A stepwise adjustment feature point matching method

Abstract

This invention relates to the field of computer vision, and more particularly to a progressively adjusted feature point matching method; including obtaining a feature point set {P} of two images to be matched using existing feature point detection operators. o 0}、{P r 0}, and obtain its corresponding feature vector set {S}. o 0}、{S r 0}; Perform feature vector matching to obtain the feature point set {P} o Each feature point in the feature point set {P} is in the feature point set {P}. r The nearest matching points of feature points {P} are selected; the feature points that satisfy the threshold and their nearest matching points are selected to obtain the initial matching point set {P}. o}、{P r The process involves updating the initial matching point set through a 2×2 region partitioning spatial transformation; updating the matching point set through a k×k, k>2 region partitioning spatial transformation; and outputting the final matching point set if the termination condition is met, otherwise continuing the iteration. This invention requires only a small number of gradual adjustments to the region partitioning size to obtain better matching feature point pairs, which is more time-efficient than traditional methods.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer vision, and particularly relates to a step-by-step adjusted feature point matching method. BACKGROUND

[0002] Feature point matching is a very important basic technology in computer vision, and has a wide range of applications in image registration, motion structure recovery, simultaneous localization and mapping, target tracking and image matching. Robust feature points can have robustness under different spatial transformations, different scales, different illuminations and other imaging conditions, and have consistency in feature descriptors, so as to establish the connection between the same name feature points in different images. Compared with the line features and region features in images, the expression form of feature points is simpler, which can be divided into two categories of corner features and spot features, such as Harris, FAST, SIFT, etc. The feature descriptor, i.e. feature vector, can be divided into two categories of binary description and real number description, such as BRIEF, SIFT, etc.

[0003] After obtaining the feature points and the feature vectors, the matching feature point pairs under a certain tolerance threshold can be found by the exhaustive method or the adjacent search method, but the matching feature point pairs at this time often contain many mis-matching points, which need to be further refined, i.e. outlier screening. The most classic and most widely used method is RANSAC, which establishes a spatial transformation model by random sampling, and constantly iteratively optimizes to find the best spatial transformation model while eliminating the feature outliers that do not meet the model constraints. At the same time, a series of improved methods of RANSAC and outlier elimination methods based on local similarity and geometric structure constraints have been developed. In addition, another new technical route is based on deep learning, which uses deep neural networks to implement feature extraction, descriptor calculation, feature point matching, homography matrix estimation and other work.

[0004] Generally, the success of feature point matching technology depends on three main factors: feature points, feature vectors, and outlier selection. Feature point detection and feature vector techniques can accurately represent scenes through high-dimensional feature analysis of local neighborhoods, and there are currently many excellent feature point detection and feature vector description algorithms. However, due to the complexity of application scenarios—illumination, repetitive structures, disparity, etc.—relying solely on feature vectors cannot achieve accurate feature point matching; therefore, the outlier selection step is particularly important. Parametric outlier selection randomly selects a small number of sample feature points from a large number of feature points to fit a spatial transformation model, and iterates until convergence. This generally yields an optimal match, but it is time-consuming and does not utilize the potential structural relationships between feature points. When disparity exists, it may get stuck in local errors. Constraint-based outlier selection often uses local similarity and geometric structure to limit feature point matching, but it is prone to the phenomenon of feature points clustering locally and in small numbers, and it is difficult to balance time and efficiency. Deep learning-based feature point matching technology, to some extent, depends on the dataset, significantly reducing its applicability to special scenarios and lacking interpretability. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a progressively adjusted feature point matching method that establishes a correspondence between input images and feature points for use in image registration, motion reconstruction, simultaneous localization and mapping, target tracking, and image matching.

[0006] A progressively adjusted feature point matching method includes the following steps:

[0007] S1. Obtain two source images I to be matched. o and reference image I r ;

[0008] S2. Obtain image I using existing feature point detection operators. o I r Feature point set {P o 0}、{P r 0}, and obtain the feature point set {P} o 0}、{P r The set of eigenvectors corresponding to 0} {S} o 0}、{S r 0};

[0009] S3. Using the feature vector set {S o 0}、{S r 0} performs feature vector matching to obtain the feature point set {P} o Each feature point in the feature point set {P} is in the feature point set {P}. r The nearest matching point in 0};

[0010] S4. Determine whether each feature point in the feature point set {P o 0} and its adjacent matching point satisfy the threshold value by ratio-test, screen the feature points and their adjacent matching points satisfying the threshold value, and obtain the initial matching point set {P o} and {P r};

[0011] S5. Update the initial matching point set {P o} and {P r} by 2x2 region division space transformation matching, and obtain the matching point set {P o}2 and {P r}2;

[0012] S6. Update the matching point set {P o} k-1 and {P r} k-1 by kxk, k>2 region division space transformation matching, and obtain the matching point set {P o} k and {P r} k ;

[0013] S7. Determine whether the stop condition is satisfied, if yes, output the final matching point set; if not, k+1 and return to step S6.

