CN115457383A

CN115457383A - Method for rapidly detecting high-orbit target of large-field telescope by using clustering algorithm

Info

Publication number: CN115457383A
Application number: CN202211002187.XA
Authority: CN
Inventors: 王松; 刘佳伟; 高澜; 邵圣祥; 张荣杰; 张福军; 马黎俊; 贾鹏; 万欣
Original assignee: 63610 Troops Of Chinese Pla
Current assignee: 63610 Troops Of Chinese Pla
Priority date: 2022-08-21
Filing date: 2022-08-21
Publication date: 2022-12-09

Abstract

The invention provides a method for rapidly detecting a high-orbit target of a large-view-field telescope by using a clustering algorithm. And then, completing the rapid detection of the high-orbit target through the steps of registration alignment of the image sequence, filtering of stars, clustering of satellite candidate targets, detection after clustering and the like. Different from the existing detection method, the method fully considers the moving characteristics of the high-orbit target in the ground-based large-view-field telescope, and introduces the clustering algorithm into the target detection process, thereby greatly improving the detection speed and improving the target detection stability.

Description

Method for rapidly detecting high-orbit target of large-field telescope by using clustering algorithm

Technical Field

The invention belongs to the field of space situation perception, and particularly relates to a method for quickly detecting a high-orbit target of a large-view-field telescope by utilizing a clustering algorithm.

Background

With the increasing use of space by human beings, the number of space spacecrafts is rapidly increased. High-orbit spaces represented by geosynchronous orbits are becoming increasingly crowded. Once collision occurs, even if the collision is a slight scratch, the satellite load is failed or the whole satellite is scrapped, and secondary collision can be caused by debris generated by collision. In order to ensure that the high-value satellite target of one party is prevented from being impacted, the operation state of the space target needs to be monitored, and space situation perception is achieved, so that timely early warning is provided for possible impact, and disasters are avoided.

Because the distance between the high-orbit target and the ground is usually in the order of tens of thousands of kilometers, the ground-based radar is difficult to realize effective detection, and even if the advanced radar can detect the high-orbit target, the high-orbit target cannot be continuously monitored in space for a long time due to high operation cost. Therefore, it is a common method in the industry to use a ground-based large-field telescope for spatial high-orbit target monitoring. The field width of the ground-based large-field telescope generally reaches more than 2 degrees, and the resolution of the shot image can reach the level of 4K or 6K. The high orbit satellite target can be effectively captured by adopting the exposure time of 2-5 seconds.

The high-orbit satellite images captured by the large-field telescope contain thousands of light spots (i.e., light sources), including several (e.g., several or tens) of high-orbit satellite targets. Without prior information, how to detect the high-orbit satellite target from the shooting result is a problem worthy of study.

Disclosure of Invention

In order to solve the problems of situation perception and monitoring of high orbit targets at present, the invention monitors high orbit artificial celestial body targets by using a foundation large-field telescope and provides a large-field telescope high orbit target rapid detection method by using a clustering algorithm, which can automatically detect the artificial celestial body targets from an observation image sequence, thereby providing technical support for space target orbit determination and collision early warning.

Considering that target light spots in a single-frame image are very similar to natural celestial body light spots, the method adopts an algorithm idea of Track-before-detect, and can quickly detect the satellite target from a continuously shot multi-frame image sequence. Different from the existing detection method, the method fully considers the moving characteristics of the high-orbit target in the ground-based large-view-field telescope, and introduces the clustering algorithm into the target detection process, thereby greatly improving the detection speed and improving the target detection stability.

