A method, system, device, medium and product for three-dimensional reconstruction of a space target

Abstract

The application discloses a three-dimensional reconstruction method, system, device, medium and product of a space target, and relates to the field of aerospace engineering.The method comprises the following steps: in the approaching section of a spacecraft, a two-dimensional image sequence of a space target is collected, and a two-dimensional Gaussian disc model with an uncertainty metric is trained; a plurality of groups of candidate orbits are generated based on spacecraft orbit dynamics, a single-view information gain is calculated by using an uncertainty two-dimensional Gaussian field, an overall orbit gain is calculated, and a candidate orbit with the maximum overall orbit gain is selected as an optimal observation orbit; a sampling image obtained by sampling the spacecraft along the optimal observation orbit is merged with the two-dimensional image sequence, and the two-dimensional Gaussian disc model with the uncertainty metric is trained again; then, discrete point filtering processing is performed, depth estimation and geometric fusion are performed based on the filtered two-dimensional Gaussian disc model, and a three-dimensional grid model of the space target is output. The application can realize high-precision reconstruction of a non-cooperative space target.

Description

Technical Field

[0001] This application relates to the field of aerospace engineering, and in particular to a method, system, device, medium, and product for three-dimensional reconstruction of space targets. Background Technology

[0002] In the field of aerospace engineering, missions such as asteroid exploration, on-orbit servicing, and space debris removal are of irreplaceable strategic significance for expanding the boundaries of human understanding of space, ensuring the safety of the near-Earth space environment, and improving the efficiency of space resource utilization. They have become core research directions in the global aerospace field. The successful execution of such missions relies heavily on accurate perception and interaction with space targets. However, due to the lack of pre-defined two-way communication protocols, achieving safe and efficient approach to unknown targets remains a key challenge restricting the improvement of mission autonomy. Spacecraft cannot directly obtain crucial information such as the target's attitude and structure and must rely on autonomous sensing systems to complete preliminary exploration tasks.

[0003] In the core of autonomous perception, target recognition, semantic segmentation, and position estimation technologies provide spacecraft with preliminary target localization and feature descriptions. 3D reconstruction technology, by processing acquired 2D image data, estimates the target's 3D geometry and spatial attitude, a prerequisite for subsequent operations such as high-precision navigation control and robotic arm grasping. Currently, 3D reconstruction of space targets faces several unique challenges: target surfaces often have sparse textures and weak features, and under single-source illumination, images contain large areas of shadow and highlight, making it difficult for traditional feature-matching-based multi-view geometric methods to extract sufficiently stable feature points. This often results in reconstruction results with issues such as holes, noise, or surface discontinuities. Furthermore, high-quality 3D reconstruction typically relies on a large number of observation images from different perspectives. Existing orbital observation schemes often use preset fixed trajectories, failing to dynamically adjust observation strategies based on the target's actual geometric characteristics and reconstruction status. This not only leads to long mission cycles and low efficiency but also may result in poor reconstruction quality in some areas due to insufficient viewpoint coverage, making it difficult to meet the demands of agile and intelligent on-orbit missions.

[0004] In recent years, differentiable rendering methods have attracted widespread attention due to their high-precision 3D reconstruction capabilities, strong robustness to sparse textures, and low inter-frame view overlap requirements, making them suitable for reconstruction tasks involving space targets. Early neural radiation field techniques required a large number of samples in the target space, and the high computational complexity of the fully connected network architecture made it difficult to meet the on-orbit computational constraints of airborne platforms. More recently, Gaussian sputtering methods have significantly improved reconstruction speed through ordered rasterization, achieving feasible on-orbit 3D reconstruction. Techniques that select subsequent observation perspectives based on real-time reconstruction performance are called active reconstruction. Although these terrestrial methods improve the efficiency of 3D reconstruction, they do not consider on-orbit motion and lighting constraints, making it difficult to achieve ideal results.

[0005] Therefore, in the face of the above problems, the complex environment of space, and the limitations of spacecraft orbits, developing an active 3D reconstruction method that can integrate on-orbit dynamic constraints and adapt to special space lighting and imaging conditions is of great significance for the autonomous mission execution of future intelligent spacecraft in non-cooperative target scenarios. Summary of the Invention

[0006] The purpose of this application is to provide a method, system, device, medium, and product for three-dimensional reconstruction of space targets, which can achieve high-precision reconstruction of non-cooperative space targets.

