Methods for detecting faint staring debris targets against a background of dense stares and long tails

Abstract

A detection method of dim and weak gazing fragment target in dense constant star long tail background is provided, which adopts rectangle fitting to complete tailing image aggregation and reduce the influence of noise on tailing information, analyzes target shielding state, classifies semi-shielding scene, realizes accurate classification of limited point features based on full connection network, extracts semi-shielding image, and performs correlation detection based on the extracted semi-shielding image to improve the correlation success probability of the target and realize effective detection and tracking of the fragment. The method can detect and track the shielded target under the premise of ensuring that the detection rate is better than 90% and the false alarm rate is lower than 10%, and effectively solves the problem that the target cannot be continuously tracked due to shielding in the previous method.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of debris target detection, in particular to a method for detecting dark and weak gazing debris targets in a dense star long-tail background. BACKGROUND

[0002] With the continuous development of civil aerospace, satellite missions are intensively launched, and space debris has become an important factor affecting the safe and reliable execution of current space missions. Early debris monitoring mainly relied on ground-based radar and astronomical telescopes. However, with the increasing number of debris, limited by observation time and field of view, in the past decade, research has gradually shifted to space-based debris monitoring.

[0003] Space-based space debris monitoring mainly relies on optical cameras to image debris targets to obtain astronomical data similar to a "star catalog". Target information is confirmed according to the difference in motion state between the star background and the target. However, as debris targets continue to increase in size to decimeter and centimeter levels, the number of targets and the complexity of the image content have gradually become a difficulty in the information extraction process. For distant small-scale dark targets, in order to achieve a sufficient target-to-background signal-to-noise ratio for detection and identification, a long-exposure mode is often used for target tracking cameras. In this mode, the gazed target forms a point target information on the image, the background stars form a long-tail shape, and the non-gazed targets form a shape with different lengths of tails.

[0004] Early space debris target detection mainly relied on long-term observation for orbit determination to determine the target. With the development of image processing technology, current image processing-based methods mainly have two categories: one is to match the camera imaging data with the star catalog to obtain the mapping relationship of the sky for each image, and then to match and exclude each star target to finally retain the debris target data information; the other is to register and compare the difference between consecutive images to distinguish target information by relying on the different temporal distribution of stars and debris targets. These two methods have good accuracy for data processing in a short-tail star background, and can balance real-time performance and accuracy by relying on inter-frame difference calculation in a high signal-to-noise ratio.

[0005] However, in a long-tail star background, as shown in the accompanying Figure 1 Due to noise interference and low signal-to-noise ratio, the star tail often presents an "unstable" state with unstable and discontinuous grayscale changes in the image, which seriously interferes with the image processing process. In addition, debris targets are often blocked by star tails, further increasing the difficulty of detecting dark and weak gazing debris targets in a long-tail star background. SUMMARY

[0006] The present disclosure provides a method for detecting dark and weak gazing debris targets in a dense star long-tail background, which can reduce the influence of noise on tail information and improve the detection performance of star long-tail blocked point targets.

[0007] The method for detecting faint staring debris targets against a background of dense stellar tails disclosed herein includes the following steps:

[0008] S1, perform rectangle fitting on the trailing image;

[0009] S2, classify the state of the target being occluded by the star's tail in the fitted image, and extract the target half-occlusion image based on the sequence joint features identified by multi-frame superposition.

[0010] S3 performs correlation detection based on the extracted target partially occluded image to achieve effective fragment detection and tracking.

[0011] Furthermore, the method for performing rectangular fitting on the trailing image in step S1 specifically includes:

[0012] Step S1.1, define the image as I N Where N is the image pixel size, two square operators of different sizes are defined for image I. N Calculate its contrast I d ;

[0013] Step S1.2, for I d Perform traditional binary annotation calculations to obtain the labeled M-target set:

[0014]

[0015] Where x,y∈[0,N-1](x,y∈Z) + ), where m represents the m-th target after being marked, m∈[0,M-1], m∈Z + ;K m This represents the maximum number of pixels for the m-th labeled target; This indicates that the m-th marked target is k-th. m Pixel coordinates, k m ∈[0,K m -1],k m ∈Z + ;

[0016] Step S1.3, for any labeled target I b (m), find its centroid I. c Position coordinates of (m):

[0017]

[0018] Using the centroid of the m-th marked target as any point within the rectangle, and with the star's tail angle ±θ as the search direction, a fitted bounding box search is performed to obtain the set of generated rectangles R(m) for the m-th marked target. R(m,q) represents the bounding box set with the centroid of the m-th marked target as the centroid of the m-th marked target.c (m) is the set of rectangles with width and length Wf and length Lf at any point, and q represents the q-th rectangle;

[0019] Step S1.4: Calculate the corresponding image region I based on R(m,q). N (R(m,q)), calculate the gray-level summation S(m,q) and histogram h(m,q) within the rectangular region, and obtain the optimal rectangular solution constraint as shown in formula (4):

[0020]

[0021] Among them, S min With S max h represents the upper and lower limits for the sum of gray levels within a rectangular area. st The standard rectangular region trailing target histogram is used. The three thresholds are all obtained statistically and are used to constrain formula (4) to obtain the optimal rectangle fit.

