WO2020206903A1 - 影像匹配方法、装置及计算机可读存储介质 - Google Patents

影像匹配方法、装置及计算机可读存储介质 Download PDF

Info

Publication number
WO2020206903A1
WO2020206903A1 PCT/CN2019/102187 CN2019102187W WO2020206903A1 WO 2020206903 A1 WO2020206903 A1 WO 2020206903A1 CN 2019102187 W CN2019102187 W CN 2019102187W WO 2020206903 A1 WO2020206903 A1 WO 2020206903A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
matching
image matching
epipolar
images
Prior art date
Application number
PCT/CN2019/102187
Other languages
English (en)
French (fr)
Inventor
王义文
王健宗
Original Assignee
平安科技(深圳)有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 平安科技(深圳)有限公司 filed Critical 平安科技(深圳)有限公司
Publication of WO2020206903A1 publication Critical patent/WO2020206903A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/22Matching criteria, e.g. proximity measures
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T17/00Three dimensional [3D] modelling, e.g. data description of 3D objects

Definitions

  • This application relates to the field of computer technology, and in particular to an image matching method, device and computer-readable storage medium.
  • Image matching refers to the process of identifying points with the same name between two or more images through a certain matching algorithm. It is an important preliminary step in image fusion, target recognition, target change detection, computer vision and other problems. It has a wide range of applications in many fields such as remote sensing, digital photogrammetry, computer vision, cartography and military applications. At present, the most effective and effective method for image matching is to judge the different points of the image by visual inspection; secondly, according to the principle of the essence of the image, which is the principle of pixels, compare the gray values of all pixels in the target area; or find the target image based on the principle of template matching The same or the most similar position to the sub-image in the search image, etc.
  • the present application provides an image matching method, device, and computer-readable storage medium, the main purpose of which is to provide a new image matching method applied to dense stereo scenes under aerial photography to improve image matching efficiency.
  • an image matching method provided by this application includes:
  • the image imaging map is generated, and the first image matching is performed on the image imaging map using the scale-invariant feature transformation method to generate the first image matching set;
  • a dense matching of all pixels between images is established, a third image matching set is generated, and three-dimensional reconstruction is performed to obtain a reconstructed scene image.
  • the present application also provides an image matching device.
  • the device includes a memory and a processor.
  • the memory stores an image matching program that can run on the processor.
  • the image matching program When executed by the processor, the steps of the image matching method described above are realized.
  • the present application also provides a computer-readable storage medium with an image matching program stored on the computer-readable storage medium, and the image matching program can be executed by one or more processors to achieve The steps of the image matching method described above.
  • the image matching method, device, and computer-readable storage medium proposed in this application generate an image imaging map based on the scene image shot by an aerial camera, and use the scale-invariant feature transformation method to perform the initial image matching on the image imaging map to generate the initial image matching set, Based on the first image matching set, generate epipolar images and calculate the degree of overlap between the epipolar images, complete the second image matching, and generate a second image matching set based on the second image matching set Establish dense matching of all pixels between images, generate a third image matching set, and perform 3D reconstruction to obtain a reconstructed scene image.
  • This application improves the efficiency of image matching, and can perform three-dimensional reconstruction of images of dense scenes under aerial photography, so as to more effectively help users to conduct analysis and research.
  • FIG. 1 is a schematic flowchart of an image matching method provided by an embodiment of this application.
  • FIG. 2 is a schematic diagram of the internal structure of an image matching device provided by an embodiment of the application.
  • Fig. 3 is a schematic diagram of modules of an image matching program in an image matching device provided by an embodiment of the application.
  • This application provides an image matching method.
  • FIG. 1 it is a schematic flowchart of an image matching method provided by an embodiment of this application.
  • the method can be executed by a device, and the device can be implemented by software and/or hardware.
  • the image matching method includes:
  • the scene images taken by aerial equipment such as drones, helicopters and other flight control systems, have a large number of images and a wide viewing angle, especially buildings, which are dense and dense. Therefore, this application first restores the overlapping image sets to their respective positions, and reconstructs the imaging map of the object.
  • the model formula for object imaging under aerial photography is used to recover overlapping scene images.
  • Position generate image imaging map.
  • model formula of the object imaging is as follows:
  • s is the scale factor
  • m is the coordinate of the image point
  • M is the coordinate of the object point (the object point and the image point are the object position and the image position in the optical imaging respectively)
  • K is the parameter matrix in the aerial photography tool, It is composed of focal length and principal point coordinates
  • R is a rotation matrix, which can be converted to approximate values according to the yaw, pitch, and roll recorded by the aerial tool’s system
  • C is the projection center
  • the position vector can be approximated directly from the longitude, latitude, and altitude recorded by the GPS of the aerial photography tool
  • I is the third-order unit matrix.
  • the imaging map of n images can be obtained.
  • G n (V n , E n), wherein, referred to as a vertex set V n, En is called an edge set (a graph is a widely used data structure.
  • the nodes in the graph are called vertices.
  • the relationship between two vertices can be represented by a pair, called an edge. If the graph represents an edge Even pairs are ordered, then the graph is called a directed graph, if the pairs representing edges are disordered, then it is called an undirected graph).
  • No set of edges E in the drawing represents the number nE E n-side is, for each table represents one image, on behalf of the E nE n-th image for subsequent image matching process performed between only one image pair nE . If the relationship between images is not considered and the exhaustive traversal strategy is used for image matching, the total number of matches is Usually n*(n-1)/2 will be much larger than nE.
  • the method of constructing the image relationship undirected graph limits the scope of image matching, can avoid blind image matching, reduce the total image matching calculation complexity from O(n 2 ) to O(n), and improve the matching calculation efficiency ; At the same time, it can effectively eliminate the interference of unrelated image pairs, fundamentally avoid mismatches caused by non-overlapping images, and improve the accuracy of matching and the robustness of reconstruction.
  • the scale-invariant feature transform (SIFT) algorithm is used for image matching.
  • SIFT scale-invariant feature transform
  • image matching if there are few matching points in the two images I i and I j , which are smaller than the threshold N 1 , it means that the overlap is small or the correlation is weak, and (I i , I j ) is removed from the set E. If the number of matching points in the two images I i and I j is greater than the threshold N 1 , then the imaging image pair is retained to generate There are n 1 E image pairs in total, and the first image matching set E 1 is generated.
  • S20 Based on the first image matching set, generate an epipolar image and calculate the degree of overlap between the epipolar images, complete the second image matching, and generate a second image matching set.
  • the epipolar image is a method of changing the search range from a two-dimensional plane image to a one-dimensional straight line during the matching process.
  • the plane formed by the shooting baseline and any ground point is called the nuclear surface
  • the intersection line between the nuclear surface and the image surface is called the nuclear line.
  • the image points with the same name must be on the epipolar line of the same name, and the image points on the epipolar line of the same name have a one-to-one correspondence.
  • an epipolar pair with the same name can be determined on a stereo image pair, then using the above-mentioned properties of the epipolar pair with the same name, the search and matching of the two-dimensional image can be transformed into the search and matching along the epipolar line.
  • the epipolar image eliminates the upper and lower parallax between the stereo images, narrows the search range, reduces the amount of matching calculations, and improves the matching accuracy, so it is of great significance for dense stereo image matching.
  • a preferred embodiment of the present application discloses a method for making and matching epipolar images for generating epipolar images and calculating the degree of overlap between the epipolar images.
  • the method includes: (a) using the SIFT algorithm to compare the After image pair point feature extraction, uniformly distributed high-precision points with the same name are obtained, the basic matrix estimation based on the RANSAC strategy is used to obtain the basic matrix of n 1 E image pairs; (b) using the basic matrix to determine each group of points with the same name Corresponding core line with the same name; (c) According to the principle that the core line must intersect at the core point, the least square method is used to determine The core point coordinates of the image pair are used to generate a quick mapping of the epipolar lines between the images according to the core point coordinates, and the epipolar line is resampled by bilinear interpolation along the epipolar line direction to complete the epipolar image production and matching regeneration There are a total of n 2 E image pairs to generate the second image matching set.
  • the epipolar image production method based on the basic matrix can avoid the problems of iterative calculation and initial value assignment when calculating the relative relationship, and it can also have good accuracy when the aerial photography uses a large angle of view.
  • a rotation matrix and a projection matrix are constructed, and the rotation matrix is divided into x, y, and z axes, each of which is among them, Is expressed as follows:
  • the least square method is used to determine the core point coordinates (x p , y p ) of the image from the calculation result of the projection matrix.
  • mapping is to directly correct the epipolar image on the central projection image.
  • the specific steps of the mapping are: derive the central projection according to the collinear condition equation, calculate the angle relationship between two adjacent epipolar lines when the epipolar lines are sampled, and complete the determination of each epipolar line on the central projection image; use The basic matrix generated above can determine each image pair The corresponding core lines are based on The nuclear point coordinates of each nuclear image in the image determine the epipolar line of the point, and complete the epipolar line correspondence between the same image pair to obtain The epipolar equation is:
  • (x p , y p ) are the coordinates of the core point calculated above, (x base , y base ) are the reference coordinates of the center projection image, and the same goes for The epipolar equation.
  • each image pair is established After the epipolar line mapping, according to the resampling rule of the bilinear interpolation method, the epipolar line image is generated and the overlap degree is calculated. The image pairs whose overlap degree of the epipolar line image is less than the threshold N 2 are discarded to generate There are n 2 E image pairs in total, and the second image matching set E 2 is obtained .
  • the two-dimensional image restores the geometric structure of the three-dimensional object surface, so further processing is required, and the following S30 is executed.
  • the preferred embodiment of the present application adopts a dense matching algorithm, and on the basis of the second image matching set E 2 , the corner detection algorithm of the least equivalent segmentation absorbing core is used to extract respectively
  • the corner points of the corner points form a set of matching points of the corner points; combined with the epipolar geometric constraints, dynamic programming algorithms and other methods to establish Dense matching of all pixels between images. Specific steps are as follows:
  • the second image matching set of corner points in an image in E 2 (a) detection is the second image matching set of corner points in an image in E 2 (a) detection.
  • the corner point is the point where the local curvature changes the most on the contour of the image. It contains important information in the image, so it is of great significance for the detail matching of the image in the image set E 2 .
  • the preferred embodiment of the present application adopts the smallest uni-value segment assimilating nucleus (SUSAN) method to detect image corners: Gaussian smoothing is performed on the input image; each pixel in the image is traversed, and Sobel is used first The operator (it is a discrete first-order difference operator, used to calculate the approximate value of a gradient of the image brightness function) to determine whether the pixel is an edge point; if it is an edge point, the loss function according to the global energy equation is the smallest To determine whether the path loss L r (p, d) of the corner point is minimized, the determination principle is as follows:
  • C(p,d) is the local loss of the path
  • L r ,min(pr) is the minimum loss of the previous step of the path, from which it can be determined whether the point is a corner point, and redundant corner points are removed; further judge if the two corners are detected If the points are adjacent, the corner points with larger L r (p,d) are removed.
  • the automatic matching of corner points can effectively distinguish the difference between image pairs based on the similarities and differences of the corner points, which is an effective means to achieve precise matching.
  • the corner point automatic matching can be divided into these steps: 1For Each point in the set of image corners is in Find matching corner points in the corresponding search area in the image, similarly, for Each point in the set of image corners is searched in the same way Find their corresponding matching points in the image, and call the intersection of these two matching point sets the initial matching point set K 1 ; 2In the initial matching point set K 1 , the pair is The corner point where the diagonal points of the image are concentrated with each other, find the matching point in the corresponding search area, calculate the similarity between this point and each candidate matching point in the search area, and select the candidate matching point with the greatest similarity as its match point.
  • the similarity calculation method adopts the gradient size similarity method: if the gradient size of a pixel is g, and the gradient of another pixel that matches it approximately obeys the normal distribution, then the two The similarity l g of each pixel is:
  • d(x) is the density function and k is the density coefficient.
  • the dense matching includes: 1 Obtain the matching point set K 2 according to the limit geometric constraint relationship
  • the polar line correspondence of the image is obtained, and the limit correspondence set K 3 is obtained .
  • the so-called limit geometric constraint relationship means that if l and l'are the two corresponding epipolar lines in the left and right images, the corresponding point of the point p on the epipolar line l in the left image in the right image must be on the epipolar line l' on.
  • the polar lines in K 3 are segmented according to gray levels. Each polar line is divided into several gray segments, and the gray values of pixels on each segment are similar. .
  • the formula for gray scale segmentation is as follows:
  • the physical meaning of the above formula is to divide the continuous points of gray value in a certain range into one section.
  • I (x t , y t ) is the gray value of the pixel (x t , y t );
  • w is the number of pixels on a gray segment, that is, the length of the gray segment;
  • T is For a certain threshold, the smaller the value, the fewer pixels are divided into a certain gray-scale segment, and the more gray-scale segments.
  • the matching effect is the best when T is set to 3.
  • the gray-level segment set of is K 4 ; 3Using dynamic programming algorithm (an optimization method to find the best matching path) to establish the correspondence between gray-level segments, and using linear interpolation to save the corresponding gray-level The corresponding relationship between the pixels is established between the segments, so that the dense matching of all the pixels between the images is realized, and the third image matching set E 3 is obtained .
  • the preset 3Dmax software can be used to reconstruct the scene to restore the three-dimensional geometry of the scene space Information to get the reconstructed image.
  • the application also provides an image matching device.
  • FIG. 2 it is a schematic diagram of the internal structure of an image matching device provided by an embodiment of this application.
  • the image matching device 1 may be a PC (Personal Computer, personal computer), or a terminal device such as a smart phone, a tablet computer, or a portable computer.
  • the image matching device 1 at least includes a memory 11, a processor 12, a communication bus 13, and a network interface 14.
  • the memory 11 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), magnetic memory, magnetic disk, optical disk, and the like.
