Space debris target detection and positioning method based on dynamic programming and multi-hypothesis testing

Abstract

The invention discloses a space debris target detection and positioning method based on dynamic programming and multi-hypothesis testing, and the method comprises the following steps: carrying out the preprocessing of an obtained space debris image, and obtaining a candidate point state set under different time frames; in the preprocessing, a denoising segmentation extraction channel and a lightweight CNN target extraction channel are divided according to a statistical local signal-to-noise ratio; obtaining a plurality of motion paths corresponding to each candidate point according to the candidate point state sets under different time frames; and for each candidate point, generating an initial hypothesis set according to the corresponding multiple motion paths, and processing the initial hypothesis set through Kalman filtering to obtain the position of the candidate point. According to the method, the detection precision of the dark and weak target is remarkably improved, the false alarm rate is reduced, the real-time performance of detection is improved, and the method is suitable for real-time monitoring scenes. The research result provides effective technical support for improving the operation safety of the spacecraft.

Description

Technical Field

[0001] This invention relates to the field of target detection and recognition technology, and more specifically to a method for detecting and locating spatial debris targets based on dynamic programming and multiple hypothesis testing. Background Technology

[0002] With the rapid development of aerospace technology and the increasing frequency of space activities, the amount of space debris is also constantly increasing. Space debris refers to non-functional man-made objects orbiting the Earth, including abandoned satellites, rocket debris, and collision debris between them. These debris travel at high speeds in orbit, posing a serious threat to spacecraft in orbit. A collision between a spacecraft and space debris can lead to damage to the spacecraft or even mission failure, causing huge economic losses and safety hazards. Therefore, space debris detection and location technology has become one of the key technologies for ensuring the safe operation of spacecraft.

[0003] However, space debris detection faces numerous challenges. Space debris is typically a faint target with low reflected light intensity, making it difficult to accurately identify against the complex cosmic background. Furthermore, the cosmic background contains a large number of densely packed stars whose brightness and distribution are similar to space debris, making it prone to false positives under low signal-to-noise ratio conditions, resulting in low detection accuracy and a high false alarm rate. Traditional space debris detection methods often fail to meet the requirements of detection accuracy and real-time performance in practical applications when faced with these complex conditions.

[0004] Currently, space debris detection technology has made some progress, with common methods including optical detection, radar detection, and image processing-based detection methods. Optical detection methods acquire image information of space debris using optical equipment such as telescopes, and then analyze and process it; however, it is greatly affected by weather and lighting conditions. Radar detection methods can provide relatively accurate distance and velocity information, but the equipment is expensive and its ability to detect small debris is limited. Image processing-based detection methods have attracted widespread attention due to their efficiency and flexibility. However, traditional image processing-based detection methods still suffer from low detection accuracy and high false alarm rates under conditions of low signal-to-noise ratio and dense stellar backgrounds.

[0005] In recent years, researchers have proposed various improved algorithms to address these challenges, but each method has its limitations. Hough transform, a classic feature extraction method, can detect straight lines and curves in images. However, it is susceptible to noise interference in complex backgrounds, resulting in low detection accuracy, high false alarm rate, and high computational cost, making it difficult to meet the requirements of real-time monitoring. Template matching matches images with predefined templates and is suitable for situations where the target shape is known. However, it is prone to false positives under low signal-to-noise ratio and background noise interference, and requires high accuracy and completeness of the template. Background subtraction highlights the target by subtracting the background image, but it requires high stability of the background and performs poorly under dynamic backgrounds or noise interference. Kalman filtering is a statistical estimation-based tracking algorithm that can predict and correct the target's motion state. However, it requires high accuracy in estimating the initial state and is prone to tracking failure under target loss or noise interference. Particle filtering is a nonlinear filtering method that represents the probability distribution of the target through a set of particles. Although it has some robustness in complex environments, it has high computational cost and is sensitive to particle degradation. Deep learning-based detection methods excel in feature extraction, but their reliance on large amounts of labeled data and complex model structures limit their application in real-time monitoring scenarios. Multi-frame accumulation methods enhance target signals by accumulating multiple frames, but they require high temporal resolution and are prone to blurring when the target is moving rapidly. Although these methods exhibit certain advantages under specific conditions, they still struggle to simultaneously meet the requirements of high accuracy, low false alarm rate, and real-time performance in low signal-to-noise ratio and complex background conditions.

[0006] Therefore, how to improve the detection accuracy and real-time performance of faint targets in space debris under complex backgrounds and reduce the false alarm rate remains a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of the above problems, the present invention provides a method for detecting and locating space debris targets based on dynamic programming and multiple hypothesis testing, so as to at least solve some of the technical problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] This invention provides a method for detecting and locating space debris targets based on dynamic programming and multiple hypothesis testing, comprising the following steps:

[0010] The acquired spatial fragment images are preprocessed to obtain candidate point state sets at different time frames; the preprocessing is divided into a denoising segmentation extraction channel and a lightweight CNN target extraction channel based on the statistical local signal-to-noise ratio.

[0011] Based on the candidate point state set under different time frames, obtain multiple motion paths corresponding to each candidate point;

[0012] For each candidate point, an initial hypothesis set is generated based on the corresponding multiple motion paths. The initial hypothesis set is then processed by Kalman filtering to obtain the position of the candidate point.

