Planetary orbiter image matching strategy constructed based on image partitioning and corresponding relation

Abstract

The invention relates to a planetary orbiter image matching strategy constructed based on image partitioning and a corresponding relation, and the strategy comprises the following steps: S1, obtaining a planetary orbiter image matching pair, and obtaining an overlapping region; s2, performing grid division on the overlapping region by adopting a row-to-column division method to obtain a final block image; s3, the block images of the left image and the right image correspond to each other, and finally block images with the corresponding relation are obtained; s4, performing matching by adopting a GPU acceleration-based feature matching algorithm to obtain matching points, and aggregating the matching points to obtain a final matching result; and S5, performing process parallel processing on the multiple groups of to-be-matched planetary orbiter image pairs, performing parallel operation from S1 to S5, performing thread parallel processing on a single group of to-be-matched planetary orbiter image pairs, and performing parallel operation from S2 to S4. Compared with the prior art, the method has the advantages that the matching precision and the matching efficiency of high-resolution planet orbiter image matching are both considered, and the like.

Description

Technical Field

[0001] The present invention relates to the field of image matching, and in particular to a planetary orbiter image matching strategy based on image segmentation and corresponding relationships. Background Art

[0002] Planetary orbiter images are an important data basis for deep space exploration and are widely used in planetary geological research, landing site selection, exploration path planning and other fields. Image matching, as a key step in photogrammetric processing, provides support for regional network adjustment, image registration and other links. However, commonly used matching methods are mainly aimed at regular small-scale images, while planetary orbiter images have the characteristics of large size, sparse texture and long strip shape. When traditional matching methods are directly used, they often face the problems of limited computing resources and insufficient algorithm carrying capacity. In addition, the redundant features in large-size images participate in matching, which will reduce matching efficiency and accuracy. Therefore, it is necessary to combine matching methods with matching strategies to effectively extract image homonymous points.

[0003] Common matching strategies use multi-scale pyramid images to extract seed points and gradually encrypt matching points. However, due to the sparse and similar feature textures of planetary orbiter images, it is difficult to detect high-quality seed points on multi-scale images, resulting in matching errors. In addition, by extracting overlapping areas of images and dividing and matching them based on circumscribed rectangles, non-overlapping areas may be introduced or overlapping areas may be omitted for complex polygonal shapes, resulting in mismatches. Although the matching strategy based on the imaging model can effectively improve the matching accuracy, its scope of application is limited due to its high reliance on the positioning accuracy of geometric information. Summary of the invention

[0004] The purpose of the present invention is to provide a planetary orbiter image matching strategy based on image segmentation and corresponding relationship in order to achieve both matching accuracy and matching efficiency for high-resolution planetary orbiter image matching.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A planetary orbiter image matching strategy based on image segmentation and corresponding relationship construction, the strategy comprising the following steps:

[0007] S1. Obtain matching pairs of planetary orbiter images and obtain overlapping areas;

[0008] S2, using a row-to-column division method to divide the overlapping area of ​​the image pair into grids, and based on the divided grids, dividing the images in the overlapping area into blocks to obtain a final block image composed of image blocks;

[0009] S3, matching the block images of the left image and the right image in the overlapping area, determining the block correspondence relationship, and finally obtaining the block images with the corresponding relationship;

[0010] S4, using a feature matching algorithm based on GPU acceleration, matching is performed based on the final corresponding block images to obtain matching points, and the matching points on all block images are aggregated to obtain the final matching result;

[0011] S5. Process parallel processing is adopted for multiple groups of planetary orbiter image pairs to be matched, and S1 to S5 are operated in parallel. Thread parallel processing is adopted for a single group of planetary orbiter image pairs to be matched, and S2 to S4 are executed in parallel.

[0012] Furthermore, the specific steps of using the row-to-column division method to grid the overlapping area of ​​the image pair are as follows:

[0013] Obtain the planetary orbiter image pair to be matched, obtain the overlapping area of ​​the projected image pair, and project the image pair into the image space through the imaging equation.