The technical scheme of the invention is as follows:

a method for quickly detecting a high-orbit target of a large-field telescope by using a clustering algorithm comprises the following steps:

step 1: acquiring continuous N frames of observation images of a space domain where a high-orbit target to be detected is located by a large-field telescope, performing photometric processing, and respectively determining the positions of light sources in respective image coordinate systems in each frame of image;

step 2: carrying out registration alignment on an image sequence consisting of N continuous observation images;

and step 3: determining a candidate target set:

step 3.1: combining the coordinates of the light source in the N frames of observation images after registration alignment under the 1 st frame of image coordinate system into a matrix M; the size of the matrix M is M × 2,m represents the total number of light sources in the N frames of images, and 2 represents two columns which are respectively X coordinates and Y coordinates;

step 3.2: setting key parameters in a clustering method, and clustering elements in the matrix M by using the clustering method; the key parameter is the distance radius r of the clustering sample _cluster And minimum number of samples n of the cluster _min ；

Step 3.3: after the clustering of the step 3.2, the light sources which are successfully clustered are light sources to be filtered, and the light sources which are not successfully clustered form a satellite candidate target set;

and 4, step 4: detecting satellite targets in the candidate target set by using a clustering algorithm:

step 4.1: clustering the light source points in the satellite candidate target set by using a clustering method, and setting r _cluster ＝2*v*t～8*v*t，n _min ＝3；

And 4.2: clustering in the step 4.1 to obtain a plurality of clustering clusters, wherein the number of points of each cluster is not less than 3; and for each cluster, obtaining a point trace which meets the straight line characteristic by adopting a straight line detection method, namely the point trace is the detected satellite target.

Further, in step 2, the following steps are adopted to perform registration alignment on the image sequence:

step 2.1: respectively acquiring the coordinates of the first N brightest light source points in the image coordinate systems of the ith frame of observation image and the (i + 1) th frame of observation image to form coordinate sets A and B;

step 2.2: respectively calculating the distance from each point in the coordinate set A to all points in the coordinate set B; for a certain point in the coordinate set A, obtaining the minimum value of the distances between the certain point and all the points in the coordinate set B and the point in the coordinate set B corresponding to the minimum value, further forming a minimum distance set by the minimum distances corresponding to all the points in the coordinate set A, and obtaining a point pair corresponding to each minimum distance;

step 2.3: counting a mode in the minimum distance set to obtain a plurality of minimum distances corresponding to the mode, and further obtaining a plurality of point pairs corresponding to the minimum distances corresponding to the mode, wherein the point pairs are matched characteristic point pairs in the ith frame observation image and the (i + 1) th frame observation image;

step 2.4: obtaining a coordinate difference value in the X direction and the Y direction by using the characteristic point pairs, namely obtaining the displacement of the ith frame of observation image and the (i + 1) th frame of observation image;

step 2.5: and overlapping the displacement between the observation images of the adjacent frames to obtain the displacement between the rest of the observation images in the image sequence and the observation image of the 1 st frame, and further realizing the registration and alignment of the rest of the observation images in the image sequence and the observation image of the 1 st frame.

Further, a clustering method DBSCAN based on density characteristics is adopted in the step 3 for clustering.

Further, the distance radius r of the cluster sample _cluster Greater than the positional deviation after alignment of the stars and less than the number of pixels of the satellite target moving across the adjacent frames.

Further, the distance radius r of the cluster sample _cluster =3; minimum number of samples n of a cluster _min ＝2。

Further, in step 3, before clustering, a boundary constraint condition is added to the matrix M, and only the light sources in the overlapping region after registration and alignment of the N observation images are clustered, and the light sources outside the overlapping region are ignored.

Further, in step 4, for a plurality of clustering clusters, a parallel computing mode is adopted to perform straight line detection.

Further, in step 4, when the number of points in the cluster is smaller than a set threshold, a traversal method is adopted to determine a straight line point trace, otherwise, a Hough transformation method is adopted to determine the straight line point trace.

The process of determining the straight line point trace by the traversal method comprises the following steps: each point in the cluster is from N frames of images in the current image sequence; selecting at least 3 points belonging to different frames from the cluster each time to form a trace point, calculating the speed according to the time sequence by combining the coordinates and the time of each point in the trace point, and if the difference of the speed vectors of each point in the trace point is less than a set threshold, determining that the trace point meets the detection requirement of the high-orbit satellite target and recording and outputting the trace point.

The process of determining the straight line point trace by the Hough transformation method comprises the following steps: sequentially carrying out XY and TX or XY and TY two-layer hough transformation linear detection on the point sets in the cluster, wherein T, X, Y respectively represent a time axis, an X axis and a Y axis; and outputting the trace points meeting the straight line characteristics.

