Complex background adaptive target detection method and system based on space-time domain fusion

Abstract

The application discloses a complex background adaptive target detection method and system based on space-time domain fusion, which comprises the following steps: performing sub-pixel matching on a corrected infrared sequence image to obtain n precision registration images; performing difference on two adjacent precision registration images in the n precision registration images to obtain a difference image; obtaining a minimum relative gray gradient filtered image; performing convolution on the minimum relative gray gradient filtered image and a preset template to obtain a neighborhood correlation target enhancement processed image; screening initial targets in the neighborhood correlation target enhancement processed image according to target morphological characteristics to obtain a candidate target set; screening the candidate target set by adopting a neighborhood decision method principle to obtain screened target points; and performing false alarm elimination on the screened target points according to interframe target similarity to obtain final target points. The application improves the detection rate of adaptive targets in a complex background and simultaneously suppresses the false alarm rate.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of infrared image processing, and particularly relates to a complex background adaptive target detection method and system based on space-time domain fusion. BACKGROUND

[0002] In recent years, with the development of space-based infrared imaging technology, the application of infrared remote sensing images is becoming mature and widespread. China has vast sea areas, and the detection and reconnaissance of sea surface targets have practical needs in civil and military fields. Infrared images have all-weather and all-day observation capabilities, and the extraction of targets using infrared remote sensing imaging technology is a key issue of current concern.

[0003] Infrared remote sensing images have low resolution, and targets often occupy only a few pixel sizes, losing geometric and texture features. In complex backgrounds, targets and noise are easily confused, and flickering and discontinuity may occur in sequence images, increasing the difficulty of infrared ship detection.

[0004] The key technical points of the current mainstream target detection algorithm can be summarized into three categories: the first category is background prediction method to suppress background influence; the second category is to use target and background information for filtering segmentation; and the third category is to use constant false alarm rate related algorithm to determine the detection threshold. Among them, the commonly used background prediction algorithm is based on the isotropic idea, which is not good for complex background. The infrared small target background prediction model based on anisotropy proposed by Guo Hongwei considers the anisotropy of the background, but the division of the complex background stationary region and non-stationary region cannot meet the high-precision background prediction demand, affecting the target detection effect; typical methods of target and background filtering include morphological Top-Hat filtering, exponential homomorphic filtering, and high-pass filtering, etc. The morphological Top-Hat filtering and exponential homomorphic filtering enhance the image contrast, but do not eliminate the complex background information. The former may cause the cloud layer brightness in the background to be enhanced, making it difficult to distinguish the target from the background. The high-pass filtering method processed image is overall dark, and it is difficult to detect target information in the cloud layer background; the constant false alarm rate related algorithm may result in too low false alarm rate, missing detection, or too high false alarm rate, low target detection accuracy. SUMMARY

[0005] The technical problem solved by the application is to overcome the shortcomings of the prior art and provide a complex background adaptive target detection method and system based on space-time domain fusion, which improves the detection rate of adaptive targets in complex backgrounds and suppresses the false alarm rate.

[0006] The objective of this invention is achieved through the following technical solution: a complex background adaptive target detection method based on spatiotemporal fusion, comprising: performing radiation uniformity correction on the original infrared sequence image to obtain a corrected infrared sequence image; performing sub-pixel matching on the corrected infrared sequence image to obtain n finely registered images; wherein, n≥2 and n is a positive integer; performing difference between two adjacent finely registered images in the n finely registered images to obtain a difference image; calculating the eight-neighbor gray-level gradient of each pixel in the difference image, and selecting the smallest eight-neighbor gray-level gradient from the eight-neighbor gray-level gradients of each pixel in the difference image; and applying the smallest eight-neighbor gray-level gradient to the difference image I. t By performing grayscale value replacement, the image I after minimum relative grayscale gradient filtering is obtained. f The image I after minimum relative gray-level gradient filtering f Convolution with a preset template yields an image with enhanced neighborhood-related targets; traversing the enhanced neighborhood-related target images I c Pixel I c (i,j) yields the adaptive detection threshold Tr(i,j), which is then compared with the enhanced image I of the neighborhood related targets. c Pixel I c Compared to (i,j), if the image I after neighborhood-related target enhancement processing... c Pixel I c If (i,j) is greater than the adaptive detection threshold Tr(i,j), then the image I after neighborhood-related target enhancement processing... c Pixel I c (i,j) represents the initial target; if the image I after enhancement of the neighboring related targets is... c Pixel I c If (i,j) is not greater than the adaptive detection threshold Tr(i,j), then the image I after neighborhood-related target enhancement processing... c Pixel I c (i,j) is the background; the image I after enhancement processing of neighborhood related targets. c The initial targets are screened based on their morphological features to obtain a candidate target set; the candidate target set is then screened to obtain the screened target points; and the screened target points are then filtered for false alarms based on the inter-frame target similarity to obtain the final target points.

