A method and system for fast ranging of a spatial moving target light comb based on a hierarchical guidance strategy

Abstract

The application relates to a method and system for fast ranging of a space mobile target light comb based on a hierarchical guiding strategy, and relates to the technical field of space mobile target ranging.The application is proposed in view of the problems that, in the traditional optical alignment before scanning of a mobile target, a preset global scanning mode is relied on, resulting in long alignment time, large energy consumption, and difficulty in meeting the demand of dynamic tracking ranging, etc.The system is divided into a laser emission scanning part, an image receiving and processing part and a processor part.The laser emission scanning part changes the laser light path by deflection of a fast mirror, and scans the reflection prism of a target surface; the image receiving and processing part reflects the image of the target reflection prism to a binocular camera through a small fast mirror, and then inputs the imaging result of the camera into an image processing module to solve the relative pose with the target; the main processor performs target motion estimation and scanning path planning according to the relative pose information, and inputs the reference path information into a fast mirror driving circuit to implement tracking control.Through a multi-sensor fusion perception algorithm, accurate target perception and positioning in a complex space environment are realized, the interference such as in-orbit platform vibration under microgravity conditions and space light condition change can be effectively coped with, and thus the robustness and accuracy of space target detection are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of space moving target ranging technology, and specifically to a method and system for rapid ranging of space moving targets using an optical comb based on a hierarchical guidance strategy. Background Technology

[0002] With the rapid development of my country's aerospace technology, spacecraft rendezvous and docking, satellite constellation formation, precision measurement of large space structures, and satellite antenna positioning are becoming increasingly common, placing higher demands on precise space ranging capabilities. Femtosecond laser optical frequency combs (or simply optical combs) are a novel broadband coherent light source. Their time-domain characteristics manifest as a series of femtosecond-level ultrashort pulses, while their frequency-domain characteristics exhibit a uniformly distributed comb-like spectrum. They possess advantages such as high stability, narrow pulse width, and high bandwidth, opening up a new technical path for high-precision absolute distance measurement and attracting widespread attention both domestically and internationally in the field of high-precision space ranging. In 2013, South Korea experimentally deployed an erbium-doped fiber optic comb light source to a near-Earth satellite, verifying its ability to operate under launch vibration, high-energy radiation, and thermal vacuum environments. In 2021, the FOKUS II dual-comb light source system developed by Germany underwent space payload testing, verifying its on-orbit self-starting capability. In 2022, the low-noise fiber optic comb light source developed by the Chinese Academy of Sciences, as a core component of the high-precision time and frequency experimental cabinet, was successfully carried on the Tiangong space station and conducted on-orbit testing, fully verifying its ability to provide high-precision time and frequency signals in the space environment.

[0003] Femtosecond laser comb ranging technology uses a comb laser as the light source. By linearly sampling two combs with a certain repetition frequency difference, it can simultaneously acquire large-scale time-of-flight information in the time domain and high-precision interferometric information in the frequency domain, thereby achieving high-precision, high-velocity absolute distance measurement. In 2018, the Karlsruhe Institute of Technology in Germany successfully achieved precise tracking of high-speed moving objects through wavelength interferometry using two microcavity combs. In 2021, Tsinghua University, through the combination of dual-comb ranging and spectral phase resolution technology, simultaneously measured the absolute distance and attitude angle information of a target, achieving a ranging accuracy of 13.7 nanometers and an angular accuracy of 0.43 microradians, demonstrating its enormous application potential in the field of high-precision distance and pose measurement in space.

[0004] The alignment speed and accuracy between the laser and the target reflecting prism directly affect the time efficiency and dynamic stability of optical comb ranging. In ground-based optical comb ranging experiments, the target's reflecting prism and the laser typically employ a fixed pre-alignment mode to improve the accuracy of static measurements. However, in space applications, the ranging platform and the target are often in relative motion, making it difficult to maintain a stable alignment. Current alignment methods require periodic global coverage of the plane containing the target reflecting prism using a pre-set multi-mode scanning strategy. However, this unguided blind scanning mechanism suffers from long scanning cycles and the tendency to miss targets, making it difficult to meet the real-time tracking and ranging requirements of dynamic targets. Therefore, how to effectively utilize the sensors on the on-orbit platform, guide and optimize the scanning path through active sensing technology, compress the search space, and construct a high-efficiency, high-precision dynamic scanning strategy is a technical challenge that urgently needs to be overcome. Summary of the Invention

[0005] The technical problem to be solved by this invention is:

[0006] This invention addresses the problems existing in the background technology by proposing a rapid laser scanning path planning method based on a hierarchical guidance strategy to achieve rapid alignment and tracking ranging of dynamic targets. The aim is to support the application of femtosecond optical comb ranging technology in the field of on-orbit precision ranging through this algorithmic innovation. This solves the problems of traditional optical alignment for moving targets relying on a preset global scanning mode, resulting in long alignment time, high energy consumption, and difficulty in meeting the needs of dynamic tracking ranging.

[0007] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0008] The system for rapid ranging of a space-moving target using an optical comb based on a hierarchical guidance strategy, as described in this invention, includes a laser emission and scanning section, an image receiving and processing section, and a processor section. The laser emission and scanning section includes a laser emitter, a large fast reflector, and a large fast reflector driving circuit. The image receiving and processing section includes a small fast reflector, a binocular camera, an image processing module, and a small fast reflector driving circuit. The laser emitter emits a laser beam for optical comb ranging, and the large fast reflector deflects the laser beam to change its direction, enabling the laser to scan the target platform and ultimately align with the reflecting prism mounted on it. The large fast reflector driving circuit provides low-level control, driving the fast reflector to accurately track the reference path given by the main processor. The small fast reflector reflects the image from the target reflecting prism to the binocular camera, and the binocular camera's imaging result is then input into the image processing module. The small fast reflector reflects external light to the binocular camera; it does not undergo large-angle deflection but is only used for fine-tuning and calibrating the reflected light path. The image processing module performs image stabilization, feature extraction and fusion on the image acquired by the binocular camera, and calculates the relative pose with the target. The main processor estimates the target motion and plans the scanning path based on the relative pose information, and inputs the reference path information into the fast-reflecting mirror drive circuit to implement tracking control.

[0009] Furthermore, the binocular camera is a D435i binocular camera; the main processor uses an NVIDIA Jetson Xavier NX high-performance edge computing module to receive relative pose information sent by the image processing module and execute trajectory prediction and path planning algorithms. A small fast-reflecting mirror reflects external light to both the binocular camera and the optical comb ranging module via a beam splitter. The image receiving and processing section also includes a lens group positioned between the beam splitter and the binocular camera. The small fast-reflecting mirror reflects external light through the beam splitter, and the reflected laser light passes through the lens group before entering the binocular camera. The lens group is used to improve optical quality, control exposure, and reduce glare.