[0014] Further, the specific operation process of step S3 includes:

[0015] S31. Establish the Euclidean distance between the feature vector sets {S o 0} and {S r 0}. Taking a feature point pair (q o , q r ) as an example, the Euclidean distance relationship between the feature vectors is represented as:

[0016]

[0017] wherein d o represents the feature vector of the source image feature point q o , and d o belongs to the feature vector set {S o 0}; d r represents the feature vector of the reference image feature point q r , and d r belongs to the feature vector set {S r 0}; o i represents the i-th descriptor information of the feature point q o , and r i represents the i-th descriptor information of the feature point q rThe information of the i-th descriptor, where n represents the descriptor dimension, i.e. the feature vector dimension;

[0018] S32. Based on the Euclidean distance relationship, a Kd-Tree is used for nearest neighbor search in the feature point set {P}. r Find the feature point set {P} in 0} o The nearest matching points of each feature point in 0}.

[0019] Furthermore, the threshold d used in step S4 t The range is (0.7, 0.9), and the initial matching point set {P} is... o}、{P r} represent the feature point set {P} o 0}、{P r A subset of 0.

[0020] Furthermore, step S5 updates the initial matching point set {P} through 2×2 region partitioning spatial transformation matching. o}、{P r},include:

[0021] S51. Find the initial matching point set {P} r The maximum and minimum x-coordinate values, maximum and minimum y-coordinate values ​​of the feature points in the array;

[0022] S52. The four coordinate values ​​obtained through S51 are in image I r Draw an overlapping rectangular area Ω above r 2 , will overlap the rectangular region Ω r 2 Divide into 2×2 blocks and label them as Ω according to the scanning order. ri 2 , i = 1, 2, 3, 4;

[0023] S53. Set the initial matching point set {P} r Each feature point in the block is assigned to a corresponding block, resulting in a block feature point set {P}. ri}2, i = 1, 2, 3, 4;

[0024] S54. Based on the initial matching point set {P} o}、{P r The feature point matching relationship between {P} will initialize the matching point set {P}. o In the block, each feature point is mapped to its corresponding block, resulting in a block feature point set {P}. oi}2, i = 1, 2, 3, 4;

[0025] S55. Calculate the set of feature points {P} corresponding to each block. ri}2、{Poi The mean coordinates of all feature points in}2 are used to obtain the mean feature point pairs (p) corresponding to each block. oi p ri )2, i = 1, 2, 3, 4;

[0026] S56. Estimate the value of the feature points from image I using the mean of the four blocks. r To image I o Similarity transformation {H S}, calculate the set of block points {P} through similarity transformation. oi Spatial transformation feature point subset {P}2 roi}2;

[0027] S57. Determine the subset of feature points {P} of spatial transformation. roi Does}2 contain any elements that do not belong to its corresponding block Ω? ri 2 If the feature points exist, proceed to step S58; otherwise, proceed directly to step S6.

[0028] S58. If it does not belong to block Ω ri 2 Feature point distance block Ω ri 2 The distance of the longer side is less than the overlapping region Ω r 2 The length is 1 / 20, or the distance from the feature point to the width of the block is less than the overlapping region Ω. r 2 If the width is 1 / 20, the feature point is retained; otherwise, the feature point and its matching feature points are deleted.

[0029] S59. Repeat step S58 until the spatial transformation feature point subset {P} is reached. roi All elements in}2 that do not belong to their corresponding blocks Ω ri 2 After all feature points have been processed, the matching point set {P} is finally updated. o}2、{P r}2.