[0007] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for three-dimensional reconstruction of a space target, comprising: During the approach phase of the spacecraft, acquire a sequence of two-dimensional images of the space target; The two-dimensional Gaussian disk model with an introduced uncertainty metric is trained based on the two-dimensional image sequence, and the uncertain two-dimensional Gaussian field and preliminary three-dimensional geometric information are output. Multiple candidate orbits are generated based on spacecraft orbital dynamics, and the single-view information gain is calculated using the uncertain two-dimensional Gaussian field for the virtual observation sequence on each candidate orbit. Based on the target spatial range determined by the preliminary three-dimensional geometric information, the geometric coverage of each candidate orbit to the spatial target is calculated, and the orbit information value score is calculated based on the single-view information gain and the geometric coverage. The total orbital gain is calculated based on the orbital information value score, fuel score, and time efficiency score, and the candidate orbit with the largest total orbital gain is selected as the optimal observation orbit. The spacecraft is controlled to sample along the optimal observation orbit to acquire a sequence of sampled images. The sampled image sequence is merged with the two-dimensional image sequence, and the two-dimensional Gaussian disk model with introduced uncertainty measure is retrained based on the merged two-dimensional image sequence to obtain the retrained two-dimensional Gaussian disk model. Discrete-point filtering is performed on the retrained two-dimensional Gaussian disk model to obtain the filtered two-dimensional Gaussian disk model. Depth estimation and geometric fusion are performed based on the filtered two-dimensional Gaussian disk model to output a three-dimensional mesh model of the spatial target, thereby realizing the three-dimensional reconstruction of the spatial target.

[0008] Secondly, this application provides a three-dimensional reconstruction system for a space target, comprising: The image acquisition module is used to acquire a sequence of two-dimensional images of space targets during the approach phase of the spacecraft. The first training module is used to train a two-dimensional Gaussian disk model with an introduced uncertainty measure based on the two-dimensional image sequence, and outputs an uncertain two-dimensional Gaussian field and preliminary three-dimensional geometric information. The single-view information gain calculation module is used to generate multiple sets of candidate orbits based on spacecraft orbital dynamics, and to calculate the single-view information gain for each set of candidate orbits using the uncertain two-dimensional Gaussian field for the virtual observation sequence. The orbital information value scoring module is used to calculate the geometric coverage of each candidate orbit to the spatial target based on the target spatial range determined by the preliminary three-dimensional geometric information, and to calculate the orbital information value score based on the single-view information gain and the geometric coverage. The optimal observation orbit determination module is used to calculate the total orbit gain based on the orbit information value score, fuel score, and time efficiency score, and select the candidate orbit with the largest total orbit gain as the optimal observation orbit. The image sampling module is used to control the spacecraft to sample along the optimal observation orbit and acquire a sequence of sampled images; The second training module is used to merge the sampled image sequence with the two-dimensional image sequence, and retrain the two-dimensional Gaussian disk model with introduced uncertainty measure based on the merged two-dimensional image sequence to obtain the retrained two-dimensional Gaussian disk model. The discrete point filtering module is used to perform discrete point filtering on the retrained two-dimensional Gaussian disk model to obtain the filtered two-dimensional Gaussian disk model. The depth estimation and geometry fusion module is used to perform depth estimation and geometry fusion based on the filtered two-dimensional Gaussian disk model, and output a three-dimensional mesh model of the space target to realize the three-dimensional reconstruction of the space target.

[0009] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for three-dimensional reconstruction of spatial targets.

[0010] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for three-dimensional reconstruction of spatial targets.

[0011] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method for three-dimensional reconstruction of spatial targets.