[0010] The laser emission scanning section uses a fast-reflecting mirror to deflect the laser beam, changing the laser path to scan the reflective prism on the target surface. The image receiving and processing section uses a small fast-reflecting mirror to reflect the image of the target's reflective prism to a binocular camera, and then inputs the camera's imaging result into the image processing module to calculate the relative pose with the target. The main processor estimates the target motion and plans the scanning path based on the relative pose information, and inputs the reference path information into the fast-reflecting mirror drive circuit for tracking control. The method for rapid ranging of a space-moving target using an optical comb based on a hierarchical guidance strategy is implemented according to the following steps:

[0011] Step 1: Target Relative Pose Acquisition Based on Multi-Sensor Fusion. Given that on-orbit platforms are frequently affected by vibration and microgravity, which can lead to image quality degradation, an image stabilization algorithm is introduced. To accurately estimate inter-frame motion, a Kalman filter and visual-IMU fusion technique are used, and the pose output is optimized by analyzing the covariance matrix, thereby suppressing drift errors and ensuring the accuracy and stability of image information. Furthermore, since the quality of RGB images is highly susceptible to changes in lighting conditions in space, a D435i camera is used to simultaneously capture both RGB and infrared images of the target, providing additional visual information even in poor lighting conditions. Subsequently, this method employs the SIFT algorithm to efficiently extract feature points from both types of images, and feature matching is used to improve the recognition accuracy and robustness of image content.

[0012] Based on this, the relative pose of the camera and the target reflecting prism was accurately calculated using the PnP algorithm, based on the image after image de-shaking and the matched feature points. Furthermore, by combining the coordinate transformation matrix between the large fast-reflecting mirror of the laser emission module and the camera, the relative pose information between the fast-reflecting mirror and the reflecting prism can be calculated. Finally, to apply this relative pose information to actual control, it is projected into the two-dimensional space containing the XY-axis mechanical scanning angles of the large fast-reflecting mirror drive device, providing a data foundation for the deflection path planning of the fast-reflecting mirror.

[0013] Step 2: Relative Pose Prediction Based on Recursive Convolutional Conditional Neural Process Algorithm. Conditional Neural Processes (CNPs) are a class of deep neural networks describing Bayesian Gaussian processes, capable of predicting the distribution of unknown sequences from contextual data sequences. First, a Convolutional Neural Network (CNN) is used as the encoder, with the historical pose sequence near the current time as the encoder input. Local temporal information is extracted through a one-dimensional convolutional network and mapped into a latent vector h. Then, a Multilayer Perceptron (MLP) is used as the decoder, with the latent vector h and the timestamp of the query pose as input, ultimately outputting the Gaussian distribution of the relative poses of the reflecting prism and the large fast mirror at future time steps.

[0014] To further enhance the model's ability to model temporal features, a recurrent neural network (RNN) architecture is introduced into the outer layer of the CNP network, enabling CNP to predict pose information in the future multiple steps recursively. Simultaneously, negative log-likelihood (NLL) is used as the loss function to ensure that the predicted distribution maximizes the likelihood of the true data, thereby improving prediction accuracy.

[0015] Then, based on the predicted trajectory point positions and distribution information, a potential field distribution map within the workspace is calculated by performing a time-decayed weighted summation on the potential fields generated for each predicted point and adding the repulsive field relative to the obstacles. When calculating the potential field for each predicted point, the Euclidean distance within the potential field space and the probability density distribution of the predicted point itself need to be considered. Simultaneously, by assigning a time-exponential decay weight to each potential field, predicted points closer to the current time receive greater weight, thus generating a time-dependent potential field map. Furthermore, considering the incident angle dead zone of the target reflecting prism, there may be areas within the workspace that cannot be effectively scanned. We set higher potential field values ​​for these areas to avoid energy loss due to invalid scanning.

[0016] Step 3: Dynamic Programming Algorithm for B-Spline Coverage Path Based on Particle Swarm Optimization. The artificial potential field map obtained in Step 2 reflects the probability of the target reflecting prism appearing in the workspace. To improve scanning efficiency, we only scan regions where the potential field is below a certain threshold. However, due to the complexity and diversity of potential field changes, these scanned regions below the threshold often exhibit irregular shapes. To simplify the complexity of coverage path planning, we first decompose the scanned region into multiple simple and non-overlapping working unit regions centered on different prediction points. These regions constitute the workspace for coverage path planning. The working units are divided according to the equipotential lines near the potential field threshold of each prediction point. However, since the potential field changes significantly between different prediction points, different equipotential lines are referenced for the division of different working units to ensure sufficient workspace for path planning.

[0017] The coverage path is constructed using spirals drawn from B-spline curves, and the planning process consists of two steps: pre-planning and online optimization. The main purpose of pre-planning is to construct irregularly spaced spirals covering the workspace. We represent the control points of the B-spline in polar coordinates and generate spirals that meet the requirements by adjusting the length coordinate spacing of control points with the same radian. In online optimization, the main indicators of the optimal coverage path are spiral density, coverage area, and path length. Spiral density ensures that areas with a high probability of target occurrence receive more detailed scanning; the coverage area of ​​the spiral should cover the work cell as much as possible; and the path length determines the scanning time—an excessively long path can lead to scanning not being completed when the target reflecting prism moves out of the work cell. Online optimization starts with the spiral control points generated by pre-planning. Using a particle swarm optimization algorithm, the polar coordinates of each control point and the number of spiral rotations are fine-tuned, and a loss function is constructed based on spiral density, path length, and coverage area. Finally, the optimal coverage path within each work cell is obtained through optimization.

[0018] However, the trajectory prediction based on the neural network in step two cannot be updated in real time. Therefore, before the prediction result is updated, the original potential field needs to be updated according to the target pose updated by the vision module, and the working units and planned paths need to be readjusted. If the original predicted path is relatively accurate, the new path planning can only fine-tune the old B-spline control points. Finally, the optimized best path is sent to the fast-reflection mirror drive module as the tracking target for scanning.

[0019] The present invention has the following beneficial technical effects:

[0020] This invention effectively solves the time-consuming and energy-intensive problem of optical alignment and scanning processes in traditional optical comb ranging. It proposes a high-efficiency, fast laser scanning coverage path planning algorithm based on a hierarchical guidance strategy for rapid ranging of optical combs (lasers) targeting moving targets in space. The hierarchical guidance strategy employed in this invention enables efficient and rapid laser scanning path planning for optical comb ranging of moving targets in space. The specific beneficial effects of this invention are as follows:

[0021] 1. Through multi-sensor fusion perception algorithm, accurate target perception and positioning in complex space environments are achieved. It can effectively cope with interference such as on-orbit platform vibration under microgravity conditions and changes in space lighting conditions, thereby greatly improving the robustness and accuracy of space target detection.