[0030] Furthermore, step S54 will initialize the matching point set {P} o In the equation, each feature point is mapped to its corresponding block to obtain the block point set {P}. oi The formula for calculating}2 is:

[0031]

[0032] Among them, P oi Represents the initial matching point set {P} othe grid number calculated by the i-th feature point in I ri 2 the vector composed of the upper left corner point, |x'| represents the length of the long side of Ω r 2 the length of the long side of Ω r 2 the length of the wide side of Ω

[0033] Further, step S6 updates the matching point set {P o} k-1 , {P r} k-1 by k×k region division space transformation matching, including:

[0034] S61. Find the maximum horizontal coordinate value, the minimum horizontal coordinate value, the maximum vertical coordinate value and the minimum vertical coordinate value of the feature points in the matching point set {P r} k-1 ;

[0035] S62. Draw an overlapping rectangular region Ω r on the image I r k by the four coordinate values obtained by S61. r k Divide the overlapping rectangular region Ω ri k into k×k blocks, and label them in scanning order as Ω 2 , i = 1, 2, …, k

[0036] S63. Assign each feature point in the matching point set {P r} k-1 to the corresponding block to obtain the block point set {P ri} k , i = 1, 2, …, k 2 ;

[0037] S64. Map each feature point in the matching point set {P o} k-1 , {P r} k-1 to the corresponding block according to the matching relationship between them to obtain the block point set {P o} k-1 , {P oi} k , i = 1, 2, …, k 2 ;

[0038] S65. Calculate the block point set {P ri} k , {Poi} k The average coordinate of all feature points in each sub-block is obtained, and the average feature point pair (p oi , p ri ) corresponding to each sub-block is obtained. k , i = 1, 2, …, k 2 ;

[0039] S66. Estimate the perspective transformation {H r} from the image I o to the image I p by using the average feature point pairs of the k x k sub-blocks, and calculate the spatial transformation feature point subset {P oi} k of the sub-block point set {P roi} k by using the perspective transformation;

[0040] S67. Determine whether there is a feature point in the spatial transformation feature point subset {P roi} k that does not belong to its corresponding sub-block Ω ri k If there is, proceed to step S68; if there is not, stop the iteration and output the final matching point set.

[0041] S68. If the feature point that does not belong to the sub-block Ω ri k is within the preset edge region, retain the feature point; otherwise, delete the feature point and its matching feature point; finally, update the matching point set {P o} k , {P r} k .

[0042] Further, the process of step S7 of determining whether the stop condition is met includes:

[0043] S71. Calculate the average spatial distance dro of all feature points in the spatial transformation feature point subset {P roi} k and the sub-block point set {P oi} k , and record that the average spatial distance dro is less than a preset threshold ds as condition one.

[0044] S72. Record that all feature points in the k x k spatial transformation feature point subset {P roi} k belong to their corresponding sub-blocks as condition two; wherein k > 2.

[0045] S73. Record that k > K, K being the maximum number of iterations, as condition three.

[0046] S74. If one of the three conditions is met, stop iteration and output the final matching feature point set.

[0047] Advantages of the present application:

[0048] The present application provides a step-by-step adjusted feature point matching method, which aims at the image registration problem under the non-parallax and small-parallax imaging condition, improves the shortcomings of the traditional image pixel-based algorithm, and accelerates the feature point registration process. The method uses region division to limit the feature point matching error, gradually increases the number of region division, improves the restriction of the region on the feature point, and estimates the spatial mapping relationship of the feature point from coarse to fine, while continuously eliminating outliers, and finally obtains the matching feature point pair. Specifically, the size of the step-by-step adjusted region division is adjusted, and only the size of the sub-region is used to limit the feature point error at each step, which can maximize the retention of as many feature point pairs as possible, and obtain more and more evenly distributed matching feature point pairs than traditional methods.

[0049] The present application only needs a small number of step-by-step adjusted region division size times to obtain good matching feature point pairs, and has higher time efficiency than traditional methods. At the same time, according to different application scenarios, the number of step-by-step adjusted region division size can be flexibly selected, so that the feature point matching can be adjusted in global accuracy, and to a certain extent, it is applicable to feature point matching in parallax scenes. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 The flowchart of the step-by-step adjusted feature point matching method of the present application;

[0051] Figure 2 The region division diagram of the embodiment of the present application. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0053] The present application provides a step-by-step adjusted feature point matching method, as shown in Figure 1 The method comprises the following steps:

[0054] S1. Input two source images I o , I r to a computer.

