Method for rapidly detecting positions and azimuth angles of multiple moving targets based on single-pixel imaging

Abstract

The invention belongs to the field of target tracking, and particularly relates to a multi-moving target position and azimuth angle rapid detection method based on single-pixel imaging, which comprises the following steps: S1, for each frame of scene image comprising a plurality of moving targets, obtaining one-dimensional projection curves of the scene image in four directions of 0 degree, 45 degrees, 90 degrees and 135 degrees; s2, dividing the scene image into N regions by using a one-dimensional projection reverse reconstruction method, wherein each region has only one moving target; s3, modulating the scene image by using a deformed geometric moment mode to obtain a centroid position coordinate of each moving target; and S4, a second-order central moment mode is obtained according to the centroid coordinates of each moving target, and the azimuth angle of each moving target is obtained after the scene image is modulated. According to the invention, rapid and high-precision detection of position coordinates and azimuth angles of a plurality of moving targets can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of moving target tracking, and in particular to a method for quickly detecting the positions and azimuths of multiple moving targets based on single-pixel imaging. Background Art

[0002] Single-pixel imaging technology is an imaging technology that modulates the target scene and uses a single-pixel detector without spatial resolution to collect the light intensity value reflected or transmitted by the scene. The image of the target scene can be restored by using the correlation between the modulated speckle and the light intensity value. Compared with the traditional two-dimensional array imaging method, single-pixel imaging technology has the characteristics of wide spectrum and high sensitivity, and can be applied to bands that array cameras cannot respond to and imaging in weak light.

[0003] At present, there are two main types of moving target tracking methods based on single-pixel imaging technology: image-based and image-free tracking methods. Image-based moving target tracking methods require multiple detections to obtain scene images, and then use image processing algorithms to obtain the position of the moving target. Image-based moving target tracking methods have the following disadvantages: first, the imaging time is long, which is not suitable for fast positioning and tracking of moving targets; second, the positioning accuracy depends on the quality of the reconstructed image. When the reconstructed image quality is poor, it may lead to large positioning errors; in addition, in practical applications, due to the large amount of image data storage, it is a challenge to track moving targets for a long time. Therefore, in recent years, the research on image-free moving target tracking methods based on single-pixel imaging has attracted more and more attention.

[0004] The imageless moving target tracking method based on single-pixel imaging does not require the acquisition of scene images and is very suitable for the recognition of moving targets. However, most of the current research is related to the research on the motion parameter acquisition method of a single moving target translation, rotation, and translation plus rotation based on single-pixel imaging. There are relatively few related researches on the positioning method of multiple translational moving targets. Therefore, there is a need for a method that can realize the position and azimuth detection of multiple moving targets based on single-pixel imaging technology. Summary of the invention

[0005] In view of the shortcomings of the prior art, in order to realize the rapid determination of the positioning and azimuth of multiple moving targets, the present invention proposes a method for detecting the position and azimuth of multiple moving targets based on single-pixel imaging, thereby promoting the development and application of single-pixel imaging technology in the coordinated control of multiple moving targets.

[0006] To achieve the above object, the present invention discloses a method for rapid detection of multiple moving target positions and azimuths based on single pixel imaging, which comprises the following steps:

[0007] S1: For each frame of a scene image including a moving target, a one-dimensional projection curve of the scene image is obtained;

[0008] For each frame of a scene image including a moving target, the scene image is modulated using the Fourier mode to obtain a first light intensity value, and the Fourier coefficient F is obtained according to the first light intensity value using the three-step phase shift method formula. i (u,v),F i (u, v) is the Fourier coefficient of the i-th frame scene image at the frequency (u, v); and according to the Fourier coefficient F i (u, v) obtains the one-dimensional projection curve of the scene image in four directions of 0°, 45°, 90°, and 135°;

[0009] S2: Divide the scene image into N regions using one-dimensional projection;

[0010] The projection inverse reconstruction method is used to obtain the regions where N moving targets are located, and the scene image is segmented according to the positions of the moving target regions. If there are N moving targets, the scene image is divided into N regions, so that there is only one moving target in each region; where N is a positive integer;

[0011] S3: Modulate the scene image using the deformed geometric moment pattern to obtain the center of mass position coordinates of each moving target;

[0012] Generate a corresponding deformation geometric moment pattern according to the coordinate range of the area where each moving target is located, use the deformation geometric moment pattern to modulate the scene image to obtain a second light intensity value, a third light intensity value, and a fourth light intensity value, obtain the centroid coordinates of the moving target relative to the area where it is located according to the second light intensity value, the third light intensity value, and the fourth light intensity value, and then convert them into the centroid coordinates of the moving target in the scene image;

[0013] S4: Obtain the corresponding azimuth angle according to the center of mass coordinates of each moving target;

[0014] According to the center of mass coordinates of the moving target in S3, the second-order central moment pattern is obtained, and the normalized difference method is used to obtain the normalized difference second-order central moment pattern for the second-order central moment pattern. The normalized difference second-order central moment pattern is then used to modulate the scene image to obtain the fifth light intensity value and the sixth light intensity value. The second-order central moment is obtained according to the fifth light intensity value and the sixth light intensity value. The covariance matrix is ​​constructed according to the second-order central moment to obtain two eigenvalues ​​and corresponding two eigenvectors of the covariance matrix, and the azimuth of the moving target is obtained according to the eigenvectors.