[0022] 2. By employing a recursive convolutional conditional neural process algorithm, the pose of a target reflecting prism over a future period was predicted. The introduction of convolutional layers and a recursive neural network architecture further enhanced the neural network's ability to extract temporal features from the data. Using negative log-likelihood as the loss function, accurate and reliable predictions were achieved over large timescales based on small sample training data.

[0023] 3. Accurate and efficient scanning path generation was achieved through a layered coverage path planning algorithm. Mapping based on an artificial potential field comprehensively considered the relative distance of predicted points, probability density distribution, time decay, and laser scanning blind zone, accurately describing the probability of the target reflecting prism's location. By introducing a working cell segmentation method, the complexity of the planning problem within a complex workspace was simplified, and path repetition was avoided. B-spline curves were used to construct the path, and pre-planning and particle swarm optimization algorithms were used to adjust the control points, improving planning efficiency and effectiveness. Attached Figure Description

[0024] Figure 1 This is a diagram of the system's components and operational structure.

[0025] Figure 2 This is a block diagram of the image processing module.

[0026] Figure 3 This is a diagram illustrating the effect of multi-sensor fusion positioning.

[0027] Figure 4 This is a diagram of the recursive convolutional conditional neural network structure.

[0028] Figure 5 The training effect diagram of the recursive convolutional conditional neural network;

[0029] Figure 6 The distribution of predicted trajectories for recursive convolutional conditional neural processes is shown in the diagram.

[0030] Figure 7 Here is a block diagram of the hierarchical guided planning algorithm structure;

[0031] Figure 8 The diagram shows the structure and reflected light path of a reflecting prism (cornerstone prism).

[0032] Figure 9 This is a diagram of the artificial potential field generated based on the trajectory prediction distribution.

[0033] Figure 10 The diagram shows the effect of the layered coverage path planning algorithm. Detailed Implementation

[0034] Combined with appendix Figure 1-10 The present invention provides a detailed description of the efficient and rapid coverage path planning system and method for space laser rapid ranging based on a hierarchical guidance strategy.

[0035] The efficient and rapid coverage path planning system based on a hierarchical guidance strategy for rapid space laser ranging described in this invention, such as... Figure 1The diagram illustrates the apparatus and operational block diagram of the system involved in this invention. The entire system is divided into a laser emission and scanning section, an image receiving and processing section, and a processor section. The laser emission and scanning section consists of a laser emitter, a large fast reflector, and a large fast reflector drive circuit. The laser emitter emits a laser beam for optical comb ranging; the large fast reflector changes the direction of the laser beam by deflecting itself, enabling the laser to scan the target platform and ultimately align with the reflecting prism mounted on it; the large fast reflector drive circuit is used for low-level control, driving the fast reflector to accurately track the reference path given by the main processor. The image receiving and processing section consists of a small fast reflector, a lens group, a D435i binocular camera, an image processing module, and a small fast reflector drive circuit. The small fast reflector reflects external light to the binocular camera and optical comb ranging module. The small fast reflector does not perform large-angle deflection, but is only used for fine-tuning and calibrating the reflected light path; the lens group improves optical quality, controls exposure, and reduces glare; the image processing module performs image stabilization, feature extraction and fusion, and relative pose calculation on the images acquired by the binocular camera. The main processor uses the NVIDIA Jetson Xavier NX high-performance edge computing module to receive relative pose information sent by the image processing module and execute related algorithms such as trajectory prediction and path planning.

[0036] To further explain the function of the large fast reflector: its basic function is to deflect a laser beam, enabling it to scan the target surface. When the laser beam scans the reflecting prism on the target surface, it is reflected back to the small fast reflector, and then reflected again by the small fast reflector to the optical comb ranging module. When the optical comb ranging module detects the reflected laser beam, it sends a command to the onboard computer. Upon receiving the command, the onboard computer controls the large fast reflector's drive circuit to stop scanning, thus keeping the laser beam aligned with the reflecting prism on the target platform. However, in reality, two spacecraft platforms in orbit cannot remain relatively stationary indefinitely. Therefore, the guidance algorithm (method section) described in this invention is needed to assist the large fast reflector in quickly aligning with the reflecting prism and shortening the time interval between two alignments. The main purpose of the system provided by this invention is to shorten the time required for the large fast reflector to deflect the laser beam and align it with the reflecting prism.

[0037] The optical comb ranging module can be implemented using existing technology; however, it is described here for the sake of completeness in the system description. The laser light reflected from the target reflecting prism is reflected by a small fast-reflecting mirror and split into two paths by a beam-splitting prism. One path passes through the lens assembly and enters the binocular camera, while the other path enters the optical comb ranging module, as shown below. Figure 1 As shown. The main objective of the system provided by this invention is to shorten the time required for a large fast-reflecting mirror to align a deflecting laser with a reflecting prism.

[0038] The implementation process of the efficient and rapid coverage path planning method based on a hierarchical guidance strategy for rapid space laser ranging described in this invention is as follows:

[0039] Step 1: Precise target perception and localization in complex spatial environments based on multi-sensor fusion sensing algorithms.

[0040] like Figure 2 As shown, the device and flowchart of the image processing module are illustrated. First, the information of the IMU and the camera is fused to compensate for the image jitter caused by the vibration of the on-orbit platform. Then, the infrared camera and the RGB camera are fused to adapt to changes in spatial illumination. Finally, the relative pose is calculated based on the coordinates of the feature points.

[0041] The IMU mounted on the camera platform can measure the platform's acceleration and angular velocity. Simultaneously, the image's motion information can be described by the x-direction translation pixels dx, the y-direction translation pixels dy, and the rotation angle da. By fusing the high-frequency components of both motion signals using a Kalman filter, the platform's vibration estimate can be obtained. Let the fused state estimate be T, then the fusion process can be expressed as:

[0042] T k|k =T k|k-1 +K k (z k -H·T k|k-1 )

[0043] Where x k|k-1 It is the predicted state, K k It is the Kalman gain, z k Here, H represents the observed values, and H is the observation matrix. Subsequently, the image pixels captured by the camera can be compensated based on the vibration estimation to eliminate image distortion caused by platform vibration.

[0044] T compensated =TT motion

[0045] The SIFT algorithm is a widely used algorithm for image feature extraction and matching, capable of extracting stable local feature points at different scales and rotation angles. Using image information acquired by a D435i binocular camera, we performed feature extraction using the SIFT algorithm to obtain feature point descriptors in the RGB and infrared images, represented as follows:

[0046] D RGB =SIFT(Image) RGB ),D IR =SIFT(Image) IR )

[0047] Next, by performing feature-level fusion on the feature points of the two, a joint feature vector set D is obtained. fusion =[D RGB D IRThis feature set integrates all feature information collected by RGB and infrared cameras, which can significantly improve the accuracy and stability of target recognition under varying spatial lighting conditions.