Advantageous effects

The method adopts the algorithm idea of first correlation and then detection (Track-before-detect), and can quickly detect the satellite target from a continuously shot multi-frame image sequence. Different from the existing detection method, the method fully considers the moving characteristics of the high-orbit target in the ground-based large-view-field telescope, and introduces a clustering algorithm into the target detection process, so that the detection speed is greatly improved, and the target detection stability is also improved.

Additional aspects and advantages of the invention will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the invention.

Drawings

The above and/or additional aspects and advantages of the present invention will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:

FIG. 1: an image sequence registration method;

FIG. 2: displacement schematic diagram in the image registration process;

FIG. 3: a schematic diagram of the distribution situation of the target straight line;

FIG. 4: light source points from the 5 frames of images after light source detection and image registration;

FIG. 5: light source points (local) from 5 frames of images after light source detection and image registration;

FIG. 6: clustering the candidate target set after filtering stars;

FIG. 7: clustering and filtering a candidate target set (local) after stars are filtered;

FIG. 8: clustering results of the candidate target set;

FIG. 9: performing XY linear detection on each cluster to obtain a result;

FIG. 10: and performing TX linear detection on each cluster.

Detailed Description

The following detailed description of embodiments of the invention is intended to be illustrative, and not to be construed as limiting the invention.

The method is used for space situation perception, the high-orbit artificial celestial body target is monitored by the aid of the large-view-field ground telescope, the clustering algorithm is introduced into a high-orbit satellite target detection process, the artificial celestial body target can be automatically detected from an observation image sequence, and target detection efficiency is greatly improved.

In this embodiment, according to the idea of association and detection, photometric processing is performed on a certain image sequence (5 frames of continuously-taken pictures of the ground-based large-field-of-view telescope), and the position of a light source in each frame of image (i.e., the coordinates of the light source in the image, using the upper left corner of the image as the origin) is determined respectively. And then, completing the rapid detection of the high-orbit target through the steps of registration alignment of the image sequence, filtering of stars, clustering of satellite candidate targets, detection after clustering and the like.

Registration alignment of image sequences:

in terms of observation distance, the distance from the artificial satellite and the natural celestial body to the ground-based telescope is far from the observation distance. When the observation direction of the telescope is unchanged, the distance difference of the observation target is reflected on the imaging plane, namely the artificial satellite moves faster. There are two observation modes, one is that the telescope observes with the fixed star, and the fixed star is consistent in the position of the front and back two frames of images in the obtained observation image sequence, and the other is that the telescope observes with the satellite, and the satellite target positions in the front and back two frames of images are consistent. Regardless of the observation mode, the shuttle movement of the artificial satellite among the natural celestial bodies can be generally summarized. To achieve this effect, it is necessary to align the natural celestial bodies in multiple images with respect to a specific frame (e.g., the first frame of the sequence), i.e., to embody the idea of correlation and detection.

The existing image registration method based on the feature points can achieve a good alignment effect, but when real-time application is further needed, the problems of large calculation amount and overlong calculation time exist, so the invention further discovers through analysis that the calculation amount of extracting the feature points (such as corner points) in a general algorithm is large, and the feature points are only used for aligning images in practice. Therefore, through further analysis, the essential characteristics of the image which can be shot by using the telescope are determined, and because a wide-field telescope image is adopted, thousands of fixed stars exist in one frame of image, and the brightness of the brightest fixed stars (such as 50 fixed stars) is much higher than that of the high-orbit satellite, the brightest fixed stars (such as 50 fixed stars) in the field of view can be directly selected, and the corresponding coordinate points are taken as the characteristic points representing the image, so that the complicated process of calculating the characteristic points of the image in a common scene is omitted.

After the feature point sets of the two frames of images are obtained, the displacement between the two feature point sets can be obtained by utilizing statistical analysis, if image rotation exists, an additional rotation angle needs to be obtained, and therefore image registration is guaranteed, but for a foundation large-field telescope, the imaging time interval is 3 seconds generally, the rotation angle between adjacent frames can be almost ignored, and therefore only linear displacement is considered here. Note that the two feature point sets do not necessarily have a one-to-one correspondence, i.e., there may be individual feature points in one of the point sets that cannot find a matching point in the other point set. However, the shooting time interval of the two frames of images before and after is short (several seconds), and most stars in the field of view exist in the two frames of images, so that most points can be guaranteed to be matched.