[0015] Preferably, in step S1, for each frame of scene image including a moving target, obtaining a one-dimensional projection curve of the scene image is specifically:

[0016] S11: Generate a series of different spatial frequencies and initial phases They are Grayscale Fourier mode of

[0017]

[0018] Among them, a is the average light intensity value of the grayscale Fourier mode, b is the contrast of the grayscale Fourier mode, x, y are the spatial coordinates of the scene image, u, v are the spatial frequencies in the discrete domain; when the size of the scene image of the moving target is M×N, the values ​​of x are 1, 2, ..., N, the values ​​of y are 1, 2, ..., M, and the values ​​of u are: The values ​​of v are is the initial phase, and its values ​​are

[0019] S12: modulating the scene image of the moving target using a binary Fourier mode to obtain a first modulated image;

[0020] S13: collecting a first light intensity value of a first modulated image using a single-pixel detector;

[0021]

[0022] Where, e is the background noise, I i (x, y) is the i-th frame scene image; is a binary Fourier mode, is the first light intensity value;

[0023] S14: Calculate the Fourier coefficient F at the frequency (u,v) using the three-step phase shift formula i (u, v):

[0024]

[0025] S15: Obtain one-dimensional projection curves of the scene image in four directions of 0°, 45°, 90°, and 135°;

[0026] The one-dimensional projection of the scene in the four directions of 0°, 45°, 90°, and 135° is equal to the one-dimensional inverse Fourier transform result of the image's two-dimensional Fourier transform corresponding to the slices in the directions of 0°, 45°, 90°, and 135° through the origin:

[0027]

[0028] Among them, θ is the angle between the projection direction and the x-axis direction, T i (θ) is the projection curve of the i-th frame scene image in the θ direction, F i -1 is the Fourier coefficient Fi The one-dimensional inverse Fourier transform of (u, v).

[0029] Preferably, in step S2, the scene image is divided into N regions using one-dimensional projection, specifically:

[0030] S21: Preliminarily obtain the regions where multiple moving targets are located by using the projection reverse reconstruction method;

[0031] According to the one-dimensional projection curve T of the scene image in four directions of 0°, 45°, 90°, and 135° i (θ) is processed to obtain the binary projection result B i (θ):

[0032]

[0033] Among them, ε is the threshold;

[0034] Using the binary projection result B i (θ) obtains projection strips in four directions, and merges them to obtain the final combined projection result of the scene image of the moving target

[0035]

[0036] in, is the projection strip of the scene image in the i-th frame in the j-direction, j = 1, 2, 3, 4, corresponding to the four directions of 0°, 45°, 90°, and 135° respectively;

[0037] Take the area where the projection strips in the four directions intersect to obtain the regional position A of multiple moving targets in the i-th frame scene image of the moving target i (x,y):

[0038]

[0039] According to the region position, the edge detection algorithm is used to obtain the number N and boundary position of the moving objects in the i-th frame scene image;

[0040] S22: performing regional segmentation on the scene image according to the regional position of the moving target;

[0041] By separating the moving targets in the horizontal and vertical directions, if there are N moving targets, the scene image is divided into N areas, so that there is only one moving target in each area.

[0042] Preferably, step S22: performing regional segmentation on the scene image according to the regional position of the moving target, specifically:

[0043] If the N moving targets can be completely separated in a single horizontal or vertical direction, then region segmentation is selected in that direction, and the position of the dividing line is the midpoint line of the boundary positions of two adjacent moving targets in that direction; if they cannot be completely distinguished in a single direction, then the direction that can divide the largest number of regions is selected for region segmentation, and the position of the dividing line is still the midpoint line of the boundary positions of adjacent moving targets that can be separated, and then region segmentation is performed in another direction.

[0044] Preferably, S3 uses a deformed geometric moment mode to modulate the scene image to obtain the center of mass position coordinates of each moving target, specifically:

[0045] S31: The method for generating the deformed geometric moment pattern is:

[0046] For a scene image of size M×N, after segmentation, the coordinate range of a moving target in the segmentation area in the x direction is (x1, x2), the coordinate range of the segmentation area in the y direction is (y1, y2), and the length of the segmentation area in the x direction is L x =x2-x1+1, the length of the segmented area in the y direction is L y =y2-y1+1, then three grayscale deformation geometric moment patterns S0, S1, S2 are generated and expressed as:

[0047]

[0048]

[0049] S32: modulating the scene image of the moving target using a deformation geometric moment pattern to obtain a second modulated image;

[0050] S33: Obtaining the center of mass position coordinates of the moving target;

[0051] A single pixel detector is used to collect the second light intensity value d0, the third light intensity value d1 and the fourth light intensity value d2 of the second modulated image:

[0052] Then the centroid coordinates (x', y') of the moving target relative to the segmented area where it is located can be expressed as:

[0053]

[0054] For the entire M×N size scene image including the moving target, the centroid coordinates (x, y) of the moving target are expressed as:

[0055]

[0056] For each moving target in the scene image, three grayscale deformation geometric moment patterns are generated according to S31, and then the centroid coordinates (x, y) of each moving target in the scene image are obtained according to S32 and S33. The centroid coordinates (x, y) of the moving target in the scene image are the position of the moving target.