[0007] In the above-mentioned adaptive target detection method for complex backgrounds based on spatiotemporal fusion, the dual-window constant false alarm rate detection method is used to traverse the pixels of the enhanced image of the neighborhood related targets to obtain the adaptive detection threshold; the neighborhood decision method is used to filter the candidate target set to obtain the filtered target points.

[0008] In the above complex background adaptive target detection method based on space-time domain fusion, the n precisely registered images are I1, I2, …In n ; the difference image is I t = I t+1 -I t ; wherein t = 1, 2, …, n-1, and t represents the serial number.

[0009] In the above complex background adaptive target detection method based on space-time domain fusion, the eight-neighborhood gray level gradient of each pixel in the difference image is obtained by the following formula:

[0010]

[0011] wherein i and j are the row and column numbers of the difference image I t , D(i, j) represents the eight-neighborhood gray level gradient of the pixel I t (i, j) in the difference image, and I t (i+w, j+v) represents the pixel DN value in the difference image I t at the coordinate position (i+w, j+v), D1(i, j), (i, j), D2(i, j), D4(i, j), D5(i, j), D6(i, j), D7(i, j), and D8(i, j) are respectively the first-neighborhood gray level gradient of each pixel, the second-neighborhood gray level gradient of each pixel, the third-neighborhood gray level gradient of each pixel, the fourth-neighborhood gray level gradient of each pixel, the fifth-neighborhood gray level gradient of each pixel, the sixth-neighborhood gray level gradient of each pixel, the seventh-neighborhood gray level gradient of each pixel, and the eighth-neighborhood gray level gradient of each pixel, wherein w = -1, 0, 1, and v = -1, 0, 1, and w and v are intermediate parameters.

[0012] In the above complex background adaptive target detection method based on space-time domain fusion, the minimum eight-neighborhood gray level gradient is obtained by the following formula:

[0013]

[0014] wherein D min (i, j) is the minimum eight-neighborhood gray level gradient, and I t (i, j) is the pixel in the difference image.

[0015] In the above complex background adaptive target detection method based on space-time domain fusion, the minimum relative gray level gradient filtered image I f is obtained by the following formula:

[0016]

[0017] wherein I f(i,j) is the minimum relative gray gradient filtered image I f DN value at coordinate position (i,j), TH is a gradient decision threshold, D min (i,j) is the minimum eight-neighborhood gray gradient, I t (i,j) is the pixel of the difference image, I t (i+w,j+v) represents the difference image I t DN value at coordinate position (i+w,j+v), (w,v|D min (i,j)) represents the minimum eight-neighborhood gray gradient direction corresponding pixel offset position, wherein w=-1, 0, 1, v=-1, 0, 1, w and v are intermediate parameters.

[0018] In the above complex background adaptive target detection method based on space-time domain fusion, the image after neighborhood-related target enhancement processing is: I c =I f *W; wherein I c is the image after neighborhood-related target enhancement processing, I f is the minimum relative gray gradient filtered image, and W is a preset template.

