Intelligent optimization sparse point cloud registration and dynamics compensation method for non-cooperative satellites

Abstract

The application discloses a method for intelligent optimization of sparse point cloud registration and dynamic compensation for non-cooperative satellites, which comprises the following steps: a service satellite flies around a target and acquires time-stamped laser point clouds; a continuous-time relative motion trajectory model based on B-spline and quaternion spline is constructed, and accurate motion compensation is performed on asynchronous point clouds to eliminate geometric distortion; a differentiable deep learning network is used to extract point cloud depth features, and a soft corresponding probability matrix is constructed to realize initial alignment with a target reference point cloud. A unified joint closed-loop optimization framework is established, the motion trajectory parameters, the depth features and the registration transformation parameters are coupled, all the parameters are simultaneously and iteratively optimized by minimizing a loss function which integrates registration errors, trajectory consistency and smoothness constraints, mutual correction and collaborative convergence of motion estimation and geometric registration are realized, and high-precision three-dimensional poses of key components of the target can be output, thereby directly supporting in-orbit refueling, maintenance and debris cleaning and the like.

Description

Technical Field

[0001] This invention belongs to the field of aerospace technology and relates to an intelligent optimized sparse point cloud registration and dynamic compensation method for non-cooperative satellites. Background Technology

[0002] With the continuous increase in the number of spacecraft in orbit and the increasing scarcity of orbital resources, on-orbit servicing (OOS) has become a key technology for extending satellite lifespan, improving mission reliability, and maintaining the safety of the space environment. Performing on-orbit refueling, component replacement, attitude correction, and health monitoring on cooperative target satellites can effectively prevent premature satellite failure, reduce total life-cycle costs, and promote the development of maintainable and upgradeable spacecraft design concepts. At the same time, the proactive removal and cleanup of failed spacecraft and space debris has become an urgent need to ensure the safety of on-orbit assets and alleviate orbital congestion.

[0003] In the aforementioned tasks, the service satellite needs to perform high-precision three-dimensional perception and positioning of the target satellite under conditions of close relative motion. Laser point cloud technology, due to its advantages such as accurate ranging, independence from lighting conditions, and direct acquisition of three-dimensional geometric information of object surfaces, has become an important means of on-orbit visual navigation and target recognition. However, in actual on-orbit service scenarios, the effective utilization of laser point clouds faces a series of severe challenges: First, the issues of temporal asynchrony and motion distortion are prominent. Serving satellites typically perform omnidirectional scanning of targets by flying around them, resulting in continuous relative translation and rotation. Each point in the laser point cloud is acquired at a different time; ignoring this temporal difference will cause deformation of the point cloud in three-dimensional space, severely reducing reconstruction and registration accuracy. Traditional methods, mostly based on static assumptions or simple linear motion compensation, struggle to adapt to the complex relative motion trajectories during the flying around process, especially in highly dynamic scenarios or those with significant changes in angular velocity, where the compensation effect is limited.

[0004] Secondly, there are inherent defects in point cloud quality. Laser point clouds obtained in space environments are typically sparse, non-uniform, and subject to occlusion and noise interference. The surface of the target satellite may contain areas of high reflectivity and low reflectivity, or its complex structure (such as deployed solar panels or antennas) may cause severe self-occlusion, further exacerbating the incompleteness of the point cloud. Traditional registration algorithms (such as Iterative Closest Point (ICP) and its variants) heavily rely on the local geometric features of the point cloud (such as normal vectors and curvature), which can easily lead to mismatches when the point cloud is sparse or features are repetitive, resulting in insufficient robustness.

[0005] Furthermore, existing learning-based methods face bottlenecks in generalization and deployment. Although deep learning has demonstrated powerful feature learning capabilities in point cloud registration, most existing networks are trained on simulations or ground datasets, making it difficult to directly transfer them to real-world space environments. In-orbit point clouds are significantly different from the distribution of training data due to factors such as sensor noise, distance attenuation, and the diversity of surface materials. In addition, onboard computing resources are severely limited, and complex deep networks often struggle to meet the requirements of real-time processing and low power consumption, thus restricting their in-orbit application.

[0006] Furthermore, the fusion of multi-source information is insufficient. Existing solutions mostly process laser point clouds independently, failing to fully integrate trajectory priors and attitude information provided by inertial measurement units (IMUs), star sensors, visual odometry, etc. When there are errors in motion priors or inaccurate time synchronization, the registration results of a single sensor are prone to instability, and there is a lack of quantification of estimation uncertainties, which cannot provide a reliable safety boundary for subsequent high-risk tasks such as robotic arm operation and docking control.