[0048] The PnP algorithm can estimate the camera pose using several known 3D coordinate points and their 2D projections onto the image. First, the 3D coordinates P of the preset calibration points on the target platform are obtained. i And obtain two-dimensional coordinates p through the camera i Then, the camera's extrinsic parameters, i.e., the rotation matrix R, are solved using direct linear transformation, EPNP, or Lvenberg-Marquardt methods. c and T c 3D coordinates P i To two-dimensional coordinates p i The projection relationship is as follows:

[0049] p i =K·[R c |T c ]·P i

[0050] Here, K is the known intrinsic parameter matrix of the camera, which mainly contains information such as the camera's focal length and principal point.

[0051] Figure 3 The positioning accuracy of RGB cameras, infrared cameras, and IMU-based visual fusion was compared. It can be seen that inherent platform vibrations cause high-frequency noise in the positioning results of both types of cameras. Furthermore, the RGB and infrared cameras are susceptible to the effects of lighting and ambient temperature, respectively, leading to errors in their independent positioning results. By combining IMU-visual anti-shake technology with RGB-IR image feature fusion, the accuracy and stability of visual positioning were improved.

[0052] Step 2: Predict trajectory information of small samples over a large time scale based on recursive convolutional conditional neural processes.

[0053] This invention proposes, as follows Figure 3 The recurrent convolutional neural network architecture shown consists of a CNP network with a convolutional layer encoder that predicts the trajectory point at the next moment based on historical trajectories, and achieves multi-step prediction of the future through the recurrent neural network framework. Wherein, S... t Pos represents the timestamp at time t. t ,Ang t These represent the position and attitude angle predicted by the network at time t, respectively. The prior data consists of timestamps and pose information from time t-1 to tM. t The mean μ and variance σ of the pose distribution at time t predicted after being jointly input into the CNP network. 2 .

[0054] The Conditional Neural Process (CNP) method combines Bayesian Gaussian processes and deep neural networks. It leverages prior knowledge and is trained using a neural network framework to achieve excellent prediction results with a small dataset and a relatively small model. The network consists of three parts: an encoder, an aggregator, and a decoder.

[0055] The encoder network maps the conditional input (historical data) to a latent spatial representation. To improve the extraction of temporal features, the historical data is first passed through a one-dimensional convolutional layer, and then the result is input into a multilayer perceptron (MLP). If the historical data is represented by D... hist ={S t-1 ,…,S t-M Pos t-1 ,…,Pos t-M ,Ang t-1 ,…,Ang t-M If we use} to represent this, then the forward propagation process of the encoder can be represented as:

[0056] Z = h θ (CONV(D hist ))=h θ (D hist *W)

[0057] Where * denotes a convolution operation, W is a one-dimensional convolution kernel, CONV(·) represents a convolutional layer, and h θ (·) represents a multilayer perceptron, θ represents the network weights, and Z is a latent space representation generated by the encoder based on the data.

[0058] An aggregator performs aggregation operations (such as weighted summation) on the encoder output Z to generate a unified feature representation C.

[0059]

[0060] The decoder is based on feature C and query set. To generate the predicted output, where the query set is the set of timestamps corresponding to the unknown sequence, the forward propagation process of the decoder can be represented as:

[0061]

[0062] Among them, Y pred The Gaussian distribution for predicting pose is represented by the mean and variance. The loss function of the CNP network can be transformed into an optimization problem, represented by the negative log-likelihood (NLL):

[0063]

[0064] Where p(·|·) is the conditional probability, representing the probability that CNP is based on historical data D.hist and query set The predicted distribution and the actual data Y true The likelihood.

[0065] Finally, the single-step convolutional CNP network was substituted into the recurrent neural network architecture, and the predicted output of the previous step was added to the conditional input of the current step to predict the trajectory in an iterative manner, which further improved the accuracy and stability of the prediction.

[0066] Figure 5 The loss function curve of the recurrent convolutional CNP network is shown. As the number of training rounds increases, the model gradually converges and the prediction accuracy gradually improves. Figure 6 The prediction performance of the recurrent convolutional CNP network is shown. The blue curve represents the conditional input consisting of 20 sets of historical poses, the red curve and the shaded area represent the Gaussian distribution of the predicted trajectory, and the green represents the real pose data. The real data are all located near the predicted data and contained within the uncertainty region, indicating that the network has good prediction accuracy.

[0067] Step 3: An efficient hierarchical coverage path planning algorithm based on particle swarm optimization.

[0068] Figure 7 The block diagram of the path planning algorithm proposed in this invention is shown. First, based on the trajectory prediction generated by the recursive convolutional CNP network, an artificial potential field is constructed based on Euclidean distance, probability distribution and time series information. Then, the optimal scanning path is generated through working unit decomposition and coverage path planning. Finally, the local potential field is dynamically updated based on visual recognition information.

[0069] Artificial potential field calculation is divided into the gravitational field U formed by the position of the moving target. att The repulsive field U formed by the movement obstacle rep and the time decay W caused by the temporal characteristics of the predicted trajectory t It consists of three parts. The purpose of the gravitational field is to attract the laser to scan towards the target location. First, the gravitational force U generated by the Euclidean distance from the workspace to the i-th prediction point is represented by a piecewise function. dist ,when To reduce potential energy and avoid excessive gravity when the target location is far away:

[0070]

[0071] Then, since the network output is a Gaussian distribution of trajectory points, a probability density is introduced as a weight on the basis of the Euclidean distance potential field to account for the impact of prediction uncertainty, specifically expressed as follows:

[0072]

[0073] Where, μi,goal and σ i,goal Let represent the mean and variance of the Gaussian distribution of target point i, respectively, and ζ be the weighting coefficient.

[0074] Next, the potential fields calculated based on n predicted points in the predicted trajectory are superimposed, while a time decay weight is introduced to ensure that the potential field of the predicted points closer to the current time has a stronger influence. The formula is expressed as:

[0075]

[0076] Where δ represents the rate at which the weight of the potential field calculated based on each prediction point decays over time.

[0077] like Figure 8 The image shows a reflective prism mounted on a target platform. It is a pyramidal prism composed of four faces: three are mutually perpendicular right-angled faces, and the fourth face is the base face, with the angle between it and the other three faces being equal. Regardless of the point on the base face from which the light rays enter, they will ultimately exit from the base face, and the exiting rays are parallel to and opposite in direction to the incident rays. This prism is widely used for laser beam ranging. However, due to limitations in its physical shape and installation location, when the laser light enters the reflective prism at an angle greater than 45°, it cannot be correctly reflected, resulting in a reduced effective scanning angle range. If scanning is still performed according to the ideal working space, it will lead to ineffective energy loss. Therefore, it is necessary to treat the failure area as an equivalent obstacle and introduce a repulsive field.