[0019] In the above complex background adaptive target detection method based on space-time domain fusion, the adaptive detection threshold Tr(i,j) is obtained by the following formula:

[0020] Tr(i,j) = m(i,j) - Φ -1 (P f )σ(i,j);

[0021] Wherein, Tr(i,j) is the adaptive detection threshold, P f is the false alarm rate, m(i,j) and σ(i,j) are the mean and standard deviation of all pixels in the decentered window of the pixel I c (i,j) of the image I c after neighborhood-related target enhancement processing.

[0022] In the above complex background adaptive target detection method based on space-time domain fusion, the candidate target set is obtained by screening the initial target in the image I c after neighborhood-related target enhancement processing according to target morphological characteristics, including: performing morphological operation on the image I c after neighborhood-related target enhancement processing to obtain a morphologically filtered image; counting the connected region area of the initial target in the image I c after neighborhood-related target enhancement processing and the corresponding position, retaining the target with an area greater than 3 pixels and less than 9 pixels according to target characteristics to obtain the candidate target set.

[0023] In the above-mentioned complex background adaptive target detection method based on space-time domain fusion, the morphological filtered image is obtained by the following formula:

[0024]

[0025] Wherein, I mod is the morphological filtered image, and H is a structure element.

[0026] A complex background adaptive target detection module based on space-time domain fusion comprises: a first module configured to perform radiation uniformity correction on an original infrared sequence image to obtain a corrected infrared sequence image; a second module configured to perform sub-pixel matching on the corrected infrared sequence image to obtain n precision registration images; wherein n is greater than or equal to 2 and is a positive integer; a third module configured to perform difference on two adjacent precision registration images in the n precision registration images to obtain a difference image; a fourth module configured to calculate eight-neighborhood gray scale gradients of each pixel in the difference image, and filter out the smallest eight-neighborhood gray scale gradient from the eight-neighborhood gray scale gradients of each pixel in the difference image; a fifth module configured to replace the gray scale values of the difference image according to the smallest eight-neighborhood gray scale gradient to obtain a minimum relative gray scale gradient filtered image; a sixth module configured to convolve the minimum relative gray scale gradient filtered image with a preset template to obtain a neighborhood correlation target enhanced processed image; a seventh module configured to traverse the pixels of the neighborhood correlation target enhanced processed image by using a double-window constant false alarm rate detection method to obtain an adaptive detection threshold, and compare the adaptive detection threshold with the pixels of the neighborhood correlation target enhanced processed image; if the pixel of the neighborhood correlation target enhanced processed image is greater than the adaptive detection threshold, the pixel of the neighborhood correlation target enhanced processed image is an initial target; if the pixel of the neighborhood correlation target enhanced processed image is not greater than the adaptive detection threshold, the pixel of the neighborhood correlation target enhanced processed image is a background; an eighth module configured to filter the initial target in the neighborhood correlation target enhanced processed image according to target morphological characteristics to obtain a candidate target set; a ninth module configured to filter the candidate target set by using a neighborhood decision method principle to obtain a filtered target point; and a tenth module configured to perform false alarm elimination on the filtered target point according to inter-frame target similarity to obtain a final target point.

[0027] Compared with the prior art, the present application has the following beneficial effects:

[0028] The application firstly carries out background registration based on sequence image information, carries out difference processing on the registered image, eliminates large-area background information, avoids poor applicability of the background prediction model, and solves the problem of complex model; then carries out background target integration processing on the difference image, so as to achieve the purpose of noise reduction and target enhancement; finally, the constant false alarm rate detection result is subjected to false alarm elimination processing in combination with the space-time domain fusion features of the target, so as to improve the target detection accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0029] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The accompanying drawings are included to provide a description of preferred embodiments and are not intended to limit the scope of the application. Furthermore, the drawings are not necessarily drawn to scale. In the drawings:

[0030] Figure 1 is a flow chart of the complex background adaptive target detection method based on space-time domain fusion provided by the embodiment of the application. DETAILED DESCRIPTION

[0031] Exemplary embodiments of the present disclosure will be described more fully hereinafter with reference to the accompanying drawings, in which exemplary embodiments of the present disclosure are shown. While several exemplary embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be embodied in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the present disclosure to those skilled in the art. It is noted that the embodiments of the present disclosure and the features in the embodiments can be combined with each other without conflict, provided that there is no conflict. The present disclosure will be described in detail below with reference to the drawings and in conjunction with the embodiments.