[0007] In summary, current point cloud sensing and registration technologies for on-orbit servicing still have significant shortcomings in areas such as motion distortion compensation, robustness to sparse noise, real-time algorithm performance, and compatibility with spaceborne platforms. There is an urgent need for a systematic solution that can tightly couple temporal motion models and deep learning features, and achieve high-precision, highly robust registration within a unified optimization framework, to support future autonomous and intelligent on-orbit servicing missions. This invention addresses these technical bottlenecks. Summary of the Invention

[0008] This invention aims to solve the core technical challenges of geometric distortion in 3D reconstruction and inaccurate positioning of key parts caused by relative motion between service satellites and cooperative target satellites during on-orbit refueling, on-orbit maintenance, attitude calibration, and debris removal. These challenges arise from the asynchronous timing of laser point cloud acquisition, complex target structures, and sparse point clouds. The invention proposes an intelligent optimized sparse point cloud registration and dynamic compensation method for non-cooperative satellites. By jointly optimizing continuous-time motion compensation and deep learning-based feature registration within a unified framework, high-precision 3D reconstruction and pose estimation of key parts of the target satellite are achieved, providing reliable data support for the precise operation of service satellites.

[0009] The technical solution of the present invention is as follows: On the one hand, this invention provides an intelligent optimized sparse point cloud registration and dynamic compensation method for non-cooperative satellites, characterized by the following steps: S1. The service satellite flies along a preset orbital trajectory, collecting temporal laser point cloud data of the target satellite's surface using an onboard laser scanning device to obtain the original point cloud set. ,in, Indicates the first The three-dimensional spatial coordinates of a laser point in the coordinate system of the serving satellite; Indicates the first The acquisition timestamps corresponding to each laser point are used to characterize the temporal characteristics of the point cloud and support subsequent motion compensation calculations; The total number of laser point clouds acquired by the airborne laser scanning device within a scanning cycle or observation time window; The index number of a single laser echo point in the point cloud, used to identify the first... One laser point was collected; S2. Based on the known orbit and attitude prior information of the target satellite, construct a continuous-time relative motion trajectory model between the service satellite and the target satellite. The trajectory model is used to continuously describe the pose changes of the target satellite relative to the serving satellite in the time domain, wherein the position... and posture It is smoothly characterized in the time domain by differentiable basis functions; S3. Based on the continuous-time relative motion trajectory model constructed in step S2, apply the model to the original point cloud set obtained in step S1. Perform motion distortion compensation: for each point in the set From its collection time Unified transformation to the selected reference time To eliminate point cloud distortion caused by relative motion, the compensation transformation formula is as follows: The corrected compensation point cloud set is obtained. ; S4. The compensated point cloud set obtained in step S3... Input a pre-trained deep learning feature extraction network to extract high-dimensional deep feature vectors representing the local and global geometric structures of the point cloud. ; S5. Using the depth features extracted in step S4, in the compensated point cloud With the pre-acquired target satellite reference point cloud Construct a differentiable soft correspondence probability matrix between them. Matrix elements Characteristic points With point The matching probability is calculated; based on this soft correspondence probability matrix, the initial rigid body transformation parameters are solved by weighted least squares and singular value decomposition. This achieves initial alignment of the point cloud; S6. The trajectory model parameters from step S2, the compensation relationship from step S3, the depth features from step S4, the soft correspondence from step S5, and the exact rigid body transformation parameters to be determined are all incorporated into a unified, differentiable joint optimization framework. By minimizing a joint loss function L, gradient optimization is used to synchronously and iteratively update all the above parameters, enabling motion compensation and geometric registration to mutually correct and converge collaboratively, ultimately outputting the optimal continuous-time trajectory. and optimal precise registration transformation ; S7. The optimal continuous-time trajectory obtained in step S6 and optimal precise registration transformation By combining and transforming the known structural model of the target satellite, high-precision three-dimensional position and attitude information of its key functional components in the service satellite coordinate system can be analyzed and output.

[0010] Furthermore, the joint loss function Defined as: , in, Based on the soft correspondence probability matrix Registration geometric error term; It is a constraint that ensures consistency between the point cloud and the trajectory model prediction after motion compensation; It is a smoothing regularization term applied to the trajectory parameters, where, For location control points, , These are the weighting coefficients.

[0011] Furthermore, the basis function smoothing representation in step S2 specifically includes: position ,in, For B-spline basis functions, Position control point; attitude Using quaternion splines Perform parameterization. This is the attitude control point.

[0012] Furthermore, in step S6, the joint optimization framework employs an alternating optimization strategy or an end-to-end gradient descent strategy for the trajectory parameters { , The registration parameters [R,t] are jointly updated, and the optimization process depends on the gradient backpropagation achieved by the differentiability of the compensation formula in step S3, the feature extraction network in step S4, and the soft correspondence calculation in step S5.

