Long-distance image multi-target recognition method for satellite constellations

Abstract

The invention discloses a long-distance image multi-target recognition method for satellite constellations, which belongs to the field of space target recognition technology. Aiming at the large number of giant constellation satellites and the existence of new and disappearing behaviors in the field of view, the invention uses satellite images taken by a space-based platform to extract four-connected regions and perform Gaussian diffusion on the regions to obtain preliminary multi-target recognition results; describes the results with a rectangular frame, obtains the grayscale histogram of each target, and obtains the likelihood function in Gaussian form; obtains a random set of color measurement values ​​of the target by calculating the prediction probability hypothesis density; describes the target state and measurement in the form of a random finite set, and estimates the satellite target state using a Gaussian mixture probability hypothesis density filter to avoid the problem of "data disaster" caused by a large number of targets and data association; considers the birth and disappearance of targets, tracks the multi-targets as a whole, and accurately and stably recognizes the multi-targets under the condition of time-varying number and state.

Description

Technical Field

[0001] The present invention relates to a space multi-target tracking method, in particular to a recognition method suitable for multiple weak imaging targets under a complex starry sky background, and belongs to the technical field of space target recognition. Background Art

[0002] With the continuous development of giant satellite constellations, the number of satellites in the constellation has increased significantly, and space monitoring and situational awareness have become increasingly important. In order to accurately obtain space situation information and improve the ability of safe and continuous monitoring, it is particularly important to identify and track multiple targets in space. Using low-orbit space-based platforms to detect low-orbit targets has the advantages of high measurement accuracy and strong concealment, and has become the main way to track and monitor space targets. However, the number of giant constellation satellites is huge, the detection range of the space-based platform is limited, and satellites are born and disappear in the field of view, making it difficult to track multiple targets. At the same time, deep space exploration tracking imaging distance is far, and the real target often appears as a weak point in the image, without contour, texture and shape features, which is difficult to distinguish from stars, resulting in detection difficulties. Accurately identifying the state of satellite targets is an urgent problem to be solved.

[0003] Among the proposed multi-target tracking methods, the previous technology (see Kong Sijie. Research on the extraction technology of faint targets under dense star background [D]. 2019.) proposed a multi-target rapid detection technology combined with background clutter suppression. The target position and track information were extracted by combining background suppression with trajectory association. However, this technology has shortcomings. On the one hand, it is necessary to accurately eliminate the complex starry sky background, which greatly increases the amount of calculation; on the other hand, data association is required to establish the correspondence between frames. When the number of targets increases, it will cause "data disaster" and the motion information of the target cannot be correctly obtained. Therefore, it is necessary to conduct research on the problem of multi-target recognition of long-distance images for satellite constellations, and provide accurate recognition methods for multiple targets under complex starry sky backgrounds, so as to achieve accurate estimation of satellite target status. Summary of the invention

[0004] In view of the problem of a large number of satellites and complex image backgrounds, the main purpose of the present invention is to provide a long-distance image multi-target recognition method for satellite constellations. The satellite images taken by the space-based platform are used to extract four-connected regions and perform Gaussian diffusion processing on the regions to obtain preliminary multi-target recognition results; the results are described with a rectangular frame, the grayscale histogram of each target is obtained, and the likelihood function in Gaussian form is obtained; the random set of color measurement values ​​of the target is obtained by calculating the predicted probability hypothesis density PHD. The target state and measurement are described in the form of a random finite set, and the target state is estimated using a Gaussian mixture probability hypothesis density filter to avoid the problem of a large number of targets and data association that easily leads to "data disasters"; considering the birth and disappearance of targets, the multiple targets are tracked as a whole, and the multiple targets are accurately and stably recognized when the number and state are time-varying.

[0005] The present invention is achieved through the following technical solutions.

[0006] The invention discloses a satellite constellation-oriented long-distance image multi-target recognition method, comprising the following steps:

[0007] Step 1: Use the frame difference method to pre-process the image taken by the navigation camera, perform corrosion operation on the processed image, and obtain the four-connected area of ​​the image, which includes the satellite four-connected area and the star four-connected area, and define the four-connected area as the quasi-target; perform Gaussian diffusion processing on the obtained quasi-target to diffuse the star point target from a small number of pixels to multiple pixels as the preliminary recognition result of multiple targets. Gaussian diffusion processing makes it easy to obtain the measurement results of the quasi-target after diffusion.