[0013] Furthermore, in the denoising segmentation extraction channel, the spatial fragment image is processed as follows:

[0014] (1) Perform nonlocal mean denoising on the acquired spatial fragment image to obtain a denoised image; wherein, for pixel (i,j) in the spatial fragment image, nonlocal mean denoising is expressed as: ;

[0015] in, N(i,j) represents the denoising result for pixel (i,j); N(i,j) represents the search window centered on pixel (i,j); (p,q) represents another group of pixels within the search window that are different from pixel (i,j); I represents the input spatial fragment image; I(p,q) represents the pixel grayscale value corresponding to pixel (p,q); The weight function is represented by P(i,j); P(p,q) represents a local block centered at pixel (i,j); Z(i,j) represents a normalization factor; and h represents a smoothing parameter used to control the noise suppression strength.

[0016] (2) Regarding the denoising results Perform adaptive threshold segmentation; specifically including:

[0017] The adaptive threshold T(i,j) for the neighborhood of pixel (i,j) is calculated as follows: ;

[0018] in, This represents the mean value of the neighborhood of pixel (i,j); This represents the standard deviation of the pixel (i,j) neighborhood; This represents an empirical coefficient used to control the false alarm rate;

[0019] Based on the adaptive threshold T(i,j), the denoising results Adaptive threshold segmentation is performed to obtain the first binary image B(i,j), which is represented as: ;

[0020] (3) Label the connected components of the first binary image B(i,j) to obtain the first connected component set; take each first connected component in the first connected component set as the first candidate target, and extract the centroid coordinates and gray-scale moment features of the first candidate target; wherein, the centroid coordinates of the first candidate target are... Represented as: ;

[0021] in, For the first The first connected component; I(i,j) represents the pixel gray value corresponding to pixel (i,j);

[0022] (4) Based on the centroid coordinates and gray-scale moment features of the first candidate target, obtain the average gray value within the region of the first candidate target; based on the average gray value, combine it with the mean of the neighborhood of pixel (i,j). and the standard deviation of the pixel (i,j) neighborhood. The first target signal-to-noise ratio of the first candidate target is obtained. , is represented as: ;

[0023] in, This represents the average gray value within the first candidate target area.

[0024] Furthermore, in the lightweight CNN target extraction channel, the spatial debris image is processed as follows:

[0025] (1) The spatial fragment image is processed by a lightweight preprocessing CNN network to obtain a second binary image;

[0026] (2) Perform connected component annotation on the second binary image to obtain a second connected component set; take each second connected component in the second connected component set as a second candidate target, and extract the centroid coordinates and gray-scale moment features of the second candidate target; wherein, the centroid coordinates of the second candidate target are... Represented as: ;

[0027] in, For the first A second connected component;

[0028] (3) Based on the centroid coordinates and gray moment characteristics of the second candidate target, the signal-to-noise ratio of the second candidate target is obtained.

[0029] Furthermore, the lightweight preprocessing CNN network consists of a four-layer downsampling encoder, a three-layer upsampling encoder, a denoising branch, and a segmentation branch.

[0030] Furthermore, the dual-channel output corresponding to the preprocessing is represented as follows: ;

[0031] in, Let represent the state of the i-th candidate point in the t-th frame, and ,in, Indicates the centroid coordinates of the candidate points; Indicates the target signal-to-noise ratio of the candidate points; This represents the set of candidate point states in frame t. This represents the centroid coordinates of the first candidate target in the denoised segmentation extraction channel output; This represents the first target signal-to-noise ratio of the denoising segmentation extraction channel output; This represents the centroid coordinates of the second candidate target output by the lightweight CNN target extraction channel; This represents the signal-to-noise ratio of the second target output in the lightweight CNN target extraction channel.

[0032] Furthermore, the step of obtaining multiple motion paths corresponding to each candidate point based on the candidate point state set under different time frames specifically includes:

[0033] (1) Motion cost function for searching the i-th candidate point from frame t-1 to frame t for the nearest suspected i-th candidate point Represented as: ;

[0034] in, This represents the state of the j-th candidate point in the (t-1)-th frame; v pred v represents the prediction speed of the i-th candidate point; obs Indicates the observation velocity of the i-th candidate point; and Both represent weighting coefficients, and ;

[0035] (2) Based on the motion cost function, obtain the state reached by the i-th candidate point. The maximum accumulated energy is recorded, and the predecessor node that achieves the maximum accumulated energy is also recorded; where:

[0036] Maximum accumulated energy Represented as: ;

[0037] Predecessor Node Represented as: ;

[0038] (3) Retain the highest energy motion paths in each frame and terminate the update of the remaining paths.