[0078]

[0079] Where D(q) is the distance to the nearest obstacle, η is the repulsive force increment, and Q * This is the threshold range of the obstacle's influence. Within this threshold range, the obstacle will generate a repulsive force; otherwise, it has no effect. Superimposing the gravitational and repulsive fields forms the final artificial potential field U(m) = U. att (m)+U rep (m).

[0080] like Figure 9As shown, an artificial potential field map generated based on trajectory prediction is displayed. Green markers represent trajectory points corresponding to past time steps, while red markers represent target positions predicted by the recurrent convolutional CNP network at future moments. The corresponding time order is the sequence of the markers in the figure (e.g., number I represents the predicted position at the next moment, and number IV represents the predicted position four time steps later). The dark red shaded area represents the invalid laser scanning area at the current moment (the laser in this area has an incident angle greater than 45° to the reflecting prism and cannot be reflected normally). From the potential field map, it can be seen that the prediction point I closest to the current moment has the lowest potential energy, indicating that this is the area where the target reflecting prism is most likely to appear. Predictions further away from the current moment, such as prediction point V, have higher potential energy, indicating that the probability of the target appearing near these areas is lower. Because it includes time weights, the potential field map needs to be dynamically updated over time. However, due to the long prediction cycle of the neural network, before the predicted trajectory is updated, local potential field optimization adjustments need to be made based on time and visual information, building upon the existing potential field. Simultaneously, to compensate for the deviation between the predicted trajectory and the true trajectory, the Gaussian Process Regression (GPR) algorithm is used. Based on the deviation between the historical predicted trajectory and the true trajectory obtained by the vision module, it predicts and compensates for the future prediction bias of the neural network, thereby improving prediction accuracy. Due to its non-parametric structure, Gaussian Process Regression has a fast prediction time, which can be used to compensate for the prediction error of the network model in real time. Moreover, its prediction results are also Gaussian distributed. Its prediction process can be expressed as:

[0081]

[0082] Where K represents the covariance matrix between different datasets, E train The training set E represents the prediction bias. test The estimated value representing the prediction bias, and the corrected prediction trajectory can be represented as:

[0083]

[0084] like Figure 10 As shown in (a), the results of segmenting working units based on artificial potential fields are illustrated. Green markers represent target locations recorded in historical data, while red markers represent future target locations predicted by the recurrent convolutional CNP network. The corresponding time order corresponds to the sequence number of the markers in the figure. The blue covered area represents the working units segmented based on the artificial potential field, and gray represents invalid laser scanning areas. To improve scanning efficiency, scanning path planning is only performed in areas where the potential field is below a specific threshold. Simultaneously, working units for path planning are segmented based on the potential field surrounding the trajectory points to avoid path repetition. The boundaries of the working units are segmented based on a pre-set potential field threshold, but B-splines are used to construct the boundary lines to ensure smoothness. Furthermore, to improve the scanning efficiency for data points far from the current time, such as… Figure 10 (a) To improve the efficiency of local path optimization for the working units to which predicted points III and IV belong, the boundary thresholds of these working units need to be increased to ensure that path pre-planning for high-potential-energy working units can cover a certain range. In practical applications of the algorithm, the number of working units corresponding to future trajectory points needs to be adjusted based on hardware computing efficiency and potential field distribution calculation results.

[0085] To achieve efficient and reliable path planning, the coverage path planning within each work unit is carried out in three steps: path pre-planning, path optimization, and dynamic replanning. Specific measures are described below:

[0086] The results of path pre-planning are as follows Figure 10 As shown in (b), based on the scope of the work unit, equally spaced, fully covered B-spline spirals were pre-planned. The purpose of pre-planning is to address irregular workspaces (such as...). Figure 10 (a) generates a uniform and fully covered B-spline spiral. During scanning, the laser will be reflected back to the small fast-reflecting mirror as long as it falls on the effective reflection area of ​​the target reflecting prism. Therefore, the coverage area of ​​each path point on the working space is the effective reflection area of ​​the reflecting prism at that time. The coverage radius r of the path point can be approximately estimated based on the distance between the target reflecting prism and the launching platform. scan .

[0087] The control points of the B-spline spiral are generated in polar coordinates, with the geometric center of each working unit as the center and the positive Y-axis as 0 radians. A control point is set at every ω radians. The shape of each spiral is then determined by… The number of control points determines the polar coordinate representation of the control points as: r(θ) = nΔr(θ), Δr(θ) ≤ r scan

[0088] Where θ represents the angle between the line connecting the control point and the geometric center of the working unit and the positive direction of the Y-axis, Δr(θ) represents the distance difference between all control points radially along angle θ, n represents the number of turns of the spiral, and the maximum radial distance from the geometric center of the working space to the boundary is r, along with the path point coverage radius. scan The ratio determines the total number of control points for the B-spline spiral within each working unit. Particle swarm optimization, such as Figure 10 As shown in (c): Although the spiral path generated in the previous step can achieve uniform coverage of the working unit, the target position is constantly changing and uncertain. Therefore, appropriate repeated scanning in areas with high probability of occurrence can improve the alignment probability. Since the artificial potential field can fully characterize the probability distribution of the target's location, weighted integration of the coverage area based on the artificial potential field intensity can increase the scanning density in high-probability areas.

[0089]

[0090] Among them, A map Let U(r,θ) be the actual physical area of ​​the working cell, and let U(r,θ) be the intensity of the artificial potential field. The constants δ1 and δ2 determine the influence of the potential field. total This represents the weighted area of ​​the work unit.

[0091] In addition to considering the coverage of the weighted area, the optimization algorithm also needs to consider the total path length and the actual overlap area. The cost function is shown below:

[0092]

[0093] Where Γ1, Γ2, Γ3 are the weights of the three cost terms, A cover This represents the area covered by all scan points along the path.