  • the memory 11 may be an internal storage unit of the image matching device 1 in some embodiments, such as a hard disk of the image matching device 1.
  • the memory 11 may also be an external storage device of the image matching device 1, such as a plug-in hard disk, a smart media card (SMC), and a secure digital (Secure Digital, SD card, Flash Card, etc.
  • the memory 11 may also include both an internal storage unit of the image matching device 1 and an external storage device.
  • the memory 11 can be used not only to store application software and various data installed in the image matching device 1, such as the code of the image matching program 01, etc., but also to temporarily store data that has been output or will be output.
  • the processor 12 may be a central processing unit (CPU), controller, microcontroller, microprocessor or other data processing chip in some embodiments, and is used to run the program code or processing stored in the memory 11 Data, such as execution of image matching program 01, etc.
  • CPU central processing unit
  • controller microcontroller
  • microprocessor or other data processing chip in some embodiments, and is used to run the program code or processing stored in the memory 11 Data, such as execution of image matching program 01, etc.
  • the communication bus 13 is used to realize the connection and communication between these components.
  • the network interface 14 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface), and is usually used to establish a communication connection between the device 1 and other electronic devices.
  • the device 1 may also include a user interface.
  • the user interface may include a display (Display) and an input unit such as a keyboard (Keyboard).
  • the optional user interface may also include a standard wired interface and a wireless interface.
  • the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode, organic light emitting diode) touch device, etc.
  • the display can also be called a display screen or a display unit as appropriate, and is used to display the information processed in the image matching device 1 and to display a visualized user interface.
  • FIG. 2 only shows the image matching device 1 with components 11-14 and the image matching program 01.
  • FIG. 1 does not constitute a limitation on the image matching device 1, and may include Fewer or more components than shown, or some combination of components, or different component arrangement.
  • the image matching program 01 is stored in the memory 11; when the processor 12 executes the image matching program 01 stored in the memory 11, the following steps are implemented:
  • the first step is to generate an image imaging map according to the scene image taken by the aerial camera, and use the scale-invariant feature transformation method to perform the initial image matching on the image imaging map to generate the initial image matching set.
  • the scene images captured by aerial equipment such as drones, helicopters and other flight control systems, have a large number of images and a wide viewing angle, especially buildings, which are dense in number. Therefore, this application first restores the overlapping image sets to their respective positions, and reconstructs the imaging map of the object.
  • the model formula for object imaging under aerial photography is used to recover overlapping scene images.
  • Position generate image imaging map.
  • model formula of the object imaging is as follows:
  • s is the scale factor
  • m is the coordinate of the image point
  • M is the coordinate of the object point (the object point and the image point are the object position and the image position in the optical imaging respectively)
  • K is the parameter matrix in the aerial photography tool, It is composed of focal length and principal point coordinates
  • R is a rotation matrix, which can be converted to approximate values according to the yaw, pitch, and roll recorded by the aerial tool’s system
  • C is the projection center
  • the position vector can be approximated directly from the longitude, latitude, and altitude recorded by the GPS of the aerial photography tool
  • I is the third-order unit matrix.
  • the imaging map of n images can be obtained.
  • G n (V n , E n), wherein, referred to as a vertex set V n, En is called an edge set (a graph is a widely used data structure.
  • the nodes in the graph are called vertices.
  • the relationship between two vertices can be represented by a pair, called an edge. If the graph represents an edge Even pairs are ordered, then the graph is called a directed graph, if the pairs representing edges are disordered, then it is called an undirected graph).
  • No set of edges E in the drawing represents the number nE E n-side is, for each table represents one image, on behalf of the E nE n-th image for subsequent image matching process performed between only one image pair nE . If the relationship between images is not considered and the exhaustive traversal strategy is used for image matching, the total number of matches is Usually n*(n-1)/2 will be much larger than nE.
  • the method of constructing the image relationship undirected graph limits the scope of image matching, can avoid blind image matching, reduce the total image matching calculation complexity from O(n 2 ) to O(n), and improve the matching calculation efficiency ; At the same time, it can effectively eliminate the interference of unrelated image pairs, fundamentally avoid mismatches caused by non-overlapping images, and improve the accuracy of matching and the robustness of reconstruction.
  • the scale-invariant feature transform (SIFT) algorithm is used for image matching.
  • SIFT scale-invariant feature transform
  • image matching if there are few matching points in the two images I i and I j , which are smaller than the threshold N 1 , it means that the overlap is small or the correlation is weak, and (I i , I j ) is removed from the set E. If the number of matching points in the two images I i and I j is greater than the threshold N 1 , then the imaging image pair is retained to generate There are n 1 E image pairs in total, and the first image matching set E 1 is generated.
  • the second step is to generate an epipolar image based on the initial image matching set and calculate the degree of overlap between the epipolar images to complete the second image matching and generate a second image matching set.
  • the above-mentioned first step is only to filter out images with no repetition or low repetition.
  • this application continues to use the epipolar image method to perform matching filtering.
  • the epipolar image is a method of changing the search range from a two-dimensional plane image to a one-dimensional straight line during the matching process.
  • the plane formed by the shooting baseline and any ground point is called the nuclear surface
  • the intersection line between the nuclear surface and the image surface is called the nuclear line.
  • the image points with the same name must be on the epipolar line of the same name, and the image points on the epipolar line of the same name have a one-to-one correspondence.
  • an epipolar pair with the same name can be determined on a stereo image pair, then using the above-mentioned properties of the epipolar pair with the same name, the search and matching of the two-dimensional image can be transformed into the search and matching along the epipolar line.
  • the epipolar image eliminates the upper and lower parallax between the stereo images, narrows the search range, reduces the amount of matching calculations, and improves the matching accuracy, so it is of great significance for dense stereo image matching.
  • a preferred embodiment of the present application discloses a method for making and matching epipolar images for generating epipolar images and calculating the degree of overlap between the epipolar images.
  • the method includes: (a) using the SIFT algorithm to compare the After image pair point feature extraction, uniformly distributed high-precision points with the same name are obtained, the basic matrix estimation based on the RANSAC strategy is used to obtain the basic matrix of n 1 E image pairs; (b) using the basic matrix to determine each group of points with the same name Corresponding core line with the same name; (c) According to the principle that the core line must intersect at the core point, the least square method is used to determine The core point coordinates of the image pair are used to generate a quick mapping of the epipolar lines between the images according to the core point coordinates, and the epipolar line is resampled by bilinear interpolation along the epipolar line direction to complete the epipolar image production and matching regeneration There are a total of n 2 E image pairs to generate the second image matching set.
  • the epipolar image production method based on the basic matrix can avoid the problems of iterative calculation and initial value assignment when calculating the relative relationship, and it can also have good accuracy when the aerial photography uses a large angle of view.
  • a rotation matrix and a projection matrix are constructed, and the rotation matrix is divided into x, y, and z axes, each of which is among them, Is expressed as follows:
  • the least square method is used to determine the core point coordinates (x p , y p ) of the image from the calculation result of the projection matrix.
  • mapping is to directly correct the epipolar image on the central projection image.
  • the specific steps of the mapping are: derive the central projection according to the collinear condition equation, calculate the angle relationship between two adjacent epipolar lines when the epipolar lines are sampled, and complete the determination of each epipolar line on the central projection image; use The basic matrix generated above can determine each image pair The corresponding core lines are based on The nuclear point coordinates of each nuclear image in the image determine the epipolar line of the point, and complete the epipolar line correspondence between the same image pair to obtain The epipolar equation is:
  • (x p , y p ) are the coordinates of the core point calculated above, (x base , y base ) are the reference coordinates of the center projection image, and the same goes for The epipolar equation.
  • each image pair is established After the epipolar line mapping, according to the resampling rule of the bilinear interpolation method, the epipolar line image is generated and the overlap degree is calculated. The image pairs whose overlap degree of the epipolar line image is less than the threshold N 2 are discarded to generate There are n 2 E image pairs in total, and the second image matching set E 2 is obtained .
  • the two-dimensional image restores the geometric structure of the three-dimensional object surface, so further processing is required, and the following third step is performed.
  • the preferred embodiment of the present application adopts a dense matching algorithm, and on the basis of the second image matching set E 2 , the corner detection algorithm of the least equivalent segmentation absorbing core is used to extract respectively
  • the corner points of the corner points form a set of matching points of the corner points; combined with the epipolar geometric constraints, dynamic programming algorithms and other methods to establish Dense matching of all pixels between images. Specific steps are as follows:
  • the second image matching set of corner points in an image in E 2 (a) detection is the second image matching set of corner points in an image in E 2 (a) detection.
  • the corner point is the point where the local curvature changes the most on the contour of the image. It contains important information in the image, so it is of great significance for the detail matching of the image in the image set E 2 .
  • the preferred embodiment of the present application adopts the smallest uni-value segment assimilating nucleus (SUSAN) method to detect image corners: Gaussian smoothing is performed on the input image; each pixel in the image is traversed, and Sobel is used first The operator (it is a discrete first-order difference operator, used to calculate the approximate value of a gradient of the image brightness function) to determine whether the pixel is an edge point; if it is an edge point, the loss function according to the global energy equation is the smallest To determine whether the path loss L r (p, d) of the corner point is minimized, the determination principle is as follows:
  • C(p,d) is the local loss of the path
  • L r , min(pr) is the minimum loss of the previous step of the path, from which it can be determined whether the point is a corner point, and redundant corner points are removed; further judge if the two corners are detected If the points are adjacent, the corner points with larger L r (p,d) are removed.
  • the automatic matching of corner points can effectively distinguish the difference between image pairs based on the similarities and differences of the corner points, which is an effective means to achieve precise matching.
  • the corner point automatic matching can be divided into these steps: 1For Each point in the set of image corners is in Find matching corner points in the corresponding search area in the image, similarly, for Each point in the set of image corners is searched in the same way Find their corresponding matching points in the image, and call the intersection of these two matching point sets the initial matching point set K 1 ; 2In the initial matching point set K 1 , the pair is The corner point where the diagonal points of the image are concentrated with each other, find the matching point in the corresponding search area, calculate the similarity between this point and each candidate matching point in the search area, and select the candidate matching point with the greatest similarity as its match point.
  • the similarity calculation method adopts the gradient size similarity method: if the gradient size of a pixel is g, and the gradient of another pixel that matches it approximately obeys the normal distribution, then the two The similarity l g of each pixel is:
  • d(x) is the density function
  • k is the density coefficient
  • the dense matching includes: 1 Obtain the matching point set K 2 according to the limit geometric constraint relationship
  • the polar line correspondence of the image is obtained, and the limit correspondence set K 3 is obtained .
  • the so-called limit geometric constraint relationship means that if l and l'are the two corresponding epipolar lines in the left and right images, the corresponding point of the point p on the epipolar line l in the left image in the right image must be on the epipolar line l' on.
  • the polar lines in K 3 are segmented according to gray levels. Each polar line is divided into several gray segments, and the gray values of pixels on each segment are similar. .
  • the formula for gray scale segmentation is as follows:
  • the physical meaning of the above formula is to divide the continuous points of gray value in a certain range into one section.
  • I (x t , y t ) is the gray value of the pixel (x t , y t );
  • w is the number of pixels on a gray segment, that is, the length of the gray segment;
  • T is For a certain threshold, the smaller the value, the fewer pixels are divided into a certain gray-scale segment, and the more gray-scale segments.
  • the matching effect is the best when T is set to 3.
  • the gray-level segment set of is K 4 ; 3Using dynamic programming algorithm (an optimization method to find the best matching path) to establish the correspondence between gray-level segments, and using linear interpolation to save the corresponding gray-level The corresponding relationship between the pixels is established between the segments, so that the dense matching of all the pixels between the images is realized, and the third image matching set E 3 is obtained .
  • the preset 3Dmax software can be used to reconstruct the scene to restore the three-dimensional geometry of the scene space Information to get the reconstructed image.
  • the image matching program may also be divided into one or more modules, and the one or more modules are stored in the memory 11 and are executed by one or more processors (in this embodiment, processing The module 12) is executed to complete the application.
  • the module referred to in the application refers to a series of computer program instruction segments capable of completing specific functions, and is used to describe the execution process of the image matching program in the image matching device.
  • FIG. 3 it is a schematic diagram of the program modules of the image matching program in an embodiment of the image matching device of the present application.
  • the image matching program can be divided into a primary matching module 10, a secondary matching module 20,
  • the three-time matching module 30 and the reconstruction module 40 are exemplary:
  • the first matching module 10 is used to generate an image imaging map according to the scene image taken by the aerial camera, and use the scale-invariant feature transformation method to perform the first image matching on the image imaging map to generate a first image matching set.
  • the secondary matching module 20 is configured to generate epipolar images based on the primary image matching set and calculate the degree of overlap between the epipolar images, complete the second image matching, and generate a second image matching set.
  • the third-order matching module 30 is configured to: based on the second-order image matching set, establish dense matching of all pixels between the images, and generate a third-order image matching set.
  • the reconstruction module 40 is configured to perform three-dimensional reconstruction according to the image matching set to obtain a reconstructed scene image.
  • an embodiment of the present application also proposes a computer-readable storage medium having an image matching program stored on the computer-readable storage medium, and the image matching program can be executed by one or more processors to implement the following operations:
  • the image imaging map is generated, and the first image matching is performed on the image imaging map using the scale-invariant feature transformation method to generate the first image matching set;
  • a dense matching of all pixels between images is established, a third image matching set is generated, and three-dimensional reconstruction is performed to obtain a reconstructed scene image.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Graphics (AREA)
  • Geometry (AREA)
  • Software Systems (AREA)
  • Image Processing (AREA)