[0057] Preferably, the step S4: obtaining the corresponding azimuth angle according to the centroid coordinates of each moving target is specifically:

[0058] S41: Generate the second-order central moment pattern by:

[0059] The center of mass coordinates of a moving target Q in the scene image is (x Q ,y Q ), the second-order central moment mode S pq Described as:

[0060] S pq =(xx Q ) p (yy Q ) q ; (14)

[0061] Among them, x Q and Q is the center of mass coordinate of the moving target Q, p+q=2, p∈{0,1,2}, q∈{0,1,2}; therefore, the second-order central moment mode S pq Including S 11 , S 20 and S 02 Three types;

[0062] Use the normalized difference method to get the second-order central moment mode of the normalized difference:

[0063]

[0064] Among them, a and b are S pq The minimum and maximum pixel grayscale values ​​in ;

[0065] S42: modulating the scene image of the moving target using a normalized difference second-order central moment mode to obtain a third modulated image;

[0066] S43: The method for obtaining the azimuth of the moving target is:

[0067] The fifth light intensity value of the third modulated image is collected using a single pixel detector and the sixth light intensity value

[0068] The second-order central moment of the moving target Q in the scene image is:

[0069]

[0070] Constructing the covariance matrix The two eigenvalues ​​λ1, λ2 of the covariance matrix T and the corresponding two eigenvectors v1, v2 are obtained; the azimuth angle θ of the moving target Q is:

[0071]

[0072] Among them, θ is the azimuth of the moving target Q, arctan() is the inverse tangent function, and v p (2) is the feature vector v p The second component, v p (1) is the feature vector v p The first component of

[0073] For each moving target in the scene image, S4 is used to obtain the azimuth angle of the corresponding moving target.

[0074] Preferably, in the step S41, in which the second-order central moment pattern is generated, the size of the second-order central moment pattern is 1.1 times the size of the moving target in the scene image.

[0075] Compared with the prior art, the present invention has the following beneficial effects:

[0076] (1) The method for rapid detection of the position and azimuth of multiple moving targets based on single-pixel imaging of the present invention can simultaneously detect the position coordinates and azimuths of multiple moving targets, thereby achieving accurate and rapid detection.

[0077] (2) The moving target position coordinates and azimuth angle results obtained by the present invention are accurate, and the number of modulations required for each frame is relatively small.

[0078] (3) By utilizing the high-speed modulation characteristics of DMD, the position and direction angle of the moving target can be detected quickly and accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 It is a flow chart of the method for rapid detection of multiple moving target positions and azimuths based on single pixel imaging of the present invention;

[0080] Figure 2 This is a schematic diagram of one-dimensional projection results in four directions obtained for a frame of scene image according to the present invention;

[0081] Figure 3 It is a schematic diagram of the segmentation of multiple drone regions obtained by the present invention for a certain frame of scene image;

[0082] Figure 4A schematic diagram of a deformation geometric moment pattern for modulating a frame of scene image according to the present invention;

[0083] Figure 5 A schematic diagram of a normalized differential second-order central moment mode for modulating a frame of scene image according to the present invention;

[0084] Figure 6 The present invention shows the positioning results of the center-of-mass motion trajectory of three UAVs and the schematic diagram of the actual center-of-mass trajectory;

[0085] Figure 7 This is a graph showing the error results of the present invention for positioning the centroid of three UAVs;

[0086] Figure 8 This is a comparison result diagram of the azimuth angle detection results of three UAVs and the actual azimuth angle, as well as an error result diagram of the azimuth angle detection of three UAVs. DETAILED DESCRIPTION

[0087] According to a method for quickly detecting the positions and azimuths of multiple moving targets based on single-pixel imaging, the embodiment of the present invention sets the moving targets as drones, and uses a single-pixel system to measure the motion information of multiple drones. The present invention is further described in detail below in conjunction with the accompanying drawings.

[0088] A fast detection method for the position and azimuth of multiple moving targets based on single-pixel imaging, such as Figure 1 As shown, the Fourier coefficients of each frame of the scene image are obtained by using a single-pixel imaging system, and the one-dimensional projection curves of the scene in four directions of 0°, 45°, 90°, and 135° are calculated; the projection reverse reconstruction method is used to preliminarily obtain the regional position of each drone, and the scene is segmented according to the regional position of the drone. The scene image is divided into several regions according to the number of drones, so that there is only one drone in each region; the scene image is modulated using a deformed geometric moment mode to obtain the coordinates of the center of mass position of each drone; according to the coordinates of the center of mass position of the drone, each drone is modulated using a second-order central moment mode to calculate the azimuth of each drone; specifically, it includes:

[0089] S1: For each frame of the scene image including the drone, a one-dimensional projection curve of the scene image is obtained.

[0090] For each frame of the scene image including the drone, the Fourier coefficients of the scene image are first obtained using the single-pixel imaging system, and then the one-dimensional projection curves of the scene image in four directions of 0°, 45°, 90°, and 135° are obtained respectively; the specific steps are as follows:

[0091] S11: Generate a series of different spatial frequencies and initial phases They are Grayscale Fourier mode of

[0092]

[0093] Where a is the average light intensity value of the grayscale Fourier mode, b is the contrast of the grayscale Fourier mode, x and y are the spatial coordinates of the scene image, u and v are the spatial frequencies in the discrete domain; assuming that the size of the scene image of the drone is M×N, the values ​​of x are 1, 2, ..., N, the values ​​of y are 1, 2, ..., M, and the values ​​of u are: The values ​​of v are is the initial phase, and its values ​​are

[0094] S12: Modulate the scene image of the drone using binary Fourier patterns.

[0095] Convert grayscale Fourier patterns to binary Fourier patterns using the Floyd-Steinberg error dithering algorithm The DMD (Digital Micromirror Devices) is used to load the binary Fourier pattern to modulate the i-th frame scene image of the drone to obtain the first modulated image of the i-th frame scene image.

[0096] S13: Collecting the total light intensity value of the first modulated image using a single-pixel detector.

