dimensional coordinate space;

[0037] According to the target three-dimensional coordinate space, conversion processing is performed through a preset surface rendering algorithm to obtain the surface mesh model.

[0038] In some embodiments, the method further comprises:

[0039] Performing a posture estimation experiment according to the target posture parameters to obtain a posture estimation experiment result;

[0040] Alternatively, a three-dimensional reconstruction experiment is performed according to the target posture parameters to obtain a three-dimensional reconstruction experiment result.

[0041] In some embodiments, constructing the posture transformation coordinates according to the range axis projection vector, the Doppler axis projection vector, the starting point three-dimensional coordinates, the rotation matrix and the offset matrix includes:

[0042] According to the distance axis projection vector, the Doppler axis projection vector, the three-dimensional coordinates of the starting point, the rotation matrix and the offset matrix, the attitude transformation coordinates are constructed by the attitude transformation coordinate calculation formula, and the attitude transformation coordinate calculation formula is:

[0043]

[0044] In the formula, is the posture transformation coordinate, R is the rotation matrix, t is the offset matrix, I(u) is the three-dimensional coordinate of the starting point, d0 is the distance axis projection vector, v0 is the Doppler axis projection vector, and u is the scene value.

[0045] On the other hand, an embodiment of the present invention provides a three-dimensional reconstruction device for a space target by integrating posture estimation, comprising:

[0046] The first module is used to obtain an inverse synthetic aperture radar image of a space target;

[0047] A second module is used to construct attitude transformation coordinates according to the inverse synthetic aperture radar image and scene value;

[0048] The third module is used to construct a position encoding strategy based on the position encoding frequency and the adjustable encoding parameters;

[0049] A fourth module is used to construct a rendering pixel value according to the posture transformation coordinates, the position encoding strategy and the sight angle direction;

[0050] A fifth module is used to calculate target posture parameters according to the rendering pixel value and the image pixel value, wherein the target posture parameters include target rotation matrix parameters or target offset matrix parameters;

[0051] The sixth module is used to perform three-dimensional reconstruction of the space target according to the target posture parameters to obtain a surface mesh model.

[0052] In another aspect, an embodiment of the present invention provides a computer device, comprising:

[0053] at least one processor;

[0054] at least one memory for storing at least one program;

[0055] When the at least one program is executed by the at least one processor, the at least one processor implements the described method.

[0056] The beneficial effects of the present invention are as follows:

[0057] The embodiment of the present invention first obtains an inverse synthetic aperture radar image of a space target, and constructs a posture transformation coordinate according to the inverse synthetic aperture radar image and a scene value, then constructs a position coding strategy according to a position coding frequency and an adjustable coding parameter, and then constructs a rendering pixel value according to the posture transformation coordinate, the position coding strategy and the sight angle direction, and calculates the target posture parameter according to the rendering pixel value and the image pixel value, and finally performs three-dimensional reconstruction of the space target according to the target posture parameter to obtain a surface mesh model, so that the three-dimensional reconstruction of the space target can be realized in combination with posture estimation, thereby improving the accuracy and applicability and reducing the calculation cost.

[0058] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0060] Figure 1 A flowchart of three-dimensional reconstruction of a space target fused with posture estimation according to an embodiment of the present invention;

[0061] Figure 2 A schematic diagram of a position coding strategy result according to an embodiment of the present invention;

[0062] Figure 3 A schematic diagram of an overall process of performing posture estimation according to an embodiment of the present invention;

[0063] Figure 4 A schematic diagram of an initial attitude of a satellite according to an embodiment of the present invention;

[0064] Figure 5 This is a schematic diagram of a satellite attitude with a satellite iteration number of 250K according to an embodiment of the present invention;

[0065] Figure 6 This is a schematic diagram of an initial posture of a space station according to an embodiment of the present invention;

[0066] Figure 7 This is a schematic diagram of a posture of a space station with 250K iterations according to an embodiment of the present invention;

[0067] Figure 8 A schematic diagram of an attitude error curve of a satellite iteration process according to an embodiment of the present invention;

[0068] Fig. 9 A schematic diagram of an attitude error curve of an iterative process of a space station according to an embodiment of the present invention;

[0069] Fig.10 A schematic diagram of a quantitative analysis result of a posture estimation error according to an embodiment of the present invention;

[0070] Fig.11 A comparison diagram of a satellite attitude optimization process according to an embodiment of the present invention;

[0071] Fig.12 This is a comparison diagram of a space station attitude optimization process according to an embodiment of the present invention;

[0072] Fig.13 A two-dimensional new view comparison diagram of a satellite joint attitude optimization process according to an embodiment of the present invention;

[0073] Fig.14 A two-dimensional new view comparison diagram of a space station joint attitude optimization process according to an embodiment of the present invention;

[0074] Fig.15 A schematic diagram of a performance evaluation result in a three-dimensional reconstruction and new view generation task according to an embodiment of the present invention;

[0075] Fig.16 A schematic diagram of the structure of a three-dimensional reconstruction device for a space target integrating posture estimation according to an embodiment of the present invention;

[0076] Fig.17 The figure is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0077] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below in conjunction with the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of the present application. They are only examples of devices and methods consistent with some aspects of the embodiments of the present application as detailed in the attached claims.

[0078] It is understood that the terms "first", "second", etc. used in this application can be used to describe various concepts in this article, but unless otherwise specified, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another concept. For example, without departing from the scope of the embodiment of the present application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the words "if" and "if" as used herein can be interpreted as "at the time of" or "when" or "in response to determination".

[0079] The terms "at least one", "multiple", "each", "any", etc. used in this application, at least one includes one, two or more, multiple includes two or more, each refers to each of the corresponding multiple, and any refers to any one of the multiple.

[0080] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0081] Before describing the embodiments of the present application in detail, some nouns and terms involved in the embodiments of the present application are first described. The nouns and terms involved in the embodiments of the present application are subject to the following explanations.

