Low-resolution star image target recognition method based on cyclic matching

Abstract

The present invention relates to a low-resolution star map target recognition method based on cyclic matching, and the method comprises the following steps: generating a navigation feature star library from a star catalog and an angular distance calculation formula; obtaining a sequence of star maps taken by a star sensor camera; calculating the three-dimensional vector of star points in the image according to camera parameters; sorting the brightness of the star points and cyclically matching them in a sliding window manner; detecting the number of successful matches of each star point and cyclically confirming them; extracting reliable stars and suspicious targets for secondary matching to identify targets; and using a sequence image target enhancement algorithm to enhance and track targets. The present invention realizes accurate recognition of stars at low camera resolution, and the cyclic matching result confirmation mechanism makes the algorithm miss alarm rate extremely low, the suspicious target secondary confirmation mechanism makes the algorithm false alarm rate extremely low, and the sequence image target enhancement algorithm makes it highly robust to noise. It can identify binary stars, and the algorithm has low time complexity and can identify multiple targets at the same time.

Description

Technical Field

[0001] The present invention relates to the field of aerospace and image processing technology, and more specifically, to a low-resolution star image target recognition method based on cyclic matching. Background Art

[0002] Star sensors often use star map matching to determine the attitude information of satellites. By matching the star points in the acquired image with the star library, the image plane coordinates are converted into celestial coordinates to determine the satellite attitude. Star map matching can also assist in identifying targets by identifying background stars, but this research is still relatively rare. The commonly used and more mature method is the triangle matching method, which uses the triangle formed by the angular distance of the star points as the matching feature. Most other algorithms are derivative algorithms of this method.

[0003] Star map matching technology has its unique advantages over traditional starry sky weak target recognition algorithms. When the target image is a point target and the motion characteristics and optical characteristics are very similar to those of the background stars, it is difficult for the recognition method based on target features to extract the target. Star map matching uses the navigation star library to identify stars to indirectly identify the target. However, the disadvantage of this method is that it requires higher positioning accuracy, especially when there are a large number of background stars when identifying weak targets, which places higher requirements on the imaging system, navigation star library and recognition algorithm.

[0004] When using traditional star map matching algorithms to identify targets, the camera needs to achieve a pixel resolution of at least 0.002°, and the algorithm has a high false alarm rate and is difficult to identify binary stars. In addition, traditional algorithms need to detect all triangles formed in the image. When the camera field of view is large and the detection limit star magnitude is high, the algorithm complexity increases exponentially. Summary of the invention

[0005] Based on this, a low-resolution star map target recognition method based on cyclic matching is provided to address the characteristics of the existing star map matching target recognition algorithm, which has high requirements on camera resolution, high false alarm rate, difficulty in identifying double stars and high computational complexity.

[0006] To solve the above problems, the present invention adopts the following technical solutions:

[0007] A low-resolution star image target recognition method based on cyclic matching, the method comprising the following steps:

[0008] Step 1: Build a navigation feature star library

[0009] 1) Download the star catalog and extract the magnitude, right ascension and declination of some stars;

[0010] 2) Calculate the angular distances between all pairs of stars to form a characteristic matrix. The angular distances are calculated in the following way:

[0011] V=(cosa·cosb, sina·cosb, sinb)

[0012] d xy =arccos(V x ·V y / |V x |·|V y |)

[0013] Where a and b are the right ascension and declination of the star respectively, and V is the three-dimensional vector of the star. x With V y is the three-dimensional vector of two different stars, d xy is the angular distance between the two stars.

[0014] 3) The angular distances in the feature matrix that are greater than the camera field of view are set to zero, and the processed sparse matrix is ​​compressed to complete the construction of the navigation star library.