[0055] S2. Obtain the source image I ofeature point set {P o 0} of the reference image I r 0} of the reference image I r 0} of the reference image I o 0} of the reference image I o 0} of the reference image I r 0} of the reference image I r 0} of the reference image I

[0056] Specifically, first, the two images I o , I r to be matched are input into a computer for rough image registration, o and r represent the serial numbers of the images; existing feature point detection operators, such as SIFT, ORB, FAST, HARRIS, etc., are used to obtain the feature point sets {P o 0}, {P r 0} corresponding to the images I o , I r , and then corresponding descriptors are used to obtain the feature vector sets {S o 0}, {S r 0} corresponding to the feature point sets {P o 0}, {P r 0}; the feature vector sets {S o 0}, {S r 0} are then subjected to two key steps, step S3 feature vector matching and step S4 initial feature point matching, to obtain a rough registration result.

[0057] S3. Feature vector matching is performed using the feature vector sets {S o 0}, {S r 0}, and a neighboring matching point for each feature point in the feature point set {P r 0} is screened out from the feature point set {P o 0}.

[0058] Specifically, the specific operation process of step S3 includes:

[0059] S31. The Euclidean distances between the feature vector sets {S o 0}, {S r 0} are established, that is, the Euclidean distances between each feature vector in the feature vector set {S o 0} and each feature vector in the feature vector set {S r 0} are calculated; taking a feature point pair (q o , q r ) as an example, the Euclidean distance relationship between the feature vectors thereof is expressed as:

[0060]

[0061] wherein d o represents the feature vector of the feature point q o on the source image, and d o belongs to the feature vector set {S o 0}; d r represents the feature vector of the feature point q r on the reference image, and d r belongs to the feature vector set {S r 0}; o i represents the i-th descriptor information of the feature point q o , and r i represents the i-th descriptor information of the feature point q r , and n represents the descriptor dimension, i.e., the feature vector dimension.

[0062] S32. Kd-Tree is used for proximity search based on the Euclidean distance relationship, and one proximity matching point is found for each feature point in the feature point set {P r 0} in the feature point set {P o 0}; and each feature point in the feature point set {P o 0} can only select one feature point in the feature point set {P r 0} that does not match other feature points in the feature point set {P o 0} as a proximity matching point.

[0063] S4. Whether each feature point in the feature point set {P o 0} and its proximity matching point satisfy the threshold value is determined by ratio-test, and the feature points and their proximity matching points satisfying the threshold value are screened to obtain the initial matching point sets {P o}, {P r}.

[0064] Specifically, in order to obtain multi-plane feature points, the range of the threshold value is relaxed, the range of the threshold value d t is set to (0.7, 0.9), i.e., d t ∈(0.7, 0.9), so as to obtain as many initial feature points as possible; and the initial matching point sets {P o}, {P r} are subsets of the feature point sets {P o 0}, {P r 0}, respectively.

[0065] S5. The initial matching point sets {P o}, {P r} are updated by 2x2 region division space transformation matching to obtain the matching point sets {P o}2, {P r}2.

[0066] Specifically, it includes the following steps:

[0067] S51. Find the initial matching point set {P} r The maximum x-coordinate value X among all feature points max Minimum x-coordinate value X min Maximum ordinate value Y max and minimum ordinate value Y min ;

[0068] S52. X obtained through S51 max X min Y max and Y min Four coordinate values ​​in image I r Draw an overlapping rectangular area Ω above r 2 Then the overlapping rectangular region Ω r 2 The coordinates of the four corner points are V0(X) min ,Y min V1(X) max ,Y min V2(X) min ,Y max ) and V3(X max ,Y max ),like Figure 2 As shown; the overlapping rectangular region Ω r 2 Divide into 2×2 blocks and label them as Ω according to the scanning order. ri 2 , i = 1, 2, 3, 4;

[0069] S53. Set the initial matching point set {P} r Each feature point in the block is assigned to a corresponding block, resulting in a block feature point set {P}. ri}2, i = 1, 2, 3, 4;

[0070] S54. Based on the initial matching point set {P} o}、{P r The feature point matching relationship between {P} will initialize the matching point set {P}. o In the matrix, each feature point is mapped to a corresponding block, resulting in a block point set {P}. oi}2, i = 1, 2, 3, 4;

[0071] Preferably, such as Figure 2 As shown, with the top-left corner V0 as the reference point, points V0 and V1 form a vector x′ = V1 - V0, and points V0 and V2 form a vector y′ = V2 - V0. The initial matching point set {P}r The coordinates of any feature point in the set {P r} are represented as (x, y), and the vector representation is V = V r -V0; step S54 maps each feature point in the initial matching point set {P o} to the corresponding sub-block to obtain the sub-block point set {P oi}2, and the calculation formula is:

[0072]

[0073] where P oi represents the sub-block number calculated by the i-th feature point in the initial matching point set {P o}, represents the distance between the i-th feature point and the sub-block Ω ri 2 , and |x'| represents the length of the long side of the sub-block Ω r 2 , and y'| represents the length of the short side of the sub-block Ω r 2 .