[0082] Neural Radiance Fields: A deep learning model for 3D scene reconstruction. It is trained to learn how the light of a scene interacts with the surface of an object, thereby generating high-quality 3D scene images.

[0083] Inverse-Synthetic-Aperture-Radar (ISAR): is a radar technology that generates high-resolution images of targets by analyzing the reflected signals of radar waves on the target. ISAR technology uses the relative motion of the radar platform during flight or movement to simulate a larger antenna aperture, thereby improving imaging resolution.

[0084] High Resolution Range Profile (HRRP): It is a radar signal processing technology that obtains detailed shape and size information of a target by analyzing the change in the reflected signal of the radar wave on the target over time. HRRP is often used in the military field to identify and classify different targets such as aircraft and ships.

[0085] Back Propagation: An algorithm used to train models in neural networks. It calculates the gradient of the loss function with respect to the network weights and uses these gradients to update the weights to minimize the prediction error.

[0086] Mean Square Error (MSE): is a commonly used indicator to measure the accuracy of a prediction model or estimation method. It is used to calculate the average of the squares of the difference between the actual value and the predicted value. The smaller the MSE, the closer the model's prediction is to the actual value, and the better the model's performance is.

[0087] Precision / Accuracy: It is used to measure the percentage of the reconstructed predicted point cloud that can be aligned with the real point cloud.

[0088] Completeness / Recall: It is used to measure the percentage of the reconstructed real point cloud that can align with the predicted point cloud.

[0089] Peak Signal to Noise Ratio: An evaluation indicator used to evaluate the difference between two images.

[0090] Structural Similarity Index Measure: It is a metric used to quantify the structural similarity between two images. It implements structural similarity calculation based on a model of the human visual system and is very sensitive to the perception of local structural changes in images.

[0091] Learned Perceptual Image Patch Similarity (LPIPS): A metric for quantifying the human-perceived similarity between two images. Unlike traditional metrics that calculate differences based on raw pixel values ​​or simple transformations, LPIPS uses deep learning to better approximate human visual perception.

[0092] Neural Radiance Field (NeRF): is a novel view synthesis method with implicit scene representation that has taken the field of computer vision by storm. As a novel view synthesis and 3D reconstruction method, NeRF models have been applied in robotics, urban mapping, autonomous navigation, virtual reality / augmented reality, and other fields.

[0093] In the related technologies, as aerospace technology continues to make breakthrough progress, the number of space targets on orbit has increased dramatically. In this context, accurately estimating the attitude information of space targets and realizing high-resolution three-dimensional imaging technology are crucial to ensuring space safety. Neural Radiance Fields is a method that realizes image rendering of any new perspective and restores the three-dimensional geometry of the target object from optical images of multiple perspectives. NeRF three-dimensional reconstruction technology requires accurate attitude information of the target object, which is very important for geometric reconstruction and new perspective image rendering effects. Accurately inferring attitude information is a key step in this technology. At present, the method of attitude estimation is limited by the extraction technology of scattering points and the accuracy of image registration. The loss of scattering points caused by the occlusion effect makes it extremely difficult to estimate the accuracy of attitude estimation, which brings great challenges to the joint attitude estimation using multi-view ISAR (Inverse-Synthetic-Aperture-Radar) images. The lack of attitude information further affects the reconstruction of three-dimensional objects. Therefore, it is of great significance to solve the problem of the loss of scattering points caused by the occlusion effect, which in turn affects the accuracy of attitude estimation. Traditional space target attitude estimation technology is mainly divided into two approaches: using attitude sensors and matching through databases. The first method using attitude sensors relies on multiple sensor data such as gyroscopes, magnetometers, sun sensors, and star trackers to achieve a comprehensive and accurate estimation of the target attitude. However, this method is only applicable to cooperative space targets equipped with corresponding sensors, and is not applicable to non-cooperative targets. The second method of matching through databases relies on a pre-built target feature database, which contains detailed information about the target, such as full-view electromagnetic scattering characteristics, ISAR images, and high-resolution range profiles (HRRP). This method estimates the target attitude by screening candidate information that matches the actually observed ISAR images or HRRP from the database. Due to its rich prior data, this method can provide extremely accurate attitude estimation results. However, this method has some limitations, such as the need to consume a lot of computing resources and the inability to perform attitude estimation on targets that are not yet included in the database. Among the existing attitude estimation methods, there are attitude estimation methods based on attitude sensors. Specifically, in order to solve the problem that MEMS inertial sensors have low accuracy and error accumulation over time, which makes it impossible to meet the requirements of long-term attitude measurement, an iterative update extended Kalman filter algorithm is proposed. However, this method requires large computing resources, and the number of iterations affects the estimation of attitude accuracy.The attitude estimation method based on prior database matching uses a software library to support the interpretation of ISAR images and realizes the function of fitting radar mapping to the actual observation sequence, which helps to reconstruct the attitude history of the target and then estimate the attitude information of the space target. However, this method requires a 3D model of the target to generate the radar mapping. If the 3D model does not completely match the geometry of the actual target, it may lead to inaccurate attitude estimation. The attitude estimation method based on feature inversion uses a robust high-level feature extraction method based on parallelogram descriptors to replace traditional scattering point features, which helps to automatically extract and estimate the attitude and geometric parameters of the target. However, this method has a strong dependence on the diversity of radar line of sight angles and has high requirements on the quality of image processing. It can be seen that the shortcomings of the existing technology include: (1) The sensor-based attitude estimation method requires the target to be equipped with a corresponding attitude sensor and the sensor information is available. This means that for space targets that do not have the required attitude sensor or whose sensor information is not available, this method cannot estimate its attitude. In addition, the information fusion redundancy of multiple sensors is high, and its complete sensor information cannot be fully utilized for attitude estimation. (2) The attitude estimation method using a priori database matching requires the establishment of a database of the target to be measured in advance, and the database can be used to accurately estimate the attitude information. However, one of the disadvantages of this method is that the estimation time is long, which requires a lot of computing time and has high computing cost. In addition, for space targets that do not exist in the database, their attitude cannot be estimated, and the applicability is low. (3) The attitude estimation method of feature inversion requires good data features. Generally, for ISAR images, their point features, line features or other physical features unique to the target are mainly used. However, for targets with more complex structures, it is difficult to obtain effective attitude estimation features, the effect is poor, and the accuracy is low. In summary, how to reduce the requirements for target information acquisition and obtain accurate attitude estimation features, and how to achieve accurate attitude estimation from ISAR image attitudes containing noise errors or unknown parameters is a technical problem worth exploring.

