A parachute three-dimensional reconstruction and parameter identification method

Abstract

This invention relates to a method for 3D reconstruction and parameter identification of parachutes. It generates a 3D physical skeleton of the parachute through fluid-structure interaction numerical simulation, constructs a virtual optical imaging link including a camera projection model based on physically based rendering technology, and generates a synthetic image and a depth ground truth map spatiotemporally aligned with the physical skeleton, forming a high-fidelity synthetic dataset. Bidirectional domain adaptation processing is performed on both the synthetic image and the real airdrop image to eliminate differences in their feature distributions. A depth-sensing network is trained using the processed synthetic image and depth ground truth map, and depth maps are obtained through inference on the processed real image. A relative 3D point cloud is reconstructed based on the inverse projection relationship of the camera projection model. The time-varying scale factor is solved using known geometric invariants as physical constraints, mapping the relative point cloud to absolute physical space and calculating aerodynamic parameters. This invention solves the problems of data scarcity and domain discrepancies in parachute measurement, achieving high-precision 3D reconstruction and parameter identification under monocular video.

Description

Technical Field

[0001] This invention relates to the field of spacecraft parachute measurement technology, specifically to a method for three-dimensional reconstruction and parameter identification of parachutes. Background Technology

[0002] In deep space exploration and manned spaceflight missions, the inflation, deployment, and stabilization of flexible aerodynamic deceleration devices such as parachutes are accompanied by severe nonlinear geometric deformation and fluid-structure interaction (FSI) effects. Accurately acquiring the dynamic three-dimensional aerodynamic shape during this process is crucial for evaluating aerodynamic drag characteristics, validating fluid-structure interaction simulation algorithms, and ensuring spacecraft landing safety.

[0003] In existing measurement technologies, traditional contact sensor methods often interfere with the flow field distribution of flexible canopies, leading to distorted measurement data. Non-contact optical measurement methods are primarily used, but they still face the following significant challenges in practical applications: (1) Limitations of traditional photogrammetry techniques: Traditional monocular or multi-view visual measurement methods usually rely on strict prior assumptions (such as assuming the parachute canopy is an axisymmetric structure or a rigid body) or are highly dependent on the matching accuracy of feature points. However, in real airdrop tests, asymmetric breathing motion, local collapse, or high-frequency wrinkles often occur on the surface of the parachute, causing the theoretical model to fail; at the same time, feature point matching in weak texture areas is difficult, which can easily cause holes or mismatches in the 3D reconstruction results, making it impossible to obtain the complete topological structure.

[0004] (2) The "Data Scarcity" Challenge of Data-Driven Methods: Although monocular depth estimation (MDE) technology based on deep learning has made significant progress in general fields such as autonomous driving and indoor scene reconstruction, its application in parachute aerodynamic measurement is still in the exploratory stage. The fundamental reason is that the performance of this type of method is highly dependent on the quality and scale of the training data. In open scenarios such as parachute airdrops, it is impossible to obtain ground truth depth values ​​as training labels using LiDAR or similar methods. Due to the lack of real training samples with accurate depth annotations, existing deep learning models struggle to learn the complex topological features unique to flexible fabrics.

[0005] In summary, there is an urgent need to provide a method for 3D reconstruction and parameter identification of parachutes to solve the problems mentioned above, such as the reliance on ideal assumptions in traditional photogrammetry and the difficulty in obtaining the true depth of airdrop tests. Summary of the Invention

[0006] The purpose of this invention is to provide a method for three-dimensional reconstruction and parameter identification of parachutes, so as to solve the problems mentioned above, such as the reliance on ideal assumptions in traditional photogrammetry and the difficulty in obtaining the true value of the depth of airdrop tests.

[0007] The above objective is achieved through the following technical solution: a method for three-dimensional reconstruction and parameter identification of a parachute, comprising the following steps: S1, Constructing the physical skeleton: Based on fluid-structure interaction numerical simulation, a three-dimensional physical skeleton of the parachute is generated to dynamically deform under dynamic flow field, and a sequence of parachute structure node coordinates that conforms to aerodynamic laws is obtained. S2, Generate a high-fidelity synthetic dataset: Construct a virtual optical imaging link containing a camera projection model based on physical rendering technology, generate a synthetic image and a corresponding depth ground truth map that are spatiotemporally aligned with the three-dimensional physical skeleton, so as to construct a high-fidelity synthetic dataset. S3, Bidirectional Domain Adaptation Processing: Feature alignment processing is performed on the synthetic images and real parachute airdrop images in the high-fidelity synthetic dataset to eliminate the feature distribution differences between the synthetic domain and the real domain, thereby achieving feature space alignment between the two and obtaining the processed synthetic image set and the processed real image set. S4, Deep Perception Model Training and Inference: Using the processed synthetic image set and its corresponding ground truth depth map, train a deep learning-based deep perception network; input the processed real image set into the trained deep perception network, and output the corresponding depth map; S5, Relative 3D point cloud reconstruction: Based on the inverse projection relationship of the camera projection model, the depth map is converted into a relative 3D point cloud with the camera coordinate system as the reference. S6, Physical Scale Recovery and Parameter Identification: Solve for the time-varying scale factor using known geometric invariants in the parachute structure as physical constraints, map the relative three-dimensional point cloud to the absolute physical space, and calculate the aerodynamic parameters of the parachute based on the mapped absolute physical space coordinates.