[0074] S55. Calculate the average coordinates of all feature points in the sub-block point set {P ri}2, {P oi}2, to obtain the average feature point pair (p oi , p ri )2 corresponding to each sub-block, i = 1, 2, 3, 4.

[0075] S56. Estimate the similarity transformation {H r} from the image I o to the image I S by the average feature point pairs of the four sub-blocks, and calculate the spatial transformation feature point subset {P oi}2 of the sub-block point set {P roi}2 by the similarity transformation.

[0076] S57. Determine whether there are feature points in the spatial transformation feature point subset {P roi}2 that do not belong to its corresponding sub-block Ω ri 2 If there are, proceed to step S58; if not, proceed directly to step S6.

[0077] S58. If the feature points that do not belong to the sub-block Ω ri 2 are closer to the long side of the sub-block Ω ri 2 than to the short side, then proceed to step S59; otherwise, proceed to step S60. r 2The length is 1 / 20, or the distance from the feature point to the width of the block is less than the overlapping region Ω. r 2 If the width is 1 / 20, the feature point is retained; otherwise, the feature point and its matching feature points are deleted.

[0078] S59. Repeat step S58 until the spatial transformation feature point subset {P} is reached. roi All elements in}2 that do not belong to their corresponding blocks Ω ri 2 After all feature points have been processed, the matching point set {P} is finally updated. o}2、{P r}2.

[0079] S6. Perform spatial transformation matching by dividing the region into k×k, k>2 regions, and update the matching point set {P}. o} k-1 、{P r} k-1 The matching point set {P} is obtained. o} k 、{P r} k ;

[0080] Specifically, it includes the following steps:

[0081] S61. Find the matching point set {P} r} k-1 The maximum and minimum x-coordinate values, maximum and minimum y-coordinate values ​​among all feature points;

[0082] S62. The four coordinate values ​​obtained through S61 are in image I r Draw an overlapping rectangular area Ω above r k , will overlap the rectangular region Ω r k Divide the data into k×k blocks and label them Ω according to the scanning order. ri k i = 1, 2, ..., k 2 ;

[0083] S63. Match the point set {P} r} k-1 Each feature point is assigned to a corresponding block, resulting in a block point set {P}. ri} k i = 1, 2, ..., k 2 ;

[0084] S64. Based on the matching point set {P} o} k-1 、{P r}k-1 The matching relationship between them will match the set of points {P} o} k-1 Each feature point is mapped to a corresponding block, resulting in a block point set {P}. oi} k i = 1, 2, ..., k 2 ;

[0085] Specifically, the mapping formula is:

[0086]

[0087] S65. Calculate the set of block points {P} corresponding to each block. ri} k 、{P oi} k The mean coordinates of all feature points in the data are used to obtain the mean feature point pairs (p) for each block. oi p ri ) k i = 1, 2, ..., k 2 ;

[0088] S66. Estimate the sum of the mean feature points of image I using k×k blocks. r To image I o Perspective transformation {H p}, and abandons the traditional RANSAC algorithm, adopting SVD decomposition to accelerate the computation; then, the set of block points {P} is calculated through perspective transformation. oi} k Spatial transformation feature point subset {P roi} k ;

[0089] S67. Determine the subset of feature points {P} of spatial transformation. roi} k Does it exist that does not belong to its corresponding block Ω? ri k If the feature points exist, proceed to step S68; otherwise, stop the iteration and output the final set of matching points.

[0090] S68. If it does not belong to block Ω ri k If a feature point is within a preset edge region, the feature point is retained; otherwise, the feature point and its matching feature points are deleted. Finally, the matching point set {P} is updated. o} k 、{P r} k .