[0094] In view of this, the embodiment of the present invention can jointly handle the tasks of attitude estimation and three-dimensional reconstruction of space targets in the presence of estimation errors or unknown parameters, and finally achieve high-precision attitude estimation and high-resolution three-dimensional reconstruction effects. Based on the attitude estimation model of the space target, the attitude parameters and the NeRF implicit neural network parameters are jointly optimized, and the mathematical process is rigorously derived, and the mathematical formula based on the back propagation algorithm is analyzed. In view of the problem of the asynchrony between the attitude parameter optimization and the implicit neural network optimization in the mathematical principle of the back propagation process, a modified weighted position encoding function is proposed to solve the problem. Compared with the prior art solutions, the method of the present invention has lower requirements for target information acquisition, and does not rely on the constructed prior information library, and is applicable to all space targets. At the same time, after the attitude estimation is completed, the three-dimensional reconstruction of the space target is performed using high-precision attitude parameters, so that the three-dimensional reconstruction of the space target can be realized in combination with the attitude estimation, which improves the accuracy and applicability and reduces the computational cost.

[0095] The three-dimensional reconstruction method of space targets with fusion posture estimation provided in the embodiment of the present application relates to the field of space target recognition technology. The three-dimensional reconstruction method of space targets with fusion posture estimation provided in the embodiment of the present application can be applied to a terminal, can also be applied to a server, and can also be software running in a terminal or a server. In some embodiments, the terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, and a car terminal, etc., but is not limited to this; the server side can be configured as an independent physical server, or a server cluster or distributed system composed of multiple physical servers, and can also be configured to provide cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and cloud servers for basic cloud computing services such as big data and artificial intelligence platforms. The server can also be a node server in a blockchain network; the software can be an application of a three-dimensional reconstruction method for space targets with fusion posture estimation, etc., but is not limited to the above forms.

[0096] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0097] The following is a detailed explanation of the embodiments of the present application in conjunction with the accompanying drawings:

[0098] Figure 1 is an optional flowchart of the method for three-dimensional reconstruction of a space target fused with posture estimation provided in an embodiment of the present application. Figure 1 The method may include but is not limited to steps S101 to S106.

[0099] Step S101, obtaining an inverse synthetic aperture radar image of a space target;

[0100] Step S102, constructing attitude transformation coordinates according to the inverse synthetic aperture radar image and the scene value;

[0101] Step S103, constructing a position coding strategy according to the position coding frequency and the adjustable coding parameters;

[0102] Step S104, constructing a rendering pixel value according to the posture transformation coordinates, the position encoding strategy and the sight angle direction;

[0103] Step S105, calculating target posture parameters according to the rendering pixel values ​​and the image pixel values, the target posture parameters including target rotation matrix parameters or target offset matrix parameters;

[0104] Step S106: Perform three-dimensional reconstruction of the space target according to the target posture parameters to obtain a surface mesh model.

[0105] Steps S101 to S106 shown in the embodiment of the present application are combined with posture estimation to achieve three-dimensional reconstruction of space targets, thereby improving accuracy and applicability and reducing computing costs.

[0106] In step S101 of some embodiments, the inverse synthetic aperture radar image of the space target may be obtained through a space database or other methods, which are not limited thereto.

[0107] In some embodiments, in step S102, constructing attitude transformation coordinates according to the inverse synthetic aperture radar image and the scene value may include but is not limited to the following steps:

[0108] Initialize the rotation matrix and offset matrix;

[0109] constructing a normal vector according to a range axis projection vector of the inverse synthetic aperture radar image and a Doppler axis projection vector of the inverse synthetic aperture radar image;

[0110] Constructing the three-dimensional coordinates of the starting point according to the range axis projection vector, the Doppler axis projection vector and the point coordinates in the inverse synthetic aperture radar image;

[0111] Construct the spatial point coordinates in the normal vector according to the normal vector, the three-dimensional coordinates of the starting point and the scene value;

[0112] The attitude transformation coordinates are constructed according to the range axis projection vector, the Doppler axis projection vector, the three-dimensional coordinates of the starting point, the rotation matrix and the offset matrix.