[0008] This invention follows the technical route of "physically driven data generation → domain adaptation to bridge differences → deep learning-based perceptual inference → physical constraint to restore scale," and in its specific implementation: First, a physical framework is constructed. Using fluid-structure interaction numerical simulation technology, the entire process of parachute inflation, deployment, and stabilization under the action of real airflow is simulated, generating a dynamic three-dimensional structural sequence that conforms to the laws of aerodynamic evolution, and obtaining nodal coordinate data that characterizes the actual aerodynamic shape of the parachute.

[0009] Secondly, a high-fidelity synthetic dataset is generated. Based on physically based rendering technology, a virtual optical imaging link containing a camera projection model is constructed. The aforementioned three-dimensional physical skeleton is projected onto a two-dimensional image plane, and a synthetic image spatiotemporally aligned with the physical skeleton and its corresponding pixel-level depth ground truth map are generated simultaneously, forming a high-fidelity synthetic dataset.

[0010] Subsequently, bidirectional domain adaptation processing is performed. Pre-defined image processing algorithms are applied to both the synthetic image and the real airdrop image to bring their feature distributions closer together, eliminating the visual differences between the synthetic and real domains and obtaining a processed image set with aligned feature spaces.

[0011] Next, the depth perception model is trained and inference is performed. Using the processed synthetic image and its ground truth depth map, a deep learning-based depth perception network is trained, enabling it to infer three-dimensional depth information from two-dimensional images. The processed real airdrop image is then input into the trained network, which outputs the corresponding depth map.

[0012] Then, the relative 3D point cloud is reconstructed. Based on the inverse projection relationship of the camera projection model, each pixel in the depth map is inversely projected into 3D space to generate a relative 3D point cloud with the camera coordinate system as the reference.

[0013] Finally, the physical scale is restored and parameters are identified. Using the known geometric features of the parachute structure (such as the top hole) as physical constraints, the scale factor that changes over time is solved, and the relative point cloud is mapped to the absolute physical space to obtain a three-dimensional point cloud with real scale. Based on this, key aerodynamic parameters such as the parachute's projected area and aerodynamic volume are calculated.

[0014] This invention establishes a dual-drive data generation and perception framework integrating fluid-structure interaction (FSI) numerical simulation and physically based rendering (PBR) technologies. Breaking away from the limitations of traditional photogrammetry methods in terms of prior assumptions, its basic idea is to no longer obtain true depth values ​​from actual measurement experiments, but rather start from the physical source. It utilizes FSI numerical simulation technology to construct a dynamic physical skeleton conforming to aerodynamic laws, and constructs a virtual optical imaging link through physically based rendering (PBR) technology, generating a high-fidelity synthetic dataset containing true depth values, thus providing accurate physical priors for the depth perception model. Secondly, a bidirectional domain adaptation strategy is introduced as a bridge for feature alignment. Addressing the problem that synthetic data cannot fully reproduce the degradation characteristics of the real environment, environmental perturbations are applied to the synthetic domain to simulate the degradation features of the real environment. The simulation method simulates real-world imaging defects and enhances the signal in the real domain to filter out sensor noise, eliminating the inconsistency between the two in the feature space distribution, thereby solving the domain difference problem between simulation data and real data. Furthermore, it utilizes a deep perception network with a global attention mechanism to learn the nonlinear mapping relationship from two-dimensional image features to three-dimensional spatial depth, achieving high-precision reconstruction of the parachute relative point cloud in real monocular video. Finally, it solves the time-varying scale factor with known structural parameters as physical constraints, mapping the reconstructed relative geometry back to absolute physical space, thereby achieving quantitative identification of key aerodynamic parameters such as parachute projected area and aerodynamic volume.

[0015] A further technical solution is that, in step S1, the parachute inflation, deployment, and stabilization process is simulated using an arbitrary Lagrange-Euler method, generating a dynamic three-dimensional mesh sequence characterizing the parachute's nonlinear geometric deformation, which serves as the three-dimensional physical skeleton sequence; in step S2, based on physical rendering technology, the diffuse reflection and scattering of light on the fabric surface are simulated by solving the ray tracing rendering equation, generating an RGB image with realistic lighting attributes as the composite image; simultaneously, a virtual camera pinhole model consistent with the intrinsic parameters of a real camera is established, and while generating the RGB image, the depth value of each pixel is calculated using ray tracing, generating a depth ground truth map pixel-level aligned with the RGB image.

[0016] This effectively solves the problems of geometric distortion caused by the lack of physical mechanism support in existing synthetic data, and the lack of optical features due to the simplicity of the rendering model.

[0017] A further technical solution involves first establishing a structural domain model and a fluid domain model for the parachute. The structural domain model is constructed based on the parachute's geometric design parameters, and the canopy is discretized into a finite element mesh. The fluid domain model includes the flow field space surrounding the parachute, with mesh refinement applied to the canopy region. Step S1, employing an arbitrary Lagrange-Euler method for fluid-structure interaction numerical simulation, includes: introducing a mesh transport velocity-corrected Navier-Stokes equation to the fluid domain to solve the mass and momentum conservation governing equations; using a Lagrange scheme to describe the nonlinear response of the parachute fabric material in the structural domain to solve the structural governing equations; and using a penalty function method to handle the coupling interface between the fluid and structural domains. By tracking the relative permeation displacement between fluid and structural elements and applying virtual coupling forces, the breathing, bulging, and wrinkling deformation of the parachute under unsteady flow fields are simulated, generating a dynamic three-dimensional physical skeleton containing the coordinate sequence of the parachute's structural nodes. This accurately reproduces the physical characteristics of the parachute's dynamic deformation throughout the entire process, providing the three-dimensional physical skeleton with a rigorous fluid-structure interaction mechanism.

