A Feature Matching Method Based on a Multi-Level Refinement Strategy

Abstract

The present invention discloses a feature matching method based on a multi-level refinement strategy, belonging to the technical field of computer vision. The present invention generates an initial correspondence by using the similarity of the local appearance of the feature descriptor in the Hamming space, combines the local image motion smoothness constraint, and uses the GMS algorithm to improve the accuracy of the initial matching; finally, the PROSAC algorithm is used to optimize the matching to obtain an accurate matching based on the global gray information in the Euclidean space. The experimental results show that the feature matching method based on the multi-level fine matching strategy of the present invention reduces the error by an average of 29.92% compared with the ORB algorithm in complex scenes with illumination changes and blurring, effectively improving the matching accuracy of feature matching in complex spaces.

Description

Technical Field

[0001] The present invention relates to a feature matching method based on a multi-level refinement strategy, belonging to the technical field of computer vision science. Background Art

[0002] Feature matching is the core technology for positioning and mapping in feature-based visual SLAM, which directly affects the accuracy, efficiency, and robustness of the SLAM system. Feature matching first extracts significant structural features with physical meanings in two images, including feature points, feature lines, etc., and then identifies and aligns the same / similar structures at the pixel level. That is, feature matching is based on feature detection and descriptor construction. In visual SLAM, various feature detections ultimately boil down to feature point detection, and common classical feature point detection methods include SIFT, SURF, and ORB, etc. Although these methods all have good performance and robustness in feature-based visual SLAM, when dealing with complex scenes, single-level feature matching methods often cannot meet the actual needs. For this reason, researchers have proposed to establish a multi-level matching strategy by making full use of various information such as the local similarity of features in the Hamming space, the local image structure in the Euclidean space, and the similarity of global gray information to eliminate ambiguities and false matches caused by insufficient information. Such methods mainly establish a preliminary matching relationship through the similarity of local descriptors of feature points, and then eliminate incorrect matches according to geometric constraints. The elimination of incorrect matches mainly uses resampling-based methods, and resampling-based methods rely heavily on the accuracy of sampling. When there are a large number of incorrect matches in the initial matching, the effectiveness of such methods is greatly reduced.

[0003] During the matching process, a large number of incorrect matches and false matches will inevitably occur. In 2019, Zhu et al. proposed GMS-RANSAC based on improved grid motion statistical features, introduced the principle of distance consistency, and used the similarity between distances to eliminate outliers, improving the accuracy but increasing the running time. When dealing with complex scenarios, single-level feature matching methods often cannot meet the actual needs. Especially in visual SLAM systems, designing a fast and accurate strategy for rejecting incorrect matches remains a difficult point. In order to improve the matching accuracy, Ye et al. combined K-Nearest Neighbor (KNN), nearest neighbor ratio, bidirectional matching, Cosine Similarity (CS), and PROSAC to propose the MP-ORB matching method. In order to solve the influence of irregular dynamic changes in brightness and contrast caused by non-uniform illumination and randomly appearing textureless areas on matching, Sun et al. proposed a multi-stage matching module composed of KNN, threshold filtering, feature vector norm, and RANSAC to eliminate mismatches. Most of the above methods believe that the quality of matching is affected by the invariance and distinctness of features in Hamming space and Euclidean space, and initial matches are established based on this. Rejecting incorrect matches mainly relies on resampling-based methods such as RANSAC or PROSAC. However, the initial matches established in this way are affected by factors such as noise, blur, and occlusion, resulting in a large number of incorrect matches. And resampling-based methods rely heavily on the accuracy of sampling. When there are a large number of incorrect matches in the initial matches, the effectiveness of such methods is greatly reduced. Summary of the Invention

[0004] In order to improve the accuracy of feature matching in complex workspaces, the present invention provides a feature matching method and device based on a multi-level refinement strategy. The technical solutions are as follows:

[0005] The first object of the present invention is to provide a feature matching method based on a multi-level refinement strategy, including:

[0006] Step 1: Obtain two images to be matched, extract the feature points of the two images respectively through the ORB algorithm, and then use the KNN algorithm to establish the data association of two given feature sets in the two images, generating a large number of one-to-two associated feature pairs;

[0007] Step 2: Through threshold filtering, convert the one-to-two feature association pairs obtained in Step 1 into one-to-one associated feature pairs by calculating the distance ratio;

[0008] Step 3: Use the GMS algorithm to remove low-quality matches and false matches after the threshold filtering operation in the solution process of Step 2;

[0009] Step 4: Use the result of Step 3 as the initial matching input for the resampling-based method, and then use the progressive consistent sampling algorithm to sort the results obtained by GMS in descending order;

[0010] Step 5: Extract samples from the result of Step 4 to solve the model parameters, and then optimize the matching of the feature matching set assumed to contain outliers to obtain the optimal matching.

[0011] Optionally, Step 1 selects the nearest neighbor matching with K = 2, which specifically includes:

[0012] Let be the sampling set of feature points in the target image I t+1 , and be the template set of feature points in the reference image I t ;

[0013] Use the Kd-tree algorithm to generate an index tree for the feature descriptors, so as to achieve the K-nearest neighbor search from to . For each feature in the set, find the feature points with the first and second smallest Hamming distances in the set, as shown in Equation (1):

[0014]

[0015] where a and b are the numbers of feature points in the two images I t+1 , I t respectively, ⊕ represents the exclusive OR operation, represent 256-bit binary vectors respectively;

[0016] For each feature point in the set, find the nearest neighbor point and the second nearest neighbor point in the set, as shown in Equation (2):

[0017]

[0018] where {p1, s1 , p 2,s1 , …, p a,s1} represents the set of nearest neighbor points corresponding to in the set, and {p 1,s2 , p 2,s2 , …, p a,s2} represents the set of second nearest neighbor points corresponding to in the set.

[0019] Optionally, Step 2 includes:

[0020] Hypothetical point to point p 1,s1 、p 1,s2 The distances are respectively Then the ratio of the nearest neighbor distance to the second nearest neighbor distance, which is the ratio of d1 and d2, is calculated as shown in Equation (3):

[0021]

[0022] When W is less than the given threshold T w at that time, in retain in the one-to-two association pair and turn it into a one-to-one association pair; on the contrary, if the ratio is greater than the threshold T w , then remove both association pairs with (p 1,s1 , p 1,s2 ).

[0023] Optionally, in step 3, the GMS algorithm is used to convert the smoothness constraint of the motion into the number of matches within the neighborhood of the statistical match pair as the basis for judging true and false matches.

[0024] Optionally, step 3 includes:[[]]

[0025] Assume that two input images are {I t , I t+1}, with {N, M} feature points respectively. Denote X = {x1, x2,..., x i ,..., x N} as the set of all match pairs from the I t image to the I t+1 image. The region {t, t + 1} is the neighborhood of the match pair x t , I t+1} in the images {I i}, and each neighborhood has {n, m} supporting match pairs; the number of match pairs within the neighborhood of each match pair is also called the neighborhood support amount, and its calculation is shown in Equation (4):

[0026]

[0027] where X i ∈X is the match subset between the neighborhood {t, t + 1} corresponding to the match pair x i , and S i is the neighborhood support amount of the match pair x i ; if the neighborhood support amount is less than the given threshold T G , then this match pair is considered an incorrect match or a low-quality match, otherwise it is recorded as a correct match.

[0028] Optionally, in step 4, the progressive consistent sampling method is used to estimate the optimal homography matrix, and inliers and outliers are distinguished, so as to further remove low-quality matches and redundant matches.