Abstract

一种影像匹配方法、装置(1)以及一种计算机可读存储介质,该方法包括:根据航拍仪拍摄的场景影像,生成影像成像图,利用尺度不变特征变换法对影像成像图进行初次影像匹配,生成初次影像匹配集(S1);基于所述初次影像匹配集,生成核线影像并计算所述核线影像之间的重叠度,完成第二次影像匹配,生成第二次影像匹配集(S2);基于所述的第二次影像匹配集,建立影像间所有像素点的密集匹配,生成第三次影像匹配集,并执行三维重建,得到重构后的场景影像(S3)。该应用于航拍下密集立体场景的新型影像匹配方案,可以提升影像匹配效率。

Description

影像匹配方法、装置及计算机可读存储介质
本申请基于巴黎公约申明享有2019年4月8日递交的申请号为CN201910274078.5、名称为“影像匹配方法、装置及计算机可读存储介质”的中国专利申请的优先权,该中国专利申请的整体内容以参考的方式结合在本申请中。
技术领域
本申请涉及计算机技术领域,尤其涉及一种影像匹配方法、装置及计算机可读存储介质。
背景技术
影像匹配是指通过一定的匹配算法在两幅或多幅影像之间识别同名点的过程。它是影像融合、目标识别、目标变化检测、计算机视觉等问题中的一个重要前期步骤,在遥感、数字摄影测量、计算机视觉、地图学及军事应用等多个领域都有着广泛的应用。目前影像匹配最行而有效的方法就是通过目测,判断影像的不同点;其次,根据影像的本质就是像素的原理,比较目标区域的所有像素灰度值;或者基于模板匹配的原理,找到目标影像和搜寻影像中的子影像相同或最相似的位置等。
以上方法都能在某些领域中切实可行,但当应用于航拍下的密集立体场景的匹配中,效果都会差强人意。由于航拍下的影像不仅数量巨大,而且每张图片的物体也很密集,人眼目测应接不暇不切实际;而利用像素匹配在对多幅影像像素做差值时,受噪声、量化误差、微小的光照变化、极小的平移等影响,都将产生较大的像素差值,影响匹配效果;而模板匹配方法应用于密集的影像中,需生成大量的匹配模块在图中进行寻找,时效性一般,同时受影像噪声影响,误匹配的概率也非常大。综合来说,由于航拍下的密集场景复杂多变,能进行三维重建可以更有效的帮助使用人员进行分析研究,而上述方法都缺少该功能。
发明内容
本申请提供一种影像匹配方法、装置及计算机可读存储介质,其主要目的在于提供一种应用于航拍下密集立体场景的新型影像匹配方法,提升影像匹配效率。
为实现上述目的,本申请提供的一种影像匹配方法,包括:
根据航拍仪拍摄的场景影像,生成影像成像图,利用尺度不变特征变换法对影像成像图进行初次影像匹配,生成初次影像匹配集;
基于所述初次影像匹配集,生成核线影像并计算所述核线影像之间的重叠度,完成第二次影像匹配,生成第二次影像匹配集;
基于所述的第二次影像匹配集,建立影像间所有像素点的密集匹配,生成第三次影像匹配集,并执行三维重建,得到重构后的场景影像。
此外,为实现上述目的,本申请还提供一种影像匹配装置,所述装置包括存储器和处理器,所述存储器上存储有可在所述处理器上运行的影像匹配程序,所述影像匹配程序被所述处理器执行时实现上述所述的影像匹配方法的步骤。
此外,为实现上述目的,本申请还提供一种计算机可读存储介质,所述计算机可读存储介质上存储有影像匹配程序,所述影像匹配程序可被一个或者多个处理器执行,以实现如上所述的影像匹配方法的步骤。
本申请提出的影像匹配方法、装置及计算机可读存储介质根据航拍仪拍摄的场景影像,生成影像成像图,利用尺度不变特征变换法对影像成像图进行初次影像匹配,生成初次影像匹配集,基于所述初次影像匹配集,生成核线影像并计算所述核线影像之间的重叠度,完成第二次影像匹配,生成第二次影像匹配集,基于所述的第二次影像匹配集,建立影像间所有像素点的密集匹配,生成第三次影像匹配集,并执行三维重建,得到重构后的场景影像。本申请提升了影像匹配效率,能够对航拍下的密集场景的影像进行三维重建,从而能够更有效的帮助使用人员进行分析研究。
附图说明
图1为本申请一实施例提供的影像匹配方法的流程示意图;
图2为本申请一实施例提供的影像匹配装置的内部结构示意图;
图3为本申请一实施例提供的影像匹配装置中影像匹配程序的模块示意 图。
本申请目的的实现、功能特点及优点将结合实施例,参照附图做进一步说明。
具体实施方式
为了使本申请的目的、技术方案及优点更加清楚明白,以下结合附图及实施例,对本申请进行进一步详细说明。应当理解,此处所描述的具体实施例仅用以解释本申请,并不用于限定本申请。基于本申请中的实施例,本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例,都属于本申请保护的范围。
本申请的说明书和权利要求书及上述附图中的术语“第一”、“第二”、“第三”、“第四”等(如果存在)是用于区别类似的对象,而不必用于描述特定的顺序或先后次序。应该理解这样使用的数据在适当情况下可以互换,以便这里描述的实施例能够以除了在这里图示或描述的内容以外的顺序实施。此外,所述“第一”、“第二”等的描述仅用于描述目的,而不能理解为指示或暗示其相对重要性或者隐含指明所指示的技术特征的数量。由此,限定有“第一”、“第二”的特征可以明示或者隐含地包括至少一个该特征。
进一步地,术语“包括”和“具有”以及他们的任何变形,意图在于覆盖不排他的包含,例如,包含了一系列步骤或单元的过程、方法、***、产品或设备不必限于清楚地列出的那些步骤或单元,而是可包括没有清楚地列出的或对于这些过程、方法、产品或设备固有的其它步骤或单元。
另外,各个实施例之间的技术方案可以相互结合,但是必须是以本领域普通技术人员能够实现为基础,当技术方案的结合出现相互矛盾或无法实现时应当认为这种技术方案的结合不存在,也不在本申请要求的保护范围之内。
本申请提供一种影像匹配方法。参照图1所示,为本申请一实施例提供的影像匹配方法的流程示意图。该方法可以由一个装置执行,该装置可以由软件和/或硬件实现。
在本实施例中,影像匹配方法包括:
S10、根据航拍仪拍摄的场景影像,生成影像成像图,利用尺度不变特征变换法对影像成像图进行初次影像匹配,生成初次影像匹配集。
由航拍仪器,如无人机、直升机等飞控***拍摄的场景影像,影像数量多且视角宽广,特别是建筑物,数量多密集度大。所以,本申请首先将存在重叠的影像集恢复出各自的位置,重新构建出物体的成像图。
本申请较佳实施例根据航拍仪器拍摄需场景影像时所记录的低精度位置、姿态信息及测区的概略高度等参数,利用航拍下物体成像的模型公式将存在重叠的场景影像恢复出各自的位置,生成影像成像图。
本申请较佳实施例中,所述物体成像的模型公式如下:
sm=KR[I-C]M,
其中,s为尺度系数;m为像点坐标,M为物点坐标(所述物点、像点分别是光学成像中的物方位置和像方位置);K为航拍工具内的参数矩阵,由焦距、像主点坐标组成;R为旋转矩阵,可根据航拍工具的***所记录的偏航(yaw)、俯仰(pitch)、侧滚(roll),然后转换得到其近似值;C为投影中心位置向量,可直接由航拍工具的GPS记录的经度(longitude)、纬度(latitude)、高(altitude)近似得到;I为3阶单位矩阵。
利用上述物体成像的模型公式可以得到n幅影像的成像图。
上述生成的n幅影像成像图,统称为影像集,将所述影像集转换为对应的无向图边集合E:G n=(V n,E n),其中,V n称为顶点集,E n称为边集(图是一种广泛应用的数据结构,图中的结点称为顶点,两个顶点之间的关系可用一个偶对来表示,称为边。如果图中代表边的偶对是有序的,那么称该图为有向图,如果代表边的偶对是无序的,则称其为无向图)。无向图边集合E中,E n代表边的数量为nE,每条表代表一个影像对,则E n代表nE个影像对,后续的影像匹配处理只需在这nE个影像对之间进行。若不考虑影像间关系,采用穷举遍历策略进行影像匹配,则总匹配数为
Figure PCTCN2019102187-appb-000001
通常n*(n-1)/2会远大于nE。因此,构建影像关系无向图的方法限定了影像匹配的范围,可以避免盲目的影像匹配,使总的影像匹配计算复杂度由O(n 2)减少到O(n),提高了匹配计算效率;同时又能有效排除非关联影像对的干扰,从根本上避免了无重叠影像产生的误匹配,提高匹配的准确率和重建的稳健性。
在无向图边集合E中,采用尺度不变特征变换(Scale-invariant feature transform,SIFT)算法进行影像匹配。在影像匹配中,若I i、I j两影像的匹配 点较少,小于阈值N 1,说明重叠较小或关联较弱,将(I i、I j)从集合E中剔除。若I i、I j两影像的匹配点数量大于阈值N 1,则保留该成像图对,生成
Figure PCTCN2019102187-appb-000002
共n 1E个影像对,产生初次影像匹配集E 1
S20、基于所述初次影像匹配集,生成核线影像并计算所述核线影像之间的重叠度,完成第二次影像匹配,生成第二次影像匹配集。
上述的S10只是过滤掉没有重复或重复度小的影像,对于具有一定重复的影像来说,本申请继续利用核线影像方法进行匹配过滤。
所述核线影像是在匹配的过程中,将搜索范围从二维平面的成像图变为一维直线的一种方法。具体来说,在密集的立体航拍中,拍摄基线与任一地面点构成的平面称为核面,核面与像面的交线称为核线。在立体像对上,同名像点一定位于同名核线上,而且同名核线对上的像点是一一对应的。因而,如果能够在立体像对上确定同名核线对,那么利用同名核线对的上述性质,就可以把二维影像的搜索和匹配转变成沿核线的搜索和匹配。核线影像消除了立体影像间的上下视差,缩小搜索范围,降低匹配计算量,提高匹配准确性,所以对于密集立体的影像匹配具有重要意义。
本申请较佳实施例揭露了一种核线影像制作和匹配方法,用于生成核线影像并计算所述核线影像之间的重叠度,该方法包括:(a)利用SIFT算法对所述
Figure PCTCN2019102187-appb-000003
影像对进行点特征提取,获取分布均匀的高精度同名点后,利用基于RANSAC策略进行基础矩阵估计,得到n 1E个影像对的基础矩阵;(b)利用所述基础矩阵确定每组同名点对应的同名核线;(c)根据核线必相交于核点的原理,采用最小二乘法确定
Figure PCTCN2019102187-appb-000004
影像对的核点坐标,根据核点坐标生成影像间核线的快速映射,并沿核线方向采用双线性内插法进行核线重采样,完成核线影像制作并匹配重新生成
Figure PCTCN2019102187-appb-000005
共n 2E个影像对,产生第二次影像匹配集。
基于基础矩阵的核线影像制作方法可以避免相对关系解算时的迭代计算以及赋初值等问题,且在航拍采用大角度视角的情况下,也可以有很好的解算精度。以下是上述步骤(c)即的具体步骤:
(1)确定核点坐标:
基于所述基础矩阵,构建旋转矩阵和投影矩阵,将所述旋转矩阵分为x、y、z轴,各为
Figure PCTCN2019102187-appb-000006
其中,
Figure PCTCN2019102187-appb-000007
的表达方式如下:
Figure PCTCN2019102187-appb-000008
并得到投影矩阵
Figure PCTCN2019102187-appb-000009
Figure PCTCN2019102187-appb-000010
其中,
Figure PCTCN2019102187-appb-000011
为相机left的相机参数,
Figure PCTCN2019102187-appb-000012
为相机right的相机参数,t left、t right分别是相机left和相机right的相机参数在x、y、z轴的分量;
根据核线必交于核点原理,由投影矩阵的计算结果,采用最小二乘法确定影像的核点坐标(x p,y p)。
(2)执行核线映射:
映射的目的是为了在中心投影影像上直接进行核线影像纠正。映射的具体步骤是:根据共线条件方程推导出中心投影,在核线采样时,计算出相邻两条核线之间的夹角关系,完成中心投影影像上每条核线的确定;利用上述生成的基础矩阵,可以确定每个影像对
Figure PCTCN2019102187-appb-000013
所对应的核线,分别根据
Figure PCTCN2019102187-appb-000014