A specific registration algorithm for the image sequence, i.e. a displacement calculation method from the second frame to the first frame of the image, is shown in fig. 1.

Firstly, respectively acquiring the coordinates of the first N brightest light source points in the first frame image and the second frame image in respective image coordinate systems to form coordinate sets A and B;

secondly, respectively calculating the distance from each point in the coordinate set A to all points in the coordinate set B; for a certain point in the coordinate set A, obtaining the minimum value of the distances between the certain point and all the points in the coordinate set B and the point in the coordinate set B corresponding to the minimum value, further forming a minimum distance set by the respective minimum distances of all the points in the coordinate set A, and obtaining a point pair corresponding to each minimum distance;

thirdly, counting the modes in the minimum distance set, taking actual measurement errors into consideration, performing mode counting after error interception on elements in the minimum distance set to obtain a plurality of minimum distances corresponding to the modes, and further obtaining a plurality of point pairs corresponding to the minimum distances corresponding to the modes, wherein the point pairs are feature point pairs matched with each other in the first frame image and the second frame image;

and finally, taking the coordinate difference value of the X direction and the Y direction by using a certain point pair in the characteristic point pairs, namely the displacement of the first frame image and the second frame image.

Therefore, the displacement of the first frame image and the second frame image, the displacement of the second frame image and the third frame image, the displacement of the third frame image and the fourth frame image, and the displacement of the fourth frame image and the fifth frame image in the 5-frame continuous shooting images of the ground-based large-field-of-view telescope can be obtained, and the displacement of the third frame image, the fourth frame image, the displacement of the fifth frame image and the first frame image can be obtained through the accumulation of the displacements.

In this way, the registration of the image sequence is realized, and in the step 3 of the flow chart, individual points which cannot be matched can be filtered out by calculating the mode, namely filtering outliers, so that the good matching of most points is ensured. The algorithm is simple and efficient, and the required calculation time can be ignored compared with the time required by detecting the position of the light source in the image.

Most fixed star targets are filtered out by using a clustering algorithm, and a candidate target set is determined

The location of the light sources in the image is mostly representative of stars, of which there are only a few high-orbit satellites. For example, a frame of image has more than 10000 light source points, but only contains 5 high-track objects. If a majority of stars can be filtered by a method, and targets are screened from the filtered satellite candidate set, the calculation efficiency can be greatly improved. Here, a clustering method is used, and clustering is performed using coordinate (XY coordinate) information of the light source. Further, repeated attempts find that the clustering method DBSCAN based on the density features can achieve a good effect. The specific operation is as follows.

Given an image sequence (5 consecutive images), using the aforementioned registration method, the image registration shift has been obtained, i.e. frame i (i)>1) The amount of translation of the light source position in the image relative to the light source position in the 1 st frame image. Accordingly, the light source positions in the image sequence are uniformly aligned to the coordinate system of the 1 st frame image. Combining the coordinates of the light source points in each frame of image in the image sequence under the coordinate system of the 1 st frame of image into a matrix M, wherein the size of the matrix M is M × 2,m and represents the total light source points in 5 frames of imagesThe number, 2, represents two columns, X and Y coordinates respectively. Clustering the matrix M by using a DBSCAN method, and setting two key parameters in the clustering as the distance radius r of a clustering sample _cluster =3 (i.e. 3 pixels) and minimum number of samples n of the cluster _min And (2). And after clustering, obtaining cluster identifications of all light source points, wherein the star light source to be filtered is successfully clustered, and the satellite candidate target set is formed after the clustering is unsuccessful.