[0015] Step 2: Obtaining a sequence of star images taken by a star sensor camera;

[0016] Step 3: Calculate the three-dimensional vector of the star point in the image according to the camera parameters;

[0017] Given the coordinates (x, y) of the star in the image obtained by the camera, where x and y are the coordinate values ​​obtained with the center of the image as the origin and f is the focal length of the camera, the corresponding three-dimensional vector v of the star is:

[0018]

[0019] Step 4: Sort the brightness of the stars and match them in sequence

[0020] 1) Sort the grayscale of the stars in the image from large to small.

[0021] 2) First, select the three brightest stars to form a characteristic triangle, use the triangle matching method to match the characteristic star library with low precision, and record the results. Calculate the three angular distances of the characteristic triangle from the three-dimensional vectors of the image star points, match the obtained angular distances with the angular distances in the characteristic star library, and if the three groups of stars matched can complete the first connection, the match is successful, and record the star number matched by the star point.

[0022] 3) Then select the next brightest stars except the brightest star to form a triangle match. Finally, the cycle matching is completed after each star point is matched three times.

[0023] Step 5: Check the number of successful matches for each star point and confirm in a loop

[0024] 1) Loop confirmation starts from the brightest star point, traverses all star points and ends at the brightest star point. If the match is successful, the number of stars involved in the match is set to be greater than or equal to one.

[0025] 2) No loop confirmation is performed at the beginning. Starting from the second round of the loop, the matching results of the first two stars are compared with the matching results of the last two stars in the previous match. If there are any two identical groups of matching results, the confirmation results are the same, and the number of successful matches of the first two stars is increased by one.

[0026] Step 6: Extract reliable stars and suspicious targets for secondary matching and target identification

[0027] 1) Star points with less than 2 successful matches are considered suspicious targets, and star points with 3 successful matches are considered reliable stars.

[0028] 2) Each suspicious target is matched with three reliable stars to form two new characteristic triangles. If both matches are successful and the cyclic confirmation is the same, the point is considered to be a star.

[0029] Step 7: Sequence image target enhancement algorithm enhances and tracks the target

[0030] 1) According to the star numbers in the matched star catalog, the stars in different images are matched, and the stars with the same numbers identified in the images are used for image registration.

[0031] 2) After registration, multiple images are subjected to AND operation to suppress the target, and then the stars are removed from the original image to obtain the target enhanced image.

[0032] 3) After removing false targets and noise in a single frame image, the least squares prediction and Kalman filter algorithm are used to track the target.

[0033] The present invention can realize accurate identification of stars under low camera resolution. The circular matching result confirmation mechanism makes the algorithm miss alarm rate extremely low. The suspicious target secondary confirmation mechanism makes the algorithm false alarm rate extremely low. The sequence image target enhancement algorithm makes it highly robust to noise. It can identify binary stars, the algorithm has low time complexity, and can identify multiple targets at the same time. Figure 6 The target recognition results are shown in the figure, where the target is marked with a box and the stars are marked with a circle. In the figure, the false target and all the stars are successfully identified. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flow chart of the low-resolution star image target recognition method based on cyclic matching of the present invention;

[0035] Figure 2 This is a schematic diagram of sliding window loop matching;

[0036] Figure 3 It is a schematic diagram of double cycle confirmation;

[0037] Figure 4 This is a schematic diagram of the target enhancement algorithm for sequence images;

[0038] Figure 5 This is one of the original binary star maps obtained;

[0039] Figure 6 This is a diagram of the star and target identification results obtained based on the present invention. DETAILED DESCRIPTION

[0040] The method of the disclosed embodiment is applied to a high-orbit satellite to identify targets in the field of view based on a sequence of star maps taken by a satellite optical camera. In one embodiment, the pixel angular resolution of the camera is 0.02°, and the main frequency of the embedded device is 100MHz. The present invention discloses a low-resolution star map target recognition method based on cyclic matching, which can accurately identify stars and targets within 1s. The specific flow chart is as follows Figure 1 As shown, the method comprises the following steps:

[0041] 1) Build a navigation feature star library

[0042] Download the SAO star catalog, and extract information about stars with magnitudes less than (x+0.5) from the catalog based on the weakest magnitude x that the satellite imaging camera can detect, including right ascension, declination, serial number, and magnitude. Extracting stars with magnitudes slightly higher than the detection limit can ensure that all stars imaged by the camera can be found in the navigation star library.