[0008] The specific implementation method of step 1 is:

[0009] Step 1.1 pre-processes the satellite images taken by the space-based platform to obtain detection results including quasi-satellites and star targets.

[0010] Distant space targets often appear as faint dots in images, occupying a small number of pixels and without contours, textures, and shape features, making it difficult to distinguish between distant space targets and stars. At the same time, due to the movement of the observation platform, the starry sky background has relative motion between different observation frames of the tracking target. Step 1.1 aims to detect the satellite to be tracked with high precision, while eliminating the influence of some stars on tracking.

[0011] The satellite images taken by the space-based platform are preprocessed using the frame difference method. The image A is obtained by subtracting adjacent frames to eliminate some star interference. The image A is eroded using the structural element B to obtain the image C containing satellites and stars. The result of eroding A with B is the set of all a's that remain in A after B is translated by a, satisfying:

[0012]

[0013] in(·) M The pixel matrix representing the image. In step 1, satellites and stars are collectively referred to as quasi-targets. The four-connected region of C is obtained to obtain the recognition result of the quasi-target, which is recorded as T i .

[0014] Step 1.2 performs Gaussian diffusion processing on the target, so that the star point target diffuses from a small number of pixels to multiple pixels as the preliminary multi-target recognition result.

[0015] Since the quasi-target extracted in step 1.1 occupies only a small number of pixels, it is not convenient to obtain measurement information. The defocus method is adopted to convert the star target T i A single pixel diffuses to multiple pixels, and the Gaussian distribution of the point spread function satisfies:

[0016]

[0017] Where E is the total energy of the target, ρ represents the energy concentration of the point spread function, and formula (2) is used to ensure that the total energy of the pixel after diffusion is close to T i The total energy of the target is calculated and the diffusion result is used as the preliminary recognition result of multiple targets to facilitate the acquisition of measurement information.

[0018] Step 2: The preliminary recognition results of the multiple targets obtained in step 1 are described by rectangular boxes respectively, and the size and position of each rectangular box are used as the state of the multiple targets; the grayscale histogram of each rectangular box is obtained, and the object is a satellite constellation. Considering the consistency of the satellite structure in the constellation, the same template grayscale model is used for different satellites, and the Bartcharia distance is used to characterize the similarity between the target and template grayscale histograms, and a Gaussian similarity function between the target and the template is constructed; the predicted probability hypothesis density PHD of the image is obtained and the peak value is collected from its product with the likelihood function as a random set of color measurement values ​​of the target.

[0019] The specific implementation method of step 2 is:

[0020] Step 2.1 describes the multi-target recognition results of step 1 with a rectangular box to obtain a grayscale likelihood function in Gaussian form.

[0021] A single target state is described by a rectangular box as shown in formula (3):

[0022] x={x i ,y i ,L i ,H i},i=1,2,...,n (3)

[0023] Where n is the number of multi-targets initially identified in step 1.2, and the center and size of the rectangular box are respectively represented by {xi ,y i} and {L i ,H i The grayscale histogram of the target is represented as p(u), the grayscale histogram of the template is represented as q(u), and the similarity function between the target and the template is characterized by the Bartcharya distance:

[0024]

[0025] The object-oriented model is a satellite constellation. Considering the consistency of the satellite structure in the constellation, the same template grayscale model is used for different satellites. The grayscale likelihood function is defined as:

[0026]

[0027] Among them, P is the current image, σ 2 is the noise variance.

[0028] Let v k (x) is the PHD at time k, v k|k-1 (x) is the predicted PHD at time k, then:

[0029]

[0030] Where L P (x) is the grayscale likelihood function, so Also v k The peak value of (x).

[0031] Step 2.2 calculates the predicted PHD of the image and collects a random set of color measurements whose peaks are taken as targets from its product with the likelihood function.

[0032] The Gaussian mixture prediction probability hypothesis density v calculated for the image at time k k|k-1 (x) is sampled, and the particles are recorded as: Among them, l i represents the integer number of the particle, w i represents the weight of the particle, then the posterior probability density v k (x) Use {x i ,l i ,φ i} is approximately replaced by, where φ i =L(x i w i ). Cluster the particles and obtain the set of measurement results at time k:

[0033] Z f,k = {Z f,1 ,Z f,2 ,...,Z f,n} (7)

[0034] The result of formula (7) is a random set of color measurements.