[0029] Optionally, step 4 includes:

[0030] Introduce an evaluation function q(u) to represent the probability that a data point becomes an inlier. The set of all N data points is denoted as U N , then for U N The calculation of sorting the data points inside is shown in Equation (5):

[0031]

[0032] For each pair of feature points in the image, the ratio β of the Euclidean distance is used to measure the quality of the feature point matching. The calculation of β is shown in Equation (6):

[0033]

[0034] where d min is the minimum Euclidean distance, and d min2 is the second minimum Euclidean distance. The smaller the β value, the greater the probability q(u) of becoming an inlier, and the better the matching quality;

[0035] Subsequently, the initial matching pairs input are sorted in descending order according to the above evaluation function. Let p{u i} represent the probability that u i is a correct match in the sorted data point subset. Make a correlation hypothesis about the correlation between this probability and the evaluation function, as shown in Equation (7):

[0036]

[0037] Denote the set of the first n data points with the largest correlation function values as U n , and sample m data points from it, denoted as set M; According to the sample quality, the sequence of the T N -th sampling from the data set U N is denoted as If the sequence is the result sorted by the evaluation function, then there is as shown in Formula (8):

[0038]

[0039] Optionally, step 5 includes: comparing the feature points with a set inlier threshold. If it is less than the threshold, it is an inlier; otherwise, it is an outlier.

[0040] Compare the obtained number of inlier points with the set threshold of the number of inlier points. If it is greater than the threshold, update the number of inlier points to the current value; otherwise, perform iteration until the optimal match is obtained. The second object of the present invention is to provide a feature matching system based on a multi-level refinement strategy, including:

[0041] An image acquisition module for acquiring an image to be matched;

[0042] A feature matching module that performs feature matching on the image acquired by the image acquisition module by using the feature matching method based on the multi-level refinement strategy described in any one of the above;

[0043] An output display module for outputting the matching result of the feature matching module.

[0044] The third object of the present invention is to provide a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the feature matching method based on the multi-level refinement strategy described in any one of the above is implemented.

[0045] The beneficial effects of the present invention are as follows:

[0046] The present invention designs a new multi-level refinement matching strategy. Based on the fact that the motion smoothness brings more matching points in the area around the matching feature points, the motion smoothness constraint of the local image is added to the multi-level refinement strategy, and combined with the fast data association of feature matching in the Hamming space, it solves the problem of low accuracy in the initial matching based on the resampling method and achieves a high precision in feature matching in complex scenarios. Description of the Drawings

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0048] Figure 1 It is a technical roadmap of a feature matching method based on a multi-level refinement strategy of the present invention.

[0049] Figure 2 It is an example diagram of the matching result in the embodiment of the present invention, where (a) is the result diagram of establishing an association by the KNN algorithm, (b) is the result diagram of association conversion by the TF algorithm, (c) is the result diagram of eliminating incorrect matches by the GMS algorithm, and (d) is the result diagram of screening by the PROSAC algorithm. Detailed Embodiments

[0050] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0051] Embodiment 1:

[0052] This embodiment provides a feature matching method based on a multi-level refinement strategy. Refer to Figure 1 , including:

[0053] Step 1: Obtain two images to be matched, extract the feature points of the two images respectively through the ORB algorithm, and then use the KNN algorithm to establish the data association of two given feature sets in the two images, generating a large number of one-to-two associated feature pairs;

[0054] Step 2: Through threshold filtering, use the calculated distance ratio to convert the one-to-two feature association pairs obtained in Step 1 into one-to-one associated feature pairs;

[0055] Step 3: Adopt the GMS algorithm to remove the low-quality matches and false matches after the threshold filtering operation in the solution process of Step 2;

[0056] Step 4: Use the result of Step 3 as the initial matching input for the resampling-based method, and then use the progressive consistent sampling algorithm to sort the results obtained by GMS in descending order;

[0057] Step 5: Extract samples from Step 4 to solve the model parameters, and then optimize the matching of the feature matching set assumed to contain outliers to obtain the optimal match.

