Rapid scene matching method for large-range search area

Abstract

The invention belongs to the technical field of scene matching, and particularly relates to a rapid scene matching method for a large-range search area, which greatly improves the scene matching speed during large-range search and solves the problem of high time consumption in initial positioning of current scene matching. According to the method, a four-dimensional descriptor is firstly designed, so that the robustness of a back-end matching strategy can be greatly improved; furthermore, on the basis of the multi-dimensional reference descriptor and the multi-dimensional real-time descriptor, a convolution coarse matching and fine matching strategy is designed, firstly, the multi-dimensional reference descriptor is segmented to perform software acceleration, coarse matching is performed on the obtained nine regions by using convolution, and fine matching is performed on the obtained nine regions; cutting an image with the same size as the real-time image on the original reference image according to the matching position for accurate matching, and finally designing a matching judgment rule to judge a matching result. According to the method, a mode of combining segmentation rough matching and fine matching is adopted for a large-range search region, so that the matching speed of the large-range region is greatly improved.

Description

Technical Field

[0001] This invention belongs to the field of scene matching technology, specifically relating to a fast scene matching method for large-scale search areas. Background Technology

[0002] Autonomous navigation technology has developed rapidly in recent years. Currently, smart munitions all have autonomous navigation requirements in satellite-denied environments. Visual scene matching navigation, with its advantages of simplicity, high reliability, low cost, and high positioning accuracy, is gradually becoming the mainstream technology for autonomous navigation. However, visual scene matching navigation is not yet widely used, primarily due to real-time requirements. Furthermore, the larger the unknown search area, the higher the matching time, especially when there are no satellites available throughout the entire process. In such cases, due to significant initial positioning errors, typically on the order of 10 km on one side, the time required to achieve the initial match using conventional scene matching methods would be catastrophic.

[0003] Therefore, how to design a fast scene matching algorithm for a large search area is an urgent problem to be solved for the large-scale application of scene matching. Summary of the Invention

[0004] (a) Technical problems to be solved The technical problem to be solved by this invention is: how to design a fast scene matching algorithm for a large search area, so that the initial matching time within a 10km search range on one side is less than 1 second, so as to meet the real-time requirements of intelligent munition autonomous navigation.

[0005] (II) Technical Solution To address the aforementioned technical problems, this invention provides a fast scene matching method for large-scale search areas, the method comprising the following steps: Step A: Input a large-scale baseline map and real-time sensing images; Step B: Extract the multidimensional benchmark descriptor DesRef from the large-scale benchmark map; Step C: Segment the multidimensional baseline descriptor; Step D: Extract multi-dimensional real-time descriptors (DesSen) from real-time perceived images; Step E: Implement the convolution of DesRef1 and DesSen using the same convolution method. Step F: Based on the result in step E, take (x1, y1) as the pixel center, and crop an image of the same size as ImgSen from ImgRef, denoted as ImgRef1; Step G: Repeat steps E and F to obtain a series of matching results at the pixel coordinates (x2, y2), ..., (x9, y9) of DesRef and the same size graphics ImgRef2, ..., ImgRef9 that are cropped from ImgRef and are the same size as ImgSen; Step H: Match ImgRef1 with ImgSen; Step I: Repeat step H, and use the HOPC method to match ImgRef2, ..., ImgRef9 with ImgSen in sequence, and count the matching result flags as vector res=[res1, res2, ..., res9], and count the matching positions as vector pos=[(r1, c1), (r2, c2), ..., (r9, c9)]; Step J: Judge the matching result in Step I. If the number of 1s in res is greater than 1, the matching is determined to be a failure. If the number of 1s in res is equal to 1, the matching is determined to be a success, and the pos result corresponding to the flag bit is the matching result. If the number of 1s in res is equal to 0, the matching is determined to be a failure.