Parameter r _cluster Representing the distance between the samples in the cluster, which should be greater than the positional deviation after alignment of the stars and less than the number of pixels the satellite target moves across the frame. Because the star positions in the continuous multi-frame images are aligned, the deviation of the same star position in the continuous 5-frame images is small (the deviation is found by tests to be generally not more than 1 pixel), the adjacent frame moving pixel number of the satellite target generally exceeds more than ten pixels (for example, the adjacent frame imaging time interval is 3 seconds, the corresponding high-orbit satellite moving angle is about 45 arc seconds, each pixel is about 2 arc seconds, and about 22.5 adjacent frame moving pixels are obtained), and the distance between the light sources in the large-field telescope is far larger than 3 pixels, so r is set here _cluster =3. Parameter n _min And setting the minimum value of the number of samples contained in each cluster representing successful clustering as 2 to represent the lowest threshold for filtering stars, so that stars can be filtered to the maximum extent.

In this embodiment, after the light source points in each image are filtered out of stars, the number of points entering the candidate target set is reduced by about one order of magnitude, for example, from 1 ten thousand points to 1 thousand points.

In addition, because a certain displacement (typical value is about 20 pixels across a frame) exists in the image registration process, although the displacement is very small (neglect detection rate is ignored) compared with the image with 4k or 6k resolution, if the displacement is not processed properly, the star filtering effect is affected, the number of candidate light collection sources is increased, and the subsequent target detection is not good. Therefore, boundary constraints can be added to the matrix M before clustering filters out stars. As shown in fig. 2, in the effect diagram obtained after the first 3 frames of images are aligned to the first frame of coordinate system, the coordinate set actually participating in the clustering is all coordinate points in the M matrix that satisfy the shadow constraint condition (i.e., the overlap region). For example, with the upper left corner of the first frame as the origin, the points in the second frame and the third frame with X-coordinate value or Y-coordinate value smaller than 0 do not participate in the clustering, and the points in the second frame and the third frame with X-coordinate value or Y-coordinate value larger than Q-w, where Q is the maximum value of X-coordinate or Y-coordinate of the first frame image, such as an image with 4k resolution, the maximum value of coordinate is 4096, and w is the displacement of the fifth frame image from the first frame image do not participate in the clustering.

Detecting satellite objects in a set of candidate objects using a clustering algorithm

After filtering fixed stars to obtain a candidate target set, detecting straight lines from XY and TX (or TY) levels sequentially by using linear movement characteristics of a high-orbit satellite target, for example, by using a Ranac method and a Hough transformation method, wherein T, X, Y respectively represent a time axis, an X axis and a Y axis, and a satellite target point trace is obtained through straight line detection. In practical application, however, we find that this kind of method needs to traverse the whole possible linear direction interval from 0 to 180 degrees, and even if the method is divided by 1 ° interval, it needs to cycle 180 times, and the size of the candidate set obtained for 5 frames of continuous images may be 5000 points (1000 per frame), so the calculation amount is still relatively large. Therefore, the invention further provides a method of clustering and detecting, thereby realizing the divide-and-conquer of the candidate target set. The method comprises the following specific steps:

clustering the candidate target set by adopting DBSCAN, and setting r _cluster ＝2*v*t，n _min =3, where v represents the estimated moving speed (unit: pixel/second) of the satellite object, and t is the inter-frame imaging time interval, i.e., r _cluster Which is 2 times the distance the satellite target moves across the frame. In practical application, the high orbit can be adjusted to 2 times to 8 times according to the situation (because the high orbit has a large elliptic orbit star besides the synchronous star). n is _min The value 3 is considered in compromise, the false alarm rate is easily increased when the value is less than 3, and missing detection is easily caused under the condition of 5 frames of continuous image sequences when the value exceeds 3. In the case of a straight line of targets, they must be grouped into clusters. As shown in case (1) in fig. 3, there are three target points arranged in a straight line (an approximate straight line formed by setting a threshold value), when r is set _cluster Not less than continuous thereinAnd when the distance between the two points is larger than the preset threshold value, the three black solid points in the graph can be clustered into a certain cluster, namely the clustering is successful. In practical cases, there may be a case where the 2 nd point is not detected in the first photometry, as shown in the case (2) and the case (3) in the figure, and it is necessary to increase r at this time _cluster To ensure that three points on the same line are clustered into one cluster. That is, consider r _cluster Optimization was adjusted between 2 and 8.