[0097] For using the initial phase A first modulated image is obtained by modulating the grayscale Fourier mode with spatial frequencies u and v, and a first total light intensity value of the first modulated image is collected by using a single pixel detector. for:

[0098]

[0099] Where, e is the background noise, I i (x, y) is the i-th frame scene image; is a binary Fourier mode, is the first light intensity value.

[0100] S14: Calculate the Fourier coefficient F at the frequency (u,v) using the three-step phase shift formula i (u, v):

[0101]

[0102] According to S11 to S14, for the i-th frame scene image of the drone, firstly, S12 is used to modulate the series of grayscale Fourier patterns in S11, and then the i-th frame scene image is obtained using the initial phase according to formula (2). The first light intensity value after the grayscale Fourier pattern with spatial frequency (u, v) is modulated, and finally formula (3) is used to obtain the Fourier coefficient F of the i-th frame scene image of the drone at the frequency (u, v): i (u, v).

[0103] S15: Obtain one-dimensional projection curves of the scene image in four directions of 0°, 45°, 90°, and 135°.

[0104] Based on the Fourier slice theorem: the Fourier transform result of the one-dimensional projection curve of a two-dimensional image in a certain direction is equivalent to the slice result of the two-dimensional Fourier transform of the image in the same direction through the origin; therefore, the one-dimensional projection of the scene in the four directions of 0°, 45°, 90°, and 135° is equal to the one-dimensional Fourier inverse transform result of the slice in the corresponding directions of 0°, 45°, 90°, and 135° of the two-dimensional Fourier transform of the image through the origin:

[0105]

[0106] Among them, θ is the angle between the projection direction and the x-axis direction, T i (θ) is the projection curve of the i-th frame scene image in the θ direction, F i -1 is the Fourier coefficient F i The one-dimensional inverse Fourier transform of (u, v).

[0107] S2: Divide the scene image into N regions using one-dimensional projection.

[0108] The projection inverse reconstruction method is used to preliminarily obtain the areas where multiple drones are located, and the scene image is segmented according to the location of the drone area. The scene image is divided into several areas according to the number of drones, so that there is only one drone in each area.

[0109] S21: Use the projection reverse reconstruction method to preliminarily obtain the areas where multiple drones are located. The specific method is as follows:

[0110] Since there is a certain contrast between the drone and the background, the one-dimensional projection value is higher in the area where the drone is located, and the one-dimensional projection value is lower in the area where there is no drone. A threshold ε is set to i (θ) is processed to obtain the binary projection result B i (θ):

[0111]

[0112] According to step S1 and expression (4), the one-dimensional projection curves of the scene in four directions of 0°, 45°, 90°, and 135° are used in this embodiment. Therefore, according to expression (5), the projection strip of the i-th frame scene image is obtained by reverse reconstruction using projection: They correspond to the four directions of 0°, 45°, 90°, and 135° in expression (4). The projection strips in the four directions are combined to obtain the final combined projection result of the drone scene image.

[0113]

[0114] Take the area where the projection strips in the four directions intersect to obtain the regional position A of multiple drones in the i-th frame scene image of the drone i (x,y):

[0115]

[0116] According to the regional position, the edge detection algorithm is used to obtain the number N and boundary position of drones in the i-th frame scene image.

[0117] S22: The specific method for performing regional segmentation on the scene image according to the regional position of the drone is:

[0118] If N drones can be completely separated in a single horizontal or vertical direction, region segmentation is performed in that direction first, and the position of the segmentation line is the midpoint line of the boundary positions of two adjacent drones in that direction; if they cannot be completely distinguished in a single direction, the direction that can divide the largest number of regions is prioritized for region segmentation, and the position of the segmentation line is still the midpoint line of the boundary positions of adjacent drones that can be separated, and then region segmentation is performed in another direction. If there are N drones, the scene image is divided into N regions, so that there is only one drone in each region.

[0119] S3: The scene image of the UAV is modulated using the deformed geometric moment pattern to obtain the center of mass position coordinates of each UAV.

[0120] S31: The method for generating the deformed geometric moment pattern is:

[0121] For a scene image of size M×N, assume that after segmentation, the coordinate range of the segmented area of ​​a certain drone in the x direction is (x1, x2), the coordinate range of the segmented area in the y direction is (y1, y2), and the length of the segmented area in the x direction is L x =x2-x1+1, the length of the segmented area in the y direction is L y=y2-y1+1, then three grayscale deformation geometric moment patterns S0, S1, S2 are generated and expressed as:

[0122]

[0123] Where AND represents the AND relationship. Taking expression (8) as an example, when i belongs to the coordinate range (x1, x2) of the segmented region in the x direction and j belongs to the coordinate range (y1, y2) of the segmented region in the y direction, the element a in the grayscale deformation geometric moment pattern S0 is ji The value of is 1; when i does not belong to the coordinate range (x1, x2) of the segmented region in the x direction, or j does not belong to the coordinate range (y1, y2) of the segmented region in the y direction, the element a in the grayscale deformation geometric moment mode S0 is ji The value of is 0.

[0124] S32: Modulating the scene image of the drone using the deformation geometric moment pattern.

[0125] The Floyd-Steinberg error dithering algorithm is used to convert the three grayscale deformation geometric moment patterns into three binary deformation geometric moment patterns S′0, S′1, and S′2 respectively, and the DMD is used to modulate the i-th frame scene image of the drone using the generated binary deformation geometric moment patterns to obtain the second modulated image.