[0035] Step 3: Describe the multi-target states and measurement values ​​in the form of random finite sets, use the Gaussian mixture probability hypothesis density GMPHD filter to estimate the satellite target states, and track multiple satellite targets under complex backgrounds based on the estimation results. There is no need to associate the trajectories of each target during tracking, thereby reducing the amount of calculation. Estimation includes two processes: prediction and update. During the prediction process, the Gaussian mixture prediction probability hypothesis density is obtained based on the posterior intensity at time k-1. During the update process, the PHD in the Gaussian mixture probability hypothesis density filter is updated using the random set of color measurement values ​​obtained in step 2, and the Gaussian components with weights greater than the threshold in the PHD are determined. The mean set of components is obtained, which is the multi-target position estimation result, thereby achieving stable tracking of long-distance multi-targets under complex starry sky background.

[0036] The transfer density of a single target is denoted as f k|k-1 (·|·), given the state x at time k-1 k-1 , then at time k, state x k The transition probability density is: k|k-1 (x k |x k-1 ).

[0037] Transfer process in space Part of it is observed, given k state x k , the observed probability density is: g k (z k |x k ).

[0038] The state of the target before time k is recorded as ζ. For target i, its state transition probability density and received observation probability density are written in Gaussian form, which are:

[0039] f k|k-1 (x k |ζ)=N(x k ; F i,k-1 ζ,Q i,k-1 ) (8)

[0040] g i,k (z k |x k )=N(z k ;H i,k x k ,R i,k ) (9)

[0041] Among them, N(y;m,P) represents the Gaussian density of vector y with mean m and variance P, and F i,k-1 is the state transfer matrix; H i,k is the observation matrix; Q i,k-1 and R i,k are the covariance matrices of process and observation noise respectively, and i=1,2,…,n is the number of targets.

[0042] Due to the limitation of the field of view, the target has three behaviors: movement, disappearance, and rebirth, which are recorded as:

[0043] 1) New state γ k (·): the strength of the random finite set of newly generated targets at time k;

[0044] 2) Movement state S k|k-1 (·): the random finite set intensity of the movement from time k-1 to the target at time k;

[0045] 3) Disappearing state At time k, the target disappears.

[0046] Then the multi-objective state set at time k is:

[0047]

[0048] Among them, Γ k is the random finite set strength set of the new target, and γ k (·) is written as a Gaussian mixture form:

[0049]

[0050] The posterior intensity at time k–1 is a Gaussian mixture of the form:

[0051]

[0052] Then the posterior intensity of target i at time k is also a Gaussian mixture:

[0053]

[0054] The PHD is updated using the random set of color measurement values ​​through the formula (13) in the GMPHD filter. After the predicted PHD is updated using the random set of color measurement values, v is obtained. k (x), according to v k (x), find the Gaussian component whose weight is greater than the threshold. The mean set of the Gaussian components is the 2D position estimation value of the satellite target, so as to achieve stable tracking of long-distance multi-targets under a complex starry sky background.

[0055] Beneficial effects:

[0056] 1. The long-distance image multi-target recognition method for satellite constellations disclosed in the present invention aims at the problem that there are a large number of satellites in giant constellations and there are new and disappearing behaviors in the field of view. The multi-target states and measurements are described in the form of random finite sets, and the Gaussian mixture probability hypothesis density (GMPHD) filter is used to estimate the satellite target states, avoiding the problem of "data disaster" easily caused by a large number of targets and data association. It can achieve stable tracking of multiple satellite targets under a complex starry sky background, and there is no need to associate the trajectory of each target during tracking, thereby reducing the amount of tracking and recognition calculations.

[0057] 2. The long-distance image multi-target recognition method for satellite constellations disclosed in the present invention adopts a defocusing method to perform Gaussian diffusion processing on point targets to address the problem that targets appear as faint points in images without features such as contours, textures and shapes. That is, by extracting four-connected regions and performing Gaussian diffusion processing on the regions, preliminary multi-target recognition results are obtained, which facilitates the acquisition of measurement information. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a schematic flow chart of a method for multi-target recognition of long-distance images for satellite constellations of the present invention;

[0059] Figure 2 is the original image taken by the space-based platform used in the simulation in the example of the present invention;

[0060] Figure 3 is a pre-processed image obtained by using the frame difference method in step 1 of the embodiment of the present invention;

[0061] Figure 4 is the multi-target preliminary recognition result image obtained by Gaussian diffusion processing in step 1 of the example of the present invention;

[0062] Figure 5 is the target grayscale histogram obtained in step 2 of the present invention;

[0063] Figure 6 is the position tracking result of the target in the example of the present invention;

[0064] Figure 7 It is the final tracking result of multiple targets in the example of the present invention, wherein: 7a) is the target horizontal coordinate estimation result, and 7b) is the target vertical coordinate estimation result. DETAILED DESCRIPTION

[0065] In order to better illustrate the purpose and advantages of the present invention, the invention is further described below with reference to the accompanying drawings and examples.