[0039] Further, the step of processing the initial hypothesis set using Kalman filtering to obtain the positions of the candidate points specifically includes:

[0040] (1) Each hypothesis in the initial hypothesis set corresponds to a spatial trajectory. For each hypothesis, Kalman filtering is used to predict the running state of candidate points in the next frame; expressed as: ;

[0041] in, This represents the process noise vector, which indicates the random disturbances experienced by the system during state transitions. Indicates process noise It follows a pattern with a mean of 0 and a covariance matrix of . The normal distribution; Let represent the running state of the candidate points in frame t predicted by Kalman filtering, and , where x t This represents the relative displacement of the candidate point on the horizontal axis in frame t; y t This represents the relative displacement of the candidate point on the vertical axis in frame t; v x v represents the relative velocity of the candidate point on the horizontal axis in frame t; y This represents the relative velocity of the candidate point on the vertical axis in frame t. F represents the running state of the candidate point in frame t-1; Q represents the process noise.

[0042] (2) Based on the predicted running state of candidate points in the next frame, construct the observation equation; expressed as: ;

[0043] Among them, z t G represents the observed values; G represents the observation matrix; R represents the observation noise covariance. The observation noise vector represents the random error introduced during the observation process; Indicates observation noise It follows a normal distribution with a mean of 0 and a covariance matrix of R;

[0044] (3) For the candidate observations in frame t, calculate the probability of matching each hypothesis; expressed as: ;

[0045] Where P(·) represents the matching probability; This represents the candidate observation value in frame t; Indicates the k-th hypothesis; Let represent the covariance matrix, and ;

[0046] (4) Retain the hypothesis that the matching probability is greater than the preset threshold. If the spatial distance between two hypotheses satisfies the preset distance condition, merge the two hypotheses into one hypothesis. Use the hypothesis pruning and merging result of the final initial hypothesis set as the final true trajectory. Obtain the position of the candidate point in each frame based on the true trajectory.

[0047] Furthermore, the preset distance condition is expressed as: ;

[0048] Among them, T i Indicates a hypothesis The corresponding spatial distance; T j Indicates a hypothesis The corresponding spatial distance; This indicates the preset distance threshold.

[0049] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method for detecting and locating space debris targets based on dynamic programming and multiple hypothesis testing, which has the following beneficial effects:

[0050] This invention improves the accuracy and reliability of space debris detection and reduces the false alarm rate by combining dynamic programming and multiple hypothesis testing with dual-channel preprocessing. It can effectively reduce the risk of spacecraft colliding with space debris, ensure the safe operation of spacecraft, and has important practical significance for maintaining the smooth progress of space missions.

[0051] The method provided by this invention also offers new ideas and approaches for solving the problem of detecting faint targets in complex backgrounds, and has certain theoretical and practical reference value for research in related fields.

[0052] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0054] Figure 1 This is a schematic diagram of the process for a space debris target detection and localization method based on dynamic programming and multiple hypothesis testing provided in an embodiment of the present invention.

[0055] Figure 2 This is a schematic diagram of a lightweight preprocessing CNN network structure provided in an embodiment of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] This invention discloses a method for detecting and locating space debris targets based on dynamic programming and multiple hypothesis testing. See [link to relevant documentation]. Figure 1 As shown, it includes the following steps:

[0058] S1. Preprocess the acquired spatial fragment images to obtain candidate point state sets under different time frames; the preprocessing is divided into a denoising segmentation extraction channel and a lightweight CNN target extraction channel based on the statistical local signal-to-noise ratio.

[0059] S2. Based on the candidate point state set under different time frames, obtain multiple motion paths corresponding to each candidate point;

[0060] S3. For each candidate point, an initial hypothesis set is generated based on the corresponding multiple motion paths. The initial hypothesis set is then processed by Kalman filtering to obtain the position of the candidate point.

[0061] Next, each of the above steps will be explained in detail.

[0062] In step S1 above, which is the preprocessing part, the acquired spatial fragment images are preprocessed to obtain candidate point state sets at different time frames. The preprocessing is divided into a denoising segmentation extraction channel and a lightweight CNN target extraction channel based on the statistical local signal-to-noise ratio. This allows for the use of a lightweight CNN for extraction in complex environments with high signal-to-noise ratios, while traditional denoising methods are used for targets with low signal-to-noise ratios to improve accuracy. Specifically:

[0063] 1. In the denoising segmentation extraction channel, the spatial fragment image is processed as follows:

[0064] (1) Perform nonlocal mean denoising on the acquired spatial fragment image to obtain a denoised image; wherein, for pixel (i,j) in the spatial fragment image, nonlocal mean denoising is expressed as: ;

[0065] in, This represents the denoising result for pixel (i,j); N(i,j) represents the search window centered on pixel (i,j) (e.g., ...). (p,q) represents another set of pixels in the search window that are different from pixel (i,j); I represents the input spatial fragment image, and , where R represents the set of real numbers; W represents the image width; H represents the image height; I(p,q) represents the pixel grayscale value corresponding to pixel point (p,q); The weight function is represented by P(i,j); P(i,j) represents a local block centered at pixel (i,j) (e.g., a block of size 1). P(p,q) represents a local block centered at pixel (p,q) (e.g., a block of size q). Z(i,j) represents the normalization factor, and h represents the smoothing parameter, used to control the noise suppression strength;

[0066] (2) Regarding the denoising results Perform adaptive threshold segmentation; specifically including:

[0067] The adaptive threshold T(i,j) for the neighborhood of pixel (i,j) is calculated as follows: ;