[0094] The control point positions of the B-spline spirals corresponding to each working unit are optimized using a hierarchical particle swarm optimization algorithm. The outer layer particle swarm optimization optimizes the number of spiral coils n. loop and the number of control points added / reduced n newp The number of particles is N outer The inner particle swarm optimization optimizes the positions of all control points P = {(r1,θ1),(r2,θ2),…,(r...} n ,θ n The number of particles is N. inner To improve optimization efficiency, the inner and outer layers have different optimization frequencies; that is, the inner layer particle swarm optimization algorithm iterates M times, while the outer layer iterates only once. Before optimization begins, the starting point for each particle's position vector is randomly set near the pre-planned path parameters in a normal distribution. The position update formula for each particle swarm is:

[0095]

[0096] in, Let represent the position vector of particle i at time t. Let represent the velocity vector of particle i at time t+1. The velocity update formula for the particle swarm in each iteration is:

[0097]

[0098] Where w is the inertia weight, c1 and c2 are acceleration constants, representing the particle's dependence on its own experience and the group's experience, respectively, and r1 and r2 are random numbers between [0,1] used to increase the randomness of the search. It is the optimal position of particle i, g t This is the globally optimal position. After multiple iterations, the particle swarm can find the optimal solution within the search space. For example... Figure 10As shown in (c), the optimized path coverage density tends to decrease radially from the predicted coordinate point towards the boundary of the working unit, which ensures that areas with a high probability of the target reflecting prism receive more scans.

[0099] Dynamic replanning: When reaching the time corresponding to the next predicted trajectory point, in addition to modifying the time decay weight of the potential field corresponding to each trajectory point, the potential field is also fine-tuned based on the actual target position and invalid scan areas calculated by the vision module. After the potential field is updated, path planning is re-performed based on the new potential field. If the predicted trajectory point position and its surrounding potential field change significantly, pre-planning is performed first, and then particle swarm optimization is performed starting from the pre-planned B-spline spiral control point; otherwise, particle swarm optimization is performed starting from the B-spline spiral control point of the previous time step.

[0100] The method described in this invention has been verified through simulation experiments and practical applications, demonstrating the technical effects claimed by this invention. It solves the technical problems proposed by this invention, and the method described in this invention, through practical application, verifies the claimed technical effects and practicality. Traditional laser alignment methods rely on global cyclic scanning, and the alignment process has considerable randomness in terms of time consumption. This invention, however, uses a binocular camera to accurately identify the target reflecting prism and, through active guidance and path planning, drives the laser to quickly approach areas with high target occurrence rates, effectively improving the efficiency of laser alignment. Furthermore, traditional methods require a complete global scan after each alignment, while this invention, based on historical data, predicts the future position and trajectory of the target reflecting prism and pre-plans the scanning path, greatly shortening the time interval between two alignments.

[0101] The method (algorithm) for rapid ranging of space moving targets based on a hierarchical guidance strategy proposed in this invention is the underlying technical core of this invention, and various products can be derived based on the algorithm.

[0102] Based on the algorithm (method) proposed in this invention, a software system for rapid ranging of a space moving target using an optical comb based on a hierarchical guidance strategy is developed using a programming language. This system has program modules corresponding to the steps of the above-mentioned technical solution, and executes the steps in the above-mentioned rapid ranging method for a space moving target using an optical comb based on a hierarchical guidance strategy when running.

[0103] The developed system (software) computer program is stored on a computer-readable storage medium. This computer program is configured to, when called by a processor, implement the steps of the aforementioned method for rapid ranging of a space-moving target using an optical comb based on a hierarchical guidance strategy. In other words, the invention is materialized on a carrier, becoming a computer program product.

[0104] A rapid ranging device for space moving targets based on a hierarchical guidance strategy is disclosed. The device includes at least one processor and a memory communicatively connected to the at least one processor. The memory stores instructions executable by the at least one processor. The instructions are executed by the at least one processor to enable the at least one processor to perform the aforementioned rapid ranging method for space moving targets based on a hierarchical guidance strategy, thereby achieving rapid ranging of space moving targets.

[0105] Various implementations of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, application-specific integrated circuits (ASICs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include: implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transferring data and instructions to the storage system, the at least one input device, and the at least one output device.

[0106] The computational programs (also referred to as programs, software, software applications, or code) of this invention include machine instructions of a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device PLD) for providing machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0107] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A system for rapid ranging of a moving space target using an optical comb based on a hierarchical guidance strategy, characterized in that, The system includes a laser emission scanning section, an image receiving and processing section, and a processor section; The laser emission and scanning section includes a laser emitter, a large fast mirror, and a large fast mirror drive circuit; the image receiving and processing section includes a small fast mirror, a binocular camera, an image processing module, and a small fast mirror drive circuit. The laser emitter is used to emit a laser beam for optical comb ranging. The deflection of the large fast mirror changes the direction of the laser beam, enabling the laser to scan the target platform and finally align with the reflecting prism mounted on it. The large fast mirror drive circuit is used for low-level control, driving the fast mirror to accurately track the reference path given by the main processor. The image of the target reflecting prism is reflected to the binocular camera by a small fast-reflecting mirror, and the imaging result of the binocular camera is then input into the image processing module. The small fast-reflecting mirror is used to reflect external light to the binocular camera. The small fast-reflecting mirror does not perform large-angle deflection, but is only used for fine adjustment and calibration of the reflected light path. The image processing module is used to perform image stabilization, feature extraction and fusion, and to calculate the relative pose with the target in the image acquired by the binocular camera. The main processor estimates the target motion and plans the scanning path based on the relative pose information, and inputs the reference path information into the fast-reflecting mirror drive circuit to implement tracking control.

2. The system for rapid ranging of a space-moving target based on a hierarchical guidance strategy using an optical comb, as described in claim 1, is characterized in that: The stereo camera is a D435i stereo camera; The main processor uses the NVIDIA Jetson Xavier NX high-performance edge computing module to receive relative pose information sent by the image processing module and execute trajectory prediction and path planning algorithms.

3. The system for rapid ranging of a space-moving target based on a hierarchical guidance strategy using an optical comb, as described in claim 1 or 2, is characterized in that: The small fast-reflecting mirror uses a beam splitter to reflect external light to the binocular camera and the optical comb ranging module.

4. The system for rapid ranging of a space-moving target based on a hierarchical guidance strategy using an optical comb, as described in claim 3, is characterized in that... The image receiving and processing section also includes a lens group set between the beam splitter and the binocular camera. The small fast-reflecting mirror reflects external light through the beam splitter, and the laser passes through the lens group before entering the binocular camera. The lens group is used to improve optical quality, control exposure, and reduce glare.