[0013] Furthermore, in the trajectory model construction in step S2 or the joint optimization in step S6, measurement data from the inertial measurement unit and star sensor are further integrated as prior constraints or loss terms to form a trajectory estimation model with tight coupling of multi-source information.

[0014] Furthermore, the elements of the soft correspondence probability matrix W Calculated using the Softmax function based on feature distance: In the formula, This is a soft correspondence probability matrix used to characterize the feature matching relationship between the compensated point cloud and the reference point cloud; Indicates the compensation point in the cloud. The point and the reference point cloud The soft matching weight or corresponding probability between points reflects the degree of similarity between them in the feature space. The first point extracted from the motion-compensated sparse point cloud by a deep neural network. The depth feature vector of each point; To extract features from a reference point cloud using the same feature extraction network The extracted first The depth feature vector of each point; The dimension of the feature vector; This is the feature distance sensitivity coefficient (temperature coefficient), used to adjust the influence of feature distance on the Softmax normalization process; when When the value is large, the algorithm pays more attention to point pairs with small feature distances, thereby enhancing the discriminative power of the matching; when When the value is small, the weight distribution tends to be smooth, which is beneficial to improving the stability of matching under conditions of high noise or sparse point cloud. Represents the Euclidean norm; This indicates that all candidate matching points in the reference point cloud are summed and normalized to ensure that the matching is applied to a fixed point cloud. ,have The Softmax weight calculation method can establish a probabilistic correspondence between the compensation point cloud and the reference point cloud in the feature space, thereby providing continuous and differentiable matching constraints for subsequent rigid body pose estimation and registration optimization.

[0015] Furthermore, in step S4, the deep learning feature extraction network is a point cloud network pre-trained with simulation data from on-orbit service scenarios, and its structure is selected from PointNet, PointNet++, DGCNN or graph convolutional networks.

[0016] Furthermore, the reference point cloud The point cloud obtained by sampling the computer-aided design model of the target satellite, or the high-precision point cloud model generated by fusing historical on-orbit scanning data.

[0017] Second, the present invention also provides a spaceborne data processing device for on-orbit service, characterized in that it includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method as described in any one of claims 1 to 8.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1) Abandoning the traditional discrete inter-frame registration or simple linear motion compensation approach, continuous-time B-splines and quaternion splines are used to model the relative motion trajectory, realizing precise time-level motion compensation for each laser point, and solving the problem of point cloud geometric distortion caused by target motion during scanning.

[0019] 2) A parallel processing flow of motion compensation and deep learning feature extraction was designed. The point cloud simultaneously performs geometric correction and deep feature description before registration, coupling the modules that are processed sequentially or in isolation in traditional methods, which significantly improves information utilization efficiency and system robustness.

[0020] 3) By jointly modeling laser point clouds, motion compensation, and deep learning features, highly robust registration of point clouds can be achieved even under conditions of relative motion, noise interference, and sparse structure. The continuous-time trajectory model effectively avoids motion deviations caused by discrete interpolation, improving the time synchronization accuracy of point clouds. Deep learning features enhance the registration's adaptability to occlusion, geometric repetition, and sparse structure, resulting in more stable and accurate localization of key areas. The joint optimization framework allows motion compensation and registration to mutually reinforce each other, achieving higher overall accuracy than traditional independent solution methods.

[0021] 4) Applicable to various tasks such as on-orbit refueling, maintenance, attitude correction and debris removal, providing unified high-precision space awareness capabilities for service satellites. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the intelligent registration and positioning method for laser point clouds for on-orbit services of cooperative satellites according to the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0024] The implementation system is deployed on the service satellite performing on-orbit service missions. Service satellite (sensor) reference frame: . Target satellite reference frame: . Carrier inertial reference frame / geocentric reference frame: .

[0025] The pose of the target relative to the service satellite at time (the transformation from to ): . Represented by a matrix: where (3×3 rotation matrix), (translation vector).

[0026] Quaternion representation: Corresponding to .

[0027] The th point obtained by laser scanning (recorded in the service satellite sensor coordinate system , with a timestamp): , where is the time, is the point coordinate (in ).

[0028] A point cloud set obtained from one scan (or one fly-around observation): .

[0029] Reference (or template) point cloud (which may be the previous frame or the model CAD): (in or depending on the situation, and hereinafter unified as ).

[0030] Sensor noise: Point measurement noise covariance .

[0031] Orbit / velocity prior (provided by ground or spacecraft state navigation): Linear velocity , angular velocity (in or ).