[0018] A further technical solution is that the feature alignment processing in step S3 is a bidirectional domain adaptation operation based on digital signal processing, specifically including: performing grayscale processing, random Gamma nonlinear transformation, Gaussian blur processing and noise injection processing on the synthesized image to degrade the synthesized domain signal; and performing bilateral filtering, histogram adaptive stretching and image segmentation background replacement processing on the real airdrop image in sequence to enhance the real domain signal.

[0019] In this way, without relying on complex adversarial networks, the synthetic data and real data can achieve convergence in distribution at the level of visual texture and signal-to-noise ratio features through the aforementioned defined digital signal processing steps. This solves the "Sim-to-Real" generalization problem, which is difficult to reduce the feature distribution differences between synthetic data and real data, causing the deep perception model to fail in cross-domain reasoning.

[0020] A further technical solution involves using Gaussian blurring and noise injection in the synthetic domain signal degradation operation to simulate dynamic blurring and scene degradation in the real imaging process; the histogram adaptive stretching in the real domain signal enhancement operation is replaced with histogram adaptive stretching in the HSV color space; and the image segmentation background is replaced by retaining the umbrella area and replacing the background with the background area of ​​the synthetic image. This more accurately simulates the characteristics of the real imaging process and further reduces the difference in feature distribution between the synthetic and real domains.

[0021] A further technical solution is that, in step S4, the depth perception network adopts a VisionTransformer-based architecture, which uses its multi-head self-attention mechanism to capture long-distance dependencies between pixels at different locations in the image, so as to infer continuous and consistent three-dimensional spatial depth information of the parachute surface from the two-dimensional image features.

[0022] A further technical solution is that the multi-head self-attention mechanism realizes global feature aggregation by calculating the correlation weight of any two pixels in the input feature map, decoupling the physical occlusion edge of the parachute from the surface texture wrinkles, and inferring the continuous and consistent spatial contour of the parachute canopy in the global range.

[0023] A further technical solution is that, in step S5, based on the inverse projection relationship of the pinhole camera projection model, the pixel coordinates and corresponding depth values ​​of each pixel in the depth map are converted into three-dimensional spatial coordinates in the camera coordinate system, and a continuous and complete parachute relative three-dimensional point cloud is obtained by reconstructing pixel by pixel.

[0024] A further technical solution is that, in step S6, the known geometric invariants are one or more reference geometric features in the parachute structure whose dimensions are known or remain unchanged during the parachute opening process; the specific steps for mapping the relative three-dimensional point cloud to the absolute physical space are: multiplying the camera coordinates of the relative three-dimensional point cloud with the time-varying scale factor at the corresponding moment to obtain the three-dimensional coordinates in the absolute physical space. In this way, the mapping method solves the inherent scale ambiguity problem in monocular vision reconstruction, achieving accurate conversion from dimensionless relative point cloud to absolute physical space. The reference geometric features include the top opening region of the parachute; calculating the aerodynamic parameters includes calculating the instantaneous projected area and aerodynamic volume of the parachute.

[0025] A further technical solution is that the specific steps for solving the time-varying scale factor in step S6 are as follows: select the parachute top hole region as the known geometric invariant and use it as a reference quantity; calculate the relative perimeter of the top hole region corresponding to the reconstructed point cloud at any time; define the instantaneous scale factor as the ratio of the reference quantity to the relative perimeter; and decompose the instantaneous scale factor into the product of the reference mapping ratio and the dynamic compensation term; the dynamic compensation term adaptively corrects the non-physical fluctuations generated by the deep sensing network; and calculate the key aerodynamic parameters of the parachute based on the mapped absolute physical space coordinates using the infinitesimal integral method.

[0026] Compared with existing technologies, the implementation of the technical solution of this invention has the following technical advantages: (1) This study solved the problem of scarce data and lack of true depth values ​​in the field of parachute measurement. Addressing the objective difficulty of obtaining true depth values ​​in real airdrop experiments, a dual-drive data generation method combining "fluid-structure interaction numerical simulation + physical rendering" was constructed. The dynamic mesh generated using the ALE algorithm ensured that the training data possessed a "physical skeleton" conforming to aerodynamic laws, while the rendered images generated using PBR technology provided the data with a realistic "visual appearance." This method enabled the acquisition of a large number of high-fidelity training samples with precise depth annotations at low cost, laying the foundation for the application of data-driven algorithms in the field of parachute measurement.

[0027] (2) Significantly improved the model's generalization reasoning ability in cross-domain scenarios of "synthetic-real". To address the domain discrepancy issue caused by differences in feature distribution despite the geometric accuracy of synthetic data, a bidirectional domain adaptation strategy based on image processing was adopted. By applying environmental degradation perturbation to the synthetic domain and signal enhancement to the real domain, the distributions of the two domains were aligned in the feature space, eliminating the semantic gap between "ideal simulation" and "complex reality". Combined with the global attention mechanism of the ViT architecture, the model is able to perform high-precision 3D topology reconstruction through reasoning.