[0126] For the traditional geometric moment method, a grayscale geometric moment mode is used, in which the grayscale of the first-order moment mode is related to the image size. For example, if the scene image size is 256×256, the grayscale of the traditional first-order moment mode is 256. Therefore, the coordinate position error of the center of gravity obtained after the existing grayscale geometric moment mode is dithered and binarized is large; the grayscale deformation geometric moment mode is used here to reduce the grayscale of the first-order moment mode: the grayscale of the first-order moment mode in the x direction is L x , the grayscale value of the first-order moment mode in the y direction is L y ,After the gray deformation geometric moment pattern is dithered and binarized, the centroid position coordinate error is reduced compared with the ,traditional method.

[0127] S33: The method for obtaining the center of mass position coordinates of each drone is:

[0128] A single-pixel detector is used to collect the total light intensity value of the second modulated image. The binary deformation geometric moment patterns S′0, S′1, and S′2 are used to modulate the i-th frame scene image respectively. The corresponding total light intensity values ​​are the second light intensity value d0, the third light intensity value d1, and the fourth light intensity value d2 respectively:

[0129]

[0130] Among them, I i(x, y) is the i-th frame scene image.

[0131] Then the centroid coordinates (x', y') of the drone relative to the segmented area where it is located can be expressed as:

[0132]

[0133] For the entire M×N size scene image including the drone, the center of mass coordinates (x, y) of the drone are expressed as:

[0134]

[0135] For each UAV in the scene image, three grayscale deformation geometric moment patterns are generated according to S31, and then the center of mass coordinates (x, y) of each UAV in the scene image are obtained according to S32 and S33. The center of mass coordinates (x, y) of the UAV in the scene image is the position of the UAV.

[0136] S4: Get the azimuth angle according to the center of mass coordinates of the UAV.

[0137] According to the center of mass coordinates of the drone in S3, each drone is modulated using the second-order central moment mode to calculate the azimuth of each drone, specifically:

[0138] S41: Generate the second-order central moment pattern by:

[0139] The center of mass position coordinates of each drone are obtained through step S3, and each drone is modulated using the second-order central moment mode; assuming that the center of mass coordinates of one of the drones Q is (x Q ,y Q ), the second-order central moment mode S pq It can be described as:

[0140] S pq =(xx Q ) p (yy Q ) q ; (14)

[0141] Among them, x Q and Q is the center of mass coordinate of the drone Q. According to step S3, p+q=2, p∈{0, 1, 2}, q∈{0, 1, 2}. Considering the permutations and combinations of p and q, the second-order central moment pattern S can be obtained. pq Including S 11 , S 20 and S 02 Three types.

[0142] In order to reduce the binarization error, we can consider reducing the size of the second-order central moment pattern, thereby reducing the grayscale. Given the size of the drone in the scene image, we can select the size of the second-order central moment pattern to be slightly larger than the size of the drone in the scene image. For example, if the size of the second-order central moment pattern is 1.1 times the size of the drone, and the size of the drone in the scene image is 100 pixels × 100 pixels, then the size of the second-order central moment pattern is x. Q and Q In addition, in order to reduce the noise caused by background light, the normalized difference method is used to obtain the second-order central moment mode of the normalized difference:

[0143]

[0144] Among them, a and b are S pq The minimum and maximum pixel grayscale values ​​in S pq+ and S pq- satisfy:

[0145]

[0146] S42: Modulating the scene image of the drone using the normalized difference second-order central moment pattern.

[0147] The Floyd-Steinberg error dithering algorithm is used to convert the normalized difference second-order central moment pattern into a binary normalized difference second-order central moment pattern S′. pq+ and S′ pq- , and use DMD to load the binary normalized difference second-order central moment mode S′ pq+ and S′ pq- Modulate the scene image of the drone. Because the second-order central moment mode S corresponding to each drone in the scene image pq Including S 11 , S 20 and S 02 There are three types, each of which uses the normalized difference method to obtain the second-order central moment pattern of the normalized difference. Therefore, each drone in the i-th frame scene image is modulated using 6 central moment patterns to obtain the third modulated image. The scene image includes N drones, so 6N modulations are required.

[0148] S43: The method for obtaining the azimuth angle of the drone is:

[0149] The second-order central moment mode S′ is obtained by binary normalized difference pq+ and S′ pq- After modulation, the total light intensity value of the third modulated image collected by the single pixel detector includes the fifth light intensity value and the sixth light intensity value for:

[0150]

[0151] Where, e is the background noise, I i (x,y) is the i-th frame scene image.

[0152] Similarly, taking the drone Q in the scene image as an example, its second-order central moment is:

[0153]

[0154] According to the relationship between the second-order central moment mode and the second-order central moment mode of the normalized difference, and the relationship between the fifth light intensity value, the sixth light intensity value and the second-order central moment mode of the normalized difference, it can be known that the second-order central moment can be obtained by the fifth light intensity value of the third modulated image. and the sixth light intensity value Get as follows:

[0155]

[0156] Constructing the covariance matrix Assume that the two eigenvalues ​​of the covariance matrix T are λ1 and λ2, and the corresponding two eigenvectors are v1 and v2; the larger eigenvalue of λ1 and λ2 is λ p , the corresponding eigenvector is v p ;Right now Through the principal component analysis theory, the azimuth angle θ of the UAV can be obtained from the eigenvector v of the covariance matrix T p The calculation results are:

[0157]

[0158] Where θ is the azimuth angle of the UAV Q, v p (2) is the feature vector v p The second component, v p (1) is the feature vector v p The first component of .