[0113] In some embodiments, the rotation matrix and the offset matrix can be initialized first, wherein the rotation matrix and the offset matrix are both initial attitude parameters, and the initial attitude parameters can be subsequently updated through the loss function to perform attitude estimation to obtain the target attitude parameters. Then, according to the range axis projection vector of the inverse synthetic aperture radar image and the Doppler axis projection vector of the inverse synthetic aperture radar image, a normal vector is constructed, wherein the calculation formula of the normal vector is: n0=d0×v0, wherein n0 is the normal vector, d0 is the range axis projection vector, and v0 is the Doppler axis projection vector. It can be understood that the range axis projection vector and the Doppler axis projection vector are fixed, and can be subsequently transformed according to the rotation matrix and the offset matrix of the attitude parameters to obtain the vector for three-dimensional reconstruction. Then, according to the range axis projection vector, the Doppler axis projection vector and the point coordinates in the inverse synthetic aperture radar image, the three-dimensional coordinates of the starting point are constructed, wherein the calculation formula of the three-dimensional coordinates of the starting point is: o0=r·d0+f d v0, where o0 is the three-dimensional coordinate of the starting point, (r, f d) is the coordinate of any point in the inverse synthetic aperture radar image. It can be understood that the three-dimensional coordinates of the starting point can be used as the starting point of the sampling coordinates in the NeRF model. Then, according to the normal vector, the three-dimensional coordinates of the starting point and the scene value, the spatial point coordinates in the normal vector are constructed, that is, the spatial coordinates of any point in the normal vector, where the calculation formula of the spatial point coordinates is: I(u)=o0+un0, where I(u) is the spatial point coordinate, and u is the scene value, that is, the number between the scene sizes. For example, assuming that the scene size is [-12, 12], then the scene value u is a number within this range. Finally, according to the distance axis projection vector, the Doppler axis projection vector, the three-dimensional coordinates of the starting point, the rotation matrix and the offset matrix, the attitude transformation coordinates are constructed through the attitude transformation coordinate calculation formula. The estimated attitude information used for three-dimensional reconstruction is noisy or completely unknown. The initialized attitude information can be corrected by the rotation matrix and the offset matrix. The coordinates sampled by the correction result are applied to the attitude transformation The coordinates after the posture transformation. The calculation formula of the posture transformation coordinates is: or, In the formula, is the attitude transformation coordinate, R is the rotation matrix, t is the offset matrix, also known as the distance Doppler direction offset of the space target from the center position, I(u) is the three-dimensional coordinate of the starting point, d0 is the distance axis projection vector, v0 is the Doppler axis projection vector, u is the scene value, Γ is the attitude transformation function, and o0 is the three-dimensional coordinate of the starting point. It can be understood that the attitude transformation function Γ first translates the initial coordinate I(u) in the initial projection plane by [d0v0]t, and then performs a rotation transformation through the rotation matrix R to obtain the transformed attitude transformation coordinate The relationship between the transformed posture transformation coordinates and the original starting point three-dimensional coordinates is the superposition of rotation and translation.

[0114] In some embodiments, in step S103, a position coding strategy is constructed according to the position coding frequency and the adjustable coding parameters, which may include but is not limited to the following steps:

[0115] Divide the position coding frequency to obtain multiple coding frequencies;

[0116] According to the coding frequency and the adjustable coding parameters, a position coding weighting function is constructed;

[0117] A position encoding strategy is constructed based on multiple position encoding weighting functions.

[0118] In some embodiments, high-frequency position encoding will produce higher gradient weights, which means that in the process of optimizing the attitude, the high-frequency details of the ISAR image will dominate the attitude optimization process. However, in the process of implicit neural network fitting three-dimensional information to render two-dimensional ISAR images, network optimization often captures the low-frequency features of the data first, and then gradually fits the high-frequency features. This learning strategy leads to the network fitting of high-frequency details in the early stage of optimization, which is not stable and reliable enough, resulting in many inaccurate and invalid gradient calculation results in the attitude optimization process dominated by high-frequency information. In order to synchronize the optimization process of attitude optimization and implicit neural network, the position encoding needs to be improved accordingly. The improved position encoding can make the dominant information relied on in the attitude optimization process gradually transition from low-frequency features to high-frequency features during the training process. In some embodiments, the position encoding frequency can be divided first to obtain multiple encoding frequencies. Exemplarily, the position encoding frequency can be divided into k=0, k=2, k=4, k=6 or k=8, where k is the encoding frequency. Then, according to the encoding frequency and the adjustable encoding parameter, the position encoding weighting function w is constructed. k (α), where α is an adjustable encoding parameter, and the position encoding weighting function is used to weight the position encoding. Finally, a position encoding strategy is constructed based on multiple position encoding weighting functions. For example, the position encoding strategy result is as follows Figure 2 As shown in the figure, the value of the position encoding weighting function increases linearly from 0 to the encoding length during the training process. k When (α) = 0, the contribution of the kth encoding frequency and its higher frequency components to the gradient is zero, and the information of posture optimization is dominated by the posture information lower than the kth encoding frequency, which better matches the process of implicit neural network fitting. It can be understood that the position encoding strategy can weight the frequency of position encoding as training progresses, and gradually increase the weight of high-frequency position encoding as training progresses.

[0119] In some embodiments, in step S104, constructing a rendering pixel value according to the posture transformation coordinates, the position encoding strategy and the sight angle direction may include but is not limited to the following steps:

[0120] According to the posture transformation coordinates and the sight angle direction, the high and low frequency information is obtained by encoding through the position encoding strategy;

[0121] The high and low frequency information is input into the implicit neural network to obtain the particle density and backscattering coefficient;

[0122] Construct a normal line based on the normal vector and the three-dimensional coordinates of the starting point;

[0123] Constructs the rendered pixel value based on the particle density, backscatter coefficient, and normal line.

[0124] In some embodiments, the coordinates and the direction of the sight angle can be transformed according to the posture, and the high and low frequency information can be obtained by encoding through the position encoding strategy, and the high and low frequency information can be input into the implicit neural network to obtain the particle density and backscattering coefficient. The mathematical formula of the implicit neural network processing process is expressed as: In the formula, F Ψ represents an implicit neural network, is the attitude transformation coordinate, is the sight angle direction, is the particle density, is the backscattering coefficient. The calculation formula for the sight angle direction is: In the formula, is the line of sight angle direction, R is the rotation matrix, and d0 is the distance axis projection vector. It can be understood that the three-dimensional space coordinates can be discretely sampled, and then the discrete coordinate points are applied with the attitude transformation function Γ to obtain the attitude transformation coordinates Then, the normal line is constructed based on the normal vector and the three-dimensional coordinates of the starting point. d ) can be obtained by cross-producting the distance axis projection vector d0 and the Doppler axis projection vector v0 to obtain the normal vector n0, which is obtained by the coordinates (r, f) of the ISAR image. d ) is calculated with d0 and v0 to obtain the three-dimensional coordinates of the starting point of the normal line, and the normal line is obtained from the three-dimensional coordinates of the starting point and the normal vector (i.e. the starting point and direction). Finally, according to the particle density and backscattering coefficient, radar imaging is performed on the normal line to construct the rendering pixel value, where the calculation formula of the rendering pixel value is: In the formula, is the rendered pixel value, u t is the upper bound of the scene size, u b is the lower bound of the scene size, ρ is the backscattering coefficient, σ is the particle density, is the attitude transformation coordinate, is the sight angle direction, j is the imaginary part of the complex number, f c is the fundamental frequency of the radar signal, c is the speed of light, R d is the function of the distance from the transformed coordinate to the radar position, u is the scene value, and c is the integral variable.