[0032] Figure 1 is a flow chart of the complex background adaptive target detection method based on space-time domain fusion provided by the embodiment of the application. As shown in Figure 1 , the method comprises the following steps:

[0033] (1) performing radiation uniformity correction on the original infrared sequence image to obtain a corrected infrared sequence image

[0034] (2) performing sub-pixel matching on the corrected infrared sequence image according to local features such as cloud edge, sea island, and coastline to obtain n precisely registered images, denoted as I1, I2, …In n .

[0035] (3) performing difference on adjacent frames of the precisely registered images I1, I2, …In n to obtain a difference image I t = I t+1 -I twhere t = 1, 2, …, n-1 represents the image frame number.

[0036] (4) Calculate the difference image I t The eight-neighborhood gray level gradient of each pixel in the image I t is denoted as D(i, j), i and j are the row and column numbers of the image I t respectively, and their value ranges are i = 1, 2, …, M and j = 1, 2, …, N respectively.

[0037] (5) Traverse all pixels in the image I t (t = 1, 2, …, n-1) and calculate the minimum relative gradient of the eight-neighborhood of the pixel I t (i, j) where i = 1, 2, …, M and j = 1, 2, …, N.

[0038] (6) Replace the gray level of the image I t according to the minimum relative gradient D min (i, j) of the eight-neighborhood of the pixel I t (i, j), to obtain the image I f filtered by the minimum relative gray level gradient.

[0039] (7) Convolve the filtered image I f with the template W = [1 1 1 1; 1 4 4 1; 1 4 4 1; 1 1 1 1] to replace the pixel DN value of the target original image with the convolution result, to obtain the image I c = I f * W after neighborhood correlation target enhancement processing.

[0040] (8) Traverse the image pixel I c (i, j) after neighborhood correlation target enhancement processing using the double-window constant false alarm rate detection method to calculate the adaptive detection threshold Tr(i, j).

[0041] (9) Use the adaptive detection threshold Tr(i, j) to perform target detection and traverse the image I c , when I c (i, j) > Tr(i, j), determine that I c (i, j) is an initial target; when I c (i, j) ≤ Tr(i, j), determine that it is background.

[0042] (10) According to the morphological characteristics of the target, screen the initial targets in the image I c , and obtain the candidate target set t from the difference image I where t = 1, 2, …, n-1 represents the image frame number after difference, h t represents the image I t The number of candidate targets.

[0043] (11) The candidate target set generated by the sequence frame image after difference is further screened based on the principle of neighborhood decision method, and the screened target points are denoted as A1, A2…A nll , and the vectors constructed by the difference image frame numbers where the feature points exist are denoted as a1, a2…a nll , where nll is the number of retained target points.

[0044] (12) For the screened target points A1, A2…A nll , false alarm elimination is performed according to the inter-frame target similarity, and the final target points are denoted as B1, B2…B all , where all is the number of final determined target points.

[0045] In step (4), the calculation method of the eight-neighborhood gray gradient of the pixel I t (i, j) is as follows:

[0046]

[0047] where i and j are the row and column numbers of the image I t , D(i, j) represents the eight-neighborhood gray gradient of the pixel I t (i, j), and I t (i+w, j+v) represents the pixel DN value of the image I t at the coordinate position (i+w, j+v), where w=-1, 0, 1, v=-1, 0, 1, and w and v are intermediate parameters.

[0048] D1(i, j), (i, j), D2(i, j), D4(i, j), D5(i, j), D6(i, j), D7(i, j), and D8(i, j) are the first neighborhood gray gradient of each pixel, the second neighborhood gray gradient of each pixel, the third neighborhood gray gradient of each pixel, the fourth neighborhood gray gradient of each pixel, the fifth neighborhood gray gradient of each pixel, the sixth neighborhood gray gradient of each pixel, the seventh neighborhood gray gradient of each pixel, and the eighth neighborhood gray gradient of each pixel, respectively.