[0028] (3) Three-dimensional aerodynamic parameter identification was achieved. By using known structural parameters as physical constraints, the inherent scale ambiguity problem of monocular vision was solved. The dimensionless relative point cloud was accurately mapped back to absolute physical space, thereby enabling the quantitative calculation of parameters such as the aerodynamic volume of the parachute. This provides more data support for evaluating the aerodynamic drag characteristics of the parachute and verifying the accuracy of fluid-structure interaction simulation, greatly improving the data utilization rate of airdrop tests and the efficiency of engineering evaluation. Attached Figure Description

[0029] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0030] Figure 1This is a flowchart illustrating a method for three-dimensional reconstruction and parameter identification of a parachute according to one embodiment of the present invention. Figure 2 This is a comparative schematic diagram of the parachute stabilization phase, inflation and deployment phase, and full-fill stabilization phase in an airdrop test according to one embodiment of the present invention. Figures (a), (d), and (g) are actual airdrop test observation images of the parachute stabilization phase, inflation and deployment phase, and full-fill stabilization phase, respectively; Figures (b), (e), and (h) are PBR composite rendering images of the parachute stabilization phase, inflation and deployment phase, and full-fill stabilization phase, respectively; and Figures (c), (f), and (i) are numerical simulation three-dimensional structural diagrams of the parachute stabilization phase, inflation and deployment phase, and full-fill stabilization phase, respectively. Figure 3 This is a comparison of the bidirectional feature alignment effect according to one embodiment of the present invention, wherein Figure (a) is the original real image; Figure (b) is the processed real image; Figure (c) is the original synthesized image; and Figure (d) is the processed synthesized image. Figure 4 To characterize Figure 3 The cosine similarity matrix of the feature distribution differences among the four groups of images; Figure 5 Figure 1 is a schematic diagram comparing cloud inference results according to one embodiment of the present invention. Figure 2 shows the point cloud results of training with the original synthetic image and inference with the original real image; Figure 3 shows the point cloud results of training with the original synthetic image and inference with the processed real image; Figure 4 shows the point cloud results of training with the processed synthetic image and inference with the original real image; Figure 5 shows the point cloud results of training with the processed synthetic image and inference with the processed real image. Detailed Implementation

[0031] The present invention will now be described in detail with reference to the accompanying drawings. This description is merely illustrative and explanatory, and should not be construed as limiting the scope of protection of the present invention. Furthermore, those skilled in the art can combine the features in the embodiments described herein and in different embodiments accordingly based on the description in this document.

[0032] The embodiments of the present invention are as follows, with reference to Figure 1 A method for three-dimensional reconstruction and parameter identification of parachutes includes the following steps: S1, Constructing the physical skeleton: Based on fluid-structure interaction numerical simulation, a three-dimensional physical skeleton of the parachute is generated to dynamically deform under dynamic flow field, and a sequence of parachute structure node coordinates that conforms to aerodynamic laws is obtained. In order to obtain the true aerodynamic shape of the flexible aerodynamic deceleration device parachute under complex flow field, this invention uses the arbitrary Lagrange-Euler method to perform fluid-structure interaction numerical simulation of the parachute inflation, deployment and stabilization process, and generates a dynamic three-dimensional mesh sequence characterizing the nonlinear geometric deformation of the parachute, which serves as the three-dimensional physical skeleton sequence. Before allowing the computer to perform calculations, the boundaries and objects of the computation need to be defined. First, a structural domain model of the parachute is established (based on the parachute's design blueprints, such as the geometry of a ring-sail parachute, the parachute is discretized into finite elements, such as triangular or quadrilateral meshes; this mesh is the prototype of the "skeleton," with each node having initial coordinates) and a fluid domain model are established (a sufficiently large wind tunnel or flow field space (such as a cylindrical computational domain) is built and divided into fluid meshes. Mesh refinement is needed in the region near the parachute to capture subtle flow field changes). The structural domain model is constructed based on the parachute's geometric design parameters, discretizing the parachute into finite element meshes. The fluid domain model contains the flow field space surrounding the parachute, with mesh refinement in the region where the parachute is located. As a flexible body with large deformation, the actual deployment process of a parachute is determined by the coupling effect between the fluid domain and the structural domain. In the fluid domain solution, the Navier-Stokes equations are modified by introducing the grid transport velocity *w* to solve the governing equations for mass and momentum conservation, thus accommodating the large-scale motion of the flow field grid. The mass and momentum conservation equations include the following formulas: (1) (2) (3) In the formula: For fluid density; For time; For fluid velocity; For grid transport speed; It is a volume force; For stress tensor; The internal energy per unit mass. The governing equations of the structural domain are described using a Lagrangian scheme to describe the nonlinear response of the fabric material: (4) In the formula: For structural density; The displacement of the structural point; For structural stress; For structural point volume forces; This is the acceleration due to gravity.