中每个核像的核点坐标确定该点的核线,完成同影像对之间的核线对应,得到
Figure PCTCN2019102187-appb-000015
的核线方程为:
Figure PCTCN2019102187-appb-000016
其中,(x p,y p)是上述计算的核点坐标,(x base,y base)为中心投影影像的基准坐标,同理得到
Figure PCTCN2019102187-appb-000017
的核线方程。
(3)产生第二次影像匹配集:
以所述核线方程为基础,建立起每个影像对
Figure PCTCN2019102187-appb-000018
的核线映射之后,按照双线性内插法的重采样规则,生成核线影像并计算重叠度,将核线影像的重叠度小于阈值N 2的影像对丢弃,生成
Figure PCTCN2019102187-appb-000019
共n 2E个影像对,得到第二次影像匹配集E 2
上述生成的第二次影像匹配集E 2虽然解决了大部分的立体匹配相似问题,达到了影像匹配的标准,但是对于战场监察、灾害情况搜救等这类密集的立体环境来说,还需要从二维影像恢复三维物体表面的几何结构,因此需进一步的处理,执行下述的S30。
S30、基于所述的第二次影像匹配集,建立影像间所有像素点的密集匹配,实现三维重建。
本申请较佳实施例采用密集匹配算法,在第二次影像匹配集E 2的基础上, 利用最小同值分割吸收核的角点检测算法分别提取
Figure PCTCN2019102187-appb-000020
的角点,形成角点的匹配点集合;结合对极几何约束、动态规划算法等方法建立起
Figure PCTCN2019102187-appb-000021
影像间所有像素点的密集匹配。具体步骤如下:
(一)检测第二次影像匹配集E 2中影像的角点。
角点是影像的轮廓线上局部曲率变化最大的点,它含有影像中重要的信息,所以对于影像集E 2中影像的细节匹配具有很大意义。本申请较佳实施例采用最小同值分割吸收核方法(smallest uni-value segment assimilating nucleus,SUSAN)对影像角点进行检测:对输入影像进行高斯平滑;遍历影像中每个像素点,先利用Sobel算子(它是一个离散的一阶差分算子,用来计算影像亮度函数的一阶梯度之近似值)判断该像素点是否为边缘点;若是边缘点,则进一步根据全局能量方程的损失函数最小化原理,判断角点的路径损失L r(p,d)是否最小化,判定原则如下:
Figure PCTCN2019102187-appb-000022
其中C(p,d)为路径的局部损失,
Figure PCTCN2019102187-appb-000023
为路径的上一步损失,L r,min(p-r)为路径上一步损失的最小损失,由此可判定该点是否为角点,并去除冗余角点;进一步判断若被检测出的两角点相邻,则去掉L r(p,d)较大的角点。经过以上步骤,可以检测出第二次影像匹配集E 2中影像对的角点。
(二)对
Figure PCTCN2019102187-appb-000024
影像对的角点进行自动匹配,得到匹配点集。
角点的自动匹配可以根据角点的异同有效区分出影像对之间的差异,是达到精准匹配的一个有效手段。角点自动匹配可分为这几步:①对于
Figure PCTCN2019102187-appb-000025
影像角点集合中的每一个点,在
Figure PCTCN2019102187-appb-000026
影像中的相应搜索区域内寻找与之相匹配的角点,相似的,对于
Figure PCTCN2019102187-appb-000027
影像角点集合中的每一个点,按同样的搜索方法在
Figure PCTCN2019102187-appb-000028
影像中寻找它们的对应匹配点,将这两个匹配点集合的交集称为初始匹配点集K 1;②在所述初始匹配点集K 1中,对在
Figure PCTCN2019102187-appb-000029
影像对角点都互相集中的角点,在相应搜索区域内寻找匹配点,计算这个点与搜索区域内每个候选匹配点的相似度,选取与之相似度最大的候选匹配点为它的匹配点。本申请较佳实施例中,所述相似度的计算方法采用梯度大小相似法:若一像素点的梯度大小为g,与之 匹配的另一像素点的梯度大小近似服从正态分布,则两个像素的相似度l g为:
Figure PCTCN2019102187-appb-000030
其中,d(x)为密度函数,k为密度系数。经过相似度的计算得到匹配点集K 2
(三)根据所述匹配点集K 2,建立起
Figure PCTCN2019102187-appb-000031
影像间所有像素点的密集匹配。
本申请较佳实施例中,所述密集匹配包括:①对匹配点集K 2根据极限几何约束关系得到
Figure PCTCN2019102187-appb-000032
影像的极线对应关系,得到极限对应关系集K 3。所谓的极限几何约束关系就是指若l和l’是左右两幅影像中的两条对应极线,则左影像中极线l上的点p在右影像中的对应点一定在极线l’上。②由生成的极限对应关系集K 3,按灰度对K 3内的极线进行分段,每条极线被分为若干灰度分段,每段上的像素点的灰度值都相近。灰度分段的公式如下:
Figure PCTCN2019102187-appb-000033
上述公式的物理意义是将灰度值在某一范围内的连续点分为一段。其中,I(x t,y t)为像素点(x t,y t)的灰度值;w为某一灰度分段上的像素点个数,即灰度分段的长度;T为某一阈值,取值越小,被划分为某一灰度分段上的像素点个数越少,灰度分段数就越多,经实验研究,T取3时匹配效果最好,生成的灰度分段集合为K 4;③利用动态规划算法(是一种寻找最佳匹配路径的优化方法)建立灰度分段之间的对应关系,利用线性插值的方法在省对应的灰度分段之间建立各像素点之间的对应关系,从而实现影像间所有像素点的密集匹配,得到第三次影像匹配集E 3
经过密集匹配后,所述第三次影像匹配集E 3中所有像素点都存在对应关系,因此可以计算出场景的景深,利用预设的3Dmax软件对场景进行重构,恢复场景空间的三维几何信息,得到重构后的影像。
本申请还提供一种影像匹配装置。参照图2所示,为本申请一实施例提供的影像匹配装置的内部结构示意图。
在本实施例中,影像匹配装置1可以是PC(Personal Computer,个人电脑),也可以是智能手机、平板电脑、便携计算机等终端设备。该影像匹配装置1至少包括存储器11、处理器12,通信总线13,以及网络接口14。
其中,存储器11至少包括一种类型的可读存储介质,所述可读存储介质包括闪存、硬盘、多媒体卡、卡型存储器(例如,SD或DX存储器等)、磁性 存储器、磁盘、光盘等。存储器11在一些实施例中可以是影像匹配装置1的内部存储单元,例如该影像匹配装置1的硬盘。存储器11在另一些实施例中也可以是影像匹配装置1的外部存储设备,例如影像匹配装置1上配备的插接式硬盘,智能存储卡(Smart Media Card,SMC),安全数字(Secure Digital,SD)卡,闪存卡(Flash Card)等。进一步地,存储器11还可以既包括影像匹配装置1的内部存储单元也包括外部存储设备。存储器11不仅可以用于存储安装于影像匹配装置1的应用软件及各类数据,例如影像匹配程序01的代码等,还可以用于暂时地存储已经输出或者将要输出的数据。
处理器12在一些实施例中可以是一中央处理器(Central Processing Unit,CPU)、控制器、微控制器、微处理器或其他数据处理芯片,用于运行存储器11中存储的程序代码或处理数据,例如执行影像匹配程序01等。
通信总线13用于实现这些组件之间的连接通信。
网络接口14可选的可以包括标准的有线接口、无线接口(如WI-FI接口),通常用于在该装置1与其他电子设备之间建立通信连接。
可选地,该装置1还可以包括用户接口,用户接口可以包括显示器(Display)、输入单元比如键盘(Keyboard),可选的用户接口还可以包括标准的有线接口、无线接口。可选地,在一些实施例中,显示器可以是LED显示器、液晶显示器、触控式液晶显示器以及OLED(Organic Light-Emitting Diode,有机发光二极管)触摸器等。其中,显示器也可以适当的称为显示屏或显示单元,用于显示在影像匹配装置1中处理的信息以及用于显示可视化的用户界面。
图2仅示出了具有组件11-14以及影像匹配程序01的影像匹配装置1,本领域技术人员可以理解的是,图1示出的结构并不构成对影像匹配装置1的限定,可以包括比图示更少或者更多的部件,或者组合某些部件,或者不同的部件布置。
在图2所示的装置1实施例中,存储器11中存储有影像匹配程序01;处理器12执行存储器11中存储的影像匹配程序01时实现如下步骤:
第一步、根据航拍仪拍摄的场景影像,生成影像成像图,利用尺度不变特征变换法对影像成像图进行初次影像匹配,生成初次影像匹配集。
由航拍仪器,如无人机、直升机等飞控***拍摄的场景影像,影像数量 多且视角宽广,特别是建筑物,数量多密集度大。所以,本申请首先将存在重叠的影像集恢复出各自的位置,重新构建出物体的成像图。
本申请较佳实施例根据航拍仪器拍摄需场景影像时所记录的低精度位置、姿态信息及测区的概略高度等参数,利用航拍下物体成像的模型公式将存在重叠的场景影像恢复出各自的位置,生成影像成像图。
本申请较佳实施例中,所述物体成像的模型公式如下:
sm=KR[I-C]M,
其中,s为尺度系数;m为像点坐标,M为物点坐标(所述物点、像点分别是光学成像中的物方位置和像方位置);K为航拍工具内的参数矩阵,由焦距、像主点坐标组成;R为旋转矩阵,可根据航拍工具的***所记录的偏航(yaw)、俯仰(pitch)、侧滚(roll),然后转换得到其近似值;C为投影中心位置向量,可直接由航拍工具的GPS记录的经度(longitude)、纬度(latitude)、高(altitude)近似得到;I为3阶单位矩阵。
利用上述物体成像的模型公式可以得到n幅影像的成像图。
上述生成的n幅影像成像图,统称为影像集,将所述影像集转换为对应的无向图边集合E:G n=(V n,E n),其中,V n称为顶点集,E n称为边集(图是一种广泛应用的数据结构,图中的结点称为顶点,两个顶点之间的关系可用一个偶对来表示,称为边。如果图中代表边的偶对是有序的,那么称该图为有向图,如果代表边的偶对是无序的,则称其为无向图)。无向图边集合E中,E n代表边的数量为nE,每条表代表一个影像对,则E n代表nE个影像对,后续的影像匹配处理只需在这nE个影像对之间进行。若不考虑影像间关系,采用穷举遍历策略进行影像匹配,则总匹配数为
Figure PCTCN2019102187-appb-000034
通常n*(n-1)/2会远大于nE。因此,构建影像关系无向图的方法限定了影像匹配的范围,可以避免盲目的影像匹配,使总的影像匹配计算复杂度由O(n 2)减少到O(n),提高了匹配计算效率;同时又能有效排除非关联影像对的干扰,从根本上避免了无重叠影像产生的误匹配,提高匹配的准确率和重建的稳健性。
在无向图边集合E中,采用尺度不变特征变换(Scale-invariant feature transform,SIFT)算法进行影像匹配。在影像匹配中,若I i、I j两影像的匹配点较少,小于阈值N 1,说明重叠较小或关联较弱,将(I i、I j)从集合E中剔除。 若I i、I j两影像的匹配点数量大于阈值N 1,则保留该成像图对,生成
Figure PCTCN2019102187-appb-000035
共n 1E个影像对,产生初次影像匹配集E 1
第二步、基于所述初次影像匹配集,生成核线影像并计算所述核线影像之间的重叠度,完成第二次影像匹配,生成第二次影像匹配集。
上述的第一步只是过滤掉没有重复或重复度小的影像,对于具有一定重复的影像来说,本申请继续利用核线影像方法进行匹配过滤。
所述核线影像是在匹配的过程中,将搜索范围从二维平面的成像图变为一维直线的一种方法。具体来说,在密集的立体航拍中,拍摄基线与任一地面点构成的平面称为核面,核面与像面的交线称为核线。在立体像对上,同名像点一定位于同名核线上,而且同名核线对上的像点是一一对应的。因而,如果能够在立体像对上确定同名核线对,那么利用同名核线对的上述性质,就可以把二维影像的搜索和匹配转变成沿核线的搜索和匹配。核线影像消除了立体影像间的上下视差,缩小搜索范围,降低匹配计算量,提高匹配准确性,所以对于密集立体的影像匹配具有重要意义。