[0014] Furthermore, the imaging equation is:

[0015]

[0016] Among them, (x, y) is the focal plane coordinate of the image point, f is the focal length, (X, U, Z) is the ground point coordinate corresponding to the image point, (X s ,Y s ,Z s ) is the coordinate of the photography center in the J2000 coordinate system, R orbiter is the rotation matrix from the camera coordinate system to the orbiter coordinate system, R Fixed is the rotation matrix from the orbiter coordinate system to the satellite-fixed coordinate system, R J2000 is the rotation matrix from the Star-fixed coordinate system to the J2000 coordinate system, and λ is the scale factor.

[0017] Furthermore, the specific steps of S2 are:

[0018] The left image of the overlapping area is divided in the row direction, an initial rectangle in the left image is set, and a polygon after division is obtained by intersecting the initial rectangle in the left image with the overlapping area, and the circumscribed rectangle of the polygon after division is used as the final row direction division result of the left image;

[0019] The right image of the overlapping area is divided in the row direction, the height of the initial rectangle in the right image is calculated based on the fixed height of the initial rectangle in the left image, a buffer is added to the height of the initial rectangle in the right image, and the initial rectangle in the right image after adding the buffer is intersected with the overlapping area to obtain a divided polygon, and the circumscribed rectangle of the divided polygon is used as the final row direction division result of the right image;

[0020] After the row division is completed, the column division is performed. In the left image, the fixed height of the initial rectangle in the left image is used as the side length to construct a square and arrange it from left to right to form a grid, and the column division result of the left image is obtained, and the block image of the left image is obtained. In the right image, the number of rows and columns of the rectangle when the column is divided in the right image is calculated based on the height of the initial rectangle in the right image, and the column division result of the right image is obtained, and the block image of the right image is obtained;

[0021] Based on the point set formed by the overlapping area, the principal components of the overlapping area are determined, the rotation angle between the overlapping areas of the left image and the right image is calculated, and then the block image of the right image is rotated based on the rotation angle, and the circumscribed rectangle of the rotated block image is used as the final block image.

[0022] Furthermore, the initial rectangle in the left image is set as follows:

[0023] In the left image, the maximum width in the image column direction is calculated based on the contour points of the overlapping area. A rectangle is constructed with the maximum width as the length and a fixed value as the height, which is called the initial rectangle in the left image. In the row direction of the left image, the initial rectangle in the left image is used to cover the overlapping area. When the last initial rectangle is laid, if the remaining number of rows is not enough to meet the height requirement of the initial rectangle in the left image, the remaining number of rows is used as the height of the initial rectangle in the left image.

[0024] Furthermore, the height of the initial rectangle in the right image is:

[0025]

[0026] Among them, Block src Y represents the fixed height of the initial rectangle in the left image, Block dst Y represents the height of the initial rectangle in the right image, GSD Lrow , GSD Rrow Represents the spatial resolution of the left image and the right image in the row direction respectively.

[0027] Furthermore, the initial rectangles in the right image after adding the buffer zone overlap each other, and the overlap size is half of the buffer zone size.

[0028] Furthermore, the number of rows and columns of the rectangle when dividing the columns in the right image is:

[0029]

[0030] Among them, blockSize src The size of the block image on the left image, Block dst X is the column dimension of the block image on the right image, GSD Rcol is the resolution of the right image in the upper column direction.

[0031] Furthermore, the rotation angle between the overlapping area of ​​the left image and the right image is:

[0032]

[0033] Where θ represents the rotation angle, v L and v R It is the eigenvector matrix on the left and right images, indicating the main direction.

[0034] Furthermore, the specific steps of S3 are:

[0035] According to the final block image of the left image, the block images are numbered from top to bottom and from left to right. The pixel with the minimum number in the row direction in the overlapping area of ​​the left image is marked as Based on the imaging model corresponding to the left image, the corresponding ground point coordinates are calculated, recorded as (B0, L0). The image point coordinates of point (B0, L0) are calculated through the imaging model of the right image and the row direction coordinates of the image point coordinates are obtained. If the image point coordinates are close to the minimum number in the row direction in the overlapping area, they are marked as The row division direction of the final block image of the right image is from top to bottom, and the column division direction is from left to right; if the image point coordinate is close to the maximum number of the row direction in the overlapping area, it is marked as The row division direction of the final block image of the right image is from bottom to top, and the column division direction is from right to left. Based on the division direction, the final block images of the left image and the right image in the overlapping area are matched, the block correspondence is determined, and the final block images with corresponding relationships are obtained.