[0049] D1(i, j), D2(i, j), D2(i, j), D4(i, j), D5(i, j), D6(i, j), D7(i, j), and D8(i, j) are in the neighborhood corresponding to the pixel I t (i, j) as follows:

[0050] [D1(i, j)] [D2(i, j)] D3(i, j) D4(i, j) I t (i, j) D5(i, j) [D6(i, j)] [D7(i, j)] [D8(i, j)]

[0051] In step (6), the minimum relative gray gradient filtered image I f is obtained by the following equation:

[0052]

[0053] where I f (i,j) is the filtered image I f , DN(i,j) is the DN value at the coordinate position (i,j), TH is the gradient decision threshold, and TH=0.5, D min (i,j) is the minimum eight-neighborhood gray gradient, (w,v|D min (i,j)) represents the eight-neighborhood minimum relative gradient direction corresponding pixel offset position, I t (i+w,j+v) is the DN value at the coordinate position (i+w,j+v) of the eight-neighborhood minimum relative gradient direction of I t (i,j). The equation indicates that when the minimum gray gradient D min (i,j) is less than or equal to the threshold value TH, I t (i,j) remains unchanged; when the minimum gray gradient D min (i,j) is greater than the threshold value TH, the DN value I t (i+w,j+v) of the minimum relative gray gradient direction corresponding pixel is used to replace I t (i,j). All pixels i=1,2,…,M,j=1,2,…,N are traversed to obtain the minimum relative gray gradient filtered image I f .

[0054] In step (8), the adaptive detection threshold calculation process is as follows:

[0055] 1) Taking the neighborhood correlation target enhanced processed image I c as the center, a double sliding window is set, and the sizes of the inner window W A and the outer window W B are 3x3 and 13x13, respectively.

[0056] 2) Each pixel I c (i,j) in the neighborhood correlation target enhanced processed image I c is traversed, and the adaptive detection threshold in the double sliding window of the pixel I -1 (i,j) is calculated as:

[0057] Tr(i,j)=m(i,j)-Φ f (P c )σ(i,j)

[0058] where Tr(i,j) is the pixel Ic adaptive detection threshold of (i,j), P f is false alarm rate, take P f = 10 -3 , m(i,j), σ(i,j) are I c (i,j) decentering window (i.e. meet I c (i,j) ∈ W B and ) mean and standard deviation of all pixels within.

[0059] In step (10), the initial target in the image I c after the neighborhood correlation target enhancement processing is screened according to the target morphological characteristics to obtain a candidate target set, including:

[0060] 1) Perform morphological operation on the image I c :

[0061]

[0062] Wherein, I mod is the morphological filtered image, H = [1 1; 1 1] is the structure element;

[0063] 2) Count the connected region area of the initial target corresponding position of the morphological filtered image I mod , according to the target characteristics, retain the target with an area greater than 3 pixels and less than 9 pixels, to obtain the morphological screened candidate target set.

[0064] In step (11), the further screening process of the candidate target set is as follows:

[0065] (111) According to the target speed setting, the neighborhood size is set to 5 pixels.

[0066] (112) Select the candidate target set with the largest number, i.e. h k = max(h t |t = 1,2,…,n-1) corresponding difference image I k (1≤k≤n-1), determine whether there is a target point in the neighborhood centered at the candidate target set position in the difference image I , if there is, record the position of the target in the difference image I k+1 , turn to determine whether there is a target point in the neighborhood of the image target point position in the difference image I , and so on; if not, expand the neighborhood range by one time, turn to determine whether there is a target point in the neighborhood of the image I k target point position in the difference image I .

[0067] (113) Regarding the above-mentioned transition determination in the difference image Image I k If the target point exists in the neighborhood of its location, record the location of that target point in the difference image. If the position in the image is not found, the neighborhood range is doubled before proceeding to the determination in the difference image. Image I k Does the target point exist in the neighborhood of the target point's location?