本申请较佳实施例揭露了一种核线影像制作和匹配方法,用于生成核线影像并计算所述核线影像之间的重叠度,该方法包括:(a)利用SIFT算法对所述
Figure PCTCN2019102187-appb-000036
影像对进行点特征提取,获取分布均匀的高精度同名点后,利用基于RANSAC策略进行基础矩阵估计,得到n 1E个影像对的基础矩阵;(b)利用所述基础矩阵确定每组同名点对应的同名核线;(c)根据核线必相交于核点的原理,采用最小二乘法确定
Figure PCTCN2019102187-appb-000037
影像对的核点坐标,根据核点坐标生成影像间核线的快速映射,并沿核线方向采用双线性内插法进行核线重采样,完成核线影像制作并匹配重新生成
Figure PCTCN2019102187-appb-000038
共n 2E个影像对,产生第二次影像匹配集。
基于基础矩阵的核线影像制作方法可以避免相对关系解算时的迭代计算以及赋初值等问题,且在航拍采用大角度视角的情况下,也可以有很好的解算精度。以下是上述步骤(c)即的具体步骤:
(1)确定核点坐标:
基于所述基础矩阵,构建旋转矩阵和投影矩阵,将所述旋转矩阵分为x、y、z轴,各为
Figure PCTCN2019102187-appb-000039
其中,
Figure PCTCN2019102187-appb-000040
的表达方式如下:
Figure PCTCN2019102187-appb-000041
并得到投影矩阵
Figure PCTCN2019102187-appb-000042
Figure PCTCN2019102187-appb-000043
其中,
Figure PCTCN2019102187-appb-000044
为相机left的相机参数,
Figure PCTCN2019102187-appb-000045
为相机right的相机参数,t left、t right分别是相机left和相机right的相机参数在x、y、z轴的分量;
根据核线必交于核点原理,由投影矩阵的计算结果,采用最小二乘法确定影像的核点坐标(x p,y p)。
(2)执行核线映射:
映射的目的是为了在中心投影影像上直接进行核线影像纠正。映射的具体步骤是:根据共线条件方程推导出中心投影,在核线采样时,计算出相邻两条核线之间的夹角关系,完成中心投影影像上每条核线的确定;利用上述生成的基础矩阵,可以确定每个影像对
Figure PCTCN2019102187-appb-000046
所对应的核线,分别根据
Figure PCTCN2019102187-appb-000047
中每个核像的核点坐标确定该点的核线,完成同影像对之间的核线对应,得到
Figure PCTCN2019102187-appb-000048
的核线方程为:
Figure PCTCN2019102187-appb-000049
其中,(x p,y p)是上述计算的核点坐标,(x base,y base)为中心投影影像的基准坐标,同理得到
Figure PCTCN2019102187-appb-000050
的核线方程。
(3)产生第二次影像匹配集:
以所述核线方程为基础,建立起每个影像对
Figure PCTCN2019102187-appb-000051
的核线映射之后,按照双线性内插法的重采样规则,生成核线影像并计算重叠度,将核线影像的重叠度小于阈值N 2的影像对丢弃,生成
Figure PCTCN2019102187-appb-000052
共n 2E个影像对,得到第二次影像匹配集E 2
上述生成的第二次影像匹配集E 2虽然解决了大部分的立体匹配相似问题,达到了影像匹配的标准,但是对于战场监察、灾害情况搜救等这类密集的立体环境来说,还需要从二维影像恢复三维物体表面的几何结构,因此需进一步的处理,执行下述的第三步。
第三步、基于所述的第二次影像匹配集,建立影像间所有像素点的密集匹配,实现三维重建。
本申请较佳实施例采用密集匹配算法,在第二次影像匹配集E 2的基础上,利用最小同值分割吸收核的角点检测算法分别提取
Figure PCTCN2019102187-appb-000053
的角点,形成角点的匹配点集合;结合对极几何约束、动态规划算法等方法建立起
Figure PCTCN2019102187-appb-000054
影像间所 有像素点的密集匹配。具体步骤如下:
(一)检测第二次影像匹配集E 2中影像的角点。
角点是影像的轮廓线上局部曲率变化最大的点,它含有影像中重要的信息,所以对于影像集E 2中影像的细节匹配具有很大意义。本申请较佳实施例采用最小同值分割吸收核方法(smallest uni-value segment assimilating nucleus,SUSAN)对影像角点进行检测:对输入影像进行高斯平滑;遍历影像中每个像素点,先利用Sobel算子(它是一个离散的一阶差分算子,用来计算影像亮度函数的一阶梯度之近似值)判断该像素点是否为边缘点;若是边缘点,则进一步根据全局能量方程的损失函数最小化原理,判断角点的路径损失L r(p,d)是否最小化,判定原则如下:
Figure PCTCN2019102187-appb-000055
其中C(p,d)为路径的局部损失,
Figure PCTCN2019102187-appb-000056
为路径的上一步损失,L r,min(p-r)为路径上一步损失的最小损失,由此可判定该点是否为角点,并去除冗余角点;进一步判断若被检测出的两角点相邻,则去掉L r(p,d)较大的角点。经过以上步骤,可以检测出第二次影像匹配集E 2中影像对的角点。
(二)对
Figure PCTCN2019102187-appb-000057
影像对的角点进行自动匹配,得到匹配点集。
角点的自动匹配可以根据角点的异同有效区分出影像对之间的差异,是达到精准匹配的一个有效手段。角点自动匹配可分为这几步:①对于
Figure PCTCN2019102187-appb-000058
影像角点集合中的每一个点,在
Figure PCTCN2019102187-appb-000059
影像中的相应搜索区域内寻找与之相匹配的角点,相似的,对于
Figure PCTCN2019102187-appb-000060
影像角点集合中的每一个点,按同样的搜索方法在
Figure PCTCN2019102187-appb-000061
影像中寻找它们的对应匹配点,将这两个匹配点集合的交集称为初始匹配点集K 1;②在所述初始匹配点集K 1中,对在
Figure PCTCN2019102187-appb-000062
影像对角点都互相集中的角点,在相应搜索区域内寻找匹配点,计算这个点与搜索区域内每个候选匹配点的相似度,选取与之相似度最大的候选匹配点为它的匹配点。本申请较佳实施例中,所述相似度的计算方法采用梯度大小相似法:若一像素点的梯度大小为g,与之匹配的另一像素点的梯度大小近似服从正态分布,则两个像素的相似度l g为:
Figure PCTCN2019102187-appb-000063
其中,d(x)为密度函数,k为密度系数。经过相似度的计算得到匹配点集K 2
(三)根据所述匹配点集K 2,建立起
Figure PCTCN2019102187-appb-000064
影像间所有像素点的密集匹配。
本申请较佳实施例中,所述密集匹配包括:①对匹配点集K 2根据极限几何约束关系得到
Figure PCTCN2019102187-appb-000065
影像的极线对应关系,得到极限对应关系集K 3。所谓的极限几何约束关系就是指若l和l’是左右两幅影像中的两条对应极线,则左影像中极线l上的点p在右影像中的对应点一定在极线l’上。②由生成的极限对应关系集K 3,按灰度对K 3内的极线进行分段,每条极线被分为若干灰度分段,每段上的像素点的灰度值都相近。灰度分段的公式如下:
Figure PCTCN2019102187-appb-000066
上述公式的物理意义是将灰度值在某一范围内的连续点分为一段。其中,I(x t,y t)为像素点(x t,y t)的灰度值;w为某一灰度分段上的像素点个数,即灰度分段的长度;T为某一阈值,取值越小,被划分为某一灰度分段上的像素点个数越少,灰度分段数就越多,经实验研究,T取3时匹配效果最好,生成的灰度分段集合为K 4;③利用动态规划算法(是一种寻找最佳匹配路径的优化方法)建立灰度分段之间的对应关系,利用线性插值的方法在省对应的灰度分段之间建立各像素点之间的对应关系,从而实现影像间所有像素点的密集匹配,得到第三次影像匹配集E 3
经过密集匹配后,所述第三次影像匹配集E 3中所有像素点都存在对应关系,因此可以计算出场景的景深,利用预设的3Dmax软件对场景进行重构,恢复场景空间的三维几何信息,得到重构后的影像。
可选地,在其他实施例中,影像匹配程序还可以被分割为一个或者多个模块,一个或者多个模块被存储于存储器11中,并由一个或多个处理器(本实施例为处理器12)所执行以完成本申请,本申请所称的模块是指能够完成特定功能的一系列计算机程序指令段,用于描述影像匹配程序在影像匹配装置中的执行过程。
例如,参照图3所示,为本申请影像匹配装置一实施例中的影像匹配程序的程序模块示意图,该实施例中,影像匹配程序可以被分割为初次匹配模块10、二次匹配模块20、三次匹配模块30以及重构模块40,示例性地:
所述初次匹配模块10用于:根据航拍仪拍摄的场景影像,生成影像成像 图,利用尺度不变特征变换法对影像成像图进行初次影像匹配,生成初次影像匹配集。
所述二次匹配模块20用于:基于所述初次影像匹配集,生成核线影像并计算所述核线影像之间的重叠度,完成第二次影像匹配,生成第二次影像匹配集。
所述三次匹配模块30用于:基于所述的第二次影像匹配集,建立影像间所有像素点的密集匹配,生成第三次影像匹配集。
所述重构模块40用于:根据所述影像匹配集,执行三维重建,得到重构后的场景影像。
上述初次匹配模块10、二次匹配模块20、三次匹配模块30和重构模块40等程序模块被执行时所实现的功能或操作步骤与上述实施例大体相同,在此不再赘述。
此外,本申请实施例还提出一种计算机可读存储介质,所述计算机可读存储介质上存储有影像匹配程序,所述影像匹配程序可被一个或多个处理器执行,以实现如下操作:
根据航拍仪拍摄的场景影像,生成影像成像图,利用尺度不变特征变换法对影像成像图进行初次影像匹配,生成初次影像匹配集;
基于所述初次影像匹配集,生成核线影像并计算所述核线影像之间的重叠度,完成第二次影像匹配,生成第二次影像匹配集;
基于所述的第二次影像匹配集,建立影像间所有像素点的密集匹配,生成第三次影像匹配集,并执行三维重建,得到重构后的场景影像。
本申请计算机可读存储介质具体实施方式与上述影像匹配装置和方法各实施例基本相同,在此不作累述。
需要说明的是,上述本申请实施例序号仅仅为了描述,不代表实施例的优劣。并且本文中的术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含,从而使得包括一系列要素的过程、装置、物品或者方法不仅包括那些要素,而且还包括没有明确列出的其他要素,或者是还包括为这种过程、装置、物品或者方法所固有的要素。在没有更多限制的情况下,由语句“包括一个……”限定的要素,并不排除在包括该要素的过程、装置、物品或者方法中还存在另外的相同要素。
通过以上的实施方式的描述,本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现,当然也可以通过硬件,但很多情况下前者是更佳的实施方式。基于这样的理解,本申请的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来,该计算机软件产品存储在如上所述的一个存储介质(如ROM/RAM、磁碟、光盘)中,包括若干指令用以使得一台终端设备(可以是手机,计算机,服务器,或者网络设备等)执行本申请各个实施例所述的方法。
以上仅为本申请的优选实施例,并非因此限制本申请的专利范围,凡是利用本申请说明书及附图内容所作的等效结构或等效流程变换,或直接或间接运用在其他相关的技术领域,均同理包括在本申请的专利保护范围内。