[0012] According to the specific embodiments provided in this application, this application has the following technical effects: 1. Feasible path planning under on-orbit dynamic constraints has been achieved: This application generates multiple candidate orbits based on spacecraft orbital dynamics. The generated candidate orbits strictly follow the natural trajectory of relative motion, ensuring that the selected "optimal observation orbit" not only maximizes reconstruction benefits but also fully complies with spacecraft dynamic constraints, which can be achieved without consuming huge amounts of fuel for non-natural maneuvers. 2. Breakthrough in reconstruction bottleneck of extreme lighting environment with single main light source in space: This application adopts a two-dimensional Gaussian disk model that introduces uncertainty measurement. This model can accurately quantify and eliminate low confidence areas and background noise caused by extreme lighting, which significantly improves the reconstruction robustness and integrity under weak texture and high dynamic range lighting conditions. 3. Achieved multi-objective global optimization of reconstruction quality, fuel consumption and mission timeliness: Constructed a comprehensive evaluation system that integrates orbital information value score, fuel score and time efficiency score, automatically selected the observation strategy with the highest total orbital gain, and minimized on-orbit resource consumption to the greatest extent while ensuring high-precision reconstruction, meeting the stringent requirements of intelligent spacecraft for agile and low-cost missions. 4. A closed-loop reconstruction system adapted to non-cooperative targets was constructed: Through a progressive strategy of "pre-training guidance - active planning - fusion retraining - filtering output", the problem of space targets having no prior information and sparse features was effectively solved, and a high-fidelity, topologically correct 3D mesh model was output, providing a reliable geometric basis for on-orbit service missions. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A flowchart illustrating a three-dimensional reconstruction method for a spatial target provided in an embodiment of this application; Figure 2 A schematic diagram of the three-dimensional spatial coordinates of the spacecraft; Figure 3 A schematic diagram of the candidate orbits; Figure 4 A schematic diagram of four orbital configurations for spacecraft; Figure 5 A comparative diagram of a real model of an aerial target and a three-dimensional mesh model constructed using the method of this application; wherein, (a) is a three-dimensional mesh model constructed using the method of this application, and (b) is a real model of an aerial target; Figure 6A comparison graph of uncertainty, information gain, and rendering error for all targets in two candidate views. Detailed Implementation

[0015] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] In one exemplary embodiment, such as Figure 1 As shown, a three-dimensional reconstruction method for a spatial target is provided. This method is executed by a computer device, specifically by a computer device such as a terminal or a server alone, or by a terminal and a server together. In this embodiment, the method is described using a server as an example, and includes the following steps S1 to S9.

[0018] S1: During the approach phase of the spacecraft, acquire a sequence of two-dimensional images of the space target.

[0019] S2: Based on the two-dimensional image sequence, train the two-dimensional Gaussian disk model with introduced uncertainty measure, and output the uncertain two-dimensional Gaussian field and preliminary three-dimensional geometric information.

[0020] S3: Generate multiple candidate orbits based on spacecraft orbital dynamics, and calculate the single-view information gain using the uncertain two-dimensional Gaussian field for the virtual observation sequence on each candidate orbit.

[0021] S4: Based on the target spatial range determined by the preliminary three-dimensional geometric information, calculate the geometric coverage of each candidate orbit to the spatial target, and calculate the orbit information value score based on the single-view information gain and the geometric coverage.

[0022] S5: Calculate the total orbital gain based on the orbital information value score, fuel score, and time efficiency score, and select the candidate orbit with the largest total orbital gain as the optimal observation orbit.

[0023] S6: Control the spacecraft to sample along the optimal observation orbit and acquire a sequence of sampled images.

[0024] S7: Merge the sampled image sequence with the two-dimensional image sequence, and retrain the two-dimensional Gaussian disk model with introduced uncertainty measure based on the merged two-dimensional image sequence to obtain the retrained two-dimensional Gaussian disk model.

[0025] S8: Perform discrete point filtering on the retrained 2D Gaussian disk model to obtain the filtered 2D Gaussian disk model.

[0026] S9: Based on the filtered two-dimensional Gaussian disk model, depth estimation and geometric fusion are performed to output a three-dimensional mesh model of the spatial target, thereby realizing the three-dimensional reconstruction of the spatial target.

[0027] By implementing steps S1 to S9 above, high-precision reconstruction of non-cooperative space targets can be achieved.

[0028] In a specific embodiment, step S1 specifically includes: During the approach phase of the spacecraft as it moves from a distant distance toward a non-cooperative space target to a pre-defined observation point, the spacecraft's onboard camera sensors acquire a sequence of two-dimensional images of the target, which serves as initial training data for subsequent 3D reconstruction. During the acquisition process, auxiliary information such as the relative position, attitude, and lighting conditions between the spacecraft and the target is recorded, providing fundamental data support for subsequent uncertainty measurement and orbit planning.