[0036] Furthermore, the specific steps of S4 are:

[0037] SiftGPU is an efficient SIFT algorithm implementation based on GPU. On GPU, grid parallel computing is used to efficiently build Gaussian difference (DoG) pyramids, accurately locate key points, and generate 128-dimensional feature descriptors. When matching block images, SiftGPU is used to determine the matching points (local matching points) between image blocks. After obtaining the matching points of the block images, the RANSAC algorithm is used to eliminate gross errors. Finally, the coordinates of all local matching points are normalized to generate the final matching results.

[0038] Furthermore, the specific steps of S5 are:

[0039] For multiple sets of planetary orbiter image pairs to be matched, multiple images are matched by process parallel processing, and for a single set of planetary orbiter image pairs to be matched, thread parallel processing is adopted.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] The present invention utilizes the "row to column" polygon overlapping area grid division method to make the range of the block image closer to the boundary, avoid introducing a large number of non-overlapping areas or missing some overlapping areas in the block process, reduce redundant features involved in matching, and introduce a feature matching algorithm based on GPU acceleration for the block image matching process, thereby achieving improved matching efficiency and matching accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a flow chart of planetary orbiter image matching of the present invention;

[0043] Figure 2 It is a schematic diagram of the overlapping area image segmentation process;

[0044] Figure 3 Optimize flow chart for multi-level matching efficiency;

[0045] Figure 4 This is the matching result diagram of three sets of orbiter images;

[0046] Figure 5 is the accuracy diagram of the matching results of each image pair in the test area, where Figure 5 (a) is the matching result of area 1. Figure 5 (b) is the matching result of area 2. DETAILED DESCRIPTION

[0047] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0048] The present invention relates to a planetary orbiter image matching strategy based on row-to-column image segmentation and fast correspondence relationship construction, specifically including the following steps: firstly, the object-side overlapping area of ​​the matching image pair is obtained through ephemeris information and projected to the image side, and the polygon overlapping area grid division method "from row to column" is used to make the range of the segmented image more closely fit the boundary, so as to avoid introducing a large number of non-overlapping areas or missing some overlapping areas in the segmentation process; in the image segmentation process, the relative position relationship in the matching image pair is determined by the extreme point in the row direction to determine the grid division order, and the correspondence relationship between the segmented images is quickly constructed based on this as an index; then, the feature matching algorithm is used to extract the same-name points for all paired segmented images, and finally a complete matching result is obtained by aggregation; in order to optimize the running efficiency of the matching process, process parallel processing is adopted between multiple image pairs, thread parallel processing is adopted for the matching process within a single group of image pairs, and a feature matching algorithm based on GPU acceleration is introduced for the segmented image matching process, so as to improve the matching efficiency. Compared with the prior art, the present invention can efficiently solve the large-scale and large-scale planetary orbiter image matching problems, and has the advantages of good stability and practicality.

[0049] The specific steps of the present invention are:

[0050] S1. Acquire a pair of planetary orbiter images to be matched;

[0051] S2, using a row-to-column division method to divide the overlapping area of ​​the image pair into grids, and based on the divided grids, dividing the images in the overlapping area into blocks to obtain a final block image composed of image blocks;

[0052] S3, matching the final block images of the left image and the right image in the overlapping area to determine the block correspondence relationship;

[0053] S4, based on the final block image and the block correspondence, determine the matching points between the image blocks to obtain the matching results;

[0054] S5. Process parallel processing is adopted for multiple groups of planetary orbiter image pairs to be matched, and S1 to S5 are operated in parallel. Thread parallel processing is adopted for a single group of planetary orbiter image pairs to be matched, and S2 to S4 are executed in parallel.

[0055] The advantage of the present invention is that it can effectively cope with high-resolution planetary orbiter image matching, overcome the matching problem of image pairs from different regions and overlapping regions of different shapes, and take into account both matching accuracy and matching efficiency. The process of the entire matching strategy is as follows Figure 1 shown.

[0056] The overlapping area of ​​the complex polygon of the matching image pair is divided into grids by the “row to column” division method, and the image in the overlapping area is divided into image blocks. The overlapping area image block process is as follows: Figure 2 shown.