[0006] In step A: Input a large-scale reference image ImgRef and a real-time sensing image ImgSen. Both are grayscale images with dimensions of M*N*1 and m*n*1, respectively. They are required to have no scale error, no perspective transformation difference, and the size of ImgRef is larger than that of ImgSen.

[0007] Step B involves extracting a multidimensional benchmark descriptor (DesRef) from a large-scale benchmark map, using the following extraction rules: (1) In this equation, e on the right side represents the natural logarithm, x on the left side represents the coordinates of each pixel in the image, and d represents the vector consisting of pixels in the 1-pixel neighborhood with x as the origin. Specifically: (2) In equation (1), D is calculated as follows: (3) K represents a 3x3 image centered at x, where each pixel undergoes the above calculation to produce 4-dimensional data. The results of each pixel's calculation are then concatenated to obtain the final image. At this point, DesRef is three-dimensional data with a size of M*N*4.

[0008] In step C, the multidimensional baseline descriptor DesRef is divided into DesRef1, DesRef2, DesRef3, ..., DesRef9. The division method is equal division with mutual overlap. The size of the overlapping area is at least larger than the size of ImgSen. The purpose of this step is to divide it into small blocks for acceleration, and at the same time, it is convenient to introduce the multi-process acceleration method of CPU calculator. The size after division is denoted as Mt*Nt*4.

[0009] In step D, the multidimensional real-time descriptor DesSen is extracted from the real-time perceived image. The method is the same as in step B, and its size is m*n*4.

[0010] Specifically, step E involves performing a convolution between DesRef1 and DesSen using the same convolution method, as detailed below: At this point, the size of M is Mt*Nt*1; the same convolution method is the same as the same convolution method in MATLAB tools; Furthermore, the coordinates corresponding to the maximum value of M are taken as the matching position, and the matching result can be obtained in advance. The matching position is the pixel coordinates of the center of DesSen on DesRef1. Further, based on the pixel coordinates of DesRef1 on DesRef in the segmentation stage in step C, the pixel coordinates (x1, y1) of the matching result on DesRef are calculated.

[0011] In step G: Repeat step E to match DesRef2, DesRef3, ..., DesRef9 with DesSen and calculate the pixel coordinates (x2, y2), ..., (x9, y9) of the matching result on DesRef; further repeat step F to extract the same size graphic as ImgSen on ImgRef based on the pixel coordinates (x2, y2), ..., (x9, y9), and denote them as ImgRef2, ..., ImgRef9 respectively.

[0012] In step H: At this point, ImgRef1, ..., ImgRef9 and ImgSen are all the same size, and as can be seen from step A, there is no scale error or perspective transformation difference between them. At this point, the HOPC method is used to match ImgRef1 and ImgSen. In the HOPC method used, the number of control points is set to 100. The matching is considered successful if the number of successfully matched control points is greater than 40. The matching result is marked as res1. If successful, it is marked as 1, and the matching result is recorded, that is, the pixel coordinates of the center of ImgSen on ImgRef1. The pixel coordinates on ImgRef1 are further calculated and marked as (r1, c1). If the matching fails, res1 is marked as 0.

[0013] In step I: Repeat step H, and use the HOPC method to match ImgRef2, ..., ImgRef9 with ImgSen in sequence, and count the matching result flags as vectors res=[res1, res2, ..., res9], and count the matching positions as vectors pos=[(r1, c1), (r2, c2), ..., (r9, c9)].

[0014] In step H, the HOPC method is used, but it can be replaced with other methods. As long as the process described in this invention is used, no matter what matching method is used in step H, it is within the scope of the claims of this invention.

[0015] (III) Beneficial Effects Compared with the prior art, the present invention has the following advantages: (1) This invention provides a fast scene matching algorithm for a large search area, which greatly improves the speed of scene matching in a large search area and provides a fast method for initial positioning of scene matching.