[0033] The nonlinear response of the parachute fabric material is described using a Lagrangian scheme for the structural domain, and the structural governing equations are solved. To achieve energy and momentum transfer between the fluid and the flexible fabric, the penalty method is used to handle the fluid-structure interaction interface (the coupling interface between the fluid domain and the structural domain). By tracking the relative permeation displacement between the fluid and structural elements, a virtual coupling force proportional to the permeation is applied at the interface. This allows for accurate simulation of the parachute's breathing, bulging, and wrinkling deformation under unsteady flow field conditions, while ensuring mass and energy conservation. A dynamic three-dimensional physical skeleton containing the coordinate sequence of the parachute's structural nodes is generated.

[0034] S2, Generate a high-fidelity synthetic dataset: Construct a virtual optical imaging link containing a camera projection model based on physical rendering technology, generate a synthetic image and a corresponding depth ground truth map that are spatiotemporally aligned with the three-dimensional physical skeleton, so as to construct a high-fidelity synthetic dataset. Based on the obtained physical skeleton, a virtual optical imaging link is constructed using physically based rendering (PBR) technology.

[0035] By solving the ray tracing rendering equation to simulate the diffuse reflection and scattering of light on the fabric surface, an RGB image with realistic lighting and shadow properties is generated as the synthesized image. The rendering equation is as follows: (5) In the formula, The emitted radiation brightness, The incident radiation brightness, It is self-illuminating. It is a two-way reflection distribution function.

[0036] To achieve accurate mapping from three-dimensional physical space to two-dimensional image space to obtain depth information, and to establish a virtual camera pinhole model consistent with the intrinsic parameters of a real camera, while generating the RGB image, ray tracing is used to calculate the depth value of each pixel to generate a depth ground truth map that is pixel-level aligned with the RGB image.

[0037] The pinhole camera projection model, based on the intrinsic parameter matrix of a real imaging device, establishes the following geometric projection relationship: (6) In the formula, For pixel coordinates, The coordinates of the physical skeleton nodes in the world coordinate system; The camera intrinsic parameter matrix includes focal length and principal point offset; This is the extrinsic parameter matrix describing the camera pose; This represents the depth scaling factor. Using this projection model, while rendering the RGB image, a ray tracing algorithm is employed to calculate the corresponding depth scaling factor for each pixel. The values ​​are used to generate pixel-aligned depth maps, providing accurate geometric supervision signals for depth-aware models.

[0038] S3, Bidirectional Domain Adaptation Processing: Feature alignment processing is performed on the synthetic images and real parachute airdrop images in the high-fidelity synthetic dataset to eliminate the feature distribution differences between the synthetic domain and the real domain, thereby achieving feature space alignment between the two and obtaining the processed synthetic image set and the processed real image set. To address the differences in feature distributions regarding texture detail and noise distribution between synthetic and real data, a bidirectional domain adaptation strategy based on digital signal processing is constructed. This strategy does not rely on physical assumptions but instead forces the synthetic and real domains to align their feature spaces through specific image processing algorithms.

[0039] Specifically, the feature alignment process is a bidirectional domain adaptation operation based on digital signal processing, including: performing grayscale processing, random Gamma nonlinear transformation, Gaussian blur processing and noise injection processing on the synthetic image in sequence to degrade the synthetic domain signal; and performing bilateral filtering, histogram adaptive stretching and image segmentation background replacement processing on the real airdrop image in sequence to enhance the real domain signal.

[0040] Signal degradation processing in the synthesis domain: To eliminate the problems of idealization and lack of real-world interference in the synthesized image, the following image degradation processing method is adopted: Grayscale Conversion: Converts an RGB three-channel image into a single-channel grayscale image, eliminating the inconsistency in color space distribution caused by color differences between rendered materials and real fabrics, allowing the model to focus on geometric structural features.

[0041] Random Gamma Correction: A non-linear mapping of image pixel intensity. ,in Random values ​​are selected within a preset range. This step simulates the differences in brightness caused by variations in light intensity and exposure time in a real-world scene.

[0042] Gaussian Blurring: This process uses a Gaussian kernel function to perform spatial convolution on an image, simulating motion blur caused by high-speed movement of objects or focusing errors in real imaging, and suppressing non-physical features such as overly sharp edges in the synthesized image.

[0043] Additive Noise Injection: Injecting Gaussian white noise or salt-and-pepper noise into the image simulates the thermal noise and shot noise of a real sensor under high gain conditions, improving the robustness of the depth perception model to low signal-to-noise ratio data.

[0044] Real-domain signal enhancement processing: To eliminate blur and noise in real images caused by the imaging process and achieve a clean, idealized image similar to the synthesized image, the real image is processed as follows: Bilateral filtering and histogram adaptive stretching are applied to real airdrop images. Bilateral filtering effectively filters out high-frequency geometric information such as the edges of the canopy folds while removing high-frequency noise from the sensor; histogram stretching expands the dynamic range of the image and enhances contrast. Simultaneously, image segmentation techniques are employed to preserve the canopy area and replace it with the background area of ​​the synthetic image, further reducing differences in feature distribution.

[0045] The Gaussian blurring and noise injection in the synthetic domain signal degradation operation are used to simulate dynamic blurring and scene degradation in the real imaging process; the histogram adaptive stretching in the real domain signal enhancement operation is the histogram adaptive stretching of the HSV color space; the image segmentation background replacement is to retain the umbrella area and replace the background with the background area of ​​the synthetic image.

[0046] Through the bidirectional processing based on the specific algorithm described above, the synthetic data and the real data tend to be consistent in terms of image statistical characteristics such as spectral distribution and gray-level histogram features, thereby significantly reducing semantic bias during cross-domain reasoning.