[0068] (114) For the step (113) of the transfer to the difference image, Image I k If the target point exists in the neighborhood of its location, record the target in the difference image. The position in the image is then determined in the difference image. medium image Does the target point exist in the neighborhood of its location? If not, remove the target point.

[0069] (115) Using the difference image I k As the starting frame, according to the different cases in steps (112), (113), and (114), respectively, [the following steps are performed] to [the following frames]. Calculate in both directions sequentially until n-1 frames of difference image calculation are completed, and record the retained target points A1, A2...A nll The vectors constructed from the differential image frame numbers are denoted as a1, a2, ... a1. nll Their vector lengths are l1, l2…l nll Then the filtered target points A1, A2...A nll The frames that appear are a1, a2...a nll The corresponding set of difference images is Where nll is the final number of target points retained.

[0070] In step (12), the inter-frame target similarity false alarm removal process is as follows:

[0071] (121) For target point A c Select the target point in frame number a c That is, difference image The target region is a window of size α = 3×3, denoted as [missing information]. Where c represents the target point number, satisfying 1≤c≤null.

[0072] (122) Calculate the difference image of the above with respect to the target point A. c Correlation coefficient:

[0073]

[0074] wherein, r c (x, y) is the correlation coefficient of the target point A c in the image x, y two frames, x, y is the difference image containing the target point A c , here the adjacent two frames are taken in turn Ar x , Ar y are the DN values of the target point A c target area in the x, y difference image respectively, Ar y are the DN average values of the target point A c target area in the x, y difference image respectively.

[0075] (123) Let c = 1, 2… null, repeat steps (121), (122), if there is 0.2 < r c (x, y) < 0.6, it is considered that A c is a false target, which is rejected. The finally determined target point is recorded as B1, B2… B all , wherein all is the number of finally determined target points.

[0076] The embodiment also provides a complex background adaptive target detection module based on space-time domain fusion, comprising:

[0077] A first module configured to perform radiation uniformity correction on an original infrared sequence image to obtain a corrected infrared sequence image;

[0078] A second module configured to perform sub-pixel matching on the corrected infrared sequence image to obtain n precision registration images; wherein n is greater than or equal to 2 and n is a positive integer;

[0079] A third module configured to perform difference on adjacent two precision registration images in the n precision registration images to obtain a difference image;

[0080] A fourth module configured to calculate eight-neighborhood gray scale gradients of each pixel in the difference image, and filter out the smallest eight-neighborhood gray scale gradient from the eight-neighborhood gray scale gradients of each pixel in the difference image;

[0081] A fifth module configured to replace the gray scale values of the difference image according to the smallest eight-neighborhood gray scale gradient to obtain a minimum relative gray scale gradient filtered image;

[0082] A sixth module configured to convolve the minimum relative gray scale gradient filtered image with a preset template to obtain a neighborhood correlation target enhanced processed image;

[0083] The seventh module is used for traversing the pixels of the image processed by the neighborhood correlation target enhancement method to obtain an adaptive detection threshold by using a double-window constant false alarm rate detection method, comparing the adaptive detection threshold with the pixels of the image processed by the neighborhood correlation target enhancement method, if the pixel of the image processed by the neighborhood correlation target enhancement method is greater than the adaptive detection threshold, the pixel of the image processed by the neighborhood correlation target enhancement method is an initial target, and if the pixel of the image processed by the neighborhood correlation target enhancement method is not greater than the adaptive detection threshold, the pixel of the image processed by the neighborhood correlation target enhancement method is background.

[0084] The eighth module is used for screening the initial targets in the image processed by the neighborhood correlation target enhancement method according to target morphological characteristics to obtain a candidate target set.

[0085] The ninth module is used for screening the candidate target set by using a neighborhood decision method principle to obtain screened target points.

[0086] The tenth module is used for performing false alarm elimination on the screened target points according to interframe target similarity to obtain final target points.