[0068] in, This represents the mean value of the neighborhood of pixel (i,j); This represents the standard deviation of the pixel (i,j) neighborhood; Represents the empirical coefficient (the value is generally in the range of...). ), used to control the false alarm rate;

[0069] Based on the adaptive threshold T(i,j), the denoising results Adaptive threshold segmentation is performed to obtain the first binary image B(i,j), which is represented as: ;

[0070] (3) Label the connected components of the first binary image B(i,j) to obtain the first set of connected components. Each connected component in the first connected component set is taken as a first candidate target, and the centroid coordinates and gray-scale moment features of the first candidate target are extracted; wherein, the centroid coordinates of the first candidate target are... Represented as: ;

[0071] in, For the first The first connected component; I(i,j) represents the pixel gray value corresponding to pixel (i,j);

[0072] (4) Based on the centroid coordinates and gray-scale moment features of the first candidate target, obtain the average gray value within the region of the first candidate target; based on the average gray value, combine the mean of the neighborhood of pixel (i,j) and the standard deviation of the pixel (i,j) neighborhood. The first target signal-to-noise ratio of the first candidate target is obtained. , is represented as: ;

[0073] in, This represents the average gray value within the first candidate target area.

[0074] 2. In the lightweight CNN object extraction channel, the spatial debris image is processed as follows:

[0075] (1) In this embodiment of the invention, a lightweight preprocessing CNN network is used to replace the traditional preprocessing (non-local mean denoising + adaptive threshold segmentation) to support dual-task output (denoising + segmentation), with a total parameter size of 1.2MB (FP32) and 0.3MB (INT8 quantization). For ease of design, the input of the lightweight CNN is limited to... For larger images, a pixel-by-pixel approach is used for processing; here, we only consider... For example. Figure 2 A lightweight preprocessing CNN network structure is presented, consisting of a four-layer downsampling encoder, a three-layer upsampling encoder, a denoising branch, and a segmentation branch, with the final output being... The binary segmentation mask is denoted as the second binary image; a pixel with a value of 1 in the binary segmentation mask indicates the presence of a target, while a value of 0 indicates the absence of a target.

[0076] Table 1 below shows the detailed architecture and parameter count of the lightweight CNN (CNN_Lite). The overall computational architecture is based on the U-Net network, but simplifications have been made to the details. This is mainly because space debris observation systems primarily use probe cameras rather than imaging cameras. Under these imaging conditions, the signal-to-noise ratio is mainly relied upon to distinguish the differences between stars, debris targets, and background noise. The target itself does not have light and dark details; it is simply a Gaussian distribution of point targets generated according to the camera's design state, and there is good similarity between different targets and between different regions of background statistical noise. Therefore, compared to the detection of area targets, large-parameter, deep networks are not necessary; shallow networks can achieve convergence during training.

[0077] Table 1. Detailed structure and parameters of lightweight preprocessing CNN

[0078] The second binary image generated by the lightweight preprocessing CNN network is represented as follows: ;

[0079] in, This represents the second binary image; cnn-lite represents a lightweight preprocessing CNN network;

[0080] (2) For the second binary image Perform connected component labeling to obtain the second set of connected components. Each connected component in the set of second connected components is taken as a second candidate target, and the centroid coordinates and gray-scale moment features of the second candidate target are extracted; wherein, the centroid coordinates of the second candidate target are... Represented as: ;

[0081] in, For the first A second connected component;

[0082] (3) Based on the centroid coordinates and gray-scale moment characteristics of the second candidate target, obtain the signal-to-noise ratio of the second candidate target. .

[0083] 3. The dual-channel output corresponding to preprocessing is represented as follows: ;

[0084] in, Let represent the state of the i-th candidate point in the t-th frame, and ,in, Indicates the centroid coordinates of the candidate points; Indicates the target signal-to-noise ratio of the candidate points; This represents the set of candidate point states in frame t. This represents the centroid coordinates of the first candidate target in the denoised segmentation extraction channel output; This represents the first target signal-to-noise ratio of the denoising segmentation extraction channel output; This represents the centroid coordinates of the second candidate target output by the lightweight CNN target extraction channel; This represents the signal-to-noise ratio of the second target output in the lightweight CNN target extraction channel.

[0085] The reason for this design in this embodiment of the invention is that traditional image processing methods are more accurate for targets with low signal-to-noise ratios (SNR), which are often difficult to label precisely, leading to training errors. For targets with high SNR, whose states are more complex, CNN-based methods are more advantageous. For targets with overlapping or similar merged results, a threshold can be set, and the result with the higher SNR can be selected as the final value.