[0032] Deep learning registration network parameters: The network is mapped as (Point to feature), or more generally Output registration transformation.

[0033] Loss functions are generally referred to as: .

[0034] The following combination Figure 1 The process shown provides a detailed description of the specific implementation details of each step in the method of the present invention.

[0035] Step 1: Sensor and Time Calibration Model Assuming laser measurement point Recording time At that moment, the measurement is of the target's position relative to the surface point of the serving satellite. Location: here Is this surface point at? The inherent coordinates on the surface (unknown). .

[0036] What is measured directly is ,and and All unknown (or (Given a priori). The goal is to recover... With / or Align and .

[0037] Step 2: Motion Model (Continuous Time Difference Representation) The common continuous-time uniform angular velocity + linear velocity model is adopted (which can be extended to polynomial or spline): right Rewritten as differential equations for translation and rotation: in This represents an antisymmetric matrix.

[0038] If the trajectory is approximated by linear interpolation / spline within the observation window, it can be represented using basis functions: in These are rotation vector parameters (such as axis angles). As the base coefficient, These are basis functions (e.g., B-spline basis functions). This representation makes... Differentiable and parameterized as a set of coefficients .

[0039] If a constant velocity model (discrete) is used, at time step : Step 3: Per-point correction Each point At the time of collection The goal is to transform the points to a unified reference time. (e.g., scan start) This is to eliminate distortion caused by the movement of the target during scanning.

[0040] Let the estimated trajectory be Compensation steps: Measuring points from Transform to At its position at the moment of measurement: Project this point back to the reference time exist The coordinates below (compensation). This is equivalent to placing the target point in... of Coordinates, mapped to of coordinate: Combined and written as: This is a formula that can be calculated directly. If Then all points are projected back to the same reference time.

[0041] Note numerical stability: Use quaternions for rotations. Perform interpolation to ensure Orthogonal; quaternion interpolation uses SLERP or B-spline. .

[0042] Step 4: Deep Learning Point Cloud Registration Module (Differentiable) Design a differentiable network To complete the registration from the compensated point cloud to the template point cloud, the structural conceptualization is divided into three steps: Feature extraction: Point cloud and Extract local / global features separately: here PointNet / PointNet++ / DGCNN, etc., can be used (abstracted as differentiable mappings). To maintain generality, define... It is a point-level feature network, followed by neighborhood pooling to form local descriptors.

[0043] Correspondence establishment (differentiable soft-assignment): Calculate the feature similarity matrix : And normalize to a probability distribution (row normalization or double normalization): Soft assignment allows the network to be trained end-to-end and enables the computation of expected correspondences: Estimation transformation (closed-form solution or network regression): Closed least squares (point-to-point): If given a pair Corresponding Rigid body transformations can be solved using SVD. Let the center of mass be: Constructing a matrix SVD decomposition Then the least squares rotation: This closed-form solution is differentiable in implementation (SVD can be backpropagated, and modern frameworks support it).

[0044] Direct regression of the network: or using a small MLP Mapping global features to SE(3) transformation parameters: (Li algebra parameters), and through exponential mapping The transformation is obtained. Both are compatible.

[0045] The estimated transformation is written as: Registration result: Transform the compensated point cloud again: Step 5: Loss Function (Training and Online Optimization) For it to be executable, provide an explicit and computable loss term and indicate the source of the gradient.

[0046] Point-to-point Chamfer Loss (symmetric): (Note the non-differentiability of min; soft-min or differentiable nearest neighbor approximation can be used.) Point-to-plane (point-to-plane) Loss (if the template contains a normal vector) ): in It is a nearest neighbor or soft-allocated index.

[0047] Corresponding consistency loss (based on soft allocation): Motion compensation regularization term (limits trajectory overfitting): in and From navigation priors, For weights.

[0048] Total loss: Regularize network parameters (e.g.) ), These are the motion trajectory parameters.

[0049] During training, Use gradient descent (Adam et al.); for trajectory parameters It can be solved simultaneously using the gradient method (end-to-end) or locally using nonlinear least squares (Levenberg–Marquardt) in the online stage.

[0050] Step Six: Joint Optimization Solution (Executable Process) Alternating Minimization is a simple and easy-to-implement method. initialization , (Pre-trained or randomized).

[0051] For iteration : Motion compensation: based on current trajectory parameters right Get compensation (See Section 3 for the formula).

[0052] Fixed trajectory, estimated registration: using right and estimate and update : The gradient is obtained through automatic differentiation.