Claims (20)

  1. 一种影像匹配方法,其特征在于,所述方法包括:
    根据航拍仪拍摄的场景影像,生成影像成像图,利用尺度不变特征变换法对影像成像图进行初次影像匹配,生成初次影像匹配集;
    基于所述初次影像匹配集,生成核线影像并计算所述核线影像之间的重叠度,完成第二次影像匹配,生成第二次影像匹配集;
    基于所述的第二次影像匹配集,建立影像间所有像素点的密集匹配,生成第三次影像匹配集,并执行三维重建,得到重构后的场景影像。
  2. 如权利要求1所述的影像匹配方法,其特征在于,根据航拍仪拍摄的场景影像,生成影像成像图,包括:
    根据航拍仪器拍摄所述场景影像时所记录的参数,包括低精度位置、姿态信息及测区的概略高度,利用航拍下物体成像的模型公式将航拍仪器拍摄的存在重叠的场景影像恢复出各自的位置,生成n幅影像成像图,其中,所述物体成像的模型公式如下:
    sm=KR[I-C]M,
    其中,s为尺度系数,m为像点坐标,M为物点坐标,K为航拍工具内的参数矩阵,R为旋转矩阵,C为投影中心位置向量,I为3阶单位矩阵。
  3. 如权利要求2所述的影像匹配方法,其特征在于,所述利用尺度不变特征变换法对影像成像图进行初次影像匹配,生成初次影像匹配集,包括:
    将所述n幅影像成像图组成的影像集转换为对应的无向图边集合E;
    在所述无向图边集合E中,采用尺度不变特征变换算法进行影像匹配,并在影像匹配中,对于影像(I i、I j)∈E,若I i、I j两影像的匹配点数量小于阈值N 1,则将(I i、I j)从所述无向图边集合中剔除,若I i、I j两影像的匹配点数量大于阈值N 1,则保留该成像图对,生成
    Figure PCTCN2019102187-appb-100001
    影像对,产生所述初次影像匹配集。
  4. 如权利要求3所述的影像匹配方法,其特征在于,基于所述初次影像匹配集,生成核线影像并计算所述核线影像之间的重叠度,完成第二次影像匹配,生成第二次影像匹配集,包括:
    (a)利用尺度不变特征变换法算法对所述
    Figure PCTCN2019102187-appb-100002
    影像对进行点特征提取,获取分布均匀的高精度同名点后,利用基于RANSAC策略进行基础矩阵估计,得到基础矩阵;
    (b)利用所述基础矩阵确定每组同名点对应的同名核线;
    (c)根据核线必相交于核点的原理,采用最小二乘法确定
    Figure PCTCN2019102187-appb-100003
    影像对的核点坐标,根据核点坐标生成影像间核线的快速映射,并沿核线方向采用双线性内插法进行核线重采样,完成核线影像制作并匹配重新生成
    Figure PCTCN2019102187-appb-100004
    影像对,产生所述第二次影像匹配集。
  5. 如权利要求4所述的影像匹配方法,其特征在于,所述(c)包括:基于所述基础矩阵,构建旋转矩阵和投影矩阵,将所述旋转矩阵分为x、y、z轴,各为
    Figure PCTCN2019102187-appb-100005
    其中,
    Figure PCTCN2019102187-appb-100006
    的表达方式如下:
    Figure PCTCN2019102187-appb-100007
    并得到投影矩阵
    Figure PCTCN2019102187-appb-100008
    Figure PCTCN2019102187-appb-100009
    其中,
    Figure PCTCN2019102187-appb-100010
    为相机left的相机参数,
    Figure PCTCN2019102187-appb-100011
    为相机right的相机参数,t left、t right分别是相机left和相机right的相机参数在x、y、z轴的分量;
    根据核线必交于核点原理,由投影矩阵的计算结果,采用最小二乘法确定影像的核点坐标(x p,y p);
    根据共线条件方程推导出中心投影,在核线采样时,计算出相邻两条核线之间的夹角关系,完成中心投影影像上每条核线的确定;
    利用上述生成的基础矩阵,可以确定每个影像对
    Figure PCTCN2019102187-appb-100012
    所对应的核线,分别根据
    Figure PCTCN2019102187-appb-100013
    中每个核像的核点坐标确定该点的核线,完成同影像对之间的核线对应,得到
    Figure PCTCN2019102187-appb-100014
    的核线方程为:
    Figure PCTCN2019102187-appb-100015
    其中,(x p,y p)是上述计算的核点坐标,(x base,y base)为中心投影影像的基准坐标,同理得到
    Figure PCTCN2019102187-appb-100016
    的核线方程;
    以所述核线方程为基础,建立起每个影像对
    Figure PCTCN2019102187-appb-100017
    的核线映射之后,按照双线性内插法的重采样规则,生成核线影像并计算重叠度,将核线影像的重叠度小于阈值N 2的影像对丢弃,生成
    Figure PCTCN2019102187-appb-100018
    影像对,得到所述第二次影像匹配集。
  6. 如权利要求5所述的影像匹配方法,其特征在于,所述基于所述的第二次影像匹配集,建立影像间所有像素点的密集匹配,包括:
    在所述第二次影像匹配集的基础上,利用最小同值分割吸收核的角点检测算法分别提取
    Figure PCTCN2019102187-appb-100019
    的角点,形成角点的匹配点集合,结合对极几何约束、动态规划算法建立起
    Figure PCTCN2019102187-appb-100020
    影像间所有像素点的密集匹配。
  7. 如权利要求5所述的影像匹配方法,其特征在于,所述基于所述的第二次影像匹配集,建立影像间所有像素点的密集匹配,生成第三次影像匹配集,包括:
    对输入影像进行高斯平滑,遍历影像中每个像素点,利用Sobel算子判断该像素点是否为边缘点,若是边缘点,则进一步根据全局能量方程的损失函数最小化原理,判断路径损失L r(p,d)是否最小化,判定原则如下:
    Figure PCTCN2019102187-appb-100021
    其中,C(p,d)为路径的局部损失,
    Figure PCTCN2019102187-appb-100022
    为路径的上一步损失,L r,min(p-r)为路径上一步损失的最小损失,由此可判定该点是否为角点,若被检测出的两角点相邻,则去掉L r(p,d)较大的角点,从而检测出第二次影像匹配集中影像对的角点;
    对于
    Figure PCTCN2019102187-appb-100023
    影像角点集合中的每一个点,在
    Figure PCTCN2019102187-appb-100024
    影像中的相应搜索区域内寻找与之相匹配的角点,将得到的两个匹配点集合的交集称为初始匹配点集K 1,在所述初始匹配点集K 1中,对在
    Figure PCTCN2019102187-appb-100025
    影像对角点都互相集中的角点,在相应搜索区域内寻找匹配点,计算这个点与搜索区域内每个候选匹配点的相似度,选取与之相似度最大的候选匹配点为它的匹配点,得到角点的匹配点集合K 2
    利用角点的匹配点集合K 2,根据极限几何约束关系得到
    Figure PCTCN2019102187-appb-100026
    影像的极线对应关系,生成极限对应关系集K 3,按灰度对K 3内的极线进行分段,每条极线被分为若干灰度分段,生成的灰度分段集合为K 4,利用动态规划算法建立灰度分段之间的对应关系,利用线性插值的方法在省对应的灰度分段之间建立各像素点之间的对应关系,从而实现影像间所有像素点的密集匹配,得到所述第三次影像匹配集。
  8. 如权利要求1至7中任意一项所述的影像匹配方法,其特征在于,所述 执行三维重建,得到重构后的场景影像包括:
    经过密集匹配后,利用所述第三次影像匹配集中所有像素点存在的对应关系,计算出场景的景深,利用预设的3Dmax软件对场景进行重构,恢复场景空间的三维几何信息,得到重构后的场景影像。
  9. 一种影像匹配装置,其特征在于,所述装置包括存储器和处理器,所述存储器上存储有在所述处理器上运行的影像匹配程序,所述影像匹配程序被所述处理器执行时实现如下步骤:
    根据航拍仪拍摄的场景影像,生成影像成像图,利用尺度不变特征变换法对影像成像图进行初次影像匹配,生成初次影像匹配集;
    基于所述初次影像匹配集,生成核线影像并计算所述核线影像之间的重叠度,完成第二次影像匹配,生成第二次影像匹配集;
    基于所述的第二次影像匹配集,建立影像间所有像素点的密集匹配,生成第三次影像匹配集,并执行三维重建,得到重构后的场景影像。
  10. 如权利要求9所述的影像匹配装置,其特征在于,根据航拍仪拍摄的场景影像,生成影像成像图,包括:
    根据航拍仪器拍摄所述场景影像时所记录的参数,包括低精度位置、姿态信息及测区的概略高度,利用航拍下物体成像的模型公式将航拍仪器拍摄的存在重叠的场景影像恢复出各自的位置,生成n幅影像成像图,其中,所述物体成像的模型公式如下:
    sm=KR[I-C]M,
    其中,s为尺度系数,m为像点坐标,M为物点坐标,K为航拍工具内的参数矩阵,R为旋转矩阵,C为投影中心位置向量,I为3阶单位矩阵。
  11. 如权利要求10所述的影像匹配装置,其特征在于,所述利用尺度不变特征变换法对影像成像图进行初次影像匹配,生成初次影像匹配集,包括:
    将所述n幅影像成像图组成的影像集转换为对应的无向图边集合E;
    在所述无向图边集合E中,采用尺度不变特征变换算法进行影像匹配,并在影像匹配中,对于影像(I i、I j)∈E,若I i、I j两影像的匹配点数量小于阈值N 1,则将(I i、I j)从所述无向图边集合中剔除,若I i、I j两影像的匹配点数量大于阈值N 1,则保留该成像图对,生成
    Figure PCTCN2019102187-appb-100027
    影像对,产生所述初次影像匹配集。
  12. 如权利要求11所述的影像匹配装置,其特征在于,基于所述初次影 像匹配集,生成核线影像并计算所述核线影像之间的重叠度,完成第二次影像匹配,生成第二次影像匹配集,包括:
    (a)利用尺度不变特征变换法算法对所述
    Figure PCTCN2019102187-appb-100028
    影像对进行点特征提取,获取分布均匀的高精度同名点后,利用基于RANSAC策略进行基础矩阵估计,得到基础矩阵;
    (b)利用所述基础矩阵确定每组同名点对应的同名核线;
    (c)根据核线必相交于核点的原理,采用最小二乘法确定
    Figure PCTCN2019102187-appb-100029
    影像对的核点坐标,根据核点坐标生成影像间核线的快速映射,并沿核线方向采用双线性内插法进行核线重采样,完成核线影像制作并匹配重新生成
    Figure PCTCN2019102187-appb-100030
    影像对,产生所述第二次影像匹配集。
  13. 如权利要求12所述的影像匹配装置,其特征在于,所述(c)包括:基于所述基础矩阵,构建旋转矩阵和投影矩阵,将所述旋转矩阵分为x、y、z轴,各为
    Figure PCTCN2019102187-appb-100031
    其中,
    Figure PCTCN2019102187-appb-100032
    的表达方式如下:
    Figure PCTCN2019102187-appb-100033
    并得到投影矩阵
    Figure PCTCN2019102187-appb-100034
    Figure PCTCN2019102187-appb-100035
    其中,
    Figure PCTCN2019102187-appb-100036
    为相机left的相机参数,
    Figure PCTCN2019102187-appb-100037
    为相机right的相机参数,t left、t right分别是相机left和相机right的相机参数在x、y、z轴的分量;
    根据核线必交于核点原理,由投影矩阵的计算结果,采用最小二乘法确定影像的核点坐标(x p,y p);
    根据共线条件方程推导出中心投影,在核线采样时,计算出相邻两条核线之间的夹角关系,完成中心投影影像上每条核线的确定;
    利用上述生成的基础矩阵,可以确定每个影像对
    Figure PCTCN2019102187-appb-100038
    所对应的核线,分别根据
    Figure PCTCN2019102187-appb-100039
    中每个核像的核点坐标确定该点的核线,完成同影像对之间的核线对应,得到
    Figure PCTCN2019102187-appb-100040
    的核线方程为:
    Figure PCTCN2019102187-appb-100041
    其中,(x p,y p)是上述计算的核点坐标,(x base,y base)为中心投影影像的基准坐标,同理得到
    Figure PCTCN2019102187-appb-100042
    的核线方程;
    以所述核线方程为基础,建立起每个影像对
    Figure PCTCN2019102187-appb-100043
    的核线映射之后,按 照双线性内插法的重采样规则,生成核线影像并计算重叠度,将核线影像的重叠度小于阈值N 2的影像对丢弃,生成
    Figure PCTCN2019102187-appb-100044
    影像对,得到所述第二次影像匹配集。
  14. 如权利要求13所述的影像匹配装置,其特征在于,所述基于所述的第二次影像匹配集,建立影像间所有像素点的密集匹配,包括:
    在所述第二次影像匹配集的基础上,利用最小同值分割吸收核的角点检测算法分别提取
    Figure PCTCN2019102187-appb-100045
    的角点,形成角点的匹配点集合,结合对极几何约束、动态规划算法建立起
    Figure PCTCN2019102187-appb-100046
    影像间所有像素点的密集匹配。
  15. 如权利要求9至14中任意一项所述的影像匹配装置,其特征在于,所述执行三维重建,得到重构后的场景影像包括:
    经过密集匹配后,利用所述第三次影像匹配集中所有像素点存在的对应关系,计算出场景的景深,利用预设的3Dmax软件对场景进行重构,恢复场景空间的三维几何信息,得到重构后的场景影像。
  16. 一种计算机可读存储介质,其特征在于,所述计算机可读存储介质上存储有影像匹配程序,所述影像匹配程序被处理器执行时实现如下步骤:
    根据航拍仪拍摄的场景影像,生成影像成像图,利用尺度不变特征变换法对影像成像图进行初次影像匹配,生成初次影像匹配集;
    基于所述初次影像匹配集,生成核线影像并计算所述核线影像之间的重叠度,完成第二次影像匹配,生成第二次影像匹配集;
    基于所述的第二次影像匹配集,建立影像间所有像素点的密集匹配,生成第三次影像匹配集,并执行三维重建,得到重构后的场景影像。
  17. 如权利要求16所述的计算机可读存储介质,其特征在于,根据航拍仪拍摄的场景影像,生成影像成像图,包括:
    根据航拍仪器拍摄所述场景影像时所记录的参数,包括低精度位置、姿态信息及测区的概略高度,利用航拍下物体成像的模型公式将航拍仪器拍摄的存在重叠的场景影像恢复出各自的位置,生成n幅影像成像图,其中,所述物体成像的模型公式如下:
    sm=KR[I-C]M,
    其中,s为尺度系数,m为像点坐标,M为物点坐标,K为航拍工具内的参数矩阵,R为旋转矩阵,C为投影中心位置向量,I为3阶单位矩阵。
  18. 如权利要求17所述的计算机可读存储介质,其特征在于,所述利用尺度不变特征变换法对影像成像图进行初次影像匹配,生成初次影像匹配集,包括:
    将所述n幅影像成像图组成的影像集转换为对应的无向图边集合E;
    在所述无向图边集合E中,采用尺度不变特征变换算法进行影像匹配,并在影像匹配中,对于影像(I i、I j)∈E,若I i、I j两影像的匹配点数量小于阈值N 1,则将(I i、I j)从所述无向图边集合中剔除,若I i、I j两影像的匹配点数量大于阈值N 1,则保留该成像图对,生成
    Figure PCTCN2019102187-appb-100047
    影像对,产生所述初次影像匹配集。
  19. 如权利要求18所述的计算机可读存储介质,其特征在于,基于所述初次影像匹配集,生成核线影像并计算所述核线影像之间的重叠度,完成第二次影像匹配,生成第二次影像匹配集,包括:
    (a)利用尺度不变特征变换法算法对所述
    Figure PCTCN2019102187-appb-100048
    影像对进行点特征提取,获取分布均匀的高精度同名点后,利用基于RANSAC策略进行基础矩阵估计,得到基础矩阵;
    (b)利用所述基础矩阵确定每组同名点对应的同名核线;
    (c)根据核线必相交于核点的原理,采用最小二乘法确定
    Figure PCTCN2019102187-appb-100049
    影像对的核点坐标,根据核点坐标生成影像间核线的快速映射,并沿核线方向采用双线性内插法进行核线重采样,完成核线影像制作并匹配重新生成
    Figure PCTCN2019102187-appb-100050
    影像对,产生所述第二次影像匹配集。
  20. 如权利要求15所述的计算机可读存储介质,其特征在于,所述执行三维重建,得到重构后的场景影像包括:
    经过密集匹配后,利用所述第三次影像匹配集中所有像素点存在的对应关系,计算出场景的景深,利用预设的3Dmax软件对场景进行重构,恢复场景空间的三维几何信息,得到重构后的场景影像。
PCT/CN2019/102187 2019-04-08 2019-08-23 影像匹配方法、装置及计算机可读存储介质 WO2020206903A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910274078.5 2019-04-08
CN201910274078.5A CN110135455B (zh) 2019-04-08 2019-04-08 影像匹配方法、装置及计算机可读存储介质