[0066] In order to verify the feasibility of the present invention, the Starlink operation images taken by the space-based platform are used, such as Figure 2As shown, the maximum value of the mixed component of the GMPHD filter is set to 30, and the survival probability and detection probability of the target are set to 0.99 and 0.98 respectively, and mathematical simulation verification is performed.

[0067] like Figure 1 As shown, the long-distance image multi-target recognition method for satellite constellations disclosed in this embodiment has the following specific implementation steps:

[0068] Step 1: Use the frame difference method to pre-process the image taken by the navigation camera, perform corrosion operation on the processed image, and obtain the four-connected area of ​​the image, which includes the satellite four-connected area and the star four-connected area, and define the four-connected area as the quasi-target; perform Gaussian diffusion processing on the obtained quasi-target to diffuse the star point target from a small number of pixels to multiple pixels as the preliminary recognition result of multiple targets. Gaussian diffusion processing makes it easy to obtain the measurement results of the quasi-target after diffusion.

[0069] The specific implementation method of step 1 is:

[0070] Step 1.1 pre-processes the satellite images taken by the space-based platform to obtain detection results including quasi-satellites and star targets.

[0071] Distant space targets often appear as faint dots in images, occupying a small number of pixels, and have no features such as outline, texture, and shape, making them difficult to distinguish from stars. At the same time, due to the movement of the observation platform, the starry sky background has relative motion between different observation frames of the tracking target. Step 1.1 aims to detect the satellite to be tracked with high precision, while eliminating the influence of some stars on tracking.

[0072] Images of Starlink in operation taken by space-based platforms, such as Figure 2 , use the frame difference method to preprocess the image, and eliminate some star interference by making differences between adjacent frames, such as Figure 3 As shown, the image obtained is recorded as A; the image A is corroded by the structural element B to obtain the image C containing satellites and stars. The result of corroding A with B is the set of all a that make B still in A after translation by a, satisfying:

[0073]

[0074] in(·) M The pixel matrix representing the image. In step 1, satellites and stars are collectively referred to as quasi-targets. The four-connected region of C is obtained to obtain the recognition result of the quasi-target, which is recorded as T i .

[0075] Step 1.2 performs Gaussian diffusion processing on the target, so that the star point target diffuses from a small number of pixels to multiple pixels as the preliminary multi-target recognition result.

[0076] Since the quasi-target extracted in step 1.1 occupies only a small number of pixels, it is not convenient to obtain measurement information. The defocus method is adopted to convert the star target T i A single pixel diffuses to multiple pixels, and the Gaussian distribution of the point spread function satisfies:

[0077]

[0078] Where E is the total energy of the target, ρ represents the energy concentration of the point spread function, which is 0.45. Formula (2) can ensure that the total energy of the pixel after diffusion is close to T i The total energy of the diffusion result is used as the preliminary recognition result of multiple targets to facilitate the acquisition of measurement information. The results are as follows Figure 4 shown.

[0079] Step 2: The preliminary recognition results of multiple targets obtained in step 1 are described by rectangular boxes respectively, and the size and position of each rectangular box are used as the state of multiple targets; the grayscale histogram of each rectangular box is obtained, and the object is a satellite constellation. Considering the consistency of the satellite structure in the constellation, the same template grayscale model is used for different satellites, and the Bartcharia distance is used to characterize the similarity between the target and template grayscale histograms, and a Gaussian similarity function between the target and the template is constructed; the predicted probability hypothesis density PHD of the image is obtained and the peak value is collected from its product with the likelihood function as a random set of color measurement values ​​of the target. The specific implementation method of step 2 is:

[0080] Step 2.1 describes the multi-target recognition results of step 1 with a rectangular box to obtain a grayscale likelihood function in Gaussian form.