[0053] Fixed network, optimized trajectory: The trajectory parameters are adjusted using the current registration error metric. Solution: LM (Levenberg–Marquardt) or Gauss–Newton can be used here because... about Mostly quadratic forms (point coordinates about) The Jacobian matrix can be explicitly calculated. See below for details on calculating Jacobian matrix elements.

[0054] Output after completion and .

[0055] Jacobi calculation (key) The trajectory parameters should be solved using nonlinear least squares. It is necessary to calculate the relationship between each residual and the value of the residual. The derivative of. Using single-point residuals. For example, among which .but and right Differentiation can be derived automatically or manually: The derivative of the rotation with respect to the parameter can be derived by a quaternion Lie algebra mapping or by using an exponential mapping (which can be computed using existing toolkits).

[0056] Step 7: Determine Implementation Details Trajectory parameterization: Translation and quaternions (using spherical linear interpolation or a square) are parameterized on the time axis using a cubic B-spline. In implementation, the degrees of freedom for quaternion interpolation are placed at control points and normalized.

[0057] Time synchronization: IMU / laser / navigation time must be synchronized to <1 ms to ensure compensation accuracy; if time skew error exists... They are then optimized together as parameters to be estimated.

[0058] Nearest neighbor search acceleration: Nearest neighbor search in registration uses KD-tree (libnabo, flann) or parallel implementation on GPUs. Soft-assignment can use local top-k matching to constrain computational complexity.

[0059] SVD backpropagation: Modern frameworks (PyTorch) have stable implementations of SVD; for greater stability, Procrustes' quaternion-based method (Horn algorithm) can be used and the corresponding derivative can be implemented.

[0060] Loss term weight suggestions (experimental baseline): Adjustments should be made based on the specific data noise level.

[0061] Learning rate: (Adam), dynamic adjustment of trajectory LM step size.

[0062] Point cloud downsampling: For real-time performance, voxel mesh filtering can be used (voxel size = 0.05 - 0.5 m depending on distance and resolution), but training still retains the full resolution for fine-tuning.

[0063] Step 8: From Point Cloud to Functional Localization After registration, accurate results are obtained. And local point cloud alignment. Location of maintenance / filling ports: exist Interface template points (CAD or historical observation) transformed to : Path planning for robotic arms / filling arms With the objective of planning the optimal grasping trajectory by combining relative velocity and collision constraints (this is the path planning module, which can be implemented with ROS MoveIt, etc.), we can achieve this.

[0064] Localization of the towed target (failed spacecraft or debris): Register the surrounding point cloud and estimate the object's center of mass and inertial attitude (PCA or by registering CAD) for docking of the gripping / dragging mechanism.

[0065] The beneficial effects of this embodiment are as follows: 1) Employing a continuous-time trajectory structure instead of a discrete-frame structure addresses the rigidity of traditional model structures: Using a continuous-time trajectory composed of B-splines and quaternions as the core structure, point cloud processing no longer relies on the discrete-frame assumption, but instead determines the true pose of each laser point based on its precise sampling time. This structure overcomes the limitations of traditional frame-by-frame modeling and eliminates the problem of scanning distortion during motion not being structurally expressible.

[0066] 2) Establishing a dual-channel parallel architecture of "motion compensation structure + deep feature structure": Traditional registration structures usually rely only on geometric ICP or only on neural networks. This invention constructs two structural links (motion compensation link and deep feature link), enabling the point cloud to have the dual structural advantages of "geometric correction + semantic features" before entering the registration module, significantly improving the coupling and information completeness of the overall structure.

[0067] 3) A differentiable soft correspondence matrix structure is used instead of the traditional hard matching structure. This invention constructs a differentiable soft correspondence matrix. This avoids the structural discontinuities of hard nearest-neighbor matching, making the entire registration process closely connected, differentiable, and optimizable. This matrix structure naturally solves the problem of easy interruption in matching in sparse and occluded point clouds.

[0068] 4) Establish a unified loss function structure, merging trajectory, feature, and geometry into a single architecture. Traditional structures are generally divided into "compensation before registration" or "registration before trajectory optimization," which suffers from structural fragmentation. This invention binds trajectory parameters, feature correspondences, and geometric constraints together through a unified structured loss function, making the system a truly integrated architecture and achieving a fully coupled closed loop in structure.

[0069] 5) Modules are structurally connected through differentiable operators, avoiding manual segmentation and structural jumps. From trajectory prediction, point cloud compensation, feature extraction, and correspondence generation to rigid body transformation solution, all modules are constructed as differentiable structures, enabling the entire system to be jointly optimized through gradient descent, thereby improving structural consistency and computational stability.