[0029] In one specific embodiment, step S2 specifically includes: The two-dimensional image sequence acquired in step S1 is input into a two-dimensional Gaussian disk model that incorporates an uncertainty metric, thereby improving the traditional 2D Gaussian reconstruction method to train the model.

[0030] Specifically, in this model, the fixed color values ​​of traditional two-dimensional (2D) Gaussian units are improved to random variables that follow a normal distribution. ,in, Let be the color mean of the i-th Gaussian pixel. To quantify the color variance of the reconstruction uncertainty in this region, the model training process primarily minimizes the KL divergence loss function based on Bayes' theorem. This is achieved by assuming that the rendered pixel colors follow a normal distribution. , The average color value obtained from rendering. This represents the variance of the pixel uncertainty obtained from the rendering.

[0031] Construct the following KL divergence loss function: in, This represents the total number of pixels. The true color value of the m-th pixel in the two-dimensional image. The average color value obtained by rendering the m-th pixel. The pixel uncertainty variance is obtained by rendering the m-th pixel.

[0032] Based on the KL divergence loss function, the difference between the model distribution and the real data distribution is minimized to optimize the geometric parameters (position, rotation, scaling), opacity, and color variance of the Gaussian disk. The detailed process involves representing the target space as a set of parameters including position, rotation, scaling, opacity, and color variance. A 3D Gaussian primitive set is defined. Differentiable rendering technology is then used to project the 3D parameters onto a 2D pixel surface. Based on the rendering results and the KL divergence gradient from actual observations, the geometry and variance distribution of each primitive are iteratively optimized. Finally, the primitive parameters converge under multi-view observation constraints. After model training, a 2D Gaussian field reflecting the uncertainty of the reconstruction reliability of each region of the target is output, while the 3D geometric information of the target and the corresponding uncertainty index are initially extracted.

[0033] In a specific embodiment, step S3 generates multiple candidate orbits based on spacecraft orbital dynamics, specifically including: using the Clohessy-Wiltshire dynamic equations, combined with a set orbital period and period coefficient, to derive the analytical solution of the spacecraft in the relative coordinate system; mapping the analytical solution at each discrete time node to obtain the three-dimensional spatial coordinates of the spacecraft; constructing a camera rotation matrix based on a pointing vector; the pointing vector being the unit direction vector from the spacecraft to the center of the space target; spatiotemporally integrating the three-dimensional spatial coordinates at the same discrete time node with the camera rotation matrix to form a discretized virtual sampling trajectory; the discretized virtual sampling trajectory being the candidate orbit.

[0034] First, the Clohessy-Wiltshire (CW) dynamic equations are used, combined with the orbital period. and period coefficient The analytical solution for the spacecraft in the relative coordinate system is derived. At each discrete time node... The three-dimensional spatial coordinates of the spacecraft are obtained through analytical demapping. ,like Figure 2 As shown. Subsequently, based on the observation mission requirement that the spacecraft camera always points towards the space target, the unit direction vector from the spacecraft to the center of the space target is defined as the pointing vector, and the corresponding camera rotation matrix is ​​constructed accordingly. By integrating the three-dimensional spatial coordinates at each discrete time node with the camera rotation matrix, a set of discretized virtual sampling trajectories containing position and attitude sequences is finally formed, which is defined as candidate trajectories. The candidate trajectories are teardrop-shaped trajectories.

[0035] In a specific embodiment, step S3, for each group of candidate orbits' virtual observation sequences, calculates the single-view information gain using the uncertain two-dimensional Gaussian field. Specifically, this includes: defining the color variance of each Gaussian element in the uncertain two-dimensional Gaussian field as the prior variance of Bayesian inference; for each group of candidate orbits' virtual observation sequences, projecting the observation rays into the uncertain two-dimensional Gaussian field using differentiable rendering technology to generate prediction data; performing a Bayesian update on the prior variance based on the prediction data to obtain the updated posterior variance; and determining the single-view information gain based on the difference between the prior variance and the posterior variance.