[0159] For each UAV in the scene image, S4 is used to obtain the azimuth angle of the corresponding UAV.

[0160] The embodiment of the present invention specifically provides a method for detecting the position and azimuth of multiple UAVs based on single-pixel imaging, and obtains the motion information of multiple moving UAVs using a single-pixel imaging system. The size of the scene image is 512 pixels × 512 pixels, wherein the scene image includes three UAVs, and the UAVs are all grayscale images with different grayscale values. The sizes of the three UAVs are 60 pixels × 60 pixels, 70 pixels × 70 pixels, and 80 pixels × 80 pixels, respectively. The UAVs perform translational motion, and the motion trajectories are infinity, circle, and ellipse, respectively, while performing random rotation motion. The scene image has a total of 360 frames. The present invention simultaneously measures the position coordinates and azimuth of the three UAVs for each frame in the scene image.

[0161] S1: For each frame of scene image, obtain a one-dimensional projection curve of the scene image.

[0162] For each frame of the scene image, the Fourier coefficients of the scene are first obtained using a single-pixel imaging system, and the one-dimensional projection curves of the scene image in the four directions of 0°, 45°, 90°, and 135° are calculated; the one-dimensional projection curves of the scene image in the four directions of 0°, 45°, 90°, and 135° are equal to the one-dimensional Fourier inverse transform results of the two-dimensional Fourier transform of the scene image corresponding to the slices in the directions of 0°, 45°, 90°, and 135° through the origin:

[0163]

[0164] Among them, θ is the angle between the projection direction and the x-axis direction, T i (θ) is the projection curve of the i-th frame scene image in the θ direction, F i -1 is the Fourier coefficient F i The one-dimensional inverse Fourier transform of (u, v).

[0165] Therefore, here we only need to sample the Fourier coefficients at the frequency (u, v) on the straight line passing through the origin and with an angle of 0°, 45°, 90°, and 135° to the x-axis.

[0166] The present invention does not need to perform full sampling. Assuming that the sampling rate is η, when η is 0.05, for a frame of scene image of 512pixel×512pixel, 512×0.05≈26 frequency points (including the common origin) are sampled in each direction. The common origin is removed, that is, 25 Fourier coefficients are sampled in each direction. According to the three-step phase shift method, each Fourier coefficient needs to be projected into 3 patterns, and there are 4 directions, that is, 25×3×4 patterns. Since the Fourier coefficient at the origin is a real value, it only needs to be sampled twice to obtain it. Therefore, each frame of scene image needs to project P1=25×3×4+2=302 patterns in step S1. Figure 2 It is a schematic diagram showing the one-dimensional projection results in four directions obtained by executing step S1 on a certain frame of scene image in the present invention. Figure 2 It only schematically shows three projections of three initial phases of a spatial frequency, which does not mean that there are only three projections in one direction.

[0167] S2: Divide the scene image into N regions using one-dimensional projection.

[0168] The projection inverse reconstruction method is used to preliminarily obtain the areas where multiple drones are located, and the scene image is segmented according to the location of the drone area. The scene image is divided into several areas according to the number of drones, so that there is only one drone in each area; Figure 3 As shown, the present invention is a schematic diagram of multiple drone region cuts obtained by executing step S2 on a certain frame of scene image.

[0169] S3: The scene image of the UAV is modulated using the deformed geometric moment pattern to obtain the center of mass position coordinates of each UAV.

[0170] like Figure 4 As shown, it is a schematic diagram of the deformation geometric moment pattern modulated by the present invention for a certain frame of scene image. Each drone needs to project 3 deformation geometric moment patterns. For the case where there are 3 moving drones, each frame of scene image needs to project P2=9 deformation geometric moment patterns in step S3.

[0171] Find the center of mass position of each drone. Figure 6 As shown in FIG. 1 , the center-of-mass motion trajectory positioning results and the actual center-of-mass trajectory schematic diagram of the three UAVs of the present invention are shown in FIG. Figure 7 As shown in the figure, it is the error result diagram of the centroid positioning of three UAVs by the present invention. It can be seen from the results that the centroid errors of most frames of UAVs 2 and 3 in the x and y directions are between -2 and 2 pixels, and the absolute value of the angle error of a few frames is slightly greater than 2 pixels, but not more than 3 pixels. The result of UAV 1 in the x direction is similar to that of UAVs 1 and 2; in the y direction, the centroid errors of most frames are between -3 and 3 pixels, and the absolute value of the centroid errors of a few frames is slightly greater than 3 pixels, but not more than 5 pixels. It can be seen that the centroid positioning result is relatively accurate.

[0172] The root mean square error (RMSE) is used to evaluate the positioning accuracy. Assume that the actual center of mass coordinates of a certain drone in the i-th frame is The detected center of mass coordinates are There are K frames of scene images in total, so the root mean square error of the drone positioning result is:

[0173]

[0174] By calculation, it is found that the root mean square errors of the centroid positioning of UAVs 1-3 using the method proposed in the present invention are 1.418, 0.996, and 0.995, respectively. It can be seen that the centroid positioning result of the present invention is relatively accurate.

[0175] S4: According to the center of mass coordinates of the UAV in S3, each UAV is modulated using the second-order central moment mode to calculate the azimuth angle of each UAV.

[0176] like Figure 5 As shown, it is a schematic diagram of the normalized differential second-order central moment pattern modulated by the present invention for a certain frame of scene image. Each drone needs to project 6 differential second-order central moment patterns. For the case where there are 3 moving drones, each frame of the scene needs to project P3=18 patterns in step S4. In summary, each frame of the scene needs to project P=P1+P2+P3=329 patterns in total.