[0070] The advantages of this invention are listed below in terms of functionality, performance, and engineering applicability: (1) It can provide high-precision, distortion-free laser point cloud reconstruction in flying scenarios. This invention solves the most critical problem in terms of functionality: the motion distortion caused by the service satellite's orbital flight is completely eliminated, making the reconstructed point cloud geometrically consistent with the target satellite, thereby ensuring the spatial positioning accuracy of subsequent maintenance, refueling, and retrieval tasks.

[0071] (2) Stable registration is maintained even under sparse point cloud and occluded point cloud conditions. By leveraging deep learning features, this invention maintains robust matching capabilities on orbital point clouds containing noise, sparse sampling, and irregular distribution, and can functionally adapt to different surface materials, complex geometries, and localized damage conditions of satellites.

[0072] (3) It has high robustness and can adapt to track disturbances, attitude errors and sensor degradation. The joint optimization framework automatically corrects minor trajectory deviations, making the performance independent of perfect trajectory priors; it can maintain high stability even in the presence of attitude jitter or low echo signals.

[0073] (4) It can directly output spatial pose information suitable for robotic arms, on-orbit maintenance, and debris dragging. This invention not only performs registration but also outputs structured rigid body transformations and target component attitudes, therefore it can be directly used for: On-orbit refueling interface positioning Faulty component gripping point location Determination of the drag point of the failed spacecraft Contact surface assessment for fragment removal task Improve task reliability and execution efficiency.

[0074] The following are usage examples: Example 1 (constant angular velocity flight + stable structure) In scenarios where the service satellite orbits the cooperative target satellite at a constant angular velocity, the performance of the proposed solution reaches its optimal state. At this time, the trajectory of the service satellite relative to the target is smooth with minimal acceleration variation, and the point clouds acquired by the lidar exhibit high continuity. The deep learning feature network can extract well-structured, low-noise spatial geometric features from this stable observation sequence, resulting in a more concentrated probability distribution of the soft correspondence matrix and virtually no matching jumps. Because the trajectory motion law is approximately linear or weakly nonlinear, the continuous-time trajectory model can accurately fit the relative motion changes, and the motion compensation module can effectively eliminate the distortion of the point cloud from line-by-line scanning, ensuring that each frame of point cloud is corrected to a unified coordinate system. Ultimately, the registration error reaches an extremely low level, making it suitable for millimeter-level positioning and structural integrity analysis of critical maintenance components (such as refueling interfaces and attitude control actuators).

[0075] Example 2 (Multi-view complex structure scanning + sparse point cloud completion) When the target satellite has a complex external structure, such as deployable antennas, extended robotic arms, or irregular cabins, and the serving satellite performs high-frequency scanning from multiple angles during its orbital flight, the solution of this invention demonstrates superior optimal performance in multi-view fusion scenarios. In this scenario, multi-angle laser point clouds continuously complete different aspects of the complex structure. The deep feature network can extract complementary geometric representations from each viewpoint and adaptively fuse them within a unified soft correspondence framework, effectively overcoming problems such as occlusion, reflection differences, and insufficient cloud density due to local defects. Because the flight process exhibits continuous acceleration and deceleration characteristics, traditional frame-by-frame registration methods are prone to cumulative errors. However, this solution uses continuous time trajectory modeling to treat the entire orbital trajectory as a holistic optimization target, minimizing parallax distortion caused by complex motion in the unified optimization process. Ultimately, this solution maintains stable accuracy in multi-view 3D reconstruction and complex part localization tasks, and is particularly suitable for cooperative satellite maintenance and diagnostic tasks involving attitude instability, complex structures, or occlusion.

[0076] Example 1.3 (Instable target with attitude drift + noisy point cloud) Even in scenarios where the target satellite exhibits attitude drift, attitude disturbances, or even slight jitter, the proposed solution maintains optimal performance, demonstrating significant advantages, particularly under conditions of high point cloud noise, weak edge reflections, and blurred structural boundaries. If the target satellite's point cloud scan data becomes unstable over time due to attitude disturbances, the proposed solution's deep feature network can extract more reliable local structural features from noise and pseudopoints. Furthermore, the probability distribution of the soft correspondence matrix automatically weakens outliers, isolated points, or outliers, resulting in a more continuous and smooth geometric matching. Simultaneously, the continuous-time trajectory model unifies and jointly optimizes the target's attitude drift and the service satellite's maneuvers, minimizing the impact of drift on the registration process. Through a global error backpropagation mechanism, the proposed solution maintains robust registration stability even under high-noise conditions, making it highly suitable for tasks such as close-range verification of failed satellites, attitude anomaly monitoring, and close-range observation of space debris. Its output position and attitude results maintain high accuracy within the mission requirements even in drift scenarios.