[0086] In step S2 above, which is the dynamic path planning part, the detection path is optimized through dynamic planning to reduce the amount of computation and improve detection efficiency. This mainly involves planning and associating the candidate point state set obtained through preprocessing calculations with information from multiple frames. Based on the possible target position relationships and state judgments between consecutive frames, the accurate trajectory of the target is finally determined. The transition formula is obtained through the definition of target point states and the construction of the transition model. The maximum probability path is obtained through energy accumulation and optimal path selection, and the path with the highest energy in each frame is retained, acquiring multiple motion paths corresponding to each candidate point. Specifically, this includes:

[0087] (1) Motion cost function for searching the i-th candidate point from frame t-1 to frame t for the nearest suspected i-th candidate point Represented as: ;

[0088] in, This represents the state of the j-th candidate point in the (t-1)-th frame; v pred Let represent the prediction speed of the i-th candidate point, and N represents the total number of candidate points, and k represents the k-th candidate point; v obs Let represent the observation velocity of the i-th candidate point, and ; and Both represent weighting coefficients, and , used to balance motion and appearance features; if the i-th candidate point in frame t is indeed i, then j = i;

[0089] (2) Based on the motion cost function, obtain the state reached by the i-th candidate point. The maximum accumulated energy is recorded, and the predecessor node that achieves the maximum accumulated energy is also recorded; where:

[0090] Let the initial energy be expressed as Maximum accumulated energy Represented as: ;

[0091] Predecessor Node Represented as: ;

[0092] (3) Retain the highest energy multiple motion paths in each frame (the maximum number of paths is set in advance), and terminate the update of the remaining paths.

[0093] The dynamic path planning component utilizes temporal information from multiple frames of images to enhance the detection capability of weak targets through energy accumulation and path optimization.

[0094] In step S3 above, which involves multiple hypothesis testing, for each candidate point, an initial hypothesis set is generated based on the corresponding multiple motion paths. This initial hypothesis set is then processed using Kalman filtering to obtain the position of the candidate point. Specifically, this includes:

[0095] (1) For each candidate point, from its corresponding Generate the initial hypothesis set in the path H K Let K represent the hypothesis corresponding to the Kth path; each hypothesis corresponds to a spatial trajectory, denoted as . , among which, T k Let represent the spatial trajectory corresponding to the hypothesis of the k-th path. Represents the spatial trajectory T k The trajectory point in frame t; for each hypothesis, use Kalman filtering to predict the running state of candidate points in the next frame; expressed as: ;

[0096] in, This represents the process noise vector, which indicates the random disturbances experienced by the system during state transitions. These disturbances may originate from uncertainties within the system or external disturbances. Indicates process noise It follows a pattern with a mean of 0 and a covariance matrix of . The noise follows a normal distribution; where the mean is 0, it means that the expected value of the process noise is 0, that is, the noise will not deviate from the system state under long-term average; the covariance matrix Q describes the correlation between the noise components and the noise intensity; a smaller Q value indicates that the process noise is smaller and the system state is more stable; a larger Q value indicates that the process noise is larger and the system state may fluctuate more. Let represent the running state of the candidate points in frame t predicted by Kalman filtering, and (Position and velocity), where x t This represents the relative displacement of the candidate point on the horizontal axis in frame t; y t This represents the relative displacement of the candidate point on the vertical axis in frame t; v x v represents the relative velocity of the candidate point on the horizontal axis in frame t; y This represents the relative velocity of the candidate point on the vertical axis in frame t. F represents the running state of the candidate point in frame t-1; F represents the state transition matrix (uniform speed model), which is a 4×4 matrix; Q represents the process noise, which is a 4×4 matrix.

[0097] The state transition matrix F is represented as: ;

[0098] in, Indicates a time interval;

[0099] (2) Based on the predicted running state of candidate points in the next frame, construct the observation equation; expressed as: ;

[0100] Among them, z t G represents the observation value; G represents the observation matrix, which describes the linear relationship between the observation value and the true state of the system; G determines how the system state is mapped to the observation space, that is, which state variables can be observed and the accuracy of the observation; for example, in a target tracking system in a two-dimensional plane, if only the position of the target can be observed but the velocity cannot be directly observed, the observation matrix will selectively extract the position component from the state vector. R represents the observation noise covariance, and ; The observation noise vector represents the random errors introduced during the observation process. These errors may originate from factors such as the accuracy limitations of the sensor itself and environmental interference. Indicates observation noise It follows a normal distribution with a mean of 0 and a covariance matrix of R. A mean of 0 means that the average value of the observation noise over multiple measurements is close to 0, i.e. there is no systematic bias. The covariance matrix R describes the correlation between the components of the observation noise and the intensity of the noise. The diagonal elements represent the variance of each observation dimension, and the off-diagonal elements represent the covariance between different observation dimensions.

[0101] (3) For the candidate observations in frame t, calculate the probability of matching each hypothesis; expressed as: ;

[0102] Where P(·) represents the matching probability; This represents the candidate observation value in frame t; Indicates the k-th hypothesis; Let represent the covariance matrix, and ;

[0103] (4) Retain the assumption that the matching probability is greater than a preset threshold, wherein the matching probability being greater than the preset threshold is expressed as: ;in This indicates a preset threshold (e.g., 0.1).

[0104] At the same time, if two assumptions and If the corresponding spatial distances satisfy the preset distance conditions, then the two hypotheses are merged into one hypothesis; the final initial hypothesis set is then obtained. Hypothetical pruning and merging results As the ultimate true trajectory; each Includes the corresponding real trajectory This allows us to obtain the exact location information of the target in each frame of the image;

[0105] The preset distance condition is expressed as follows: ;

[0106] Among them, T i Indicates a hypothesis The corresponding spatial distance; T j Indicates a hypothesis The corresponding spatial distance; This indicates the preset distance threshold.