[0057] The specific steps of S1 and S2 are:

[0058] (1) Overlapping area acquisition

[0059] For the orbiter image pair to be matched, the corresponding prior information (e.g., ephemeris information, metadata) is used to obtain the imaging range of the image in the object space, and then the overlapping area of ​​the image pair is calculated. In order to facilitate the subsequent grid division and image segmentation, it is projected to the image space through the imaging equation. Among them, the planetary orbiter image is usually a linear array push-broom image with the characteristics of multi-center imaging. The imaging model can be constructed based on the collinear condition equation through the internal and external orientation elements of the sensor:

[0060]

[0061] Among them, (x, y) is the focal plane coordinate of the image point, f is the focal length, (X, Y, Z) is the ground point coordinate corresponding to the image point, (X s ,Y s ,Z s ) is the coordinate of the photography center in the J2000 coordinate system. Orbiter is the rotation matrix from the camera coordinate system to the orbiter coordinate system, R Fixed is the rotation matrix from the orbiter coordinate system to the satellite-fixed coordinate system, R J2000 is the rotation matrix from the Star-fixed coordinate system to the J2000 coordinate system, and λ is the scale factor.

[0062] (2) Row division of overlapping areas

[0063] The overlapping area is first divided in the row direction, and the division is performed in the overlapping area of ​​the left image and the right image respectively. The block setting of the left image is used as the standard, and the relevant parameter settings of the right image are dynamically adjusted according to the difference in prior information between the two images to ensure the accuracy and consistency of the division results. The specific process of division is as follows: First, in the left image, the maximum width in the image column direction is calculated according to the contour points of the overlapping area. With this as the length, and a fixed value as the height to build a rectangle, called the initial rectangle. In the row direction, the overlapping area is covered with the initial rectangle. When the last initial rectangle is laid, if the remaining number of rows is not enough to meet the height requirement of the rectangle, the remaining number of rows is used as the height of the rectangle. The divided polygons are obtained by intersecting these rectangles with the overlapping area, and the circumscribed rectangle of each divided polygon is used as the final row direction division result, called the "row division rectangle".

[0064] In the right image, the process of row division is similar to that of the left image, but it is related to factors such as resolution difference and positioning deviation between the image pairs. The number of rows in the "row division rectangle" of the right image needs to be adjusted according to the resolution difference between the two images and the fixed number of rows in the left image to ensure the rationality and correspondence of the division results. The calculation method is as follows:

[0065]

[0066] Among them, Block src Y represents the fixed height of the initial rectangle in the left image, Block dst Y represents the height of the initial rectangle in the right image. GSD Lrow , GSD Rrow Represents the spatial resolution of the left image and the right image in the row direction respectively.

[0067] Due to a certain positioning deviation between the left and right images, when constructing the "row division rectangle" on the right image, a buffer zone of 100m is added to the height of the rectangle, which makes the "row division rectangles" overlap, and the overlap size is half of the buffer size.

[0068] (3) Column partitioning based on row partitioning results

[0069] After the row division is completed, the column division is performed. In the left image, the height of the row division rectangle is used as the side length to construct a square, and it is arranged from left to right to form a grid to achieve the column division. If the number of columns remaining in the division process is not enough to form a complete square, the remaining columns are used to construct a rectangle to complete the division.

[0070] In the right image, the column division is related to the resolution difference between the matching image pairs, the rotation angle between the overlapping areas of the left and right images, and the positioning deviation of the images. In order to compensate for the positioning deviation in the column direction, a buffer is also added in the column division of the right image. The number of rows and columns of the rectangle during column division is calculated according to the following formula:

[0071]

[0072] Based on the point set formed by the overlapping area of ​​each image in the matching image pair, the principal components of the overlapping area on the two images are determined, and the rotation angle between the overlapping areas of the left image and the right image can be calculated. Then, the entire image block is rotated to achieve angular alignment, as shown in the following formula:

[0073]

[0074] Among them, (x i ,y i ) stores the set of points that make up the overlapping area, θ represents the rotation difference of the overlapping area between the two images, [x′ ,y′] T It is the coordinate of the block image on the right image after the rotation relationship is restored.

[0075] After restoring the rotation relationship, the bounding rectangle of the rotated block image is used to represent the final block image range, which facilitates subsequent image block matching and enhances the stability of the buffer zone.