[0125] In some embodiments, in step S105, the target posture parameters are calculated according to the rendering pixel values ​​and the image pixel values, which may include but is not limited to the following steps:

[0126] Construct a mean square error loss function based on the set of rendering pixel values, image pixel values ​​and sample coordinate points;

[0127] A high dynamic scattering coefficient loss function is constructed based on the rendering pixel value, the image pixel value, the sign function and the number of samples involved in the loss calculation;

[0128] Construct a joint optimization loss function based on the mean square error loss function, the high dynamic scattering coefficient loss function and hyperparameters;

[0129] Constructing an objective function by jointly optimizing a loss function, rotation matrix parameters, offset matrix parameters, and implicit neural network parameters;

[0130] The rotation matrix and offset matrix are updated through the objective function to obtain the target posture parameters.

[0131] In some embodiments, given the rendered pixel values ​​and the real image pixel values ​​of the ISAR image, an objective function for optimizing the model parameters can be constructed. The joint optimization loss function adopted in this embodiment is the MSE mean square error loss function L MSE and high dynamic scattering coefficient loss function Combined. The mean square error loss function can be constructed based on the set of rendering pixel values, image pixel values ​​and sample coordinate points. The calculation formula of the mean square error loss function is: Where, L MSE is the mean square error loss function, N c The number of samples involved in the loss calculation, is the rendered pixel value, is the image pixel value, is the attitude transformation coordinate, is a set of sample coordinate points. Then, according to the rendering pixel value, image pixel value, sign function and the number of samples involved in the loss calculation, a high dynamic scattering coefficient loss function is constructed, where the calculation formula of the high dynamic scattering coefficient loss function is: In the formula, is the high dynamic scattering coefficient loss function, sg(·) is the sign function, and ε is a very small number, which is used to prevent the denominator from being 0. Then, based on the mean square error loss function, the high dynamic scattering coefficient loss function and the hyperparameters, a joint optimization loss function is constructed, where the calculation formula of the joint optimization loss function is: Where L is the joint optimization loss function and λ is a hyperparameter. Then, the objective function is constructed based on the joint optimization loss function, rotation matrix parameters, offset matrix parameters and implicit neural network parameters. The calculation formula of the objective function is: In the formula, is the optimized target rotation matrix parameter, is the optimized target offset matrix parameter, is the optimized implicit neural network parameter, w is the parameter of the rotation matrix, v is the parameter of the offset matrix, Ψ is the implicit neural network parameter, R is the rotation matrix, t is the offset matrix, I c is the ISAR image sampling point. It can be understood that in the entire joint optimization process, the learnable parameter variables are the attitude parameters and And the implicit neural network parameters Ψ. Finally, the rotation matrix and the offset matrix are updated through the objective function to obtain the target attitude parameters. Among them, the target attitude parameters include the target rotation matrix parameters or the target offset matrix parameters. It can be understood that in each iterative optimization process, the gradient information of the attitude parameters [w, v] and Ψ is calculated differentiably according to the objective function, and then the attitude parameters and implicit neural network parameters are optimized using advanced gradient optimization algorithms. For the optimization of attitude parameters, after obtaining the gradient information of the corresponding parameters, the gradient is returned through the back propagation algorithm to correct the parameters.

[0132] In some embodiments, in step S106, three-dimensional reconstruction of the space target is performed according to the target posture parameters to obtain a surface mesh model, which may include but is not limited to the following steps:

[0133] Initialize 3D voxels;

[0134] Extracting a mask image from the inverse synthetic aperture radar image according to the target posture parameters;

[0135] Back-projecting multiple mask images into the initial three-dimensional coordinate space to obtain the target three-dimensional coordinate space;

[0136] According to the target three-dimensional coordinate space, the surface mesh model is obtained by converting it through a preset surface drawing algorithm.

[0137] In some embodiments, the correction of attitude parameters is conducive to more precise three-dimensional reconstruction. The three-dimensional reconstruction uses a back-projection three-dimensional imaging algorithm. The main idea of ​​the back-projection algorithm is to first extract the mask image of the target from the ISAR image, and then back-project the mask image into three-dimensional coordinates that satisfy the projection relationship. The ISAR images of multiple perspectives are projected into three-dimensional coordinates in sequence and accumulated, and finally the three-dimensional coordinates with the highest accumulation result are subjected to the Marching Cubes algorithm to obtain the three-dimensional reconstruction result of the target. The three-dimensional voxels can be initialized first. For example, in order to implement the back-projection algorithm, the three-dimensional voxels used to reconstruct the three-dimensional structure of the target can be initialized in advance. And set the coordinates of any three-dimensional voxel to zero. The three-dimensional voxel is used to store the cumulative result of the reverse projection of the two-dimensional image. Then, according to the target attitude parameters, the mask image is extracted from the inverse synthetic aperture radar image, wherein the method for extracting the mask image can adopt the threshold method, that is, when the scattering intensity of the target in the ISAR image is greater than the threshold ∈, it is considered that the pixel belongs to the scattering information of the target, and the pixel value is set to 1, otherwise the pixel value is set to 0. The mask image is used to represent the target information at the corresponding position of the ISAR image, and is used for subsequent three-dimensional reconstruction to determine the voxel coordinate domain. Then, multiple mask images are reversely projected into the initial three-dimensional coordinate space to obtain the target three-dimensional coordinate space. After determining the voxel coordinate domain, in order to obtain the surface mesh information of the three-dimensional reconstruction result, the preset surface drawing algorithm can be set to the Marching Cubes algorithm, and the binary three-dimensional voxels are transformed and processed according to the target three-dimensional coordinate space using the Marching Cubes algorithm to obtain the surface mesh model, that is, to complete the three-dimensional reconstruction. It can be understood that the Marching Cubes (MC, voxel-level reconstruction method) algorithm is a classic algorithm in the surface drawing algorithm.