[0107] This multi-hypothesis testing section combines a multi-hypothesis testing mechanism to verify the detected target multiple times, reducing the false alarm rate and improving detection accuracy. Finally, by setting multiple hypothetical targets and dynamically adjusting the hypotheses based on their motion characteristics and observation data, the true target is determined.

[0108] Next, to verify the effectiveness of the space debris target detection and localization method based on dynamic programming and multiple hypothesis testing provided in the above embodiments of the present invention, experiments were conducted on several publicly available space debris observation datasets. These datasets cover different signal-to-noise ratio conditions and background complexities, enabling a comprehensive evaluation of the algorithm's performance.

[0109] 1. Dataset Description:

[0110] (1) International Space Debris Observation Dataset (ISDOD);

[0111] Features: The dataset includes various signal-to-noise ratio (SNR) conditions (low SNR: SNR<3, medium SNR: 3≤SNR<6, high SNR: SNR≥6), covering different background noise and target characteristics. The images have high resolution, complex backgrounds, and contain a large number of stars and noise points.

[0112] Annotation information: Each image in the dataset is annotated with the location information of the space debris target, including the centroid coordinates and motion trajectory of the target.

[0113] (2) Chinese Space Debris Observation Dataset (CSDOD);

[0114] Features: The dataset includes a variety of observation conditions, has a wide signal-to-noise ratio range, and its background noise and target characteristics are similar to those of the ISDOD dataset, but some images have lower resolution and different background noise distributions.

[0115] Annotation information: Each image in the dataset is annotated with the target's location information, including centroid coordinates and motion trajectory.

[0116] (3) Synthetic Space Debris Dataset (SSDD)

[0117] Features: The dataset includes various signal-to-noise ratio conditions, and background noise and target characteristics are controllable, enabling the simulation of different observation environments. The image resolution and background complexity are similar to those of real datasets, but the distribution and motion characteristics of targets are more uniform.

[0118] Annotation information: Each image in the dataset is annotated with the target's location information, including centroid coordinates and motion trajectory.

[0119] 2. Experimental Results:

[0120] In the experiments, the method provided in this embodiment of the invention was compared with several other common space debris detection methods, including traditional Hough transform-based detection methods, template matching methods, background subtraction methods, Kalman filter tracking methods, and deep learning-based detection methods. All methods were run in the same hardware environment to ensure the fairness and comparability of the experimental results.

[0121] The hardware configuration used in the experiment is as follows:

[0122] Computer model: Dell Precision T7920 workstation;

[0123] Processor: Intel Xeon Gold 6240, 2.6GHz, 18 cores and 36 threads;

[0124] Graphics card: NVIDIA Quadro RTX 6000, 24 GB GDDR6 video memory;

[0125] Memory: 128 GB DDR4 2666 MHz;

[0126] Operating system: Ubuntu 20.04 LTS;

[0127] Programming environment: Python 3.8, CUDA 11.2, cuDNN 8.1;

[0128] This hardware configuration can meet the training and inference requirements of the lightweight CNN in this paper, while also efficiently running other traditional methods and deep learning-based detection methods, ensuring the reliability of experimental results.

[0129] The test results of different methods on different datasets are shown in Tables 2-4 below.

[0130] Table 2. Test Results of the ISDOD Dataset

[0131] Detection methods Detection accuracy (%) False alarm rate (%) Average execution time (ms / frame) Hough transform 65.2 30.5 120 Template matching 68.4 28.7 150 Background subtraction method 63.5 32.1 100 Kalman filtering 70.1 25.4 80 Deep learning methods 72.5 24.3 200 Method of the present invention 98.2 1.8 70

[0132] Table 3. CSDOD Dataset Test Results

[0133] Detection methods Detection accuracy (%) False alarm rate (%) Average time elapsed (ms / frame) Hough transform 64.8 31.2 115 Template matching 67.9 29.5 145 Background subtraction method 62.7 33.0 95 Kalman filtering 69.5 26.0 75 Deep learning methods 71.8 25.0 190 Method of the present invention 97.8 2.0 68

[0134] Table 4 Test results of the SSDD dataset

[0135] Detection methods Detection accuracy (%) False alarm rate (%) Average execution time (ms / frame) Hough transform 66.5 30.0 125 Template matching 70.1 28.0 155 Background subtraction method 64.2 31.8 105 Kalman filtering 71.2 24.5 85 Deep learning methods 73.5 23.8 210 This article's method 98.5 1.5 72

[0136] 3. Analysis Results:

[0137] The test results from the different datasets above show that the method of this invention performs excellently on the International Space Debris Observation Dataset (ISDOD), the Chinese Space Debris Observation Dataset (CSDOD), and the Synthetic Space Debris Dataset (SSDD). It significantly outperforms other traditional methods and deep learning-based methods in terms of detection accuracy, false alarm rate, and real-time performance. Specifically:

[0138] (1) In terms of detection accuracy:

[0139] On the ISDOD dataset, the detection accuracy of the method of this invention reaches 98.2%, which is at least 25.7 percentage points higher than other methods.