[0081] A single target state is described by a rectangular box as shown in formula (3):

[0082] x={x i ,y i ,L i ,H i},i=1,2,...,n (3)

[0083] The center and size of the rectangular box are represented by {x i ,y i} and {L i ,H i} Determined, the example of the present invention identified 22 objects, including satellites and stars, and the grayscale histogram was obtained for each rectangular box area. The results are as follows Figure 5As shown in the figure, the region contains 10 pixels in total, 80% of which have grayscale values ​​below 50 and 20% of which have grayscale values ​​above 150. The features are obvious, which facilitates the acquisition of measurement information. The grayscale histogram of the target is represented as p(u), and the grayscale histogram of the template is represented as q(u). The similarity function between the target and the template is measured using the Bartholin distance:

[0084]

[0085] The object of the present invention is a satellite constellation. Considering the consistency of the satellite structure in the constellation, the same template grayscale model is used for different satellites, and the grayscale likelihood function is defined as:

[0086]

[0087] Among them, P is the current image, x is the state of the target, σ 2 is the noise variance, which is set to 0.1.

[0088] Let v k (x) is the PHD at time k, v k|k-1 (x) is the predicted PHD at time k, then:

[0089]

[0090] Where L P (x) is the grayscale likelihood function, so Also v k The peak value of (x).

[0091] Step 2.2 calculates the predicted PHD of the image and collects a random set of color measurements whose peaks are taken as targets from its product with the likelihood function.

[0092] The Gaussian mixture prediction probability hypothesis density v calculated for the image at time k k|k-1 (x) is sampled, and the particles are recorded as: Among them, l i represents the integer number of the particle, w i represents the weight of the particle, then the posterior probability density v k (x) can be used with {x i ,l i ,φ i} is approximately replaced by, where φ i =L(x i w i ). Cluster the particles and obtain the set of measurement results at time k:

[0093] Z f,k = {Z f,1 ,Z f,2 ,...,Zf,n} (7)

[0094] The result of formula (7) is a random set of color measurements.

[0095] Step 3: Describe the multi-target states and measurement values ​​in the form of random finite sets, use the Gaussian mixture probability hypothesis density GMPHD filter to estimate the satellite target states, and track multiple satellite targets in a complex background based on the estimation results. There is no need to associate the trajectory of each target during tracking, thereby reducing the amount of calculation. Estimation includes two processes: prediction and update. During the prediction process, the Gaussian mixture prediction probability hypothesis density is obtained based on the posterior intensity at time k-1. During the update process, the PHD in the Gaussian mixture probability hypothesis density filter is updated using the random set of color measurement values ​​obtained in step 2, and the Gaussian components in the PHD with weights greater than the threshold are determined. The mean set of components is obtained, which is the multi-target position estimation result.

[0096] The transfer density of a single target is denoted as f k|k-1 (·|·), given the state x at time k-1 k-1 , then at time k, state x k The transition probability density is: k|k-1 (x k |x k-1 ).

[0097] Transfer process in space Part of it is observed, given k state x k , the observed probability density is: g k (z k |x k ).

[0098] The state of the target at time k is recorded as x, and its previous state is recorded as ζ. For target i, its state transition probability density and the received observation probability density are written in Gaussian form, which are:

[0099] f k|k-1 (x k |ζ)=N(x k ; F i,k-1 ζ,Q i,k-1 ) (8)

[0100] g i,k (z k |x k )=N(z k ;H i,k x k ,R i,k ) (9)

[0101] Among them, N(y;m,P) represents the Gaussian density of vector y with mean m and variance P, and F i,k-1 is the state transfer matrix; H i,k is the observation matrix; Q i,k-1 and R i,k are the covariance matrices of process and observation noise respectively, i=1,2,…,n is the target number, and n is 6 in the present invention.

[0102] Due to the limitation of the field of view, the target has three behaviors: movement, disappearance, and rebirth, which are recorded as:

[0103] 1) New state γ k (·): the strength of the random finite set of newly generated targets at time k,

[0104] 2) Movement state S k|k-1 (·): The random finite set intensity of the movement from time k-1 to the target at time k,

[0105] 3) Disappearing state At time k, the target disappears.