[0077] Example 4 (Long-distance inspection with weak echoes + sparse point cloud modeling) When the service satellite conducts preliminary inspections at a distance of tens of meters from the target satellite, the proposed solution still achieves optimal 3D geometric reconstruction and relative position inference despite the low intensity of the laser echo, extremely sparse point cloud, and high noise level. In this weak signal environment, the structural semantics learned by the deep neural network enhances the features of the low-density point cloud, enabling the model to still identify key components such as the main fuselage structure, solar panel roots, and antennas. Simultaneously, the soft-correspondence probabilistic mechanism automatically assigns smaller weights to low-quality echoes, ensuring they do not affect the overall registration results. The continuous-time trajectory compensation mechanism is particularly effective for long-distance scanning, as even minute trajectory errors can cause point cloud deformation. This solution optimizes trajectory errors, registration errors, and compensation errors in a unified manner, allowing even sparse point clouds to generate structurally complete spatial models. This optimal state is highly suitable for early long-distance risk assessment, pre-approach spatial structure identification, and preliminary orbital service planning.

[0078] Using Example 1.5 (Cooperative Satellite Mission under Known Target Model Conditions) In collaborative satellite missions, if prior information such as the target satellite's main structural parameters, size range, and typical interface locations is known, the proposed solution exhibits extremely high optimal performance under this "model-based prior" condition. Since the prior model serves as a reference source for deep features, the network can converge to the true correspondence more quickly and accurately, significantly accelerating the registration optimization process. Simultaneously, the known orbit and attitude range provide strong constraints for continuous-time trajectory optimization, making motion compensation more accurate and reliable. In such scenarios, this solution can not only complete high-precision registration in a short time but also rapidly locate critical operational components, such as propellant loading ports, mechanical docking points, and solar panel locking mechanisms. This provides the most stable and efficient intelligent sensing capabilities for tasks such as on-orbit refueling, component replacement, optical window cleaning, and on-orbit maintenance, making it the optimal operating mode for collaborative satellite maintenance missions.

[0079] This invention may also include the following extensions: Multi-sensor fusion: Measurements from IMU, star sensors, visual cameras, etc., can be added as trajectory priors to improve compensation accuracy.

[0080] Online processing mode: Deep networks employ a distillation compression model, enabling service satellites to run in real time on low-power FPGAs / GPUs.

[0081] Scale estimation version: If the scanner has calibration error, the joint optimization can incorporate the solution of the scale factor.

[0082] Enhanced robustness to occlusion: The loss function is augmented with a Huber kernel or a Cauchy kernel to mitigate the impact of outliers.

[0083] After implementing this method, key components of the target satellite can achieve millimeter-level positioning accuracy and 0.1°–0.2° attitude accuracy in typical mission scenarios, meeting the requirements of various on-orbit servicing missions such as refueling and docking, mechanical maintenance, and debris removal. Through joint optimization of depth features and trajectory compensation, sparse point cloud registration errors can be significantly reduced, improving the reliability and safety of on-orbit servicing missions.

Claims

1. A method for intelligent optimization of sparse point cloud registration and dynamic compensation for non-cooperative satellites, characterized in that, Includes the following steps: S1. The service satellite flies along a preset orbital trajectory, collecting temporal laser point cloud data of the target satellite's surface using an onboard laser scanning device to obtain the original point cloud set. ,in, Indicates the first The three-dimensional spatial coordinates of a laser point in the coordinate system of the serving satellite; Indicates the first The acquisition timestamps corresponding to each laser point are used to characterize the temporal characteristics of the point cloud and support subsequent motion compensation calculations; The total number of laser point clouds acquired by the airborne laser scanning device within a scanning cycle or observation time window; The index number of a single laser echo point in the point cloud, used to identify the first... One laser point was collected; S2. Based on the known orbit and attitude prior information of the target satellite, construct a continuous-time relative motion trajectory model between the service satellite and the target satellite. The trajectory model is used to continuously describe the pose changes of the target satellite relative to the serving satellite in the time domain, wherein the position... and posture It is smoothly characterized in the time domain by differentiable basis functions; S3. Based on the continuous-time relative motion trajectory model constructed in step S2, apply the model to the original point cloud set obtained in step S1. Perform motion distortion compensation: for each point in the set From its collection time Unified transformation to the selected reference time To eliminate point cloud distortion caused by relative motion, the compensation transformation formula is as follows: The corrected compensation point cloud set is obtained. ; S4. The compensated point cloud set obtained in step S3... Input a pre-trained deep learning feature extraction network to extract high-dimensional deep feature vectors representing the local and global geometric structures of the point cloud. ; S5. Using the depth features extracted in step S4, in the compensated point cloud With the pre-acquired target satellite reference point cloud Construct a differentiable soft correspondence probability matrix between them. Matrix elements Characteristic points With point The matching probability is calculated; based on this soft correspondence probability matrix, the initial rigid body transformation parameters are solved by weighted least squares and singular value decomposition. This achieves initial alignment of the point cloud; S6. The trajectory model parameters from step S2, the compensation relationship from step S3, the depth features from step S4, the soft correspondence from step S5, and the exact rigid body transformation parameters to be determined are all incorporated into a unified, differentiable joint optimization framework. By minimizing a joint loss function L, gradient optimization is used to synchronously and iteratively update all the above parameters, enabling motion compensation and geometric registration to mutually correct and converge collaboratively, ultimately outputting the optimal continuous-time trajectory. and optimal precise registration transformation ; S7. The optimal continuous-time trajectory obtained in step S6 and optimal precise registration transformation By combining and transforming the known structural model of the target satellite, high-precision three-dimensional position and attitude information of its key functional components in the service satellite coordinate system can be analyzed and output.