[0138] In some embodiments, the overall process of performing posture estimation is as follows: Figure 3 As shown in the figure, based on the spatial target attitude estimation model, the attitude parameters and the parameters of the NeRF implicit neural network can be jointly optimized. The attitude parameters will affect the three-dimensional spatial coordinate position corresponding to the pixels of the sampling distance profile line. The coordinate information and direction information corresponding to the attitude parameters are sent to the weighted position encoding function to extract high- and low-frequency information, and then the encoded results are input into the implicit neural network. The implicit neural network outputs the corresponding backscattering coefficient and particle density. Then, the radar differentiable rendering formula is applied to render the image pixel value and calculate the loss with the real image pixel value. Finally, the gradient information of the learnable parameters is calculated according to the backpropagation principle for optimization and update, and the optimized attitude parameters are obtained to achieve attitude estimation. In the process of gradient backpropagation, the weighted position encoding function will correct the gradient of the high-frequency information, thereby aligning the attitude parameter optimization process with the implicit neural network optimization process.

[0139] In some embodiments, the method further comprises:

[0140] According to the target posture parameters, a posture estimation experiment is performed to obtain the posture estimation experiment results;

[0141] Alternatively, a three-dimensional reconstruction experiment is performed according to the target posture parameters to obtain a three-dimensional reconstruction experiment result.

[0142] In some embodiments, an attitude estimation experiment can be performed based on the target attitude parameters to obtain attitude estimation experiment results. The attitude estimation experiment results may include the initialization attitudes of the satellite and the space station under the condition of 3° attitude error initialization and the attitude results of 250K iterations. The initial attitude of the satellite is as follows: Figure 4 As shown, the satellite's attitude with 250K iterations is as follows Figure 5 As shown, the initial attitude of the space station is Figure 6 As shown, the posture of the space station with 250K iterations is as follows Figure 7 As shown. In these figures, the red box represents the geometric position of the true attitude, while the blue box represents the spatial position of the final converged attitude, and the red line between the two boxes represents the error distance of the attitude. From the results, under the condition of 3° attitude error initialization, the present embodiment can achieve the task of joint attitude estimation and three-dimensional reconstruction. The satellite may still have certain errors in the final result, mainly because there are fewer high-frequency details of its geometric structure, which is not convenient for high-precision attitude estimation. However, for space targets with complex geometric structures such as space stations, the attitude of the final result is almost completely consistent with the true attitude, which shows that the method of this embodiment can achieve high-precision attitude estimation for space targets with complex geometric structures. In addition, under 3° attitude error initialization, the attitude error curve of the satellite iteration process is as shown in the figure Figure 8 As shown in Figure 2, the attitude error curve of the space station iteration process is as follows: Fig. 9 As shown. Among them, Optimization Step represents the number of iterations, Rotation Error represents the rotation error, which is the blue line, and Translation Error represents the offset error, which is the red line. With the increase in the number of iterations, the rotation error and offset error in the attitude estimation process both show a trend of first increasing and then gradually decreasing. This change trend clearly demonstrates the effectiveness of the method of this embodiment in reducing the attitude estimation error. In the early stage of training, since the NeRF optimization did not converge, the erroneous attitude information calculated affected the attitude estimation results, thereby causing the error to increase, while in the later stage of training, the parameters of each part gradually converged. At the same time, the quantitative analysis results of the attitude estimation error of the method of this embodiment are shown in the figure. Fig.10 As shown, the rotation error and offset error under different attitude error initialization conditions are quantitatively shown. It can be seen that in the satellite and space station datasets, the initialization strategy with lower error can converge to a lower attitude error result, because the low error attitude can accelerate the convergence speed of NeRF 3D reconstruction, so that it can correctly guide the attitude estimation optimization process in the early stage of optimization. Due to its simple geometric structure, the satellite dataset has a larger rotation error and offset error than the space station dataset under the same initialization conditions. The outlier probability of the method in this embodiment is 0%, which proves the effectiveness and robustness of the method.