[0022] Furthermore, the square operator in step S1.1 is taken as 3×3 and 7×7 respectively.

[0023] Furthermore, in step S2, the state in which the target is obscured by the star's tail is divided into three categories, including:

[0024] Case 1: More than 2 / 3 is obscured, meaning only one line of the star trail is visible from the debris target. In this case, the debris target and the star trail overlap.

[0025] Scenario 2: The obstruction is between 2 / 3 and 1 / 5, meaning that more than two rows of debris targets are exposed above the star's tail, at which point a noticeable bulge is formed at the edge of the tail;

[0026] Scenario 3: The obstruction is less than 1 / 5, and the debris target only has partial contact with the star's tail.

[0027] Furthermore, in step S2, the method for extracting the target partially occluded image specifically includes:

[0028] A small fully connected convolutional network is used. Based on the input of three consecutive frames of data, the edge "protrusion" information after the rectangle is fitted is used as a suspected target for fully connected convolutional calculation to finally determine whether it is a fragment target.

[0029] Furthermore, the structure of the small fully connected convolutional network includes:

[0030] Number of layers Convolutional kernel Input Output 1 3×3×16 8×8(×3) 8×8×16 2 3×3×16×32 8×8×16 8×8×32 3 3×3×32×64 8×8×32 4×4×64 4 3×3×64×128 4×4×64 2×2×128 5 3×3×128×256 2×2×128 1×1×256 6 1×256 1×1×256 1×1

[0031] Here, (×3) at the input end represents the extended network, which is used for three consecutive frames of data input.

[0032] Compared with the prior art, the beneficial effects of this disclosure are:

[0033] (1) To address the problem of unstable and discontinuous pixel grayscale changes in long trailing images, rectangular fitting is used to aggregate the trailing images and reduce the impact of noise on the trailing information.

[0034] (2) Analyze the target occlusion status, statistically classify the semi-occluded scene, and accurately classify it using finite point features based on a fully connected network to extract the semi-occluded image;

[0035] (3) To address the problem of unsustainable inter-frame association caused by star trails obscuring point targets, association detection is performed based on the extracted half-occluded images to improve the success rate of target association and achieve effective fragment detection and tracking. Attached Figure Description

[0036] The above and other objects, features and advantages of this disclosure will become more apparent from the more detailed description of exemplary embodiments of this disclosure taken in conjunction with the accompanying drawings, in which the same reference numerals generally represent the same components.

[0037] Figure 1 Examples of problems in stellar tail detection;

[0038] Figure 2 A flowchart illustrating an exemplary embodiment of this disclosure;

[0039] Figure 3 Schematic diagram of rectangular fitting, (a) general tail fitting, (b) broken tail fitting;

[0040] Figure 4 This is an illustration of a rectangle expression;

[0041] Figure 5 Histogram features of rectangular regions in the trailing image;

[0042] Figure 6 The fitting result is for the trailing rectangle;

[0043] Figure 7 A diagram illustrating the classification of fragment target occlusion situations;

[0044] Figure 8 The grayscale distribution of the obscured fragments;

[0045] Figure 9 Six common scenarios of debris obscuring objects from stellar tails (left: single-frame grayscale image; right: three consecutive frames superimposed).

[0046] Figure 10 The features of the occluded fragment target image (single-frame grayscale image and three consecutive superimposed images). Detailed Implementation

[0047] Preferred embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While preferred embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure will be thorough and complete, and will fully convey the scope of the present disclosure to those skilled in the art.

[0048] This disclosure provides a method for detecting faint staring debris targets against a background of dense stellar tails. Figure 2 The flowchart according to an exemplary embodiment of this disclosure mainly includes the following steps:

[0049] Step 1: Tailing Aggregation Based on Rectangle Fitting

[0050] Due to noise interference and low signal-to-noise ratio, stellar tails often appear "unstable" in images. These states can severely interfere with image processing, mainly including three types (such as...). Figure 1 (As shown): The middle part of the trail breaks; the end of the trail adheres to other trails; the edge of the trail diverges and is difficult to determine. Traditional image processing methods have difficulty eliminating these three problems.