[0016] (2) Numerous experimental results show that the algorithm has a high running speed. When the real-time image is 400*400 pixels and the reference image is 10000*10000 pixels, the matching speed reaches 0.54s, which meets the needs of military navigation. Attached Figure Description

[0017] Figure 1 This is an overall flowchart of the present invention; Figure 2 This is an example of the multidimensional descriptor computation process. Detailed Implementation

[0018] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0019] To address the aforementioned technical problems, this invention provides a fast scene matching method for large-scale search areas, the method comprising the following steps: Step A: Input a large-scale baseline map and real-time sensing images; Step B: Extract the multidimensional benchmark descriptor DesRef from the large-scale benchmark map; Step C: Segment the multidimensional baseline descriptor; Step D: Extract multi-dimensional real-time descriptors (DesSen) from real-time perceived images; Step E: Implement the convolution of DesRef1 and DesSen using the same convolution method. Step F: Based on the result in step E, take (x1, y1) as the pixel center, and crop an image of the same size as ImgSen from ImgRef, denoted as ImgRef1; Step G: Repeat steps E and F to obtain a series of matching results at the pixel coordinates (x2, y2), ..., (x9, y9) of DesRef and the same size graphics ImgRef2, ..., ImgRef9 that are cropped from ImgRef and are the same size as ImgSen; Step H: Match ImgRef1 with ImgSen; Step I: Repeat step H, and use the HOPC method to match ImgRef2, ..., ImgRef9 with ImgSen in sequence, and count the matching result flags as vector res=[res1, res2, ..., res9], and count the matching positions as vector pos=[(r1, c1), (r2, c2), ..., (r9, c9)]; Step J: Judge the matching result in Step I. If the number of 1s in res is greater than 1, the matching is determined to be a failure. If the number of 1s in res is equal to 1, the matching is determined to be a success, and the pos result corresponding to the flag bit is the matching result. If the number of 1s in res is equal to 0, the matching is determined to be a failure.

[0020] In step A: Input a large-scale reference image ImgRef and a real-time sensing image ImgSen. Both are grayscale images with dimensions of M*N*1 and m*n*1, respectively. They are required to have no scale error, no perspective transformation difference, and the size of ImgRef is larger than that of ImgSen.

[0021] Step B involves extracting a multidimensional benchmark descriptor (DesRef) from a large-scale benchmark map, using the following extraction rules: (1) In this equation, e on the right side represents the natural logarithm, x on the left side represents the coordinates of each pixel in the image, and d represents the vector consisting of pixels in the 1-pixel neighborhood with x as the origin. Specifically: (2) In equation (1), D is calculated as follows: (3) K represents a 3x3 image centered at x, where each pixel undergoes the above calculation to produce 4-dimensional data. The results of each pixel's calculation are then concatenated to obtain the final image. At this point, DesRef is three-dimensional data with a size of M*N*4.

[0022] In step C, the multidimensional baseline descriptor DesRef is divided into DesRef1, DesRef2, DesRef3, ..., DesRef9. The division method is equal division with mutual overlap. The size of the overlapping area is at least larger than the size of ImgSen. The purpose of this step is to divide it into small blocks for acceleration, and at the same time, it is convenient to introduce the multi-process acceleration method of CPU calculator. The size after division is denoted as Mt*Nt*4.

[0023] In step D, the multidimensional real-time descriptor DesSen is extracted from the real-time perceived image. The method is the same as in step B, and its size is m*n*4.

[0024] Specifically, step E involves performing a convolution between DesRef1 and DesSen using the same convolution method, as detailed below: At this point, the size of M is Mt*Nt*1; the same convolution method is the same as the same convolution method in MATLAB tools; Furthermore, the coordinates corresponding to the maximum value of M are taken as the matching position, and the matching result can be obtained in advance. The matching position is the pixel coordinates of the center of DesSen on DesRef1. Further, based on the pixel coordinates of DesRef1 on DesRef in the segmentation stage in step C, the pixel coordinates (x1, y1) of the matching result on DesRef are calculated.