And obtaining a plurality of clusters after clustering, wherein the number of points of each cluster is not less than 3. Conventional methods such as Hough transform can be used to derive points in each cluster that satisfy an approximately rectilinear distribution. Since the number of points of each cluster is greatly reduced relative to the whole candidate set at this time, generally reduced by at least 2-3 orders of magnitude, and the amount of calculation is reduced as a whole by combining parallel calculation of a plurality of clusters.

In order to further improve the calculation speed, different straight line detection methods can be adopted according to the number of points in the cluster. In this embodiment, when the number of points in a cluster is less than 8, a traversal method can be adopted to quickly determine the trace of the point; otherwise, a Hough transform method is adopted.

The method comprises the following steps: each point in the cluster is from 5 frames of images of the current sequence, and each time, one point is selected from each frame of image to form a point trace. And calculating the speed according to the sequence from the first frame to the last frame by combining the coordinates and the time of each point, and if the speed vectors (the direction and the module value) are very close to each other (reach a certain threshold), considering that the trace of the point accords with the detection requirement of the high-orbit target and recording and outputting the trace. Therefore, the high-orbit satellite target can be quickly detected in a traversing mode.

Hough transform method: and performing XY and TX (or TY) two-layer hough transformation linear detection on each cluster point set, and finally outputting a target point trace meeting the requirement. Firstly, point traces which are distributed in an approximate straight line on an XY plane are obtained through XY detection, then, the point traces which also meet the straight line characteristics on a TX (or TY) plane are obtained through TX (or TY) detection again, and the point traces are recorded and output.

Outputting the result

The output result needs to be noted that, because the displacement amount is superimposed on each frame image in the sequence image registration process, before the output result is output, the translation amount needs to be removed from the finally determined target coordinates, so as to obtain the specific coordinates of the target in each frame image, that is, the target position in each original image. And combining the right ascension and declination data corresponding to the image center recorded in the original image and the angle value represented by one pixel of the image to obtain the right ascension and declination values corresponding to the target. And thus can be used for result verification and target tracking.

Analysis of results

The following observations were made from a telescope of some type with an image resolution of 4096 x 4096. And taking continuously shot 5 frames of images to detect the target. The following results are obtained through operations of light source detection, image registration, clustering, constant star filtering and the like. The average processing time per frame does not exceed 2 seconds.

The results after light source detection and image registration are shown in fig. 4. The figure includes all light sources for 5 frames of image, including stars and artificial celestial bodies, with a total of 85008 light source points, averaging about 17000 per frame. Local details as shown in fig. 5, the visible images are registered and the light source point line darkening is the effect of light source superposition from multiple frames. The lighter line color represents that the light source is only from a certain frame of image.

Most of the sidereal targets are filtered by using a clustering algorithm, and a candidate target set is determined as shown in FIG. 6, and local conditions are shown in FIG. 7. The total number of light sources for the candidate target set is 3236, averaging about 647 per frame. The 85008 light sources before filtering the stars are suddenly reduced to 3236 light sources, and the clustering method can be seen to obtain a good effect on filtering the star targets.

Satellite targets are detected in the candidate target set by using a clustering algorithm, and 285 clustering clusters are obtained in total, as shown in fig. 8, 1187 candidate points (previously 3236).

And performing XY linear detection firstly and then TX linear detection on each cluster. The process can be processed in parallel across multiple clusters. The results obtained after XY straight line detection are shown in fig. 9, and the results obtained after TX straight line detection are shown in fig. 10. The results show that there are 5 high-orbit satellite targets in the final detection image.

The final position results of the 5 high-orbit satellite targets can be output through displacement (image registration displacement before compensation).

Although embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are exemplary and not to be construed as limiting the present invention, and that those skilled in the art may make variations, modifications, substitutions and alterations within the scope of the present invention without departing from the spirit and scope of the present invention.