[0051] To ensure the accuracy of subsequent detection, it is essential to first eliminate the impact of the aforementioned issues on the stellar tails in the image. Considering that the angle and length of the stellar tails can be predicted, a simplified rectangle fitting method is used to perform contour search on the tail target, ultimately forming an edge rectangle with the smallest error.

[0052] The exemplary rectangular contour fitting algorithm mainly includes the following four steps:

[0053] Step 1.1 (Contrast Segmentation Calculation): Define the image as I N Where N is the image pixel size, 3×3 and 7×7 square operators are defined, and I is defined. d For image I N The contrast calculation results, T d This is the segmentation threshold. For any x, y ∈ [0, N-1], x, y ∈ Z + The formula for calculating its contrast is as follows:

[0054]

[0055] Step 1.2 (Binary Labeling): For I d Perform traditional binary annotation calculations to obtain the labeled M-target set:

[0056]

[0057] Where m represents the m-th target after labeling, m∈[0,M-1], m∈Z+ ;K m This represents the maximum number of pixels for the m-th labeled target; This indicates that the m-th marked target is k-th. m Pixel coordinates, k m ∈[0,K m -1],k m ∈Z + .

[0058] Step 1.3 (Generation of trailing rectangle): For any labeled target I b (m), find its centroid I. c Position coordinates of (m):

[0059]

[0060] The grayscale weighting was not used in formula (3) because the target width is only 3 to 4 pixels, which is greatly affected by noise, and weighting would cause more calculation errors. The centroid of the m-th marked target is taken as any point inside the rectangle, and the search direction is taken as the star tail angle ±θ. The fitted rectangle search is performed.

[0061] The fitting process is as follows Figure 3 As shown, W f =4, L f =26, therefore the maximum number of rectangles is 4W. f L f =416, the cyan dot is the centroid of the binary labeled region. When the rotation angle is θ, I... c (m) is a rectangle with its lower left corner as shown in the image. Figure 4 As shown, the coordinates of the lower left corner of the matrix when θ = 0 can be calculated from any point within the rectangle. Then, the coordinates of the four corners of the rectangle are obtained, thus yielding the set of generated rectangles R(m) for the m-th marked target. R(m,q) represents the set of rectangles generated by I... c (m) represents the width and length of any point, respectively W f and L f The set of rectangles, where q represents the q-th rectangle.

[0062] Step 1.4 (Tail Rectangle Fitting): Based on R(m,q), the corresponding image region is obtained as IN(R(m,q)). Then, the gray-level summation S(m,q) and histogram h(m,q) within the rectangular region are calculated. Therefore, the optimal rectangle solution constraint is shown in formula (4).

[0063]

[0064] Among them, S min With S max h represents the upper and lower limits for the sum of gray levels within a rectangular area. stThe standard rectangular region trailing target histogram is used. The three thresholds are all obtained statistically and are used to constrain formula (4) to obtain the optimal rectangle fit. Figure 5 Histogram features of rectangular regions in trailing images are given; Figure 6 The optimal rectangle fitting result is given.

[0065] Step 2: Clustering and Recognition of Partially Occluded Targets Based on Edge Features

[0066] After tail aggregation calculations, noise interference on stellar tails and the regions between them is eliminated. However, during debris target detection, situations where debris targets are obscured by stellar tails still occur. This disclosure classifies possible obscuration scenarios based on different characteristics of debris targets under obscuration:

[0067] Considering the width W of the stellar tail f =4, fragment target size L d =3, so when occlusion occurs, there are 3 cases based on the size of the occlusion, such as Figure 7 As shown:

[0068] Scenario 1: More than 2 / 3 is obscured, meaning only one line of the star's tail is visible. In this case, the debris target and the star's tail overlap, and although there is local brightness accumulation, it is difficult to distinguish them well due to noise interference.

[0069] Case 2: The obscuration is between 2 / 3 and 1 / 5, with more than two rows of star trails exposed. At this time, obvious bulges are formed at the edge of the trails, and the bulges have changes in brightness, which are clearly different from noise bulges. A large number of samples need to be generated to extract statistical features for debris target extraction.

[0070] Case 3: The obstruction is less than 1 / 5, and the debris target only has partial contact with the star's tail. This type of obstruction is relatively easy to distinguish based on rectangular fitting.

[0071] The distribution of the obscured targets is as follows Figure 8 The figure shows a Gaussian distribution, while the statistics for stellar tails are mainly based on a Rayleigh distribution. Therefore, in cases 2 and 3, debris targets can be distinguished based on the different distributions.