[0025] In step G: Repeat step E to match DesRef2, DesRef3, ..., DesRef9 with DesSen and calculate the pixel coordinates (x2, y2), ..., (x9, y9) of the matching result on DesRef; further repeat step F to extract the same size graphic as ImgSen on ImgRef based on the pixel coordinates (x2, y2), ..., (x9, y9), and denote them as ImgRef2, ..., ImgRef9 respectively.

[0026] In step H: At this point, ImgRef1, ..., ImgRef9 and ImgSen are all the same size, and as can be seen from step A, there is no scale error or perspective transformation difference between them. At this point, the HOPC method is used to match ImgRef1 and ImgSen. In the HOPC method used, the number of control points is set to 100. The matching is considered successful if the number of successfully matched control points is greater than 40. The matching result is marked as res1. If successful, it is marked as 1, and the matching result is recorded, that is, the pixel coordinates of the center of ImgSen on ImgRef1. The pixel coordinates on ImgRef1 are further calculated and marked as (r1, c1). If the matching fails, res1 is marked as 0.

[0027] In step I: Repeat step H, and use the HOPC method to match ImgRef2, ..., ImgRef9 with ImgSen in sequence, and count the matching result flags as vectors res=[res1, res2, ..., res9], and count the matching positions as vectors pos=[(r1, c1), (r2, c2), ..., (r9, c9)].

[0028] In step H, the HOPC method is used, but it can be replaced with other methods. As long as the process described in this invention is used, no matter what matching method is used in step H, it is within the scope of the claims of this invention.

[0029] Example 1 The specific steps in this embodiment are as follows: A. Input a large-scale reference image ImgRef and a real-time sensing image ImgSen. Both are grayscale images with dimensions M*N*1 and m*n*1 respectively. They are required to have no scale error, no perspective transformation difference, and ImgRef is larger than ImgSen. Considering that there are currently various methods for scale correction and perspective transformation correction of the original image, and that scale error and perspective transformation difference are not the endpoint of this invention and scene matching method, the above requirements proposed in this invention are reasonable. B. Extracting multidimensional baseline descriptors (DesRef) from large-scale baseline maps, with the following extraction rules: (1) In this equation, e on the right side represents the natural logarithm, x on the left side represents the coordinates of each pixel in the image, and d represents the vector consisting of pixels in the 1-pixel neighborhood with x as the origin. Specifically: (2) In equation (1), D is calculated as follows: (3) K represents a 3x3 image centered at x, where each pixel undergoes the above calculation to produce 4-dimensional data. The results of each pixel's calculation are then concatenated to obtain the final image. At this point, DesRef is three-dimensional data with dimensions M*N*4.

[0030] C. The multidimensional baseline descriptor DesRef is divided into DesRef1, DesRef2, DesRef3, ..., DesRef9. The division method is equal division and mutual overlap. The size of the overlapping area is at least larger than the size of ImgSen. The purpose of this step is to divide into small blocks for acceleration, and at the same time facilitate the introduction of multi-process acceleration methods of calculators such as CPU. In this invention, the size after division is denoted as Mt*Nt*4. D. Extract multidimensional real-time descriptors DesSen from real-time perceived images, referring to step B, where the size is m*n*4. E. Implement the convolution of DesRef1 and DesSen using the same convolution method, as follows: At this point, the size of M is Mt*Nt*1. The same convolution method is used in MATLAB.