Publications (1)

Publication Number Publication Date
WO2020206903A1 true WO2020206903A1 (zh) 2020-10-15

Family

ID=67569487

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/102187 WO2020206903A1 (zh) 2019-04-08 2019-08-23 影像匹配方法、装置及计算机可读存储介质

Country Status (2)

Country Link
CN (1) CN110135455B (zh)
WO (1) WO2020206903A1 (zh)

Cited By (21)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111160377A (zh) * 2020-03-07 2020-05-15 深圳移动互联研究院有限公司 一种自带密钥机制的图像采集***及其佐证方法
CN112233228A (zh) * 2020-10-28 2021-01-15 五邑大学 基于无人机的城市三维重建方法、装置及存储介质
CN112381864A (zh) * 2020-12-08 2021-02-19 兰州交通大学 一种基于对极几何的多源多尺度高分辨率遥感影像自动配准技术
CN112446951A (zh) * 2020-11-06 2021-03-05 杭州易现先进科技有限公司 三维重建方法、装置、电子设备及计算机存储介质
CN112509109A (zh) * 2020-12-10 2021-03-16 上海影创信息科技有限公司 一种基于神经网络模型的单视图光照估计方法
CN113096168A (zh) * 2021-03-17 2021-07-09 西安交通大学 一种结合sift点和控制线对的光学遥感图像配准方法及***
CN113741510A (zh) * 2021-07-30 2021-12-03 深圳创动科技有限公司 一种巡检路径规划方法、装置以及存储介质
CN113867410A (zh) * 2021-11-17 2021-12-31 武汉大势智慧科技有限公司 一种无人机航拍数据的采集模式识别方法和***
CN113963132A (zh) * 2021-11-15 2022-01-21 广东电网有限责任公司 一种等离子体的三维分布重建方法及相关装置
CN114140575A (zh) * 2021-10-21 2022-03-04 北京航空航天大学 三维模型构建方法、装置和设备
CN114332349A (zh) * 2021-11-17 2022-04-12 浙江智慧视频安防创新中心有限公司 一种双目结构光边缘重建方法、***及存储介质
CN114419116A (zh) * 2022-01-11 2022-04-29 江苏省测绘研究所 一种基于网匹配的遥感影像配准方法及其***
CN114758151A (zh) * 2022-03-21 2022-07-15 辽宁工程技术大学 一种结合线特征与三角网约束的序列影像密集匹配方法
CN114972536A (zh) * 2022-05-26 2022-08-30 中国人民解放军战略支援部队信息工程大学 一种航空面阵摆扫式相机定位和标定方法
CN115063460A (zh) * 2021-12-24 2022-09-16 山东建筑大学 一种高精度自适应同名像素插值与优化方法
CN115661368A (zh) * 2022-12-14 2023-01-31 海纳云物联科技有限公司 一种图像匹配方法、装置、服务器及存储介质
CN116596844A (zh) * 2023-04-06 2023-08-15 北京四维远见信息技术有限公司 一种航飞质量检查方法、装置、设备及存储介质
CN116597184A (zh) * 2023-07-11 2023-08-15 中国人民解放军63921部队 最小二乘影像匹配方法
CN116612067A (zh) * 2023-04-06 2023-08-18 北京四维远见信息技术有限公司 航飞质量检查方法、装置、设备和计算机可读存储介质
CN117664087A (zh) * 2024-01-31 2024-03-08 中国人民解放军战略支援部队航天工程大学 垂轨环扫式卫星影像核线生成方法、***及设备
CN118070434A (zh) * 2024-04-22 2024-05-24 天津悦鸣腾宇通用机械设备有限公司 一种汽车零部件的工艺信息模型构建方法及***

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110135455B (zh) * 2019-04-08 2024-04-12 平安科技(深圳)有限公司 影像匹配方法、装置及计算机可读存储介质
CN111046906B (zh) * 2019-10-31 2023-10-31 中国资源卫星应用中心 一种面状特征点可靠加密匹配方法和***
CN112866504B (zh) * 2021-01-28 2023-06-09 武汉博雅弘拓科技有限公司 一种空三加密方法和***
CN114742869B (zh) * 2022-06-15 2022-08-16 西安交通大学医学院第一附属医院 基于图形识别的脑部神经外科配准方法及电子设备

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110090337A1 (en) * 2008-02-01 2011-04-21 Imint Image Intelligence Ab Generation of aerial images
CN104751451A (zh) * 2015-03-05 2015-07-01 同济大学 基于无人机低空高分辨率影像的密集点云提取方法
CN105847750A (zh) * 2016-04-13 2016-08-10 中测新图(北京)遥感技术有限责任公司 基于地理编码的无人机视频影像实时显示的方法及装置
CN106023086A (zh) * 2016-07-06 2016-10-12 中国电子科技集团公司第二十八研究所 一种基于orb特征匹配的航拍影像及地理数据拼接方法
CN108759788A (zh) * 2018-03-19 2018-11-06 深圳飞马机器人科技有限公司 无人机影像定位定姿方法及无人机
CN110135455A (zh) * 2019-04-08 2019-08-16 平安科技(深圳)有限公司 影像匹配方法、装置及计算机可读存储介质

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104915965A (zh) * 2014-03-14 2015-09-16 华为技术有限公司 一种摄像机跟踪方法及装置
CN107492127B (zh) * 2017-09-18 2021-05-11 丁志宇 光场相机参数标定方法、装置、存储介质和计算机设备

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110090337A1 (en) * 2008-02-01 2011-04-21 Imint Image Intelligence Ab Generation of aerial images
CN104751451A (zh) * 2015-03-05 2015-07-01 同济大学 基于无人机低空高分辨率影像的密集点云提取方法
CN105847750A (zh) * 2016-04-13 2016-08-10 中测新图(北京)遥感技术有限责任公司 基于地理编码的无人机视频影像实时显示的方法及装置
CN106023086A (zh) * 2016-07-06 2016-10-12 中国电子科技集团公司第二十八研究所 一种基于orb特征匹配的航拍影像及地理数据拼接方法
CN108759788A (zh) * 2018-03-19 2018-11-06 深圳飞马机器人科技有限公司 无人机影像定位定姿方法及无人机
CN110135455A (zh) * 2019-04-08 2019-08-16 平安科技(深圳)有限公司 影像匹配方法、装置及计算机可读存储介质

Cited By (34)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111160377A (zh) * 2020-03-07 2020-05-15 深圳移动互联研究院有限公司 一种自带密钥机制的图像采集***及其佐证方法
CN112233228B (zh) * 2020-10-28 2024-02-20 五邑大学 基于无人机的城市三维重建方法、装置及存储介质
CN112233228A (zh) * 2020-10-28 2021-01-15 五邑大学 基于无人机的城市三维重建方法、装置及存储介质
CN112446951A (zh) * 2020-11-06 2021-03-05 杭州易现先进科技有限公司 三维重建方法、装置、电子设备及计算机存储介质
CN112446951B (zh) * 2020-11-06 2024-03-26 杭州易现先进科技有限公司 三维重建方法、装置、电子设备及计算机存储介质
CN112381864A (zh) * 2020-12-08 2021-02-19 兰州交通大学 一种基于对极几何的多源多尺度高分辨率遥感影像自动配准技术
CN112509109A (zh) * 2020-12-10 2021-03-16 上海影创信息科技有限公司 一种基于神经网络模型的单视图光照估计方法
CN113096168A (zh) * 2021-03-17 2021-07-09 西安交通大学 一种结合sift点和控制线对的光学遥感图像配准方法及***
CN113096168B (zh) * 2021-03-17 2024-04-02 西安交通大学 一种结合sift点和控制线对的光学遥感图像配准方法及***
CN113741510A (zh) * 2021-07-30 2021-12-03 深圳创动科技有限公司 一种巡检路径规划方法、装置以及存储介质
CN114140575A (zh) * 2021-10-21 2022-03-04 北京航空航天大学 三维模型构建方法、装置和设备
CN113963132A (zh) * 2021-11-15 2022-01-21 广东电网有限责任公司 一种等离子体的三维分布重建方法及相关装置
CN114332349A (zh) * 2021-11-17 2022-04-12 浙江智慧视频安防创新中心有限公司 一种双目结构光边缘重建方法、***及存储介质
CN113867410B (zh) * 2021-11-17 2023-11-03 武汉大势智慧科技有限公司 一种无人机航拍数据的采集模式识别方法和***
CN114332349B (zh) * 2021-11-17 2023-11-03 浙江视觉智能创新中心有限公司 一种双目结构光边缘重建方法、***及存储介质
CN113867410A (zh) * 2021-11-17 2021-12-31 武汉大势智慧科技有限公司 一种无人机航拍数据的采集模式识别方法和***
CN115063460A (zh) * 2021-12-24 2022-09-16 山东建筑大学 一种高精度自适应同名像素插值与优化方法
CN114419116A (zh) * 2022-01-11 2022-04-29 江苏省测绘研究所 一种基于网匹配的遥感影像配准方法及其***
CN114419116B (zh) * 2022-01-11 2024-04-09 江苏省测绘研究所 一种基于网匹配的遥感影像配准方法及其***
CN114758151A (zh) * 2022-03-21 2022-07-15 辽宁工程技术大学 一种结合线特征与三角网约束的序列影像密集匹配方法
CN114758151B (zh) * 2022-03-21 2024-05-24 辽宁工程技术大学 一种结合线特征与三角网约束的序列影像密集匹配方法
CN114972536B (zh) * 2022-05-26 2023-05-09 中国人民解放军战略支援部队信息工程大学 一种航空面阵摆扫式相机定位和标定方法
CN114972536A (zh) * 2022-05-26 2022-08-30 中国人民解放军战略支援部队信息工程大学 一种航空面阵摆扫式相机定位和标定方法
CN115661368A (zh) * 2022-12-14 2023-01-31 海纳云物联科技有限公司 一种图像匹配方法、装置、服务器及存储介质
CN115661368B (zh) * 2022-12-14 2023-04-11 海纳云物联科技有限公司 一种图像匹配方法、装置、服务器及存储介质
CN116596844A (zh) * 2023-04-06 2023-08-15 北京四维远见信息技术有限公司 一种航飞质量检查方法、装置、设备及存储介质
CN116612067B (zh) * 2023-04-06 2024-02-23 北京四维远见信息技术有限公司 航飞质量检查方法、装置、设备和计算机可读存储介质
CN116596844B (zh) * 2023-04-06 2024-03-29 北京四维远见信息技术有限公司 一种航飞质量检查方法、装置、设备及存储介质
CN116612067A (zh) * 2023-04-06 2023-08-18 北京四维远见信息技术有限公司 航飞质量检查方法、装置、设备和计算机可读存储介质
CN116597184B (zh) * 2023-07-11 2023-09-22 中国人民解放军63921部队 最小二乘影像匹配方法
CN116597184A (zh) * 2023-07-11 2023-08-15 中国人民解放军63921部队 最小二乘影像匹配方法
CN117664087A (zh) * 2024-01-31 2024-03-08 中国人民解放军战略支援部队航天工程大学 垂轨环扫式卫星影像核线生成方法、***及设备
CN117664087B (zh) * 2024-01-31 2024-04-02 中国人民解放军战略支援部队航天工程大学 垂轨环扫式卫星影像核线生成方法、***及设备
CN118070434A (zh) * 2024-04-22 2024-05-24 天津悦鸣腾宇通用机械设备有限公司 一种汽车零部件的工艺信息模型构建方法及***

Also Published As

Publication number Publication date
CN110135455A (zh) 2019-08-16
CN110135455B (zh) 2024-04-12

Similar Documents

Publication Publication Date Title
WO2020206903A1 (zh) 影像匹配方法、装置及计算机可读存储介质
TWI777538B (zh) 圖像處理方法、電子設備及電腦可讀儲存介質
CN107705333B (zh) 基于双目相机的空间定位方法及装置
US11928800B2 (en) Image coordinate system transformation method and apparatus, device, and storage medium
US11521311B1 (en) Collaborative disparity decomposition
CA2826534C (en) Backfilling points in a point cloud
WO2015135323A1 (zh) 一种摄像机跟踪方法及装置
EP3274964B1 (en) Automatic connection of images using visual features
US9286539B2 (en) Constructing contours from imagery
CN112686877B (zh) 基于双目相机的三维房屋损伤模型构建测量方法及***
CN111127524A (zh) 一种轨迹跟踪与三维重建方法、***及装置
US20160163114A1 (en) Absolute rotation estimation including outlier detection via low-rank and sparse matrix decomposition
CN115439607A (zh) 一种三维重建方法、装置、电子设备及存储介质
WO2021244161A1 (zh) 基于多目全景图像的模型生成方法及装置
WO2023024393A1 (zh) 深度估计方法、装置、计算机设备及存储介质
WO2021004416A1 (zh) 一种基于视觉信标建立信标地图的方法、装置
KR101593316B1 (ko) 스테레오 카메라를 이용한 3차원 모델 재구성 방법 및 장치
CN116129037B (zh) 视触觉传感器及其三维重建方法、***、设备及存储介质
CN112150518B (zh) 一种基于注意力机制的图像立体匹配方法及双目设备
US8509522B2 (en) Camera translation using rotation from device
WO2021142843A1 (zh) 图像扫描方法及装置、设备、存储介质
CN112634366A (zh) 位置信息的生成方法、相关装置及计算机程序产品
CN112002007A (zh) 基于空地影像的模型获取方法及装置、设备、存储介质
Budianti et al. Background blurring and removal for 3d modelling of cultural heritage objects
JP6156922B2 (ja) 三次元データ生成装置、三次元データ生成方法、及びプログラム

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19924318

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19924318

Country of ref document: EP

Kind code of ref document: A1