[0043] Among the extracted stars, the three-dimensional vector V of all stars is calculated using the following formula:

[0044] V=(cosa*cosb,sina*cosb,sinb)

[0045] Where a is the star's right ascension and b is the star's declination.

[0046] Then calculate the angular distance d between any two stars xy The feature matrix is ​​constructed as follows:

[0047] d xy =arccos(V x ·V y / |V x |·|V y |)

[0048] Where V x With V y for two different star vectors.

[0049] Due to the limitation of the camera field of view, the angular distance between two stars in the field of view cannot be greater than the field of view angle. The angular distance in the feature matrix that is greater than the field of view angle will not be matched. Therefore, the angular distance in the feature matrix that is greater than 19.5° of the camera field of view angle is set to zero, and the matrix is ​​converted into an upper triangular matrix.

[0050] Calculation of memory occupied before feature matrix compression: angular distance data is stored in double format, each data size is 8Byte, the number of sixth-magnitude stars in the star catalog is about 5000, and the matrix size is 5000*5000, so the memory used is 5000*5000*0.5*8Byte, about 100MB.

[0051] The processed sparse matrix is ​​compressed and stored. Specifically, the non-zero elements in the matrix can be extracted in the form of vectors, and the vector content is (star number 1, star number 2, inter-star angular distance). The extracted vectors are recombined to obtain the final feature matrix.

[0052] Calculation of memory occupied by feature matrix after compression: When the camera field of view is 19.5°×19.5°, the number of non-zero elements in the sparse matrix is ​​approximately 832000, and the memory used is 832000*3*8Byte, which is approximately 20MB.

[0053] In the process of constructing the feature star library, it is necessary to maintain high numerical accuracy as much as possible to ensure that the accuracy of the final result can meet the requirements of subsequent processing.

[0054] 2) Obtain the sequence of star images taken by the star sensor camera

[0055] The camera parameters and experimental parameters used in this example are as follows:

[0056] Focal length 16mm, pixel size 5.5um, focal plane size 5.5mm*5.5mm, image resolution 1000*1000, field of view size 19.5°×19.5°, pixel resolution 0.02°

[0057] The field of view points to the positive direction of the x-axis in the wgs84 coordinate system, the row direction of the image is parallel to the z-axis direction of the wgs84 coordinate system, and the column direction of the image is parallel to the y-axis direction of the wgs84 coordinate system

[0058] The camera detects a limit magnitude of 6. There are about 90 stars in the image and a false target point is added manually.

[0059] 3) Calculate the three-dimensional vector of the star point in the image based on the camera parameters

[0060] Use the weighted centroid method to extract the row X and column Y of the sub-pixel centroid of the star point as follows:

[0061]

[0062] Where f(x,y) is the grayscale value of the star pixel, x and y are the row and column of the star pixel, and the star coordinates (x,y) with the center of the image as the origin are obtained as follows:

[0063] x=Y–500、y=500–X

[0064] It is known that the coordinates of a star in the image obtained by the camera are (x, y), and f is the focal length of the camera. The star vector corresponding to the star is:

[0065]

[0066] 4) Sort the brightness of the stars and match them in sequence

[0067] Sort the grayscale of the star points in the image from large to small, and establish the corresponding relationship between the sorted star points and the original star points