[0106] Then the multi-objective state set at time k is:

[0107]

[0108] Among them, Γ k is the random finite set strength set of the new target, and γ k (·) is written as a Gaussian mixture form:

[0109]

[0110] Assume that the posterior intensity at time k–1 is a Gaussian mixture of the form:

[0111]

[0112] Then the posterior intensity of target i at time k is also a Gaussian mixture:

[0113]

[0114] The PHD is updated using the random set of color measurements through the Gaussian mixture probability hypothesis density (GMPHD) filter formula (13). After the predicted PHD is updated using the random set of color measurements, v is obtained. k (x), according to v k (x), find the Gaussian component whose weight is greater than the threshold. The mean set of these Gaussian components is the 2D position estimation value of the satellite target. The multi-target detection result of this embodiment is as follows: Figure 6 The corresponding horizontal coordinate estimation results are shown as Figure 7(a) shows the ordinate estimation result. Figure 7 As shown in (b), during the recognition process, the number of targets gradually changes from 6 to 3. The present invention can accurately recognize the status of all targets at any time.

[0115] At this point, the recognition of multiple targets in long-distance images required for space situational awareness has been completed.

[0116] The specific description above further illustrates the purpose, technical solutions and beneficial effects of the invention in detail. It should be understood that the above is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A long-distance image multi-target recognition method for satellite constellations, characterized by: The following steps are included: Step 1: Use the frame difference method to preprocess the image taken by the navigation camera, perform corrosion operation on the processed image, and obtain the four-connected area of ​​the image, which includes the satellite four-connected area and the star four-connected area, and define the four-connected area as the quasi-target; perform Gaussian diffusion processing on the obtained quasi-target to diffuse the star point target from a small number of pixels to multiple pixels as the preliminary recognition result of multiple targets; Gaussian diffusion processing is used to facilitate the acquisition of the measurement results of the quasi-target after diffusion; Step 2: The preliminary recognition results of the multiple targets obtained in step 1 are described by rectangular boxes respectively, and the size and position of each rectangular box are used as the state of the multiple targets; the grayscale histogram of each rectangular box is obtained, and the object is a satellite constellation. Considering the consistency of the satellite structure in the constellation, the same template grayscale model is used for different satellites, and the Bartcharia distance is used to characterize the similarity between the target and template grayscale histograms, and a Gaussian similarity function between the target and the template is constructed; the predicted probability hypothesis density PHD of the image is obtained and the peak value is collected from its product with the likelihood function as a random set of color measurement values ​​of the target; Step 3: Describe the multi-target states and measurement values ​​in the form of random finite sets, use Gaussian mixture probability hypothesis density GMPHD filter to estimate the satellite target states, and track multiple satellite targets under complex backgrounds based on the estimation results. When tracking, there is no need to associate the trajectory of each target, thereby reducing the amount of calculation; the estimation includes two processes: prediction and update; During the prediction process, the Gaussian mixture prediction probability hypothesis density is obtained based on the posterior strength at time k-1; During the updating process, the random set of color measurement values ​​obtained in step 2 is used to update the PHD in the Gaussian mixture probability hypothesis density filter, and the Gaussian components in the PHD whose weights are greater than the threshold are determined. The mean set of components is obtained, which is the multi-target position estimation result, thereby achieving stable tracking of long-distance multi-targets under a complex starry sky background.

2. The satellite constellation-oriented long-distance image multi-target recognition method according to claim 1, characterized in that: The specific implementation method of step 1 is: Step 1.1 pre-processes the satellite images taken by the space-based platform to obtain detection results including quasi-satellites and star targets; Distant space targets often appear as weak dots in images, occupying a small number of pixels and without contour, texture and shape features, making it difficult to distinguish between distant space targets and stars. At the same time, due to the movement of the observation platform, the starry sky background of the tracking target has relative motion between different observation frames. Step 1.1 aims to detect the satellite to be tracked with high precision, while eliminating the influence of some stars on tracking. The satellite images taken by the space-based platform are preprocessed using the frame difference method. The image A is obtained by subtracting adjacent frames to eliminate some star interference. The image A is eroded using the structural element B to obtain the image C containing satellites and stars. The result of eroding A with B is the set of all a's that remain in A after B is translated by a, satisfying: in(·) M The pixel matrix representing the image. In step 1, satellites and stars are collectively referred to as quasi-targets. The four-connected region of C is obtained to obtain the recognition result of the quasi-target, which is recorded as T i ; Step 1.2 performs Gaussian diffusion processing on the target, so that the star point target diffuses from a small number of pixels to a large number of pixels, which is used as the preliminary recognition result of multiple targets; Since the quasi-target extracted in step 1.1 occupies only a small number of pixels, it is not convenient to obtain measurement information. The defocus method is adopted to convert the star target T i A single pixel diffuses to multiple pixels, and the Gaussian distribution of the point spread function satisfies: Where E is the total energy of the target, ρ represents the energy concentration of the point spread function, and formula (2) is used to ensure that the total energy of the pixel after diffusion is close to T i The total energy of the target is calculated and the diffusion result is used as the preliminary recognition result of multiple targets to facilitate the acquisition of measurement information.