[0047] S4, Deep Perception Model Training and Inference: Using the processed synthetic image set and its corresponding ground truth depth map, train a deep learning-based deep perception network; input the processed real image set into the trained deep perception network, and output the corresponding depth map; To address the characteristics of parachute flexible fabric surfaces lacking significant geometric corners and exhibiting complex high-frequency wrinkles and self-occlusion during the closing and unfolding stages, this embodiment constructs a depth perception network based on the Vision Transformer (ViT) architecture as the core inference engine for 3D reconstruction. Through its multi-head self-attention mechanism, it captures long-distance dependencies between pixels at different locations in the image, thereby inferring continuous and consistent 3D spatial depth information of the parachute surface from 2D image features.

[0048] Compared to traditional convolutional neural networks (CNNs), which are limited by their local receptive field and struggle to establish effective spatial relationships in weakly textured regions, this invention utilizes the multi-head self-attention (MSA) mechanism in the ViT architecture to capture long-range geometric dependencies on the parachute surface. This mechanism aggregates global features by calculating the correlation weights between any two pixels in the input feature map. (7) In the formula, , , These are the query vector, key vector, and value vector, respectively. The scaling factor is used. This architecture enables the network to effectively decouple "physical occlusion edges" from "surface texture wrinkles," inferring a continuous and consistent spatial contour of the canopy globally.

[0049] S5, Relative 3D point cloud reconstruction: Based on the inverse projection relationship of the camera projection model, the depth map is converted into a relative 3D point cloud with the camera coordinate system as the reference. Based on the inverse projection relationship of the pinhole camera projection model according to formula (6), each pixel in the depth map is... The pixel coordinates and corresponding depth values ​​are converted into three-dimensional spatial coordinates in the camera coordinate system, and a continuous and complete parachute relative three-dimensional point cloud is obtained by reconstructing pixel by pixel.

[0050] S6, Physical Scale Recovery and Parameter Identification: Solve for the time-varying scale factor using known geometric invariants in the parachute structure as physical constraints, map the relative three-dimensional point cloud to the absolute physical space, and calculate the aerodynamic parameters of the parachute based on the mapped absolute physical space coordinates.

[0051] Since the 3D point cloud reconstructed by monocular vision only has relative geometric scale (scale ambiguity), this embodiment utilizes the geometric invariants in the flexible fabric structure as physical constraints to establish a mapping relationship from relative space to absolute physical space. The known geometric invariants are one or more reference geometric features in the parachute structure whose dimensions are known or remain unchanged during parachute opening. The specific steps for mapping the relative 3D point cloud to absolute physical space are as follows: multiply the camera coordinates of the relative 3D point cloud with the time-varying scale factor at the corresponding moment to obtain the 3D coordinates in the absolute physical space.

[0052] The specific steps for solving the time-varying scale factor are as follows: Select the parachute top hole region as the known geometric invariant and use it as a reference quantity; calculate the relative perimeter of the top hole region corresponding to the reconstructed point cloud at any time; define the instantaneous scale factor as the ratio of the reference quantity to the relative perimeter; and decompose the instantaneous scale factor into the product of the reference mapping ratio and the dynamic compensation term; the dynamic compensation term adaptively corrects the non-physical fluctuations generated by the deep sensing network; calculate the key aerodynamic parameters of the parachute based on the mapped absolute physical space coordinates using the infinitesimal integral method.

[0053] In one specific embodiment, a stable structure during the parachute deployment process, such as the top hole region, is selected as a reference quantity. At any time Calculate the relative perimeter of the corresponding feature regions in the reconstructed point cloud. Define instantaneous scale factor : (8) From the perspective of physics-model coupling, this scaling factor can be further decomposed into the baseline mapping ratio. With dynamic compensation items The product of: (9) Among them, dynamic compensation item It can adaptively correct minute non-physical fluctuations caused by different lighting or deformation conditions in the depth perception model. Based on this scale factor, the relative point cloud coordinates... Mapped to physical coordinates Furthermore, the instantaneous projected area and aerodynamic volume of the parachute are calculated using the infinitesimal integral method, thereby achieving quantitative identification of aerodynamic parameters.

[0054] To better understand the technical solution of the present invention, a specific implementation method is provided. Taking a typical annular parachute fluid-structure interaction simulation model as an example, the present invention is used to perform three-dimensional reconstruction and parameter identification of the parachute during the airdrop test.

[0055] Step 1: A full-size annular parachute is selected for numerical simulation. Based on the theoretically designed geometry, a simulation model is established. A cylindrical flow field is chosen as the fluid computation domain, and the flow is refined in the area where the parachute canopy is located to ensure a good simulation effect of the canopy deformation. The environmental conditions when the parachute is released during the simulated airdrop test are selected as the initial conditions for the free flow, thereby ensuring that the entire process of the stabilization phase of the release, the inflation and deployment phase, and the full-filling and stabilization phase can be simulated. The ALE algorithm in the numerical simulation solver is used to solve the dynamic mesh sequence of the entire parachute opening and descent process. By comparing the drag coefficients of each stage, the realism and credibility of the simulation are ensured.