[0068] First, select the three brightest stars to form a characteristic triangle, use the triangle matching method to match it with the characteristic star library, and record the results according to the corresponding relationship between the star points. Calculate the three angular distances of the characteristic triangle from the three-dimensional vectors of the image star points. Due to the limited resolution of the camera, the high-precision angular distances calculated are meaningless, and the obtained angular distances are accurate to 0.1°. After the angular distances in the characteristic star library are also accurate to 0.1°, the obtained angular distances are matched with the angular distances in the characteristic star library. If the three groups of stars matched can be connected end to end, the match is successful. Due to the low matching accuracy, the angular distance between the two stars can be matched to thousands of characteristic angular distances in the star library. There are an average of about 25 groups of angular distances that are successfully connected end to end, which can match an average of about 25 groups of triangles in the star library. At the same time, the serial numbers of the multiple groups of stars matched are recorded. The following table shows the serial numbers of the stars corresponding to some successful angular distances that are matched end to end for a certain characteristic triangle.

[0069]

[0070] Then, the second brightest stars except the brightest star are selected to form a triangle match, and then the sliding window method is used to shift and select stars until each star point is matched three times and the cycle matching is completed. If the number of star points is n, the total number of matches required is n, and the number of matches required by the traditional algorithm is It can be seen that the number of matches of this algorithm is much smaller than that of the traditional matching recognition algorithm.

[0071] Matching analysis: The number of angular distances in the feature star catalog is about 5000*5000*0.5=12500000. If the angular distance accuracy is 0.1°, the number of possible angular distances is 180*10=1800. The number of all possible combinations of three angular distances of matching triangles is 1800^3=5.832*10^9, and the number of matching triangle combinations in the feature star catalog is 5000^3=1.25*10^11. The number of all possible combinations of three angular distances is much smaller than the number of matching triangle combinations in the feature star catalog, so a triangle in the star map can be matched to multiple triangles in the star library.

[0072] 5) Check the number of successful matches for each star point and confirm it in a loop

[0073] The loop confirmation starts from the brightest star point, traverses all star points and ends at the brightest star point. If the match is successful, the number of stars participating in the match is set to >= 1. No loop confirmation is performed at the beginning. Starting from the second round of the loop, the matching results of the first two stars are compared with the matching results of the last two stars in the last match. If there are any two identical groups of matching results, the confirmation results are the same, and the number of successful matches of the first two stars is increased by 1. This algorithm can ensure that each star point participates in two loop confirmations, and the brightest and second brightest points can complete the second confirmation in the last loop, and the maximum number of successful matches is three. Since the number of loops is greater than the number of stars, a remainder operation is required to ensure that the matches are connected head to tail. The specific method is to subtract one from the number of loops and divide it by the number of star points and take the remainder, and then add one to the remainder.

[0074] Cyclic confirmation can greatly reduce the missed alarm rate. If 30 sets of triangles in the star library are matched twice, the target missed alarm probability is calculated as follows: the number of possible combinations of star pairs is 5000*5000, and the cyclic confirmation missed alarm rate is 30*30 / 5000*5000=0.0036%. The traditional matching recognition algorithm considers a star as a single successful match, so the false alarm rate is high.

[0075] 6) Extract reliable stars and suspicious targets for secondary matching and target identification

[0076] Star points with less than 2 successful matches are considered suspicious targets, and star points with 3 successful matches are considered reliable stars. Each suspicious target is reorganized into two characteristic triangles with three reliable stars. If both matches are successful and the cycle confirmation is the same, the point is considered to be a star, otherwise the point is the final identification target.

[0077] Since there is a target in the image, the star points adjacent to the target brightness can only be successfully matched once with a high probability, so a second match is required to exclude the stars that are considered suspicious targets.

[0078] 7) Sequence image target enhancement algorithm enhances and tracks the target

[0079] Using the identified stars in the image to perform sequence image registration, the stars in different images are matched according to the star serial number. This method has extremely high robustness in finding feature points. After registration, multiple images are ANDed to suppress the target, and then the stars are removed from the original image to obtain the target enhanced image. This method can remove false targets in a single frame image, so it is highly robust to noise and reduces the false alarm rate of the algorithm again. Using Kalman filtering and least squares prediction algorithms, the target trajectory is associated by predicting the target trajectory and searching with a window at the predicted position.