[0036] The uncertain two-dimensional Gaussian field in step S2 is defined as a Bayesian prior probability field in three-dimensional space. This field explicitly calibrates the reconstruction confidence of each region on the surface of the space target through spatial variance distribution. The color variance of each Gaussian element in the prior probability field is extracted as the prior variance. For each set of virtual observation sequences of candidate orbits, differentiable rendering technology is used to project the observation rays onto the two-dimensional Gaussian field (i.e., the prior probability field) to generate prediction data. Then, Bayesian updates are performed based on the prediction data to obtain the posterior variance. The sum of the differences between the prior variance and the posterior variance is used as the single-view information gain to measure the contribution of the view to reducing reconstruction uncertainty.

[0037] The detailed calculation process involves accumulating virtual observation sequences on candidate orbits. All observed rays The reduction in the variance of the covered Gaussian elements is: in, For the Gausky element through which the ray passes, and These are the Gaussian variances before and after the Bayesian update, i.e., the prior variance and the posterior variance, respectively.

[0038] In one specific embodiment, step S4 specifically includes: Calculate the geometric coverage of each candidate orbit to the space target, i.e., the percentage of arc length covered by the candidate orbit on a unit sphere centered on the space target; and obtain the orbit information value score by fusing single-view information gain and geometric coverage through a weighted allocation method. The track information value score reflects the contribution to reconstruction quality.

[0039] in, Virtual observation sequence Single-view information gain, For geometric coverage, This indicates a normalization operation. The weighting coefficients are the information gain of a single view. is the weighting coefficient for geometric coverage.

[0040] In one specific embodiment, step S5 specifically includes: Based on step S4, and considering the spacecraft's flight conditions, fuel consumption constraints and observation time gain constraints are introduced according to orbital mechanics formulas. That is, the pulse size required for orbital transfer is calculated using the CW dynamics equations. And establish fuel rating = Time efficiency rating = .

[0041] Specifically, the CW dynamic equations are used to describe the motion state transition process of the spacecraft relative to the target spacecraft, and the state transition matrix is ​​as follows: in, and The initial relative position and velocity before the maneuver. and Time elapsed after maneuver Terminal status, This is the CW state transition matrix based on the target orbital angular velocity. The matrix consists of four... The block matrix is ​​composed of the following physical meanings: Describe the impact of the initial position on the terminal position. Describe the effect of initial velocity on terminal position. Describe the effect of the initial position on the terminal velocity. Describe the effect of initial velocity on terminal velocity. If the spacecraft is required to reach a specified velocity within a given time... Reach the given desired position within. By extracting the position transfer relationship and performing inverse calculation, the magnitude of the velocity pulse required for the orbital transfer can be obtained: Meanwhile, based on the time scaling factor set for the observation task and the spacecraft's basic flight cycle The total observation time for completing the current flyby mission was calculated. : Fuel consumption calculated above and total observation time .

[0042] Set reconstruction quality weights based on task requirements. Fuel weight , Finally, the total orbital gain of each candidate orbital is calculated. : Finally, by comparing all candidate orbitals The candidate orbit with the largest total orbital gain is selected as the optimal observation orbit.

[0043] In one specific embodiment, step S7 specifically includes: The sampled image sequence acquired by the spacecraft along the optimal observation trajectory is merged with the two-dimensional image sequence from step S1 and used as multi-view training data. This data is then re-input into the two-dimensional Gaussian disk model, which incorporates an uncertainty metric, and trained again to supplement the geometric information of the unobserved areas of the space target, thereby improving the overall reconstruction accuracy of the space target.

[0044] In one specific embodiment, step S8 specifically includes: To address redundant points in the model caused by complex lighting conditions in the space environment (such as strong reflections and deep shadows), a discrete point filtering method is adopted. Specifically, this involves scale-constrained filtering, which statistically analyzes the scaling vectors of all 2D Gaussian disks. Remove spindles whose dimensions exceed a preset threshold. A diverging disk is used to remove oversized false faces used to fill the background; distance and density filtering, based on the spatial distribution of the point cloud, removes objects whose distance from the centroid of the neighborhood exceeds a threshold. Isolated Gaussian points; color threshold filtering to identify the mean of color attributes. Gaussian disks that are close to the spatial background (extremely low brightness values) are removed; uncertainty pruning is performed, using the variance map obtained from training, to remove predictions with variances exceeding a threshold. The region is defined as unreliable geometry, and the Gaussian volume at the corresponding location is removed to ensure the determinism of the reconstruction result.