[0177] like Figure 8 As shown in FIG. 1 , the comparison result of the azimuth angle detection of three UAVs and the actual azimuth angle of the present invention is shown, as well as the error result diagram of the azimuth angle detection of three UAVs. Figure 8 It can be seen that the azimuth error of UAV 3 is between -5° and 5°; the azimuth error of UAV 2 is mostly between -5° and 10°, and the angle error of two frames exceeds 10°; the azimuth error of UAV 1 is mostly between -5° and 15°, and the angle error of one frame is slightly more than 15°. The root mean square error of angle measurement is also used to evaluate the measurement accuracy: Assume that the actual azimuth angle of a UAV in the i-th frame is θ i The detected azimuth is β i , there are K frames of scene images, then the root mean square error RMSE of the azimuth detection result of the UAV is θ Defined as:

[0178]

[0179] The root mean square errors of the azimuth angles of UAVs 1-3 obtained by calculation are 6.29, 3.46, and 1.74, respectively. It can also be seen that the azimuth angle results of the present invention are relatively accurate.

[0180] The present invention has the following beneficial effects: the method for detecting the position and azimuth of multiple moving UAVs based on single-pixel imaging can realize rapid detection of the position coordinates and azimuths of multiple moving UAVs; the position coordinates and azimuths of the moving UAVs calculated by the present invention are relatively accurate, and the number of detection speckles (i.e., patterns) required for each frame is relatively small; the high-speed modulation characteristics of the DMD can be used to realize rapid and high-precision detection of the position and azimuth of the UAV, which is convenient for further controlling the flight direction and path of the UAV.

[0181] The embodiments described above are only descriptions of the preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for rapid detection of the position and azimuth of multiple moving targets based on single-pixel imaging, characterized in that: It includes the following steps: S1: For each frame of a scene image including multiple moving targets, a one-dimensional projection curve of the scene image is obtained; For each frame of the scene image, the scene image is modulated using the Fourier mode to obtain a first light intensity value, and the Fourier coefficient F is obtained according to the first light intensity value using the three-step phase shift method formula. i (u,v),F i (u, v) is the Fourier coefficient of the i-th frame scene image at the frequency (u, v); and according to the Fourier coefficient F i (u, v) obtains the one-dimensional projection curve of the scene image in four directions of 0°, 45°, 90°, and 135°; S2: Divide the scene image into N regions using one-dimensional projection; The projection inverse reconstruction method is used to obtain the regions where N moving targets are located, and the scene image is segmented according to the positions of the moving target regions. If there are N moving targets, the scene image is divided into N regions, so that there is only one moving target in each region; where N is a positive integer; S3: Modulate the scene image using the deformed geometric moment pattern to obtain the center of mass position coordinates of each moving target; Generate a corresponding deformation geometric moment pattern according to the coordinate range of the area where each moving target is located, use the deformation geometric moment pattern to modulate the scene image to obtain a second light intensity value, a third light intensity value, and a fourth light intensity value, obtain the centroid coordinates of the moving target relative to the area where it is located according to the second light intensity value, the third light intensity value, and the fourth light intensity value, and then convert them into the centroid coordinates of the moving target in the scene image; S4: Obtain the corresponding azimuth angle according to the center of mass coordinates of each moving target; According to the center of mass coordinates of the moving target in S3, the second-order central moment pattern is obtained, and the normalized difference method is used to obtain the normalized difference second-order central moment pattern for the second-order central moment pattern. The normalized difference second-order central moment pattern is then used to modulate the scene image to obtain the fifth light intensity value and the sixth light intensity value. The second-order central moment is obtained according to the fifth light intensity value and the sixth light intensity value. The covariance matrix is ​​constructed according to the second-order central moment to obtain two eigenvalues ​​and corresponding two eigenvectors of the covariance matrix, and the azimuth of the moving target is obtained according to the eigenvectors.

2. The method for rapid detection of multiple moving target positions and azimuths based on single pixel imaging according to claim 1, characterized in that: In step S1, for each frame of the scene image including the moving target, the one-dimensional projection curve of the scene image is obtained as follows: S11: Generate a series of different spatial frequencies and initial phases They are Grayscale Fourier mode of Among them, a is the average light intensity value of the grayscale Fourier mode, b is the contrast of the grayscale Fourier mode, x, y are the spatial coordinates of the scene image, u, v are the spatial frequencies in the discrete domain; when the size of the scene image of the moving target is M×N, the values ​​of x are 1, 2, ..., N, the values ​​of y are 1, 2, ..., M, and the values ​​of u are: The values ​​of v are is the initial phase, and its values ​​are S12: modulating the scene image of the moving target using a binary Fourier mode to obtain a first modulated image; S13: collecting a first light intensity value of a first modulated image using a single-pixel detector; Where, e is the background noise, I i (x, y) is the i-th frame scene image; is a binary Fourier mode, is the first light intensity value; S14: Calculate the Fourier coefficient F at the frequency (u,v) using the three-step phase shift formula i (u,v): S15: Obtain one-dimensional projection curves of the scene image in four directions of 0°, 45°, 90°, and 135°; The one-dimensional projection of the scene in the four directions of 0°, 45°, 90°, and 135° is equal to the one-dimensional inverse Fourier transform of the image's two-dimensional Fourier transform corresponding to the slices in the directions of 0°, 45°, 90°, and 135° through the origin: Among them, θ is the angle between the projection direction and the x-axis direction, T i (θ) is the projection curve of the i-th frame scene image in the θ direction, F i -1 is the Fourier coefficient F i The one-dimensional inverse Fourier transform of (u,v).