Claims

1. A method for quickly detecting a high-orbit target of a large-field telescope by using a clustering algorithm is characterized by comprising the following steps of: the method comprises the following steps:

and 2, step: carrying out registration alignment on an image sequence consisting of N continuous observation images;

and 3, step 3: determining a candidate target set:

step 3.1: combining the coordinates of the light sources in the N frames of observation images after registration alignment under the 1 st frame of image coordinate system into a matrix M; the size of the matrix M is M × 2,m represents the total number of light sources in the N frames of images, and 2 represents two columns which are respectively X coordinates and Y coordinates;

Step 4.2: clustering in the step 4.1 to obtain a plurality of cluster clusters, wherein the number of points of each cluster is not less than 3; and for each cluster, obtaining the point trace which meets the linear characteristics by adopting a linear detection method, namely the detected satellite target.

2. The method for rapidly detecting the high-orbit target of the large-field telescope by using the clustering algorithm as claimed in claim 1, wherein the method comprises the following steps: in step 2, the following steps are adopted to register and align the image sequence:

step 2.2: respectively calculating the distance from each point in the coordinate set A to all points in the coordinate set B; for a certain point in the coordinate set A, obtaining the minimum value of the distances between the certain point and all the points in the coordinate set B and the point in the coordinate set B corresponding to the minimum value, further forming a minimum distance set by the respective minimum distances of all the points in the coordinate set A, and obtaining a point pair corresponding to each minimum distance;

step 2.3: counting the mode in the minimum distance set to obtain a plurality of minimum distances corresponding to the mode, and further obtaining a plurality of point pairs corresponding to the minimum distances corresponding to the mode, wherein the point pairs are matched characteristic point pairs in the ith frame observation image and the (i + 1) th frame observation image;

3. The method for rapidly detecting the high-orbit target of the large-field telescope by using the clustering algorithm according to claim 1, is characterized in that: and 3, clustering by adopting a clustering method DBSCAN based on density characteristics.

4. The method for rapidly detecting the high-orbit target of the large-field telescope by using the clustering algorithm according to claim 1 or 3, characterized in that: distance radius r of the cluster sample _cluster Greater than the positional deviation after alignment of the fixed stars and less than the number of pixels the satellite target moves across adjacent frames.

5. The method for rapidly detecting the high-orbit target of the large-field telescope by using the clustering algorithm as claimed in claim 4, wherein the method comprises the following steps: distance radius r of the cluster sample _cluster =3; minimum number of samples n of a cluster _min ＝2。

6. The method for rapidly detecting the high-orbit target of the large-field telescope by using the clustering algorithm according to claim 1, is characterized in that: in step 3, before clustering, a boundary constraint condition is added to the matrix M, only the light sources in the overlapping region after registration and alignment of the N frames of observation images are clustered, and the light sources outside the overlapping region are ignored.

7. The method for rapidly detecting the high-orbit target of the large-field telescope by using the clustering algorithm according to claim 1, is characterized in that: and 4, performing linear detection on the plurality of clustering clusters in a parallel computing mode.

8. The method for rapidly detecting the high-orbit target of the large-field telescope by using the clustering algorithm according to claim 1, is characterized in that: and 4, determining a straight line point trace by adopting a traversal method when the number of points in the clustering cluster is less than a set threshold, otherwise determining the straight line point trace by adopting a Hough transformation method.

9. The method for rapidly detecting the high-orbit target of the large-field telescope by using the clustering algorithm according to claim 8, is characterized in that: the process of determining the straight line trace points by the traversal method comprises the following steps: each point in the cluster is from N frames of images in the current image sequence; selecting at least 3 points belonging to different frames from the cluster to form a point trace, calculating the speed according to the time sequence by combining the coordinates and the time of each point in the point trace, and if the speed vector difference of each point in the point trace is less than a set threshold, determining that the point trace meets the detection requirement of the high-orbit satellite target and recording and outputting the point trace.

10. The method for rapidly detecting the high-orbit target of the large-field telescope by using the clustering algorithm according to claim 8, is characterized in that: the process of determining the straight line point trace by the Hough transformation method comprises the following steps: sequentially carrying out XY and TX or XY and TY two-layer hough transformation linear detection on the point sets in the cluster, wherein T, X, Y respectively represent a time axis, an X axis and a Y axis; and outputting the trace points meeting the straight line characteristics.