[0072] Figure 9 Six common scenarios of stellar tails occluding debris are presented, covering three occlusion scenarios: ab represents scenario 1, cd represents scenario 2, and ef represents scenario 3. For each scenario, the left side shows a single-frame grayscale image, and the right side shows a superimposed image of three consecutive frames. Figure 10 The features of the occluded fragment target image are presented. It can be seen that: although there are feature differences when occluded, the number of feature pixels is small, and they are difficult to detect quantitatively due to noise interference, so no distinguishing rules can be formed; although the statistical distribution differences are obvious, the number of fragment target pixels is small, and it is impossible to extract them directly based on the statistical distribution differences.

[0073] However, after multiple frames are superimposed, the grayscale features of the debris target are quite different from those of the star trail. Even when the target is completely obscured in case 1, people can still distinguish them based on consecutive frame images. That is, the target can be extracted by using the sequence joint features formed by multiple frame images.

[0074] Preferably, this disclosure employs a small fully connected convolutional network (FCN) to extract targets under occlusion conditions. The network is trained using occluded fragment target samples under different conditions. Based on three consecutive frames of data input, the edge "protrusion" information after rectangle fitting is used as a suspected target for FCN calculation, ultimately determining whether it is a fragment target. An exemplary small FCN design is shown in Table 1, containing 6 convolutional layers. The final output result is 0 for non-fragment targets and 1 for fragment targets.

[0075] Table 1. Small FCN Network Structure

[0076]

[0077]

[0078] Note: (×3) indicates an extended network used for continuous data input across three consecutive frames.

[0079] Step 3 addresses the issue of unsustainable inter-frame correlation caused by star trails obscuring target points. Correlation detection is performed based on the extracted partially obscured images to improve the success rate of target correlation and achieve effective fragment detection and tracking.

[0080] The above technical solutions are merely exemplary embodiments of the present invention. For those skilled in the art, based on the application methods and principles disclosed in the present invention, it is easy to make various types of improvements or modifications, and not limited to the methods described in the specific embodiments of the present invention. Therefore, the methods described above are merely preferred and not restrictive.

Claims

1. A method for detecting dark and weak gazing debris targets in a dense constant star long tail background, comprising the following steps: S1. Rectangular fitting of the tailing image; S2. Classifying the state of the target being obscured by the star tail in the fitted image, identifying the sequence joint features based on multi-frame superposition, and extracting the target semi-obscured image; S3. Correlation detection based on the extracted target semi-obscured image to realize effective detection and tracking of the debris; In the step S2, the state of the target being obscured by the star tail is divided into three categories, including: Case 1: Obstruction more than 2 / 3, i.e. only one row of the debris target is exposed to the star tail, at this time the debris target is superimposed with the star tail; Case 2: Obstruction between 2 / 3 and 1 / 5, i.e. more than two rows of the debris target are exposed to the star tail, at this time a clear protrusion is formed at the edge of the tail; Case 3: Obstruction below 1 / 5, the debris target only has local contact with the star tail; In the step S2, the method for extracting the target semi-obscured image specifically includes: Using a small fully connected convolutional network, based on the input of three consecutive frames of data, the edge "protrusion" information after rectangular fitting is used as a suspected target for fully connected convolution calculation, and finally it is determined whether it is a debris target.

2. The method of claim 1, wherein, The method for rectangular fitting of the tailing image in step S1 specifically includes: Step S1.1, define the image as where N is the image pixel size, define two square operators of different sizes, on the image Compute its contrast ; Step S1.2, for Perform traditional binary annotation calculations to obtain the labeled M-target set: wherein, , , m represents the mth target after marking, , ; represents the maximum number of pixels of the mth marked target; represents the coordinate of the mth pixel of the mth marked target, , , ; Step S1.3, for any one of the labeled targets , the position coordinates of its centroid are calculated. Take the centroid of the mth marker target as an arbitrary point in the rectangle, and take the star tail angle As the search direction, the fitting rectangular frame search is carried out to obtain the generated rectangular set of the mth marker target , Indicate the set of rectangles with Wf and Lf as the width and length of the rectangle respectively, and q represents the qth rectangle an arbitrary point Step S1.4, according to corresponding image region is calculated as , the gray sum in the rectangular region is calculated and the histogram , the optimal rectangular solution constraint is obtained as shown in equation (4): (4) wherein, with is the upper and lower limit of the sum of the gray scale in the rectangular region, is the standard rectangular region tail target histogram, and the three thresholds are all obtained by statistics, which are used to constrain the optimal rectangular fitting in formula (4).

3. The method of claim 2, wherein, In the step S1.1, the square operators are respectively taken as 3x3 and 7x7.

Images