[0056] Step 2: Based on the mesh topology sequence obtained in Step 1, construct a virtual imaging link using a Physically Based Rendering (PBR) engine. Configure the intrinsic parameter matrix of the virtual camera to be consistent with that of the camera used in the real experiment, and set the virtual light source according to the light intensity distribution and texture distribution of the airdrop environment to generate an RGB composite image and its corresponding depth ground truth map. For example... Figure 2 As shown in the figure, this graph compares data from three different stages of the entire parachute opening process.

[0057] contrast Figure 2 As seen in each stage, the synthesized image exhibits a very high degree of geometric consistency with the real image. During the stabilization stage, the canopy, physically constrained by the drawstrings, presents a "bulb-like" shape, with the sail converging inwards to create high-frequency wrinkles and self-blocking. During the inflation and deployment stage, as the drawstrings are released, the air inlet gradually opens to the same width as the ring, forming a vertical cylindrical structure. In the full-float stabilization stage, the canopy structure is fully deployed, eliminating large-scale wrinkles and stacking, and presenting a regular hemispherical geometric surface.

[0058] Step 3: To reduce the distribution differences between pixel domains, "composite domain realization" and "real domain idealization" processes were implemented, such as... Figure 3 As shown, for synthetic data, Gaussian noise and nonlinear illumination perturbations are introduced to force the model to learn robustness to imaging defects; for real data, bilateral filtering and HSV adaptive stretching are used to effectively filter out sensor thermal noise and remove complex sky backgrounds while preserving high-frequency geometric information at the edge of the umbrella canopy. This bidirectional feature alignment strategy makes the signal-to-noise ratio and feature distribution of real data significantly closer to the ideal synthetic domain.

[0059] To quantify the effect of domain difference, a VIT network encoder with frozen parameters is used to extract high-dimensional implicit features of the images. The feature distribution differences of the four sets of images are verified using cosine similarity as the metric. Figure 4 The cosine similarity matrix is ​​shown.

[0060] Step 4: Using the synthetic dataset obtained from image processing in Step 3, train a ViT-based depth perception network. Utilize ViT's global self-attention mechanism to teach it to infer the patterns of 3D wrinkles from 2D image textures. Input the processed real-world airdrop video frames into the trained network, which outputs a single-channel depth map. A relative 3D point cloud is then generated using a pinhole camera inverse projection model. Figure 5 This demonstrates the real-world point cloud inference results after training and inference on different images. Figure 5 It can be seen that among them Figure 5 (d) The model output shows that the hemispherical structure is the most complete, with a smooth and continuous surface and no outlier noise. The ViT architecture model with deep inference, combined with the bidirectional feature alignment strategy, can perform generalized inference of real images and achieve accurate capture of the parachute shape in real airdrop tests.

[0061] Step 5: During the opening process of the parasail, the top hole rope is clearly visible throughout and the structural changes are minimal, serving as a reference benchmark. Extract the inference point cloud obtained in step 4 and calculate the relative perimeter of the corresponding region. Calculate the instantaneous scale factor Using this factor, the relative point cloud coordinates Mapped to physical coordinates The projected area is calculated using the infinitesimal integral method. with aerodynamic volume .

[0062] This invention integrates fluid-structure interaction numerical simulation and physical rendering techniques to construct a high-fidelity synthetic dataset containing accurate aerodynamic shape and depth truth values. It also combines a two-way domain adaptation strategy to eliminate the feature distribution differences between the synthetic domain and the real domain, thereby enabling the three-dimensional reconstruction of parachutes and the identification of three-dimensional aerodynamic parameters using only monocular video.

[0063] It should be noted that the above embodiments are illustrative of the invention and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. Unless otherwise specified, the steps in the above embodiments should not be construed as limiting the order of execution.

[0064] For those skilled in the art, various improvements and modifications can be made without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention.

Claims

1. A method for three-dimensional reconstruction and parameter identification of a parachute, characterized in that, Includes the following steps: S1, Constructing the physical skeleton: Based on fluid-structure interaction numerical simulation, a three-dimensional physical skeleton of the parachute is generated to dynamically deform under dynamic flow field, and a sequence of parachute structure node coordinates that conforms to aerodynamic laws is obtained. S2, Generate a high-fidelity synthetic dataset: Construct a virtual optical imaging link containing a camera projection model based on physical rendering technology, generate a synthetic image and a corresponding depth ground truth map that are spatiotemporally aligned with the three-dimensional physical skeleton, so as to construct a high-fidelity synthetic dataset. S3, Bidirectional Domain Adaptation Processing: Feature alignment processing is performed on the synthetic images and real parachute airdrop images in the high-fidelity synthetic dataset to eliminate the feature distribution differences between the synthetic domain and the real domain, thereby achieving feature space alignment between the two and obtaining the processed synthetic image set and the processed real image set. S4, Deep Perception Model Training and Inference: Using the processed synthetic image set and its corresponding ground truth depth map, train a deep learning-based deep perception network; input the processed real image set into the trained deep perception network, and output the corresponding depth map; S5, Relative 3D point cloud reconstruction: Based on the inverse projection relationship of the camera projection model, the depth map is converted into a relative 3D point cloud with the camera coordinate system as the reference. S6, Physical Scale Recovery and Parameter Identification: Solve for the time-varying scale factor using known geometric invariants in the parachute structure as physical constraints, map the relative three-dimensional point cloud to the absolute physical space, and calculate the aerodynamic parameters of the parachute based on the mapped absolute physical space coordinates.