[0143] In some embodiments, a 3D reconstruction experiment is performed based on the target attitude parameters to obtain a 3D reconstruction experiment result. The 3D reconstruction experiment result may include a 3D reconstruction result of a satellite and a space station. Fig.11 As shown in the figure, the space station attitude optimization process is compared with Fig.12 As shown. The three-dimensional results obtained by directly performing NeRF optimization at a 3° attitude error have a large number of noise points and three-dimensional outliers, and the result will have a three-dimensional "expansion" effect. Through the joint attitude optimization and three-dimensional reconstruction method, the "expansion" effect can be eliminated and a complete and clear three-dimensional result of the space target can be reconstructed. It can be very obvious from the "3° attitude error optimization" result that the high-precision three-dimensional reconstruction result after the joint attitude optimization is obtained. After the random attitude error is optimized, the accurate three-dimensional structure of the space target can also be reconstructed completely and with high resolution. Intuitively, the performance is almost the same as that of the case with no attitude error. The joint attitude optimization and three-dimensional reconstruction method can achieve three-dimensional reconstruction results comparable to those with no attitude error. In addition, the two-dimensional new view comparison of the satellite joint attitude optimization process is shown in Figure 2. Fig.13 As shown in Figure 2, the new two-dimensional view of the space station joint attitude optimization process is compared with Fig.14 As shown. Under a 3° posture error, the image generated by the new view will be blurred. The root cause of this phenomenon is the "expansion" effect in the 3D reconstruction results. When the 3D structure "expands", the generated new view will also be correspondingly blurred, resulting in a decrease in image quality. In the process of generating new views after joint posture optimization, the "expansion" effect is effectively suppressed due to the significant improvement in the structural accuracy of the 3D reconstruction. The optimized posture information can more accurately reflect the true 3D structure, so that the real details of the object can be reproduced when generating new views, avoiding image blur caused by the "expansion" effect. At the same time, the performance evaluation results in the 3D reconstruction and new view generation tasks are shown in the figure below. Fig.15As shown. It can be seen that the results of joint posture optimization have achieved a high performance level in all indicators. Compared with the case with posture error, the optimized results have achieved a significant improvement in performance. Although these optimized results are slightly inferior to the ideal case without posture error in performance, they can achieve a performance level close to the ideal state. This result indirectly proves that the proposed method can still effectively complete the 3D reconstruction task in the presence of posture errors. When performing the task of new view generation, posture estimation and 3D reconstruction are required at the same time. When the posture and 3D structure tend to converge in the later stage of model training, the backscatter coefficient information of the spatial target gradually begins to converge. This leads to a slight decrease in the "PSNR" and "SSIM" performance indicators of the new view generation task compared to the performance with a 3° posture error, but the "LPIPS" indicator that focuses more on the overall semantics is superior. Among them, Acc represents accuracy, Comp represents completeness, Overall represents the overall, and F-score * Represents the harmonic mean of precision, F-score + The harmonic mean of integrity, PSNR stands for Peak Signal to Noise Ratio, SSIM stands for Structural Similarity Index Measure, and LPIPS stands for * Learned Perceptual Image Patch Similarity (LPIPS) + Learning to represent completeness-aware image patch similarity.

[0144] In some embodiments, in practical applications, space targets are divided into cooperative parties and non-cooperative parties. For non-cooperative parties, the corresponding sensor information cannot be obtained, which is very fatal to the sensor-based attitude estimation method. In addition, it is difficult for the non-cooperative party's prior information library to construct a large amount of prior information, and matching with the observation information requires a large amount of computing resources, and the cost of use is high. It is impossible to estimate the attitude information of the target without building an information library. In view of this, the method of this embodiment can be applied to both cooperative targets and non-cooperative targets, and the attitude information of space targets with unknown three-dimensional information structure and unknown attitude information can be estimated with high precision while performing three-dimensional reconstruction. This embodiment models the attitude estimation parameters and NeRF implicit neural network parameters, optimizes the model through the joint loss function, and derives the process of its back propagation to verify its differentiability, and optimizes the convergence of joint training by weighted correction of the position encoding according to the differential result, thereby improving the accuracy and applicability. In addition, this embodiment can accurately estimate the attitude of the target and perform three-dimensional reconstruction in multi-view ISAR images containing noise interference or unknown parameters, and construct a three-dimensional information database based on NeRF. This embodiment models the range-Doppler coordinate mapping of attitude transformation and the differentiable neural rendering of ISAR imaging, and proves the differentiability of the entire model back-propagation process through derivation. This embodiment implements the strategy of weighted correction of position encoding. The weighted function of position encoding can adjust the gradient of high-frequency information to ensure that the optimization of attitude parameters is synchronized with the optimization process of implicit neural network.

[0145] The beneficial effects of implementing the embodiments of the present invention include: the embodiments of the present invention first obtain an inverse synthetic aperture radar image of a space target, and construct a posture transformation coordinate according to the inverse synthetic aperture radar image and a scene value, then construct a position coding strategy according to the position coding frequency and an adjustable coding parameter, and then construct a rendering pixel value according to the posture transformation coordinate, the position coding strategy and the sight angle direction, and calculate the target posture parameter according to the rendering pixel value and the image pixel value, and finally perform three-dimensional reconstruction of the space target according to the target posture parameter to obtain a surface mesh model, so that the three-dimensional reconstruction of the space target can be realized in combination with posture estimation, thereby improving the accuracy and applicability and reducing the calculation cost.

[0146] like Fig.16 As shown, an embodiment of the present invention further provides a three-dimensional reconstruction device for a space target integrating posture estimation, comprising:

[0147] The first module 801 is used to obtain an inverse synthetic aperture radar image of a space target;

[0148] The second module 802 is used to construct attitude transformation coordinates according to the inverse synthetic aperture radar image and the scene value;

[0149] The third module 803 is used to construct a position coding strategy according to the position coding frequency and the adjustable coding parameters;

[0150] The fourth module 804 is used to construct a rendering pixel value according to the posture transformation coordinates, the position encoding strategy and the sight angle direction;

[0151] The fifth module 805 is used to calculate the target posture parameters according to the rendering pixel value and the image pixel value, and the target posture parameters include target rotation matrix parameters or target offset matrix parameters;

[0152] The sixth module 806 is used to perform three-dimensional reconstruction of the space target according to the target posture parameters to obtain a surface mesh model.

[0153] The contents of the above method embodiments are all applicable to the present device embodiments. The functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0154] like Fig.17 As shown, an embodiment of the present invention further provides a computer device, including:

[0155] at least one processor 901;

[0156] At least one memory 902, used to store at least one program;

[0157] When at least one program is executed by at least one processor, the at least one processor implements Figure 1 The method shown.

[0158] The contents of the above method embodiments are all applicable to the present device embodiments. The functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0159] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the rights of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by a person skilled in the art without departing from the scope and essence of the present invention should be within the scope of the rights of the present invention.