[0045] In one specific embodiment, step S9 specifically includes: Depth estimation is performed based on the filtered two-dimensional Gaussian disk model. A 3D mesh structure of the space target is generated through geometric fusion. After mesh smoothing, a high-precision 3D mesh model of the space target is output, realizing high-precision 3D reconstruction of the space target.

[0046] To verify the effectiveness and generalization ability of the proposed method, Blender simulation and local anechoic chamber experiments were conducted. For the four simulation models "CloudSat", "Epoxi", "Kepler", and "Sentinel", multi-view images were generated using Blender. Simultaneously, an anechoic chamber was constructed to simulate the space environment for data acquisition, and a laser scanner with an accuracy of 0.02 mm was used to acquire their 3D models. The resolution of both the simulated and acquired images was 1920×1080. The orbital altitude of the space target was set to 1000 km, corresponding to an orbital period of 6307 s. In the experiment, four orbital configurations were defined for the spacecraft, denoted as period coefficients. Corresponding orbital period During the orbital phase, the target and spacecraft exhibit the same spin motion, with a spin period equal to the orbital period of 6307 seconds. No spin occurs during the approach phase. The spacecraft's orbital motion is generally a teardrop-shaped trajectory, such as... Figure 3 As shown, the experiment designs four orbital configurations for the spacecraft, such as... Figure 4 As shown.

[0047] Spacecraft observation of a target typically involves two phases: the approach phase and the orbiting phase. 3D reconstruction of the target is initially performed using images from the approach phase. During this phase, the target is observed from only one direction, which presents a significant challenge for 3D reconstruction. The results are as follows... Figure 5 As shown in (a) and (b), by introducing uncertainty into the Gaussian disk and using discrete-point filters for geometric optimization, the method in this application suppresses mesh redundancy caused by the black background. This facilitates the separation of the target for uncertainty assessment under any candidate view. The uncertainty estimation effect is as follows: Figure 6 As shown, the uncertainty estimation of the method in this application is basically consistent with the actual information gain and the high-value region of the rendering error under the two candidate views, which shows the rationality of the uncertainty estimation of the method in this application.

[0048] Numerically, the average F1 score of the method in this application is 47.69% higher than that of the traditional 2DGS, while the reconstruction time is only 8.3% longer than that of 2DGS.

[0049] In selecting the optimal observation orbit, the method in this application takes into account the fuel consumption and flight time issues that exist in actual fly-around observation missions. For the four orbits... The three-dimensional reconstruction, fuel consumption, and flyby time fraction of each candidate orbit are calculated separately. Since the weight of each component should be determined according to mission-specific requirements, this application uses the target "Cloudsat" as an example to illustrate the selection of orbital configurations under different weighting schemes. The optimal observation orbit varies with the weighting scheme; for example, when fuel efficiency is emphasized, it becomes... When prioritizing 3D reconstruction, .

[0050] The optimal observation orbit selected above ( The image is added to the nearest segment image, and the merged image is used to train the model. The proposed method is compared with 2DGS, NeuS, and Colmap. Experimental results show that 2DGS cannot distinguish between the target and the background and has significant redundancy; NeuS, due to its sparse texture and similar color to the background, fails to reconstruct target details; Colmap, based on geometry matching, generates relatively coarse geometry for such sparsely textured targets. The proposed method, by introducing uncertainty into a Gaussian disk and applying discrete point filtering, produces better results. The F-1 score is also compared with other methods, achieving a 63% improvement compared to traditional 2DGS. In terms of computation time, the average time of the proposed method is 11.2 minutes, while 2DGS is 10 minutes, an improvement of 11.2%. Notably, NeuS has an average computation time of 6.5 hours, and Colmap is 1 hour, both significantly higher than the Gaussian reconstruction method used in this application.

[0051] Based on the same inventive concept, this application also provides a system for implementing the three-dimensional reconstruction method of the space target involved above. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more embodiments of the three-dimensional reconstruction system of space targets provided below can be found in the limitations of the three-dimensional reconstruction method of space targets above, and will not be repeated here.