[0031] Furthermore, the coordinates corresponding to the maximum value of M are taken as the matching position, and the matching result can be obtained in advance. The matching position is the pixel coordinates of the center of DesSen on DesRef1. Further, based on the pixel coordinates of DesRef1 on DesRef in the segmentation stage in step C, the pixel coordinates (x1, y1) of the matching result on DesRef are calculated. F. Based on the results in step E, with (x1, y1) as the pixel center, crop an image of the same size as ImgSen from ImgRef, and denote it as ImgRef1; G. Repeat step E to match DesRef2, DesRef3, ..., DesRef9 with DesSen and calculate the pixel coordinates (x2, y2), ..., (x9, y9) of the matching result on DesRef; further repeat step F to extract the same size graphic as ImgSen from ImgRef based on the pixel coordinates (x2, y2), ..., (x9, y9), and denote them as ImgRef2, ..., ImgRef9 respectively; H. At this point, ImgRef1, ..., ImgRef9 and ImgSen are all the same size, and as can be seen from step A, there is no scale error or perspective transformation difference between them. Therefore, the HOPC method is used to match ImgRef1 and ImgSen. In the HOPC method used in this invention, the number of control points is set to 100. A successful match is considered to have more than 40 successfully matched control points. The matching result is marked as res1. If successful, it is marked as 1, and the matching result, i.e., the pixel coordinates of the center of ImgSen on ImgRef1, is recorded. The pixel coordinates on ImgRef1 are further calculated and marked as (r1, c1). If the match fails, res1 is marked as 0. I. Repeat step H, using the HOPC method to match ImgRef2, ..., ImgRef9 with ImgSen in sequence, and count the matching result flags as vectors res=[res1, res2, ..., res9], and the matching positions as vectors pos=[(r1, c1), (r2, c2), ..., (r9, c9)]; J. Evaluate the matching results in I. If the number of 1s in res is greater than 1, the match is considered a failure. If the number of 1s in res is equal to 1, the match is considered a success, and the pos result corresponding to the flag is the matching result. If the number of 1s in res is equal to 0, the match is considered a failure.

[0032] Example 2 The flowchart of this embodiment is as follows: Figure 1 As shown, the specific steps are as follows: A. Input a large-scale reference image ImgRef and a real-time sensing image ImgSen. Both are grayscale images with dimensions M*N*1 and m*n*1 respectively. They are required to have no scale error, no perspective transformation difference, and ImgRef is larger than ImgSen. Considering that there are currently various methods for scale correction and perspective transformation correction of the original image, and that scale error and perspective transformation difference are not the endpoint of this invention and scene matching method, the above requirements proposed in this invention are reasonable. B. Extracting multidimensional baseline descriptors (DesRef) from large-scale baseline maps, with the following extraction rules: (1) In this equation, e on the right side represents the natural logarithm, x on the left side represents the coordinates of each pixel in the image, and d represents the vector consisting of pixels in the 1-pixel neighborhood with x as the origin. Specifically: (2) In equation (1), D is calculated as follows: (3) K represents a 3x3 image centered at x, where each pixel undergoes the above calculation to produce 4-dimensional data. The results of each pixel's calculation are then concatenated to obtain the final image. At this point, DesRef is three-dimensional data with dimensions M*N*4.

[0033] The specific calculation process is as follows: Figure 2 As shown, assuming the size of ImgRef is 5*5*1, at pixel x=(3,3), a 3*3 neighborhood is taken as follows. Figure 2 As shown above, this is the point at which point... When d = (-1, 0), like Figure 2 As shown in the upper left, the calculation process of formula (3) is as follows: = (95-63) 2 +(42-95) 2 +(12-42) 2 + (39-54) 2+(21-39) 2 +(16-21) 2 +(36-44) 2 +(27-36) 2 +(11-27) 2 Furthermore, when d = (1, 0), (0, 1), and (0, -1), like Figure 2 As shown in the upper right, lower left, and lower right diagrams, after using the same calculation method, we obtain a 4-dimensional D, and then obtain the 4-dimensional descriptor DesRef according to equation (1).

[0034] C. The multidimensional baseline descriptor DesRef is divided into DesRef1, DesRef2, DesRef3, ..., DesRef9. The division method is equal division and mutual overlap. The size of the overlapping area is at least larger than the size of ImgSen. The purpose of this step is to divide into small blocks for acceleration, and at the same time facilitate the introduction of multi-process acceleration methods of calculators such as CPU. In this invention, the size after division is denoted as Mt*Nt*4. D. Extract multidimensional real-time descriptors DesSen from real-time perceived images, referring to step B, where the size is m*n*4. E. Implement the convolution of DesRef1 and DesSen using the same convolution method, as follows: At this point, the size of M is Mt*Nt*1. The same convolution method is used in MATLAB.