[0091] Specifically, when performing 3×3 region partitioning spatial transformation matching, the S68 preset edge region refers to the region that does not belong to the block Ω.ri 3 Feature point distance block Ω ri 3 The distance of the longer side is less than the overlapping region Ω r 3 1 / 25 of the longer side, or the distance of the feature point from the wider side of the block is less than the overlapping region Ω. r 3 1 / 25 of the width; when performing 4×4 and subsequent region segmentation spatial transformation matching, the S68 preset edge region refers to the region where the distance from the feature point to the long side of the segment is less than the overlapping region Ω. r k 1 / 30 of the long side, or the distance of the feature point from the wide side of the block is less than the overlapping region Ω. r k 1 / 30 of the width.

[0092] S7. Determine if the stopping condition is met. If it is, output the final set of matching points; otherwise, increment k+1 and return to step S6.

[0093] Specifically, step S7, which determines whether the stopping condition is met, includes:

[0094] S71. Calculate the subset of feature points {P} of spatial transformation. roi} k With the set of partition points {P oi} k The average spatial distance dro of all feature points in the data is defined as condition one, where the average spatial distance dro is less than a preset threshold ds.

[0095] S72. Transform a subset {P} of k×k spatial feature points. roi} k The case in which all feature points belong to their corresponding blocks is denoted as condition two.

[0096] S73. Let k > K, where K is the maximum number of iterations, denoted as condition three;

[0097] S74. If any one of the three conditions is met, stop the iteration and output the final set of matching points.

[0098] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A progressively adjusted feature point matching method, characterized in that, Includes the following steps: S1. Obtain two source images I to be matched. o and reference image I r ; S2. Obtain image I using existing feature point detection operators. o I r Feature point set {P o 0}、{P r 0}, and obtain the feature point set {P} o 0}、{P r The set of eigenvectors corresponding to 0} {S} o 0}、{S r 0}; S3. Using the feature vector set {S o 0}、{S r 0} performs feature vector matching to obtain the feature point set {P} o Each feature point in the feature point set {P} is in the feature point set {P}. r The nearest matching point in 0}; S4. Determine the feature point set {P} using the ratio-test. o For each feature point in {0} and its neighboring matching points, check if they meet the threshold. Filter out the feature points and their neighboring matching points that meet the threshold to obtain the initial matching point set {P}. o }、{P r }; S5. Update the initial matching point set {P} by performing spatial transformation matching through 2×2 region partitioning. o }、{P r }, thus obtaining the matching point set {P} o }2、{P r }2, including: S51. Find the initial matching point set {P} r The maximum and minimum x-coordinate values, maximum and minimum y-coordinate values ​​among all feature points; S52. The four coordinate values ​​obtained through S51 are in reference image I r Draw an overlapping rectangular area Ω above r 2 , will overlap the rectangular region Ω r 2 Divide into 2×2 blocks and label them as Ω according to the scanning order. ri 2 , i＝1,2,3,4； S53. Set the initial matching point set {P} r Each feature point in the block is assigned to a corresponding block, resulting in a block feature point set {P}. ri }2, i = 1, 2, 3, 4; S54. Based on the initial matching point set {P} o }、{P r The feature point matching relationship between {P} will initialize the matching point set {P}. o In the matrix, each feature point is mapped to a corresponding block, resulting in a block point set {P}. oi }2, i = 1, 2, 3, 4; S55. Calculate the set of block points {P} corresponding to each block. ri }2、{P oi The mean coordinates of all feature points in}2 are used to obtain the mean feature point pairs (p) corresponding to each block. oi p ri )2, i = 1, 2, 3, 4; S56. Estimate the value of the feature points from image I using the mean of the four blocks. r To image I o Similarity transformation {H S }, calculate the set of block points {P} through similarity transformation. oi Spatial transformation feature point subset {P}2 roi }2; S57. Determine the subset of feature points {P} of spatial transformation. roi Does}2 contain any elements that do not belong to its corresponding block Ω? ri 2 If the feature points exist, proceed to step S58; otherwise, proceed directly to step S6. S58. If it does not belong to block Ω ri 2 Feature point distance block Ω ri 2 The distance of the longer side is less than the overlapping region Ω r 2 The length is 1 / 20, or the distance from the feature point to the width of the block is less than the overlapping region Ω. r 2 If the width is 1 / 20, the feature point is retained; otherwise, the feature point and its matching feature points are deleted. S59. Repeat step S58 until the spatial transformation feature point subset {P} is reached. roi All elements in}2 that do not belong to their corresponding blocks Ω ri 2 After all feature points have been processed, the matching point set {P} is updated. o }2、{P r }2; S6. Perform spatial transformation matching by dividing the region into k×k, k>2 regions, and update the matching point set {P}. o } k-1 、{P r } k-1 The matching point set {P} is obtained. o } k 、{P r } k ,include: S61. Find the matching point set {P} r } k-1 The maximum and minimum x-coordinate values, maximum and minimum y-coordinate values ​​among all feature points; S62. The four coordinate values ​​obtained through S61 are in image I r Draw an overlapping rectangular area Ω above r k , will overlap the rectangular region Ω r k Divide the data into k×k blocks and label them Ω according to the scanning order. ri k , i=1,2,…,k 2 ; S63. Match the set of points {P} r } k-1 Each feature point is assigned to a corresponding block, resulting in a block point set {P}. ri } k i = 1, 2, ..., k 2 ; S64. Based on the matching point set {P} o } k-1 、{P r } k-1 The matching relationship between them will match the set of points {P} o } k-1 Each feature point is mapped to a corresponding block, resulting in a block point set {P}. oi } k i = 1, 2, ..., k 2 ; S65. Calculate the set of block points {P} corresponding to each block. ri } k 、{P oi } k The mean coordinates of all feature points in the data are used to obtain the mean feature point pairs (p) for each block. oi p ri ) k i = 1, 2, ..., k 2 ; S66. Estimate the sum of the mean feature points of image I using k×k blocks. r To image I o Perspective transformation {H p }, calculate the set of block points {P} through perspective transformation. oi } k Spatial transformation feature point subset {P roi } k ; S67. Determine the subset of feature points {P} of spatial transformation. roi } k Does it exist that does not belong to its corresponding block Ω? ri k If the feature points exist, proceed to step S68; otherwise, stop the iteration and output the final set of matching points. S68. If it does not belong to block Ω ri k If a feature point is within a preset edge region, the feature point is retained; otherwise, the feature point and its matching pair feature points are deleted. Finally, the matching point set {P} is updated. o } k 、{P r } k ; S7. Determine if the stopping condition is met. If it is, output the final set of matching points; otherwise, increment k+1 and return to step S6.