2. The method for three-dimensional reconstruction and parameter identification of parachutes according to claim 1, characterized in that, In step S1, the arbitrary Lagrange-Euler method is used to perform fluid-structure interaction numerical simulation of the parachute inflation, deployment and stabilization process, generating a dynamic three-dimensional mesh sequence characterizing the nonlinear geometric deformation of the parachute, which serves as the three-dimensional physical skeleton sequence. In step S2, based on physical rendering technology, the diffuse reflection and scattering of light on the fabric surface are simulated by solving the ray tracing rendering equation to generate an RGB image with realistic light and shadow attributes as the composite image; at the same time, a virtual camera pinhole model consistent with the intrinsic parameters of a real camera is established. While generating the RGB image, the depth value of each pixel is calculated using ray tracing to generate a depth ground truth map that is pixel-level aligned with the RGB image.

3. The method for three-dimensional reconstruction and parameter identification of parachutes according to claim 2, characterized in that, First, a structural domain model and a fluid domain model of the parachute are established. The structural domain model is constructed based on the geometric design parameters of the parachute and the canopy is discretized into a finite element mesh. The fluid domain model contains the flow field space surrounding the parachute and the mesh is refined in the region where the canopy is located. In step S1, the numerical simulation of fluid-structure interaction using the arbitrary Lagrange-Euler method includes: introducing the grid transport velocity to modify the Navier-Stokes equations in the fluid domain, and solving the mass conservation and momentum conservation control equations. The nonlinear response of the parachute fabric material is described by the Lagrangian scheme for the structural domain, and the structural control equations are solved. The penalty function method is used to handle the coupling interface between the fluid domain and the structural domain. By tracking the relative permeation displacement of the fluid unit and the structural unit and applying virtual coupling force, the breathing, bulging and wrinkling deformation of the parachute under unsteady flow field are simulated, and a dynamic three-dimensional physical skeleton containing the coordinate sequence of the parachute structural nodes is generated.

4. The method for three-dimensional reconstruction and parameter identification of a parachute according to any one of claims 1 to 3, characterized in that, The feature alignment process in step S3 is a bidirectional domain adaptation operation based on digital signal processing, specifically including: performing grayscale processing, random Gamma nonlinear transformation, Gaussian blur processing and noise injection processing on the synthetic image to degrade the synthetic domain signal; and performing bilateral filtering, histogram adaptive stretching and image segmentation background replacement processing on the real airdrop image to enhance the real domain signal.

5. The method for three-dimensional reconstruction and parameter identification of parachutes according to claim 4, characterized in that, The Gaussian blurring and noise injection in the synthetic domain signal degradation operation are used to simulate dynamic blurring and scene degradation in the real imaging process; the histogram adaptive stretching in the real domain signal enhancement operation is the histogram adaptive stretching of the HSV color space; the image segmentation background replacement is to retain the umbrella area and replace the background with the background area of ​​the synthetic image.

6. The method for three-dimensional reconstruction and parameter identification of a parachute according to claim 1, characterized in that, In step S4, the depth perception network adopts a Vision Transformer-based architecture, which uses its multi-head self-attention mechanism to capture long-distance dependencies between pixels at different locations in the image, so as to infer continuous and consistent three-dimensional spatial depth information of the parachute surface from the two-dimensional image features.

7. The method for three-dimensional reconstruction and parameter identification of parachutes according to claim 6, characterized in that, The multi-head self-attention mechanism achieves global feature aggregation by calculating the correlation weights of any two pixels in the input feature map, decoupling the physical occlusion edges of the parachute from surface texture wrinkles, and inferring a continuous and consistent spatial contour of the parachute canopy globally.

8. The method for three-dimensional reconstruction and parameter identification of a parachute according to claim 3, characterized in that, In step S5, based on the inverse projection relationship of the pinhole camera projection model, the pixel coordinates and corresponding depth values ​​of each pixel in the depth map are converted into three-dimensional spatial coordinates in the camera coordinate system, and a continuous and complete parachute relative three-dimensional point cloud is obtained by reconstructing pixel by pixel.

9. The method for three-dimensional reconstruction and parameter identification of a parachute according to claim 1, characterized in that, In step S6, the known geometric invariant is one or more reference geometric features in the parachute structure whose dimensions are known or remain unchanged during the parachute opening process; The specific steps for mapping the relative three-dimensional point cloud to the absolute physical space are as follows: multiply the camera coordinates of the relative three-dimensional point cloud with the time-varying scale factor at the corresponding time to obtain the three-dimensional coordinates in the absolute physical space.

10. The method for three-dimensional reconstruction and parameter identification of a parachute according to claim 1, characterized in that, The specific steps for solving the time-varying scale factor in step S6 are as follows: select the parachute top hole region as the known geometric invariant and use it as a reference quantity; calculate the relative perimeter of the top hole region corresponding to the reconstructed point cloud at any time; define the instantaneous scale factor as the ratio of the reference quantity to the relative perimeter; and decompose the instantaneous scale factor into the product of the reference mapping ratio and the dynamic compensation term; the dynamic compensation term adaptively corrects the non-physical fluctuations generated by the deep perception network. The key aerodynamic parameters of the parachute are calculated based on the mapped absolute physical space coordinates using the infinitesimal integral method.

Images