[0076] The steps of S3 are:

[0077] Before matching the block images, it is necessary to determine the correspondence between the blocks, which will provide preliminary geometric constraints, effectively filter out redundant feature points, and reduce the number of invalid matches.

[0078] The division direction of the left image is fixed: from top to bottom, and from left to right. The image blocks can be numbered and the corresponding relationship can be determined by indexing the position of the image blocks. Most planetary orbiters are imaged in near-polar orbits. Since the flight direction of the orbiter is always along the north-south direction, there is a certain regularity in the overlapping areas of the orbiter image pairs. Specifically, the extreme values ​​of latitude in the overlapping area usually correspond to the extreme values ​​of the pixel coordinates in the row direction of the image, that is, the top or bottom of the overlapping area, while the extreme values ​​of longitude usually correspond to the extreme values ​​of the pixel coordinates in the column direction of the image, that is, the left or right side of the overlapping area. Therefore, the division direction of the right image can be determined by a single extreme point in the row direction, thereby further clarifying the corresponding relationship between image blocks.

[0079] Assume that the pixel with the minimum value in the row direction in the overlapping area on the left image is marked as The corresponding ground point coordinates are calculated by the imaging model corresponding to the left image, and are recorded as (B0, L0). Then, the image point coordinates corresponding to the object point are calculated by the imaging model of the right image and its row direction coordinates are obtained. At this time, two situations may occur: one is that the image point is located near the minimum value of the row direction in the overlapping area, which is marked as This indicates that the subsequent row division on the right image should be performed from top to bottom; the other is that the image point is located near the maximum value of the row direction in the overlapping area, marked as This indicates that the row division should be done from bottom to top. Due to the characteristics of the near-polar orbit, the division in the column direction is related to the row direction. If the row division on the right image is done from top to bottom, the column division will be done from left to right; otherwise, both directions will be reversed.

[0080] In S4, perform: fast feature matching between block images

[0081] SiftGPU is an efficient SIFT algorithm implementation based on GPU. On GPU, grid parallel computing is used to efficiently build Gaussian difference (DoG) pyramids, accurately locate key points, and generate 128-dimensional feature descriptors. When matching block images, SiftGPU is used to determine the matching points (local matching points) between image blocks. After obtaining the matching points of the block images, the RANSAC algorithm is used to eliminate gross errors. Finally, the coordinates of all local matching points are normalized to generate the final matching results.

[0082] S5. The steps for the entire matching process are as follows:

[0083] The present invention starts from multiple levels of the matching process to further improve the efficiency of large-scale image matching. Figure 3 As shown in the figure, the proposed efficiency optimization method combines GPU acceleration and distributed parallel technology, and mainly includes three modules: multi-node image pair matching based on Message Passing Interface (MPI) process parallelism, image segmentation and block correspondence relationship establishment based on Open Multi-Processing (OpenMP) thread parallelism, and block image feature matching based on SiftGPU and OpenMP parallelism. Image segments, image pairs, and image pairs within a region are regarded as threads, processes, and tasks, respectively. Among them, multiple threads constitute a process, and a matching task consists of multiple processes.

[0084] In the present invention, the processes between multiple groups of matching pairs are parallel. When image matching is carried out for a certain area, there are often multiple image pairs. The matching of each image pair is regarded as an independent process, and MPI is used to realize inter-process communication. Each matching task uses the file name as an index for message transmission to realize parallel matching of multiple image pairs. Initialize n nodes, select the first node as the main node, responsible for dynamic task scheduling; nodes 2 to n are used as working nodes to execute the parallel tasks of each process. In the main node, n-1 matching tasks are assigned to the working nodes in the form of messages. In each working node, a pair of image matching tasks are performed. When the matching task is completed, each working node will send a "complete" message to the main node, and the main node will assign new matching tasks to the idle working nodes until all image pairs in the area are matched. This dynamic task allocation mechanism enables real-time optimization of resource utilization between nodes, thereby improving load balancing and maximizing computing efficiency. The threads within the image matching pair are parallel. Inside each working node, a pair of image matching is carried out, that is, a process is run. The steps of image segmentation, establishing the correspondence between segmented images, and feature matching of segmented images are carried out respectively. Each step is executed serially. Each step involves related processing for multiple segmented images, which can be regarded as multiple thread tasks. For each thread task, OpenMP is used to parallelize the task and assign it to multiple CPU cores, making full use of computing resources and giving full play to the CPU's capabilities. When all threads complete their tasks, it means that the image matching task is completed.