[0140] On the CSDOD dataset, the detection accuracy of the method of this invention is 97.8%, which is at least 29.9 percentage points higher than other methods.

[0141] On the SSDD dataset, the detection accuracy of the method of this invention is 98.5%, which is at least 25.0 percentage points higher than other methods.

[0142] (2) In terms of false alarm rate:

[0143] On the ISDOD dataset, the false alarm rate of the method of this invention is only 1.8%, which is at least 28.7 percentage points lower than other methods.

[0144] On the CSDOD dataset, the false alarm rate of the method of this invention is 2.0%, which is at least 27.5 percentage points lower than other methods.

[0145] On the SSDD dataset, the false alarm rate of the method of this invention is 1.5%, which is at least 28.5 percentage points lower than other methods.

[0146] (3) In terms of real-time performance:

[0147] On the ISDOD dataset, the average execution time of the method of this invention is 70ms / frame, which is the shortest among all methods.

[0148] On the CSDOD dataset, the average execution time of the method of this invention is 68ms / frame, which is also better than other methods.

[0149] On the SSDD dataset, the average execution time of the method of this invention is 72ms / frame, which still maintains a high level of real-time performance.

[0150] In summary, addressing the shortcomings of existing space debris detection technologies under low signal-to-noise ratio and dense stellar background conditions, this invention provides a space debris target detection and localization method based on dynamic programming and multiple hypothesis testing. Experimental verification on the International Space Debris Observation Dataset (ISDOD), the Chinese Space Debris Observation Dataset (CSDOD), and the Synthetic Space Debris Dataset (SSDD) demonstrates that this method exhibits excellent performance in detection accuracy, false alarm rate, and real-time performance, significantly outperforming traditional Hough transform, template matching, background subtraction, Kalman filtering, and deep learning-based detection methods. This invention employs a dual-channel preprocessing strategy, combining traditional image processing methods with a lightweight CNN, selecting appropriate preprocessing methods for different signal-to-noise ratio conditions to improve the accuracy of the preprocessing stage. Dynamic programming is used to optimize the detection path, reducing computational load and improving detection efficiency. A multiple hypothesis testing mechanism generates and verifies hypotheses for multiple paths output by dynamic programming, further reducing the false alarm rate and improving detection accuracy. This invention not only represents technological innovation but also has significant practical implications, effectively reducing the risk of spacecraft collisions with space debris and ensuring the safe operation of spacecraft. Based on the method of this invention, by further optimizing the network structure of the lightweight CNN, more feature extraction and path optimization methods can be explored to further improve detection accuracy and real-time performance, and the method can be applied to a wider range of scenarios, such as deep space exploration and satellite formation.

[0151] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0152] 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 method for detecting and locating space debris targets based on dynamic programming and multiple hypothesis testing, characterized in that, Includes the following steps: The acquired spatial fragment images are preprocessed to obtain candidate point state sets at different time frames; the preprocessing is divided into a denoising segmentation extraction channel and a lightweight CNN target extraction channel based on the statistical local signal-to-noise ratio. Based on the candidate point state set under different time frames, obtain multiple motion paths corresponding to each candidate point; For each candidate point, an initial hypothesis set is generated based on the corresponding multiple motion paths. The initial hypothesis set is then processed by Kalman filtering to obtain the position of the candidate point.

2. The space debris target detection and localization method based on dynamic programming and multiple hypothesis testing according to claim 1, characterized in that, In the denoising segmentation extraction channel, the spatial fragment image is processed as follows: (1) Perform nonlocal mean denoising on the acquired spatial fragment image to obtain a denoised image; wherein, for pixel (i,j) in the spatial fragment image, nonlocal mean denoising is expressed as: ; in, N(i,j) represents the denoising result for pixel (i,j); N(i,j) represents the search window centered on pixel (i,j); (p,q) represents another group of pixels within the search window that are different from pixel (i,j); I represents the input spatial fragment image; I(p,q) represents the pixel grayscale value corresponding to pixel (p,q); The weight function is represented by P(i,j); P(p,q) represents a local block centered at pixel (i,j); Z(i,j) represents a normalization factor; and h represents a smoothing parameter used to control the noise suppression strength. (2) Regarding the denoising results Perform adaptive threshold segmentation; specifically including: The adaptive threshold T(i,j) for the neighborhood of pixel (i,j) is calculated as follows: ; in, This represents the mean value of the neighborhood of pixel (i,j); This represents the standard deviation of the pixel (i,j) neighborhood; This represents an empirical coefficient used to control the false alarm rate; Based on the adaptive threshold T(i,j), the denoising results are... Adaptive threshold segmentation is performed to obtain the first binary image B(i,j), which is represented as: ; (3) Label the connected components of the first binary image B(i,j) to obtain the first connected component set; take each first connected component in the first connected component set as the first candidate target, and extract the centroid coordinates and gray-scale moment features of the first candidate target; wherein, the centroid coordinates of the first candidate target are... Represented as: ; in, For the first The first connected component; I(i,j) represents the pixel gray value corresponding to pixel (i,j); (4) Based on the centroid coordinates and gray-scale moment features of the first candidate target, obtain the average gray value within the region of the first candidate target; based on the average gray value, combine it with the mean of the neighborhood of pixel (i,j). and the standard deviation of the pixel (i,j) neighborhood. The first target signal-to-noise ratio of the first candidate target is obtained. , is represented as: ; in, This represents the average gray value within the first candidate target area.