2. The feature point matching method with progressive adjustment according to claim 1, characterized in that, The specific operation process of step S3 includes: S31. Establish the eigenvector set {S o 0}、{S r The Euclidean distance between 0 and 0, expressed as feature point pairs (q) o ,q r Taking as an example, the Euclidean distance relationship between its feature vectors is expressed as follows: Where, d o Represents feature point q on the source image o eigenvectors, and d o Belonging to the eigenvector set {S o 0};d r Represents feature point q on the reference image r eigenvectors, and d r Belonging to the eigenvector set {S r 0};o i Representing feature point q o The information of the i-th descriptor, r i Representing feature point q r The information of the i-th descriptor, where n represents the descriptor dimension, i.e. the feature vector dimension; S32. Based on the Euclidean distance relationship, a Kd-Tree is used for nearest neighbor search in the feature point set {P}. r Find the feature point set {P} in 0} o A neighboring matching point for each feature point in 0}.

3. The feature point matching method with progressive adjustment according to claim 1, characterized in that, The threshold d used in step S4 t The range is (0.7, 0.9), and the initial matching point set {P} is... o }、{P r } represent the feature point set {P} o 0}、{P r A subset of 0.

4. The feature point matching method with progressive adjustment according to claim 1, characterized in that, Step S54 will initialize the matching point set {P} o In the equation, each feature point is mapped to its corresponding block to obtain the block point set {P}. oi The formula for calculating}2 is: Among them, P oi Represents the initial matching point set {P} o The grid number calculated from the i-th feature point in} Represents the relationship between the i-th feature point and the block Ω. ri 2 The vector formed by the top-left corner points, |x'|, represents the block Ω. r 2 The length of the longer side, |y'| represents the block Ω. r 2 Wide side mold length.

5. The feature point matching method with progressive adjustment according to claim 1, characterized in that, Step S7 determines whether the stopping condition is met, including: S71. Calculate the subset of feature points {P} of spatial transformation. roi } k With the set of partition points {P oi } k The average spatial distance dro of all feature points in the data is defined as condition one, where the average spatial distance dro is less than a preset threshold ds. S72. Transform a subset {P} of k×k spatial feature points. roi } k Condition 2 is defined as the case where all feature points belong to their corresponding blocks; where k>2. S73. Let k > K, where K is the maximum number of iterations, denoted as condition three; S74. If any one of the three conditions is met, stop the iteration and output the final set of matching points.

Images