[0087] The application firstly performs background registration based on sequence image information, performs differential processing on the registered image, eliminates large-area background information, avoids the problems of poor applicability of a background prediction model, complex model and the like, performs background target integration processing on the differential image, achieves the purpose of noise reduction and target enhancement, and finally combines target space-time domain fusion features to perform false alarm elimination processing on the constant false alarm rate detection result, thereby improving the target detection accuracy.

[0088] Although the application has been disclosed as above with reference to the preferred embodiments, it is not intended to limit the application, and any person skilled in the art can make possible changes and modifications to the technical solutions of the application by using the disclosed methods and technical contents without departing from the spirit and scope of the application, therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the application, which does not depart from the content of the technical solutions of the application, belongs to the protection scope of the technical solutions of the application.

Claims

1. A complex background adaptive target detection method based on space-time domain fusion, characterized in that The method comprises the following steps: radiometric uniformity correction is performed on the original infrared sequence image to obtain a corrected infrared sequence image; Sub-pixel matching is performed on the corrected infrared sequence image to obtain n precision registration images; wherein n≥2 and n is a positive integer; to differencing the two adjacent fine-registered images in the fine-registered image sequence to obtain a difference image; eight-neighbor gray scale gradients of each pixel in the difference image are calculated, and the minimum eight-neighbor gray scale gradient is selected from the eight-neighbor gray scale gradients of each pixel in the difference image; gray value replacement is performed on the difference image according to the minimum eight-neighbor gray scale gradient to obtain a minimum relative gray scale gradient filtered image; convolution is performed on the minimum relative gray scale gradient filtered image and a preset template to obtain a neighborhood correlation target enhanced image; adaptive detection threshold is obtained by traversing pixels of the neighborhood correlation target enhanced image, and the adaptive detection threshold is compared with the pixels of the neighborhood correlation target enhanced image; if the pixel of the neighborhood correlation target enhanced image is greater than the adaptive detection threshold, the pixel of the neighborhood correlation target enhanced image is an initial target; if the pixel of the neighborhood correlation target enhanced image is not greater than the adaptive detection threshold, the pixel of the neighborhood correlation target enhanced image is background; candidate target set is obtained by screening the initial target in the neighborhood correlation target enhanced image according to target morphological characteristics; screened target points are obtained by screening the candidate target set; final target points are obtained by false alarm elimination according to inter-frame target similarity.

2. The method according to claim 1, wherein the method is characterized by: The adaptive detection threshold is obtained by traversing the pixels of the neighborhood correlation target enhanced image using a double-window constant false alarm rate detection method. The screened target points are obtained by screening the candidate target set using a neighborhood decision method.

3. The method of claim 1, wherein the method is based on a spatio-temporal fusion based complex background adaptive target detection method. The eight-neighbor gray scale gradient of each pixel in the difference image is obtained by the following formula: wherein are respectively the row, column numbers of the difference image represents a pixel of the difference image eight-neighborhood gray level gradient represents a pixel of the difference image at coordinate position pixel DN value are respectively a first neighborhood gray level gradient of each pixel, a second neighborhood gray level gradient of each pixel, a third neighborhood gray level gradient of each pixel, a fourth neighborhood gray level gradient of each pixel, a fifth neighborhood gray level gradient of each pixel, a sixth neighborhood gray level gradient of each pixel, a seventh neighborhood gray level gradient of each pixel, an eighth neighborhood gray level gradient of each pixel, wherein , and are intermediate parameters.​ 4. The method according to claim 3, wherein the method is characterized by: The minimum eight-neighbor gray scale gradient is obtained by the following formula: ； wherein, is the smallest eight-neighborhood gray level gradient, is the pixel of the difference image.