5. A method for rapid ranging of a moving spatial target using an optical comb based on a hierarchical guidance strategy, characterized in that: the method includes: Step 1: Precise target perception and localization in complex spatial environments based on multi-sensor fusion sensing algorithms An image de-shake algorithm is introduced to ensure the quality of the acquired spatial moving target images. Kalman filter and visual-IMU fusion technology are used, and the pose output is optimized by analyzing the covariance matrix to suppress drift error. The D435i camera is used to capture the RGB image and infrared image of the target simultaneously. The SIFT algorithm is used to efficiently extract feature points in these two images, and feature matching is used to improve the recognition accuracy and robustness of image content. Based on the image after image de-shaking and the matched feature points, the relative pose between the camera and the target reflecting prism is accurately calculated using the PnP algorithm. By combining the coordinate transformation matrix between the large fast reflecting mirror of the laser emission module and the camera, the relative pose information between the fast reflecting mirror and the reflecting prism can be calculated. Based on the calculated three-dimensional relative pose, the mechanical scanning angle required to drive the large fast reflecting mirror to deflect along its XY axis is calculated, providing a data basis for the deflection path planning of the fast reflecting mirror. Step 2: Predicting trajectory information from small samples over large timescales based on recursive convolutional conditional neural processes First, a convolutional neural network (CNN) is used as the encoder, with the historical pose sequence near the current time as the encoder input. Local temporal information is extracted through a one-dimensional convolutional network and mapped into a latent vector h. Then, a multilayer perceptron (MLP) is used as the decoder, with the latent vector h and the timestamp of the pose to be queried as input, and finally outputs the Gaussian distribution of the relative poses of the reflecting prism and the large fast mirror at future time. A recurrent neural network (RNN) architecture is introduced into the outer layer of the CNP network, enabling CNP to predict pose information for multiple future steps in a recursive manner. At the same time, negative log-likelihood (NLL) is used as the loss function to ensure that the predicted distribution can maximize the likelihood of the real data, thereby improving the accuracy of the prediction. Then, based on the predicted location and distribution information of the trajectory points, the potential field distribution map in the workspace is calculated by performing a time-decayed weighted summation on the potential field generated from each predicted point and adding the repulsive force field relative to the obstacle. When calculating the potential field of each predicted point, the Euclidean distance in the potential field space and the probability density distribution of the predicted point itself need to be considered. At the same time, by setting a time exponential decay weight for each potential field, the predicted points that are closer to the current time are given a greater weight, thereby generating a time-dependent potential field map. Considering that the target reflecting prism has an incident angle dead zone, there may be areas in the workspace that cannot be effectively scanned. A higher potential field value is set for these areas to avoid energy loss caused by invalid scanning. Step 3: Dynamic Programming of B-Spline Coverage Path Based on Particle Swarm Optimization Algorithm The artificial potential field map obtained in step two is used to reflect the probability of the target reflecting prism appearing in the workspace. It is only scanned in areas where the potential field is below a specific threshold. First, the scanning area is decomposed into multiple simple and non-overlapping working unit areas centered on different prediction points. These areas constitute the workspace covering the path planning. The working units are divided according to the potential field threshold equipotential lines near each prediction point. The division of different working units refers to different equipotential lines. The coverage path is constructed using spirals drawn from B-spline curves. The planning process consists of two steps: pre-planning and online optimization. Pre-planning constructs irregularly spaced spirals that cover the workspace, representing the control points of the B-spline in polar coordinates. Spirals that meet the requirements are generated by adjusting the length coordinate spacing of control points with the same radian. In online optimization, the main indicators of the optimal coverage path are spiral density, coverage area, and path length. Spiral density ensures that areas with a high probability of target occurrence receive more detailed scanning. The coverage area of ​​the spiral should cover the work cell as much as possible. The starting point of online optimization is the spiral control points generated by pre-planning. By using the particle swarm optimization algorithm, the polar coordinates of each control point and the number of spiral loops are fine-tuned. A loss function is constructed based on spiral density, path length, and coverage area, ultimately optimizing the optimal coverage path within each work cell. Before the prediction results are updated, the original potential field needs to be updated based on the target pose updated by the vision module, and the working units and planned paths need to be readjusted. If the original predicted path is relatively accurate, the new path planning can only fine-tune the old B-spline control points; finally, the optimized best path is sent to the fast-reflection mirror driver module for scanning operation as the tracking target.