Claims

1. A method for three-dimensional reconstruction of space targets by integrating posture estimation, characterized in that: The following steps are involved: Acquire inverse synthetic aperture radar images of space targets; constructing attitude transformation coordinates according to the inverse synthetic aperture radar image and the scene value; Construct a position encoding strategy based on the position encoding frequency and adjustable encoding parameters; Constructing a rendering pixel value according to the posture transformation coordinates, the position encoding strategy and the sight angle direction; Calculating target posture parameters according to the rendered pixel values ​​and the image pixel values, wherein the target posture parameters include target rotation matrix parameters or target offset matrix parameters; According to the target posture parameters, the space target is three-dimensionally reconstructed to obtain a surface mesh model.

2. The method according to claim 1, characterized in that The step of constructing attitude transformation coordinates according to the inverse synthetic aperture radar image and the scene value includes: Initialize the rotation matrix and offset matrix; constructing a normal vector according to a range axis projection vector of the inverse synthetic aperture radar image and a Doppler axis projection vector of the inverse synthetic aperture radar image; Constructing the three-dimensional coordinates of the starting point according to the range axis projection vector, the Doppler axis projection vector and the point coordinates in the inverse synthetic aperture radar image; Constructing the spatial point coordinates in the normal vector according to the normal vector, the three-dimensional coordinates of the starting point and the scene value; The posture transformation coordinates are constructed according to the range axis projection vector, the Doppler axis projection vector, the three-dimensional coordinates of the starting point, the rotation matrix and the offset matrix.

3. The method according to claim 1, characterized in that: The position coding strategy is constructed according to the position coding frequency and the adjustable coding parameters, including: Divide the position coding frequency to obtain multiple coding frequencies; Constructing a position coding weighting function according to the coding frequency and the adjustable coding parameter; The position encoding strategy is constructed according to a plurality of the position encoding weighting functions.

4. The method according to claim 2, characterized in that: The step of constructing a rendering pixel value according to the posture transformation coordinates, the position encoding strategy and the sight angle direction includes: According to the posture transformation coordinates and the sight angle direction, encoding processing is performed through the position encoding strategy to obtain high and low frequency information; Inputting the high and low frequency information into an implicit neural network to obtain particle density and backscattering coefficient; Constructing a normal line according to the normal vector and the three-dimensional coordinates of the starting point; The rendering pixel value is constructed according to the particle density, the backscattering coefficient and the normal line.

5. The method according to claim 2, characterized in that: The step of calculating target posture parameters according to the rendered pixel values ​​and the image pixel values ​​comprises: Constructing a mean square error loss function according to the rendering pixel value, the image pixel value and a set of sample coordinate points; Constructing a high dynamic scattering coefficient loss function according to the rendering pixel value, the image pixel value, the sign function and the number of samples involved in loss calculation; Constructing a joint optimization loss function according to the mean square error loss function, the high dynamic scattering coefficient loss function and hyperparameters; Constructing an objective function according to the joint optimization loss function, rotation matrix parameters, offset matrix parameters and implicit neural network parameters; The rotation matrix and the offset matrix are updated by the objective function to obtain the target posture parameters.

6. The method according to claim 1, characterized in that The three-dimensional reconstruction of the space target is performed according to the target posture parameters to obtain a surface mesh model, including: Initialize 3D voxels; extracting a mask image from the inverse synthetic aperture radar image according to the target posture parameter; Back-projecting the plurality of mask images into an initial three-dimensional coordinate space to obtain a target three-dimensional coordinate space; According to the target three-dimensional coordinate space, conversion processing is performed through a preset surface rendering algorithm to obtain the surface mesh model.

7. The method according to claim 1, characterized in that The method further comprises: Performing a posture estimation experiment according to the target posture parameters to obtain a posture estimation experiment result; Alternatively, a three-dimensional reconstruction experiment is performed according to the target posture parameters to obtain a three-dimensional reconstruction experiment result.

8. The method according to claim 2, characterized in that: The step of constructing the posture transformation coordinates according to the range axis projection vector, the Doppler axis projection vector, the three-dimensional coordinates of the starting point, the rotation matrix and the offset matrix includes: According to the distance axis projection vector, the Doppler axis projection vector, the three-dimensional coordinates of the starting point, the rotation matrix and the offset matrix, the attitude transformation coordinates are constructed by the attitude transformation coordinate calculation formula, and the attitude transformation coordinate calculation formula is: In the formula, is the posture transformation coordinate, R is the rotation matrix, t is the offset matrix, I(u) is the three-dimensional coordinate of the starting point, d0 is the distance axis projection vector, v0 is the Doppler axis projection vector, and u is the scene value.

9. A three-dimensional reconstruction device for space targets integrating posture estimation, characterized in that: include: The first module is used to obtain an inverse synthetic aperture radar image of a space target; A second module is used to construct attitude transformation coordinates according to the inverse synthetic aperture radar image and scene value; The third module is used to construct a position encoding strategy based on the position encoding frequency and the adjustable encoding parameters; A fourth module is used to construct a rendering pixel value according to the posture transformation coordinates, the position encoding strategy and the sight angle direction; A fifth module is used to calculate target posture parameters according to the rendering pixel value and the image pixel value, wherein the target posture parameters include target rotation matrix parameters or target offset matrix parameters; The sixth module is used to perform three-dimensional reconstruction of the space target according to the target posture parameters to obtain a surface mesh model.

10. A computer device, characterized in that: include: at least one processor; at least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method according to any one of claims 1 to 8.

Citation Information

Cited By

  • Static three-dimensional light field content generation method and system

    CN119603437A

  • A static three-dimensional light field content generation method and system

    CN119603437B

  • Space target three-dimensional reconstruction method and system based on ISAR-visible light double-branch fusion neural radiation field

    CN120107457A