[0085] The following is a specific experiment:

[0086] (1) Experimental data

[0087] In order to verify the matching strategy proposed in this paper, the Lunar Reconnaissance Orbiter Narrow Angle Camera (LRO NAC) and the Mars Reconnaissance Orbiter Context Camera (MRO CTX) images were selected as experimental data. The experimental data consists of two parts. The first part is three groups of representative discrete image pairs with different overlapping area shapes; the other part is a set of images covering the Apollo 17 landing area (Apollo 17) and the area near the Mawrth Vallis on Mars. 18 and 21 images were selected from the two areas, respectively, forming 20 and 37 stereo pairs. The single-group image pair matching experiment is mainly used to verify the matching quality based on the evaluation index; the multi-image matching experiment in the region is mainly used to evaluate the stability and effectiveness of the proposed matching strategy.

[0088] (2) Experimental results

[0089] The matching results are evaluated by combining qualitative and quantitative methods. Quantitative evaluation indicators include running time, distribution quality, correct matching rate, effective matching pairs, etc. In addition, a visualization example of the matching effect is used to conduct a more comprehensive qualitative analysis.

[0090] 1) Running time: This is an indicator used to evaluate the efficiency of completing the matching. The total running time of all test images is counted, and the average value of each test image pair is given.

[0091] 2) Distribution quality: The evaluation of the matching point distribution quality is based on the geometric characteristics of the Delaunay triangle. The specific calculation method is:

[0092]

[0093] Where n is the total number of triangles, A i and max(J i ) represent the area and maximum angle of the ith triangle respectively. D is a quantitative description of the geometric uniformity of the triangle. The smaller the value of D, the more uniform the distribution of points in the region.

[0094] 3) Matching accuracy based on transformation model: For each local block image pair, the matching points are used to estimate its transformation model. Considering the resolution difference and pixel aspect ratio between planetary orbit images, the threshold is set to 2 pixels. Then, the Euclidean distance between matching points is introduced as the second constraint. The distance of the correct matching points should be relatively consistent, so the median of the distance of all matching point pairs is selected as the threshold. Considering the geometric distortion and resolution difference of the image, 10% of the threshold is set as the tolerance. If the matching point pair meets the above two conditions at the same time, the matching point pair is considered correct. And the matching accuracy is calculated based on this.

[0095] 4) Matching accuracy based on back-projection: Using the rigorous imaging model after regional network adjustment, the triangulation method is used to calculate the 3D spatial coordinates corresponding to each set of matching point pairs. The corresponding projection coordinates are calculated by back-projection of the 3D coordinates, and the difference between the original coordinates and the back-projection coordinates is compared. The vector sum of the difference is the back-projection residual of the matching point pair. If the difference between the matching point and the back-projection point is less than the set threshold, the match is considered correct, the correct point pairs are counted, and the matching accuracy is calculated.

[0096] 5) Valid matching pairs:

[0097] When the number of matching points exceeds 50 and the matching accuracy is greater than 50%, the matching image pair is considered a valid match.

[0098] Figure 4The qualitative matching results of three representative image pairs with different overlapping area shapes and different imaging areas are shown. From the results, it can be seen that the strategy proposed in the present invention extracts a large number of evenly distributed matching points by fully dividing the overlapping area, showing excellent matching results. The experimental results show that the matching strategy of the present invention can effectively handle the orbiter matching problem in different areas and different polygonal overlapping areas, and obtain high-quality matching results.

[0099] Table 1 summarizes the quantitative statistical results of the matching strategy proposed in the present invention. Through the performance of various indicators, it can be concluded that the matching time of all test image pairs is less than 1 minute; the correct matching rate is more than 85%, the number of valid matching points is more than 10,000, and the distribution characteristic value of the matching points is between 1.48 and 14.77. From a quantitative point of view, the distribution of points fluctuates, but the qualitative results show that the distribution of matching points in the overlapping area is relatively uniform, which may be related to the presence of shadow areas in the overlapping area and the absence of matching points in the area, as well as the differences in the shapes of the overlapping areas of different image pairs.