[0035] Furthermore, the coordinates corresponding to the maximum value of M are taken as the matching position, and the matching result can be obtained in advance. The matching position is the pixel coordinates of the center of DesSen on DesRef1. Further, based on the pixel coordinates of DesRef1 on DesRef in the segmentation stage in step C, the pixel coordinates (x1, y1) of the matching result on DesRef are calculated. F. Based on the results in step E, with (x1, y1) as the pixel center, crop an image of the same size as ImgSen from ImgRef, and denote it as ImgRef1; G. Repeat step E to match DesRef2, DesRef3, ..., DesRef9 with DesSen and calculate the pixel coordinates (x2, y2), ..., (x9, y9) of the matching result on DesRef; further repeat step F to extract the same size graphic as ImgSen from ImgRef based on the pixel coordinates (x2, y2), ..., (x9, y9), and denote them as ImgRef2, ..., ImgRef9 respectively; H. At this point, ImgRef1, ..., ImgRef9 and ImgSen are all the same size, and as can be seen from step A, there is no scale error or perspective transformation difference between them. Therefore, the HOPC method is used to match ImgRef1 and ImgSen. In the HOPC method used in this invention, the number of control points is set to 100. A successful match is considered to have more than 40 successfully matched control points. The matching result is marked as res1. If successful, it is marked as 1, and the matching result, i.e., the pixel coordinates of the center of ImgSen on ImgRef1, is recorded. The pixel coordinates on ImgRef1 are further calculated and marked as (r1, c1). If the match fails, res1 is marked as 0. I. Repeat step H, using the HOPC method to match ImgRef2, ..., ImgRef9 with ImgSen in sequence, and count the matching result flags as vectors res=[res1, res2, ..., res9], and the matching positions as vectors pos=[(r1, c1), (r2, c2), ..., (r9, c9)]; J. Evaluate the matching results in I. If the number of 1s in res is greater than 1, the match is considered a failure. If the number of 1s in res is equal to 1, the match is considered a success, and the pos result corresponding to the flag is the matching result. If the number of 1s in res is equal to 0, the match is considered a failure.

[0036] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A fast scene matching method for a large search area, characterized in that, The method includes the following steps: Step A: Input a large-scale baseline map and real-time sensing images; Step B: Extract the multidimensional benchmark descriptor DesRef from the large-scale benchmark map; Step C: Segment the multidimensional baseline descriptor; Step D: Extract multi-dimensional real-time descriptors (DesSen) from real-time perceived images; Step E: Implement the convolution of DesRef1 and DesSen using the same convolution method. Step F: Based on the result in step E, take (x1, y1) as the pixel center, and crop an image of the same size as ImgSen from ImgRef, denoted as ImgRef1; Step G: Repeat steps E and F to obtain a series of matching results at the pixel coordinates (x2, y2), ..., (x9, y9) of DesRef and the same size graphics ImgRef2, ..., ImgRef9 that are cropped from ImgRef and are the same size as ImgSen; Step H: Match ImgRef1 with ImgSen; Step I: Repeat step H, and use the HOPC method to match ImgRef2, ..., ImgRef9 with ImgSen in sequence, and count the matching result flags as vector res=[res1, res2, ..., res9], and count the matching positions as vector pos=[(r1, c1), (r2, c2), ..., (r9, c9)]; Step J: Judge the matching result in Step I. If the number of 1s in res is greater than 1, the matching is considered to have failed; if the number of 1s in res is equal to 1, the matching is considered to have succeeded, and the pos result corresponding to the flag bit is the matching result. If the number of 1s in res is 0, then the match is considered to have failed.