3. The satellite constellation-oriented long-distance image multi-target recognition method according to claim 2, characterized in that: The specific implementation method of step 2 is: Step 2.1 describes the multi-target recognition result of step 1 with a rectangular frame to obtain a grayscale likelihood function in Gaussian form; A single target state is described by a rectangular box as shown in formula (3): x={x i ,y i ,L i ,H i },i=1,2,...,n (3) Where n is the number of multi-targets initially identified in step 1.2, and the center and size of the rectangular box are respectively represented by {x i ,y i } and {L i ,H i Determine; the grayscale histogram of the target is represented as p(u), the grayscale histogram of the template is represented as q(u), and the similarity function between the target and the template is characterized by the Bartcharya distance: The object-oriented model is a satellite constellation. Considering the consistency of the satellite structure in the constellation, the same template grayscale model is used for different satellites. The grayscale likelihood function is defined as: Among them, P is the current image, σ 2 is the noise variance; Let v k (x) is the PHD at time k, v k|k-1 (x) is the predicted PHD at time k, then: Where L P (x) is the grayscale likelihood function, so Also v k The peak value of (x); Step 2.2 calculates the predicted PHD of the image and collects the peak value from its product with the likelihood function as a random set of color measurement values ​​of the target; The Gaussian mixture prediction probability hypothesis density v calculated for the image at time k k|k-1 (x) is sampled, and the particles are recorded as: Among them, l i represents the integer number of the particle, w i represents the weight of the particle, then the posterior probability density v k (x) Use {x i ,l i ,φ i } is approximately replaced by, where φ i =L(x i w i ); Cluster the particles and obtain the set of measurement results at time k: WITH f,k ={Z f,1 ,WITH f,2 ,...,WITH f,n } (7) The result of formula (7) is a random set of color measurements.

4. The satellite constellation-oriented long-distance image multi-target recognition method as claimed in claim 3, characterized in that: Step 3 is implemented as follows: The transfer density of a single target is denoted as f k|k-1 (·|·), given the state x at time k-1 k-1 , then at time k, state x k The transition probability density is: k|k-1 (x k |x k-1 ); Transfer process in space Part of it is observed, given k state x k , the observed probability density is: g k (z k |x k ); The state of the target before time k is recorded as ζ. For target i, its state transition probability density and received observation probability density are written in Gaussian form, which are: f k|k-1 (x k |ζ)=N(x k ;F i,k-1 g,Q i,k-1 ) (8) g i,k (z k |x k )=N(z k ;H i,k x k ,R i,k ) (9) Among them, N(y;m,P) represents the Gaussian density of vector y with mean m and variance P, and F i,k-1 is the state transfer matrix; H i,k is the observation matrix; Q i,k-1 and R i,k are the covariance matrices of process and observation noise respectively, i=1,2,…,n is the number of targets; Due to the limitation of the field of view, the target has three behaviors: movement, disappearance, and rebirth, which are recorded as: 1) New state γ k (·): the strength of the random finite set of newly generated targets at time k; 2) Movement state S k|k-1 (·): the random finite set intensity of the movement from time k-1 to the target at time k; 3) Disappearing state The target disappears at time k; Then the multi-objective state set at time k is: Among them, Γ k is the random finite set strength set of the new target, and γ k (·) is written as a Gaussian mixture form: The posterior intensity at time k–1 is a Gaussian mixture of the form: Then the posterior intensity of target i at time k is also a Gaussian mixture: The PHD is updated using the random set of color measurement values ​​through the formula (13) in the GMPHD filter. After the predicted PHD is updated using the random set of color measurement values, v is obtained. k (x), according to v k (x), find the Gaussian component whose weight is greater than the threshold. The mean set of the Gaussian components is the 2D position estimation value of the satellite target, so as to achieve stable tracking of long-distance multi-targets under a complex starry sky background.

Images