[0100] Table 1 Quantitative results of three groups of orbiter image matching

[0101]

[0102] The correct matching rate statistics of all stereo pairs in the two test areas are as follows: Figure 5 The strategy proposed in the present invention is used to match multiple groups of images in two areas, and the matching results of each pair of images are obtained and counted. It can be concluded that the correct matching rate of each group of image pairs is higher than 75%, which shows that the strategy can extract high-precision matching points and show good stability in the task of matching multiple images.

[0103] Table 2 shows the quantitative results of matching multiple sets of images in the region. It takes less than 20 seconds to complete the matching of a pair of orbiter images. The average correct matching rate of the entire image exceeds 80%, and the number of effective matching pairs exceeds 70%. Due to the unstable lighting conditions at the lunar South Pole compared to the Mavos Valley on Mars, the matching accuracy and effectiveness are affected. Compared with the separate matching of three sets of orbiter images, the experimental results have significantly improved efficiency, which is mainly due to the use of thread parallelism between multiple sets of images, which fully utilizes computing resources and greatly improves matching efficiency.

[0104] Table 2 Evaluation index statistics of image matching results in the test area

[0105]

[0106]

[0107] In view of the difficulty of matching large-scale high-resolution planetary orbiter images, the present invention proposes a planetary orbiter image matching strategy based on row-to-column image segmentation and rapid establishment of corresponding relationships, which mainly includes "row-to-column" polygon overlapping area image segmentation, rapid construction of corresponding relationships based on row-direction extreme points, and multi-level matching efficiency optimization. This strategy can be used to segment overlapping areas into images that fit the boundaries more closely, effectively avoiding the introduction of non-overlapping areas or the omission of overlapping areas during the segmentation process; at the same time, the order can be quickly determined during the image segmentation process, and the corresponding relationship between the segmented images can be established accordingly; by combining GPU acceleration with distributed parallel technology, the efficiency of the matching process can be optimized. Experimental verification was carried out using lunar and Mars orbiter image data, and the results show that the strategy proposed in the present invention can effectively solve the planetary orbiter image matching problem and achieve efficient and accurate extraction of orbiter image connection points.

[0108] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A planetary orbiter image matching strategy based on image segmentation and correspondence relationship, characterized in that: The method comprises the following steps: S1. Obtain matching pairs of planetary orbiter images and obtain overlapping areas; S2, using a row-to-column division method to divide the overlapping area of ​​the image pair into grids, and based on the divided grids, dividing the images in the overlapping area into blocks to obtain a final block image composed of image blocks; S3, matching the block images of the left image and the right image in the overlapping area, determining the block correspondence relationship, and finally obtaining the block images with the corresponding relationship; S4, using a feature matching algorithm based on GPU acceleration, matching is performed based on the final corresponding block images to obtain matching points, and the matching points on all block images are aggregated to obtain the final matching result; S5. Process parallel processing is adopted for multiple groups of planetary orbiter image pairs to be matched, and S1 to S5 are operated in parallel. Thread parallel processing is adopted for a single group of planetary orbiter image pairs to be matched, and S2 to S4 are executed in parallel.

2. A planetary orbiter image matching strategy based on image segmentation and corresponding relationship construction according to claim 1, characterized in that: The specific steps of gridding the overlapping area of ​​the image pair using the row-to-column division method are as follows: Obtain the planetary orbiter image pair to be matched, obtain the overlapping area of ​​the projected image pair, and project the image pair into the image space through the imaging equation.

3. The planetary orbiter image matching strategy based on image segmentation and corresponding relationship construction according to claim 2, characterized in that: The imaging equation is: Among them, (x, y) is the focal plane coordinate of the image point, f is the focal length, (X, Y, Z) is the ground point coordinate corresponding to the image point, (X s ,Y s ,Z s ) is the coordinate of the photography center in the J2000 coordinate system, R Orbiter is the rotation matrix from the camera coordinate system to the orbiter coordinate system, R Fixed is the rotation matrix from the orbiter coordinate system to the satellite-fixed coordinate system, R J2000 is the rotation matrix from the Star-fixed coordinate system to the J2000 coordinate system, and λ is the scale factor.