[0080] The present invention realizes accurate identification of stars under low camera resolution, and overcomes the defects of the traditional star map matching target recognition algorithm that requires high-resolution imaging, cannot identify binary stars, and has a high false alarm rate. The cyclic matching result confirmation mechanism makes the algorithm's false alarm rate extremely low, the suspicious target secondary confirmation mechanism makes the algorithm's false alarm rate extremely low, and the sequence image target enhancement algorithm makes it highly robust to noise. It can identify binary stars, the algorithm has low time complexity, and can identify multiple targets at the same time.

Claims

1. A low-resolution star image target recognition method based on cyclic matching, characterized in that: The following steps are involved: 1). Build a navigation feature star library to store angular distance information. The steps are as follows: (1) Download the star catalog and extract information such as magnitude, right ascension and declination of some stars; (2) Calculate the angular distances between all pairs of stars to form a characteristic matrix, where the angular distances are calculated in the following way: V=(cosa·cosb, sina·cosb, sinb) d xy arccos(V x ·V y / |V x |·|V y |) Where a and b are the right ascension and declination of the star respectively, V is the three-dimensional vector of the star, V x With V y is the three-dimensional vector of two different stars, d xy is the angular distance between the two stars; (3) The angular distances in the feature matrix that are larger than the camera's field of view are set to zero, and the processed sparse matrix is ​​compressed to complete the construction of the navigation star library; 2). Obtain the sequence of star images taken by the star sensor camera; 3). Calculate the three-dimensional vector of the star point in the image according to the camera parameters as follows: Given the coordinates (x, y) of the star in the image obtained by the camera, where x and y are the coordinate values ​​obtained with the center of the image as the origin and f is the focal length of the camera, the corresponding three-dimensional vector v of the star is: 4) Sort the brightness of the stars and then perform cyclic matching in a sliding window manner. The steps are as follows: (1) Sort the grayscale of the star points in the image from large to small; (2) First, select the three brightest stars to form a characteristic triangle, use the triangle matching method to match it with the characteristic star library, and record the results; The triangle matching method is as follows: the three angular distances of the characteristic triangle are calculated from the three-dimensional vectors of the image star points, and the obtained angular distances are matched with the angular distances in the characteristic star library. If the three groups of stars matched can complete the first connection, the match is successful, and the star number matched by the star point is recorded; (3) Then, the second-brightest stars except the brightest star are selected in turn to form a triangle match, and finally the cycle matching is completed after each star point is matched three times; 5). Check the number of successful matches for each star point and confirm it in a loop. The steps are as follows: (1) Loop confirmation starts from the brightest star point, traverses all star points and then ends at the brightest star point. If the match is successful, the number of stars involved in the match is set to >= 1; (2) No circular confirmation is performed at the beginning. Starting from the second round of the cycle, the matching results of the first two stars are compared with the matching results of the last two stars in the previous matching. If there are any two identical groups of matching results, the confirmation results are the same, and the number of successful matches of the first two stars is increased by 1; 6) Extract reliable stars and suspicious targets for secondary matching and target identification. The steps are as follows: (1) Star points with less than 2 successful matches are considered suspicious targets, and star points with 3 successful matches are considered reliable stars; (2) Each suspicious target is matched with three reliable stars to form two new characteristic triangles. If both matches are successful and the cyclic confirmation is the same, the point is considered to be a star; 7). The sequence image target enhancement algorithm tracks the target. The steps are as follows: (1) Match the stars in different images according to the star numbers in the matched star catalog, and use the stars with the same numbers identified in the images for image registration; (2) After registration, multiple images are subjected to AND operation to suppress the target, and then the stars are removed from the original image to obtain the target enhanced image.

Images