5. The method of claim 1, wherein the method is based on a spatio-temporal fusion based complex background adaptive target detection method. Minimum relative gray scale gradient filtered image This is obtained by the equation: ； wherein, is the filtered image with minimum relative gray level gradient DN value at the position, is the gradient decision threshold, is the minimum eight-neighborhood gray level gradient, is the pixel DN value of the difference image, denotes the difference image at the coordinate position pixel DN value, denotes the minimum eight-neighborhood gray level gradient direction corresponding pixel offset position, wherein, , and is the intermediate parameter.

6. The complex background adaptive target detection method based on space-time domain fusion according to claim 1, characterized in that: The neighborhood correlation target enhanced image is: ； wherein, is the image after the neighborhood-related target enhancement processing, is the image after the minimum relative gray gradient filtering, is the preset template.

7. The method of claim 1, wherein the method further comprises: Adaptive detection threshold By the following equation: ； wherein, is an adaptive detection threshold, is a false alarm rate, , are, respectively, the image pixels of the neighborhood of the target enhanced image of all pixels in the annulus window.

8. The complex background adaptive target detection method based on space-time domain fusion according to claim 1, characterized in that: Images after neighborhood correlation target enhancement processing The candidate target set is obtained according to target morphological features Image after neighborhood correlation target enhancement processing A morphological operation is performed to obtain a morphological filtered image. Statistical morphological filtering image and neighborhood related target enhancement processing image The initial target corresponding position of the connected area in the image after the statistical morphological filtering and the neighborhood related target enhancement processing is determined. According to the target characteristics, the target with an area greater than 3 pixels and less than 9 pixels is retained to obtain a candidate target set.

9. The method according to claim 8, wherein the method is characterized by: The morphologically filtered image is obtained by the following formula: ； wherein is the morphologically filtered image, is the structuring element.

10. A complex background adaptive target detection system based on space-time domain fusion, characterized in that The method comprises the following steps: The first module is configured to perform radiometric uniformity correction on the original infrared sequence image to obtain a corrected infrared sequence image. The second module is configured to perform sub-pixel matching on the corrected infrared sequence image to obtain a n precision registration images; wherein n≥2 and n is a positive integer. The third module is configured to The difference between two adjacent fine registration images in the fine registration image is differentiated to obtain a difference image. The fourth module is configured to calculate eight-neighbor gray scale gradients of each pixel in the difference image, and select the minimum eight-neighbor gray scale gradient from the eight-neighbor gray scale gradients of each pixel in the difference image. The fifth module is configured to perform gray value replacement on the difference image according to the minimum eight-neighbor gray scale gradient to obtain a minimum relative gray scale gradient filtered image. The sixth module is configured to perform convolution on the minimum relative gray scale gradient filtered image and a preset template to obtain a neighborhood correlation target enhanced image. The seventh module is configured to traverse pixels of the neighborhood correlation target enhanced image to obtain an adaptive detection threshold, and compare the adaptive detection threshold with the pixels of the neighborhood correlation target enhanced image; if the pixel of the neighborhood correlation target enhanced image is greater than the adaptive detection threshold, the pixel of the neighborhood correlation target enhanced image is an initial target; if the pixel of the neighborhood correlation target enhanced image is not greater than the adaptive detection threshold, the pixel of the neighborhood correlation target enhanced image is background. The eighth module is configured to screen the initial target in the neighborhood correlation target enhanced image according to target morphological characteristics to obtain a candidate target set. The ninth module is configured to screen the candidate target set to obtain screened target points. The tenth module is configured to perform false alarm elimination on the screened target points according to inter-frame target similarity to obtain final target points. The adaptive detection threshold is obtained by traversing the pixels of the neighborhood correlation target enhanced image using a double-window constant false alarm rate detection method. The screened target points are obtained by screening the candidate target set using a neighborhood decision method. The eight-neighbor gray scale gradient of each pixel in the difference image is obtained by the following formula: The minimum eight-neighbor gray scale gradient is obtained by the following formula: The neighborhood correlation target enhanced image is: The morphologically filtered image is obtained by the following formula: The ninth module is configured to screen the candidate target set to obtain screened target points. The tenth module is configured to eliminate false alarms from the screened target points according to inter-frame target similarity to obtain final target points.

Images