3. The method for rapid detection of multiple moving target positions and azimuths based on single pixel imaging according to claim 1, characterized in that: In step S2, the scene image is divided into N regions using one-dimensional projection, specifically: S21: Preliminarily obtain the regions where multiple moving targets are located by using the projection reverse reconstruction method; According to the one-dimensional projection curve T of the scene image in four directions of 0°, 45°, 90°, and 135° i (θ) is processed to obtain the binary projection result B i (θ): Among them, ε is the threshold; Using the binary projection result B i (θ) obtains projection strips in four directions, and merges them to obtain the final combined projection result of the scene image of the moving target in, is the projection strip of the scene image in the i-th frame in the j-direction, j = 1, 2, 3, 4, corresponding to the four directions of 0°, 45°, 90°, and 135° respectively; Take the area where the projection strips in the four directions intersect to obtain the regional position A of multiple moving targets in the i-th frame scene image of the moving target i (x,y): According to the region position, the edge detection algorithm is used to obtain the number N and boundary position of the moving objects in the i-th frame scene image; S22: performing regional segmentation on the scene image according to the regional position of the moving target; By separating the moving targets in the horizontal and vertical directions, if there are N moving targets, the scene image is divided into N areas, so that there is only one moving target in each area.

4. The method for rapid detection of multiple moving target positions and azimuths based on single pixel imaging according to claim 3 is characterized in that: Step S22: Segment the scene image into regions according to the region position of the moving target, specifically: If the N moving targets can be completely separated in a single horizontal or vertical direction, then region segmentation is selected in that direction, and the position of the dividing line is the midpoint line of the boundary positions of two adjacent moving targets in that direction; if they cannot be completely distinguished in a single direction, then the direction that can divide the largest number of regions is selected for region segmentation, and the position of the dividing line is still the midpoint line of the boundary positions of adjacent moving targets that can be separated, and then region segmentation is performed in another direction.

5. The method for rapid detection of multiple moving target positions and azimuths based on single pixel imaging according to claim 1, characterized in that: The S3 modulates the scene image using the deformed geometric moment mode to obtain the center of mass position coordinates of each moving target, specifically: S31: The method for generating the deformed geometric moment pattern is: For a scene image of size M×N, after segmentation, the coordinate range of a moving target in the segmented area in the x direction is (x1, x2), the coordinate range of the segmented area in the y direction is (y1, y2), and the length of the segmented area in the x direction is L x =x2-x1+1, the length of the segmented area in the y direction is L y =y2-y1+1, then three grayscale deformation geometric moment patterns S0, S1, S2 are generated and expressed as: S32: modulating the scene image of the moving target using a deformation geometric moment pattern to obtain a second modulated image; S33: Obtaining the center of mass position coordinates of the moving target; A single pixel detector is used to collect the second light intensity value d0, the third light intensity value d1 and the fourth light intensity value d2 of the second modulated image: Then the centroid coordinates (x', y') of the moving target relative to the segmented area where it is located can be expressed as: For the entire M×N size scene image including the moving target, the centroid coordinates (x, y) of the moving target are expressed as: For each moving target in the scene image, three grayscale deformation geometric moment patterns are generated according to S31, and then the centroid coordinates (x, y) of each moving target in the scene image are obtained according to S32 and S33. The centroid coordinates (x, y) of the moving target in the scene image are the position of the moving target.

6. The method for rapid detection of multiple moving target positions and azimuths based on single pixel imaging according to claim 1, characterized in that: S4: obtaining the corresponding azimuth according to the centroid coordinates of each moving target, specifically: S41: Generate the second-order central moment pattern by: The center of mass coordinates of a moving target Q in the scene image is (x Q ,y Q ), the second-order central moment mode S pq Described as: S pq (xx) Q ) p (yy Q ) q ; (14) Among them, x Q and Q is the center of mass coordinate of the moving target Q, p+q=2, p∈{0,1,2}, q∈{0,1,2}; therefore, the second-order central moment mode S pq Including S 11 , S 20 and S 02 Three types; Use the normalized difference method to get the second-order central moment mode of the normalized difference: Among them, a and b are S pq The minimum and maximum pixel grayscale values; S42: modulating the scene image of the moving target using a normalized difference second-order central moment mode to obtain a third modulated image; S43: The method for obtaining the azimuth of the moving target is: The fifth light intensity value of the third modulated image is collected using a single pixel detector and the sixth light intensity value The second-order central moment of the moving target Q in the scene image is: Constructing the covariance matrix The two eigenvalues ​​λ1, λ2 of the covariance matrix T and the corresponding two eigenvectors v1, v2 are obtained; the azimuth angle θ of the moving target Q is: Among them, θ is the azimuth of the moving target Q, arctan() is the inverse tangent function, and v p (2) is the feature vector v p The second component, v p (1) is the feature vector v p The first component of For each moving target in the scene image, S4 is used to obtain the azimuth angle of the corresponding moving target.

7. The method for rapid detection of multiple moving target positions and azimuths based on single pixel imaging according to claim 6, characterized in that: In the S41, in generating the second-order central moment pattern, the size of the second-order central moment pattern is 1.1 times the size of the moving target in the scene image.

8. The method for rapid detection of multiple moving target positions and azimuths based on single pixel imaging according to any one of claims 1 to 7, characterized in that: The moving target is a drone.

Images