[0052] In one exemplary embodiment, a three-dimensional reconstruction system for a space target is provided, including the following modules.

[0053] The image acquisition module is used to acquire a sequence of two-dimensional images of space targets during the approach phase of the spacecraft.

[0054] The first training module is used to train a two-dimensional Gaussian disk model with an introduced uncertainty metric based on the two-dimensional image sequence, and outputs an uncertain two-dimensional Gaussian field and preliminary three-dimensional geometric information.

[0055] The single-view information gain calculation module is used to generate multiple sets of candidate orbits based on spacecraft orbital dynamics, and to calculate the single-view information gain for each set of candidate orbits using the uncertain two-dimensional Gaussian field for the virtual observation sequence.

[0056] The orbital information value scoring module is used to calculate the geometric coverage of each candidate orbit to the spatial target based on the target spatial range determined by the preliminary three-dimensional geometric information, and to calculate the orbital information value score based on the single-view information gain and the geometric coverage.

[0057] The optimal observation orbit determination module is used to calculate the total orbit gain based on the orbit information value score, fuel score, and time efficiency score, and select the candidate orbit with the largest total orbit gain as the optimal observation orbit.

[0058] The image sampling module is used to control the spacecraft to sample along the optimal observation orbit and acquire a sequence of sampled images.

[0059] The second training module is used to merge the sampled image sequence with the two-dimensional image sequence, and retrain the two-dimensional Gaussian disk model with introduced uncertainty measure based on the merged two-dimensional image sequence to obtain the retrained two-dimensional Gaussian disk model.

[0060] The discrete point filtering module is used to perform discrete point filtering on the retrained two-dimensional Gaussian disk model to obtain the filtered two-dimensional Gaussian disk model.

[0061] The depth estimation and geometry fusion module is used to perform depth estimation and geometry fusion based on the filtered two-dimensional Gaussian disk model, and output a three-dimensional mesh model of the space target to realize the three-dimensional reconstruction of the space target.

[0062] In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments. The computer device may be a server or a terminal. The computer device includes a processor, a memory, an input / output interface (I / O), and a communication interface. The processor, memory, and I / O are connected via a system bus, and the communication interface is connected to the system bus via the I / O interface. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device stores data to be processed. The I / O interface of the computer device is used for exchanging information between the processor and external devices. The communication interface of the computer device is used for communicating with an external terminal via a network connection. When the computer program is executed by the processor, it implements the steps in the above-described method embodiments.

[0063] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0064] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0065] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0066] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0067] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0068] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0069] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for three-dimensional reconstruction of a spatial target, characterized in that, include: During the approach phase of the spacecraft, acquire a sequence of two-dimensional images of the space target; The two-dimensional Gaussian disk model with an introduced uncertainty metric is trained based on the two-dimensional image sequence, and the uncertain two-dimensional Gaussian field and preliminary three-dimensional geometric information are output. Multiple candidate orbits are generated based on spacecraft orbital dynamics, and the single-view information gain is calculated using the uncertain two-dimensional Gaussian field for the virtual observation sequence on each candidate orbit. Based on the target spatial range determined by the preliminary three-dimensional geometric information, the geometric coverage of each candidate orbit to the spatial target is calculated, and the orbit information value score is calculated based on the single-view information gain and the geometric coverage. The total orbital gain is calculated based on the orbital information value score, fuel score, and time efficiency score, and the candidate orbit with the largest total orbital gain is selected as the optimal observation orbit. The spacecraft is controlled to sample along the optimal observation orbit to acquire a sequence of sampled images. The sampled image sequence is merged with the two-dimensional image sequence, and the two-dimensional Gaussian disk model with introduced uncertainty measure is retrained based on the merged two-dimensional image sequence to obtain the retrained two-dimensional Gaussian disk model. Discrete-point filtering is performed on the retrained two-dimensional Gaussian disk model to obtain the filtered two-dimensional Gaussian disk model. Depth estimation and geometric fusion are performed based on the filtered two-dimensional Gaussian disk model to output a three-dimensional mesh model of the spatial target, thereby realizing the three-dimensional reconstruction of the spatial target.

2. The three-dimensional reconstruction method for a space target according to claim 1, characterized in that, In the two-dimensional Gaussian disk model that introduces uncertainty measurement, the fixed color values ​​of the traditional two-dimensional Gaussian elements are improved to random variables that follow a normal distribution.