2. The intelligent optimized sparse point cloud registration and dynamic compensation method for non-cooperative satellites according to claim 1, characterized in that, The joint loss function Defined as: , in, Based on the soft correspondence probability matrix Registration geometric error term; It is a constraint that ensures consistency between the point cloud and the trajectory model prediction after motion compensation; It is a smoothing regularization term applied to the trajectory parameters, where, For location control points, , These are the weighting coefficients.

3. The intelligent optimized sparse point cloud registration and dynamic compensation method for non-cooperative satellites according to claim 1, characterized in that, The basis function smoothing representation in step S2 specifically refers to: position ,in, For B-spline basis functions, Position control point; attitude Using quaternion splines Perform parameterization. This is the attitude control point.

4. The intelligent optimized sparse point cloud registration and dynamic compensation method for non-cooperative satellites according to claim 1, characterized in that, In step S6, the joint optimization framework employs an alternating optimization strategy or an end-to-end gradient descent strategy to optimize the trajectory parameters { , The registration parameters [R,t] are jointly updated, and the optimization process depends on the gradient backpropagation achieved by the differentiability of the compensation formula in step S3, the feature extraction network in step S4, and the soft correspondence calculation in step S5.

5. The intelligent optimized sparse point cloud registration and dynamic compensation method for non-cooperative satellites according to claim 1, characterized in that, In the trajectory model construction in step S2 or the joint optimization in step S6, measurement data from the inertial measurement unit and star sensor are further integrated as prior constraints or loss terms to form a trajectory estimation model with tight coupling of multi-source information.

6. The intelligent optimized sparse point cloud registration and dynamic compensation method for non-cooperative satellites according to claim 1, characterized in that, The elements of the soft correspondence probability matrix W Calculated using the Softmax function based on feature distance: In the formula, This is a soft correspondence probability matrix used to characterize the feature matching relationship between the compensated point cloud and the reference point cloud; Indicates the compensation point in the cloud. The point and the reference point cloud The soft matching weight or corresponding probability between points reflects the degree of similarity between them in the feature space. The first point extracted from the motion-compensated sparse point cloud by a deep neural network. The depth feature vector of each point; To extract features from a reference point cloud using the same feature extraction network The extracted first The depth feature vector of each point; The dimension of the feature vector; This is the feature distance sensitivity coefficient (temperature coefficient), used to adjust the influence of feature distance on the Softmax normalization process; when When the value is large, the algorithm pays more attention to point pairs with small feature distances, thereby enhancing the discriminative power of the matching; when When the value is small, the weight distribution tends to be smooth, which is beneficial to improving the stability of matching under conditions of high noise or sparse point cloud. Represents the Euclidean norm; This indicates that all candidate matching points in the reference point cloud are summed and normalized to ensure that the matching is applied to a fixed point cloud. ,have .

7. The intelligent optimized sparse point cloud registration and dynamic compensation method for non-cooperative satellites according to claim 1, characterized in that, In step S4, the deep learning feature extraction network is a point cloud network pre-trained with simulation data from on-orbit service scenarios, and its structure is selected from PointNet, PointNet++, DGCNN or graph convolutional networks.

8. The intelligent optimized sparse point cloud registration and dynamic compensation method for non-cooperative satellites according to claim 1, characterized in that, The reference point cloud The point cloud obtained by sampling the computer-aided design model of the target satellite, or the high-precision point cloud model generated by fusing historical on-orbit scanning data.

9. A spaceborne data processing device for on-orbit servicing, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the steps of the method as described in any one of claims 1 to 8.

Images