6. The method for rapid ranging of a space-moving target using an optical comb based on a hierarchical guidance strategy according to claim 5, characterized in that: The precise target perception and localization in complex spatial environments based on multi-sensor fusion sensing algorithms described in Step 1 is specifically implemented as follows: First, IMU and camera information are fused to compensate for image jitter caused by on-orbit platform vibration. Then, infrared and RGB cameras are fused to adapt to changes in spatial illumination. Finally, the relative pose is calculated based on the coordinates of the feature points. The IMU mounted on the camera platform can measure the platform's acceleration and angular velocity. Simultaneously, the image's motion information is described by the x-direction translation pixels dx, the y-direction translation pixels dy, and the rotation angle da. By fusing the high-frequency components of both motion signals using a Kalman filter, a vibration estimate of the platform can be obtained. Let the fused state estimate be T, then the fusion process can be expressed as: in It is a predicted state. It is Kalman gain. Here, H represents the observed values, and H is the observation matrix. Subsequently, the image pixels captured by the camera can be compensated based on the vibration estimation to eliminate image distortion caused by platform vibration. Using image information acquired by the D435i binocular camera, feature extraction was performed using the SIFT algorithm to obtain feature point descriptors in the RGB and infrared images, which are represented as follows: Next, by performing feature-level fusion on the feature points of the two, a joint feature vector set is obtained. This feature set integrates all feature information collected by RGB and infrared cameras to improve the accuracy and stability of target recognition under varying spatial lighting conditions. The PnP algorithm estimates the camera pose using several known 3D coordinate points and their 2D projections onto the image: First, it obtains the 3D coordinates of preset calibration points on the target platform. And obtain two-dimensional coordinates through the camera Then, the camera's extrinsic parameters, i.e., the rotation matrix, are solved using methods such as direct linear transformation, EPNP, and Lvenberg-Marquardt. and 3D coordinates to two-dimensional coordinates The projection relationship is as follows: in, It is the known intrinsic parameter matrix of the camera, which mainly contains the camera's focal length and principal point information; The small-sample, large-time-scale trajectory information prediction based on recursive convolutional conditional neural processes described in step two is specifically implemented as follows: A proposed method uses a CNP network with a convolutional encoder to predict the trajectory point at the next moment based on historical trajectories, and implements multi-step prediction of the future through a recurrent neural network framework; where, This represents the timestamp at time t. Let represent the position and attitude angle predicted by the network at time t, respectively, and the prior data be... The timestamp and pose information at a given moment, in conjunction with The mean of the pose distribution at time t predicted after being jointly input into the CNP network and variance ; A CNP network with a convolutional encoder consists of three parts: an encoder, an aggregator, and a decoder. The encoder's role is to map historical data to a latent space representation. First, the historical data is passed through a one-dimensional convolutional layer, and then the result is input into a multilayer perceptron (MLP). If the historical data uses... If we represent it as follows, then the forward propagation process of the encoder can be expressed as: in, This represents the convolution operation. It is a one-dimensional convolution kernel. Represents a convolutional layer. Represents a multilayer perceptron. Z represents the network weights, and Z is a latent space representation generated by the encoder based on the data. The aggregator performs an aggregation operation on the Z output by the encoder to generate a unified feature representation C: The decoder is based on feature C and query set. To generate the predicted output, where the query set is the set of timestamps corresponding to the unknown sequence, the forward propagation process of the decoder can be represented as: in, The Gaussian distribution for predicting pose is represented by the mean and variance; the loss function of the CNP network is transformed into an optimization problem, represented by the negative log-likelihood (NLL): in, Conditional probability represents the probability of CNP based on historical data. and query set Predicted distribution vs. actual data The likelihood; Finally, the single-step convolutional CNP network is substituted into the architecture of the recurrent neural network, and the predicted output of the previous step is added to the conditional input of the current step to predict the trajectory in an iterative manner. The dynamic programming of B-spline coverage path based on particle swarm optimization algorithm described in step three is specifically implemented as follows: First, based on the trajectory prediction generated by the recursive convolutional CNP network, an artificial potential field is constructed based on Euclidean distance, probability distribution, and time series information. Then, the optimal scanning path is generated through working cell decomposition and coverage path planning. Finally, the local potential field is dynamically updated based on visual recognition information. Artificial potential field calculation is divided into the gravitational field formed by the position of the moving target. Repulsive field formed by movement obstacles and time decay caused by the temporal characteristics of the predicted trajectory It consists of three parts; the gravitational field is used to attract the laser to scan towards the target position. First, the gravitational force generated by the Euclidean distance from the workspace to the i-th prediction point is represented by a piecewise function. ,when To reduce potential energy and avoid excessive gravity when the target location is far away: Based on the Euclidean distance potential field, a probability density is introduced as a weight to account for the impact of prediction uncertainty, specifically expressed as follows: in, and Let represent the mean and variance of the Gaussian distribution at target point i, respectively. These are the weighting coefficients; Next, based on the predicted trajectory The potential fields calculated at each prediction point are superimposed, and a time decay weight is introduced to ensure that the potential field of prediction points closer to the current time has a stronger influence. The formula is expressed as: in, This represents the rate at which the weights of the potential field calculated based on each prediction point decay over time. A repulsive field is introduced into an artificial potential field to drive the laser scanning path away from a specific area. When the laser is incident on a reflecting prism at an angle greater than 45°, it cannot be correctly reflected along the original optical path. The scanning action is invalid in this case and needs to be avoided in path planning. By treating these specific areas as obstacles, the repulsive field can be described as follows: in, The distance to the nearest obstacle. It is an increase in repulsive force. This is the threshold range of the obstacle's influence. Within this threshold range, the obstacle will generate a repulsive force; otherwise, it has no effect. Superimposing the gravitational field and the repulsive field creates the final artificial potential field. ; To compensate for the deviation between the predicted trajectory and the actual trajectory, the Gaussian process regression (GPR) algorithm is used. Based on the deviation between the historical predicted trajectory and the actual trajectory obtained by the vision module, the algorithm predicts and compensates for the future prediction deviation of the neural network, thereby improving the prediction accuracy. Gaussian process regression is used to compensate for the prediction error of a network model in real time. Its prediction process can be expressed as follows: in, This represents the covariance matrix between different datasets. The training set representing the prediction bias. The estimated value representing the prediction bias, and the corrected prediction trajectory can be represented as: To improve scanning efficiency, scanning is only performed on regions where the potential field is below a specific threshold. Simultaneously, working units for path planning are divided according to the potential field surrounding the trajectory points to avoid path duplication. The boundaries of the working units are determined by a pre-set potential field threshold, and B-splines are used to construct boundary lines to ensure smoothness. Furthermore, to improve efficiency in local path optimization at future time steps, path pre-planning is required for trajectory points at more distant time steps. Therefore, the boundary threshold of these working units needs to be increased to ensure that path pre-planning for high-potential regions can cover a certain range. Considering computational efficiency and potential field distribution, working unit division is only performed on trajectory points in the next four steps, and coverage paths are planned within each working unit.

7. The method for rapid ranging of a space-moving target using an optical comb based on a hierarchical guidance strategy according to claim 6, characterized in that: coverage path planning within each working unit is performed in three steps: path pre-planning, path optimization, and dynamic replanning, as described in detail below: Path pre-planning: Pre-planning is used to generate uniform and fully covered spirals in an irregular workspace. During scanning, the area projected onto the angular space by the target reflecting prism can be regarded as the coverage area of ​​each path point. The coverage radius of the scanning point can be approximately estimated based on the distance between the target and the transmitter. The control points of the B-spline spiral are generated in polar coordinates, each... If a control point is set for the radian, the shape of each spiral is determined by... The control points are determined by a number of control points, and their polar coordinates are expressed as follows: in, Indicates angle, express The radial distance difference between the control points corresponding to the angle, where n represents the number of turns of the spiral, is the maximum radial distance from the center of the workspace to the boundary. The ratio determines; Particle swarm optimization: By integrating the coverage area using a weighted average based on the potential field strength, the scanning density in high-probability regions can be increased. in, This represents the original area of ​​the working unit. For potential field strength, The constant determines the influence of the potential field. The area of ​​the working unit after probability density weighting; The total path length and actual overlap area in the optimization algorithm are denoted by the following cost function: in, The weights of the three cost terms, This represents the area covered by all scan points along the path. The control point positions are optimized using a hierarchical particle swarm optimization algorithm, with the outer layer particle swarm optimizing the number of spiral coils. and the number of newly added control points The inner particle swarm optimizes the position of the control points. The inner and outer layers are optimized with different frequencies to improve optimization efficiency; the starting points of the optimization vectors of N particles are randomly set near the pre-planned path parameters in the form of a normal distribution, and the particle swarm position update formula is: in, Let represent the position vector of particle i at time t. Let i represent the velocity vector of particle i at time t+1; the velocity update formula for the particle swarm is: in, It is inertial weight. These are acceleration constants, which respectively control the particle's dependence on its own experience and the collective experience. It is a random number between [0,1], used to increase the randomness of the search. It is the optimal position for particle i. It is the globally optimal position; after multiple iterations, the particle swarm can find the optimal solution within the search space. Dynamic replanning: When the time corresponding to the next trajectory point is reached, in addition to modifying the time decay weight of the potential field corresponding to each trajectory point, the potential field is also fine-tuned based on the target position and invalid scan area position calculated by the vision module. After the potential field is updated, the path planning is re-performed based on the new potential field. If the predicted trajectory point position and its surrounding potential field change significantly, pre-planning is performed first, and then particle swarm optimization is performed starting from the pre-planned B-spline spiral control point. Otherwise, particle swarm optimization is performed starting from the B-spline spiral control point of the previous time step.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of the method for rapid ranging of a space-moving target optical comb based on a hierarchical guidance strategy as described in any one of claims 5-7.

Images