3. The three-dimensional reconstruction method for a spatial target according to claim 1, characterized in that, The loss function of the two-dimensional Gaussian disk model that introduces uncertainty measurement during training is the KL divergence loss function constructed based on Bayes' theorem.

4. The three-dimensional reconstruction method for a space target according to claim 1, characterized in that, Multiple candidate orbits are generated based on spacecraft orbital dynamics, specifically including: Using the Clohessy-Wiltshire dynamic equations, combined with the set orbital period and period coefficient, the analytical solution of the navigation spacecraft in the relative coordinate system is derived. The analytical solution is mapped at each discrete time point to obtain the three-dimensional spatial coordinates of the spacecraft; The camera rotation matrix is ​​constructed based on the pointing vector; the pointing vector is the unit direction vector from the spacecraft to the center of the space target. The three-dimensional spatial coordinates at the same discrete time node are spatiotemporally integrated with the camera rotation matrix to form a discretized virtual sampling trajectory; the discretized virtual sampling trajectory is a candidate trajectory.

5. The three-dimensional reconstruction method for a space target according to claim 1, characterized in that, For each set of candidate orbits, the single-view information gain is calculated using the uncertain two-dimensional Gaussian field, specifically including: The color variance of each Gaussian element in the uncertain two-dimensional Gaussian field is defined as the prior variance of Bayesian inference; For each set of candidate orbits, the observation rays are projected onto the uncertain two-dimensional Gaussian field using differentiable rendering technology to generate prediction data. Based on the predicted data, the prior variance is updated using Bayesian method to obtain the updated posterior variance. The information gain of a single view is determined based on the difference between the prior variance and the posterior variance.

6. The three-dimensional reconstruction method for a space target according to claim 1, characterized in that, The formula for calculating the total orbital gain is as follows: in, For total orbital gain, To score the value of track information, For fuel rating, Rate time efficiency. To reconstruct quality weights, For fuel weight, As time weight, Virtual observation sequence Single-view information gain, For geometric coverage, This indicates a normalization operation. The weighting coefficients are the information gain of a single view. is the weighting coefficient for geometric coverage.

7. A three-dimensional reconstruction system for a spatial target, characterized in that, include: The image acquisition module is used to acquire a sequence of two-dimensional images of space targets during the approach phase of the spacecraft. The first training module is used to train a two-dimensional Gaussian disk model with an introduced uncertainty measure based on the two-dimensional image sequence, and outputs an uncertain two-dimensional Gaussian field and preliminary three-dimensional geometric information. The single-view information gain calculation module is used to generate multiple sets of candidate orbits based on spacecraft orbital dynamics, and to calculate the single-view information gain for each set of candidate orbits using the uncertain two-dimensional Gaussian field for the virtual observation sequence. The orbital information value scoring module is used to calculate the geometric coverage of each candidate orbit to the spatial target based on the target spatial range determined by the preliminary three-dimensional geometric information, and to calculate the orbital information value score based on the single-view information gain and the geometric coverage. The optimal observation orbit determination module is used to calculate the total orbit gain based on the orbit information value score, fuel score, and time efficiency score, and select the candidate orbit with the largest total orbit gain as the optimal observation orbit. The image sampling module is used to control the spacecraft to sample along the optimal observation orbit and acquire a sequence of sampled images; The second training module is used to merge the sampled image sequence with the two-dimensional image sequence, and retrain the two-dimensional Gaussian disk model with introduced uncertainty measure based on the merged two-dimensional image sequence to obtain the retrained two-dimensional Gaussian disk model. The discrete point filtering module is used to perform discrete point filtering on the retrained two-dimensional Gaussian disk model to obtain the filtered two-dimensional Gaussian disk model. The depth estimation and geometry fusion module is used to perform depth estimation and geometry fusion based on the filtered two-dimensional Gaussian disk model, and output a three-dimensional mesh model of the space target to realize the three-dimensional reconstruction of the space target.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement a three-dimensional reconstruction method for a spatial target according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the three-dimensional reconstruction method of the space target as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the three-dimensional reconstruction method of the space target as described in any one of claims 1-6.

Images