2. The fast scene matching method for a large search area as described in claim 1, characterized in that, Step A: Input a large-scale reference image ImgRef and a real-time sensing image ImgSen. Both are grayscale images with dimensions of M*N*1 and m*n*1, respectively. They are required to have no scale error, no perspective transformation difference, and the size of ImgRef is larger than that of ImgSen.

3. The fast scene matching method for a large search area as described in claim 2, characterized in that, Step B: Extract multidimensional benchmark descriptors DesRef from the large-scale benchmark map, according to the following rules: （1） In this equation, e on the right side represents the natural logarithm, x on the left side represents the coordinates of each pixel in the image, and d represents the vector consisting of pixels in the 1-pixel neighborhood with x as the origin. Specifically: （2） In equation (1), D is calculated as follows: （3） K represents a 3x3 image centered at x, where each pixel undergoes the above calculation to produce 4-dimensional data. The results of each pixel's calculation are then concatenated to obtain the final image. At this point, DesRef is three-dimensional data with a size of M*N*4.

4. The fast scene matching method for a large search area as described in claim 3, characterized in that, Step C: The multidimensional baseline descriptor DesRef is divided into DesRef1, DesRef2, DesRef3, ..., DesRef9. The division method is equal division and mutual overlap. The size of the overlapping area is at least larger than the size of ImgSen. The purpose of this step is to divide it into small blocks for acceleration, and at the same time facilitate the introduction of multi-process acceleration methods of CPU calculator. The size after division is denoted as Mt*Nt*4.

5. The fast scene matching method for a large search area as described in claim 4, characterized in that, Step D: Extract the multidimensional real-time descriptor DesSen from the real-time perceived image. The method is the same as in step B, and its size is m*n*4.

6. The fast scene matching method for a large search area as described in claim 5, characterized in that, Step E: Implement the convolution of DesRef1 and DesSen using the same convolution method, as follows: At this point, the size of M is Mt*Nt*1; the same convolution method is the same as the same convolution method in MATLAB tools; Furthermore, the coordinates corresponding to the maximum value of M are taken as the matching position, and the matching result can be obtained in advance. The matching position is the pixel coordinates of the center of DesSen on DesRef1. Further, based on the pixel coordinates of DesRef1 on DesRef in the segmentation stage in step C, the pixel coordinates (x1, y1) of the matching result on DesRef are calculated.

7. The fast scene matching method for a large search area as described in claim 6, characterized in that, Step G: Repeat step E to match DesRef2, DesRef3, ..., DesRef9 with DesSen and calculate the pixel coordinates (x2, y2), ..., (x9, y9) of the matching result on DesRef; further repeat step F to extract the same size graphic as ImgSen on ImgRef based on the pixel coordinates (x2, y2), ..., (x9, y9), and denote them as ImgRef2, ..., ImgRef9 respectively.

8. The fast scene matching method for a large search area as described in claim 7, characterized in that, Step H: At this point, ImgRef1, ..., ImgRef9 and ImgSen are all the same size, and as can be seen from step A, there is no scale error or perspective transformation difference between them. At this point, the HOPC method is used to match ImgRef1 and ImgSen. In the HOPC method used, the number of control points is set to 100. The matching is considered successful if the number of successfully matched control points is greater than 40. The matching result is marked as res1. If successful, it is marked as 1, and the matching result is recorded, that is, the pixel coordinates of the center of ImgSen on ImgRef1. The pixel coordinates on ImgRef1 are further calculated and marked as (r1, c1). If the matching fails, res1 is marked as 0.

9. The fast scene matching method for a large search area as described in claim 8, characterized in that, Step I: Repeat step H, and use the HOPC method to match ImgRef2, ..., ImgRef9 with ImgSen in sequence, and count the matching result flags as vectors res=[res1, res2, ..., res9], and count the matching positions as vectors pos=[(r1, c1), (r2, c2), ..., (r9, c9)].

10. The fast scene matching method for a large search area as described in claim 1, characterized in that, The HOPC method is used in step H.

Images