4. The planetary orbiter image matching strategy based on image segmentation and corresponding relationship construction according to claim 3, characterized in that: The specific steps of S2 are: The left image of the overlapping area is divided in the row direction, an initial rectangle in the left image is set, and a polygon after division is obtained by intersecting the initial rectangle in the left image with the overlapping area, and the circumscribed rectangle of the polygon after division is used as the final row direction division result of the left image; The right image of the overlapping area is divided in the row direction, the height of the initial rectangle in the right image is calculated based on the fixed height of the initial rectangle in the left image, a buffer is added to the height of the initial rectangle in the right image, and the initial rectangle in the right image after adding the buffer is intersected with the overlapping area to obtain a divided polygon, and the circumscribed rectangle of the divided polygon is used as the final row direction division result of the right image; After the row division is completed, the column division is performed. In the left image, the fixed height of the initial rectangle in the left image is used as the side length to construct a square and arrange it from left to right to form a grid, and the column division result of the left image is obtained, and the block image of the left image is obtained. In the right image, the number of rows and columns of the rectangle when the column is divided in the right image is calculated based on the height of the initial rectangle in the right image, and the column division result of the right image is obtained, and the block image of the right image is obtained; Based on the point set formed by the overlapping area, the principal components of the overlapping area are determined, the rotation angle between the overlapping areas of the left image and the right image is calculated, and then the block image of the right image is rotated based on the rotation angle, and the circumscribed rectangle of the rotated block image is used as the final block image.

5. The planetary orbiter image matching strategy based on image segmentation and correspondence relationship construction according to claim 4, characterized in that: Set the initial rectangle in the left image as follows: In the left image, the maximum width in the image column direction is calculated based on the contour points of the overlapping area. A rectangle is constructed with the maximum width as the length and a fixed value as the height, which is called the initial rectangle in the left image. In the row direction of the left image, the initial rectangle in the left image is used to cover the overlapping area. When the last initial rectangle is laid, if the remaining number of rows is not enough to meet the height requirement of the initial rectangle in the left image, the remaining number of rows is used as the height of the initial rectangle in the left image.

6. The planetary orbiter image matching strategy based on image segmentation and corresponding relationship construction according to claim 5, characterized in that: The height of the initial rectangle in the right image is: Among them, Block src Y represents the fixed height of the initial rectangle in the left image, Block dst Y represents the height of the initial rectangle in the right image, GSD Lrow , GSD Rrow Represents the spatial resolution of the left image and the right image in the row direction respectively.

7. The planetary orbiter image matching strategy based on image segmentation and corresponding relationship construction according to claim 6, characterized in that: The initial rectangles in the right image after adding the buffer zone overlap each other, and the overlap size is half of the buffer zone size.

8. The planetary orbiter image matching strategy based on image segmentation and corresponding relationship construction according to claim 7, characterized in that: The number of rows and columns of the rectangle when dividing the columns in the right image is: Among them, BlockSize src The size of the block image on the left image, Block dst X is the column dimension of the block image on the right image, GSD Rcol is the resolution of the right image in the upper column direction.

9. The planetary orbiter image matching strategy based on image segmentation and corresponding relationship construction according to claim 8, characterized in that: The rotation angle between the overlapping area of ​​the left and right images is: Where θ represents the rotation angle, v L and v R It is the eigenvector matrix on the left and right images, indicating the main direction.

10. The planetary orbiter image matching strategy based on image segmentation and corresponding relationship construction according to claim 1, characterized in that: The specific steps of S3 are: According to the final block image of the left image, the block images are numbered from top to bottom and from left to right. The pixel with the minimum number in the row direction in the overlapping area of ​​the left image is marked as Based on the imaging model corresponding to the left image, the corresponding ground point coordinates are calculated, recorded as (B0, L0). The image point coordinates of point (B0, L0) are calculated through the imaging model of the right image and the row direction coordinates of the image point coordinates are obtained. If the image point coordinates are close to the minimum number in the row direction in the overlapping area, they are marked as The row division direction of the final block image of the right image is from top to bottom, and the column division direction is from left to right; if the image point coordinate is close to the maximum number of the row direction in the overlapping area, it is marked as The row division direction of the final block image of the right image is from bottom to top, and the column division direction is from right to left. Based on the division direction, the final block images of the left image and the right image in the overlapping area are matched, the block correspondence is determined, and the final block images with corresponding relationships are obtained.

Images