3. The space debris target detection and localization method based on dynamic programming and multiple hypothesis testing according to claim 1, characterized in that, In the lightweight CNN object extraction channel, the spatial debris image is processed as follows: (1) The spatial fragment image is processed by a lightweight preprocessing CNN network to obtain a second binary image; (2) Perform connected component annotation on the second binary image to obtain a second connected component set; take each second connected component in the second connected component set as a second candidate target, and extract the centroid coordinates and gray-scale moment features of the second candidate target; wherein, the centroid coordinates of the second candidate target are... Represented as: ; in, For the first A second connected component; (3) Based on the centroid coordinates and gray moment characteristics of the second candidate target, the signal-to-noise ratio of the second candidate target is obtained.

4. The space debris target detection and localization method based on dynamic programming and multiple hypothesis testing according to claim 3, characterized in that, The lightweight preprocessing CNN network consists of four downsampling encoders, three upsampling encoders, one denoising branch, and one segmentation branch.

5. The space debris target detection and localization method based on dynamic programming and multiple hypothesis testing according to claim 1, characterized in that, The dual-channel output corresponding to the preprocessing is represented as follows: ; in, Let represent the state of the i-th candidate point in the t-th frame, and ,in, Indicates the centroid coordinates of the candidate points; Indicates the target signal-to-noise ratio of the candidate points; This represents the set of candidate point states in frame t. This represents the centroid coordinates of the first candidate target in the denoised segmentation extraction channel output; This represents the first target signal-to-noise ratio of the denoising segmentation extraction channel output; This represents the centroid coordinates of the second candidate target output by the lightweight CNN target extraction channel; This represents the signal-to-noise ratio of the second target output in the lightweight CNN target extraction channel.

6. The space debris target detection and localization method based on dynamic programming and multiple hypothesis testing according to claim 1, characterized in that, The step of obtaining multiple motion paths corresponding to each candidate point based on the candidate point state set under different time frames specifically includes: (1) Motion cost function for searching the i-th candidate point from frame t-1 to frame t for the nearest suspected i-th candidate point Represented as: ; in, This represents the state of the j-th candidate point in the (t-1)-th frame; v pred v represents the prediction speed of the i-th candidate point; obs Indicates the observation velocity of the i-th candidate point; and Both represent weighting coefficients, and ; (2) Based on the motion cost function, obtain the state reached by the i-th candidate point. The maximum accumulated energy is recorded, and the predecessor node that achieves the maximum accumulated energy is also recorded; where: Maximum accumulated energy Represented as: ; Predecessor Node Represented as: ; (3) Retain the highest energy motion paths in each frame and terminate the update of the remaining paths.

7. The space debris target detection and localization method based on dynamic programming and multiple hypothesis testing according to claim 1, characterized in that, The step of processing the initial hypothesis set using Kalman filtering to obtain the positions of the candidate points specifically includes: (1) Each hypothesis in the initial hypothesis set corresponds to a spatial trajectory. For each hypothesis, Kalman filtering is used to predict the running state of candidate points in the next frame; expressed as: ; in, This represents the process noise vector, which indicates the random disturbances experienced by the system during state transitions. Indicates process noise It follows a pattern with a mean of 0 and a covariance matrix of . The normal distribution; Let represent the running state of the candidate points in frame t predicted by Kalman filtering, and , where x t This represents the relative displacement of the candidate point on the horizontal axis in frame t; y t This represents the relative displacement of the candidate point on the vertical axis in frame t; v x v represents the relative velocity of the candidate point on the horizontal axis in frame t; y This represents the relative velocity of the candidate point on the vertical axis in frame t. F represents the running state of the candidate point in frame t-1; Q represents the process noise. (2) Based on the predicted running state of candidate points in the next frame, construct the observation equation; expressed as: ; Among them, z t G represents the observed values; G represents the observation matrix; R represents the observation noise covariance. The observation noise vector represents the random error introduced during the observation process; Indicates observation noise It follows a normal distribution with a mean of 0 and a covariance matrix of R; (3) For the candidate observations in frame t, calculate the probability of matching each hypothesis; expressed as: ; Where P(·) represents the matching probability; This represents the candidate observation value in frame t; Indicates the k-th hypothesis; Let represent the covariance matrix, and ; (4) Retain the hypothesis that the matching probability is greater than the preset threshold. If the spatial distance between two hypotheses satisfies the preset distance condition, merge the two hypotheses into one hypothesis. Use the hypothesis pruning and merging result of the final initial hypothesis set as the final true trajectory. Obtain the position of the candidate point in each frame based on the true trajectory.

8. The space debris target detection and localization method based on dynamic programming and multiple hypothesis testing according to claim 7, characterized in that, The preset distance condition is expressed as follows: ; Among them, T i Indicates a hypothesis The corresponding spatial distance; T j Indicates a hypothesis The corresponding spatial distance; This indicates the preset distance threshold.

Images