[0058] Embodiment 2:

[0059] This embodiment provides a feature matching method based on a multi-level refinement strategy. As Figure 1 shown, it includes the following steps:

[0060] Step 1: Obtain two images to be matched, extract the feature points of the two images respectively through the ORB algorithm, and then use the k-nearest neighbor (KNN) search algorithm to quickly establish the data association of two given feature sets in the two images, generating a large number of one-to-two associated feature pairs;

[0061] Let be the sampling set of the feature points in the target image I t+1 , be the template set of the feature points in the reference image I t . Use the Kd-tree algorithm to generate an index tree for the feature descriptors, so as to quickly implement the K-nearest neighbor search from to . Since K = 2 is taken in the present invention, so for each feature on the set, find The feature points with the first and second smallest collective Hamming distances are as shown in Equation (1).

[0062]

[0063] Among them, a and b are the numbers of feature points in two images I t+1 、I t respectively, ⊕ represents the exclusive OR operation, represent 256-bit binary vectors respectively.

[0064] Therefore, for each feature point in the set the nearest neighbor point and the second nearest neighbor point can be found in the set as shown in Equation (2).

[0065]

[0066] Among them, {p 1,s1 , p 2,s1 , …, p a,s1} represents the corresponding nearest neighbor point set in the set, {p 1,s2 , p 2,s2 , …, p a,s2} represents the corresponding second nearest neighbor point set in the set.

[0067] Step 2: Convert the one-to-two feature association pairs obtained in Step 1 into one-to-one associated feature pairs by threshold filtering (TF) using the calculated distance ratio;

[0068] Assume that the distances from point to points p 1,s1 , p 1,s2 are respectively Then the ratio of the nearest neighbor distance to the second nearest neighbor distance is the ratio of d1 and d2. The calculation method is as shown in Equation (3).

[0069]

[0070] When W is less than the given threshold T w , in the one-to-two association pair, is retained and becomes a one-to-one association pair. On the contrary, if the ratio is greater than the threshold T w , then and the two association pairs between (p 1,s1 , p 1,s2 ) are all removed.

[0071] Step 3: Use the GMS algorithm to remove low-quality matches and false matches after the TF operation in the solution process of Step 2;

[0072] Assume that the two input images are {I t , I t+1}, with {N, M} feature points respectively. Denote X = {x1, x2, …, x i , …, x N} as the set of all matching pairs from image I t to image I t+1 . The region {t, t + 1} is the neighborhood of the matching pair x t , I t+1} in the images {I i}, and each neighborhood has {n, m} supporting matching pairs (excluding the original matching pair). The number of matching pairs within the neighborhood of each matching pair is also called the neighborhood support amount, and its calculation is shown in Equation (4);

[0073]

[0074] where X i ∈ X is the matching subset between the neighborhood {t, t + 1} corresponding to the matching pair x i , and S i is the neighborhood support amount of the matching pair x i . The above formula represents the support amount excluding the matching pair x i itself. Therefore, the S Figure 2 in it can be calculated as S i = 2, S j = 0.

[0075] Step 4: Use the result of Step 3 obtained by GMS as the initial matching input for the resampling-based method, and then use the Progressive Sample Consensus (PROSAC) algorithm to sort the results obtained by GMS in descending order;

[0076] According to the assumption of the PROSAC algorithm, the higher the similarity of data points, the greater the possibility of being an inlier. For this reason, an evaluation function q(u) is introduced to represent the probability of a data point becoming an inlier. The set of all N data points is denoted as U N , and the calculation of sorting the elements in U N is shown in Equation (5).

[0077]

[0078] For each pair of feature points in the image, the ratio β of the Euclidean distance is used to measure the quality of feature point matching. The calculation of β is shown in Equation (6).

[0079]

[0080] Among them, d min is the minimum Euclidean distance, d min2 It is the second smallest Euclidean distance. The smaller the β value is, the greater the probability of becoming an interior point q(u), and the better the matching quality is.

[0081] Then, the initial matching pairs of the input are sorted in descending order according to the above evaluation function, and p{u i} means that in the sorted data point subset u i For the probability of correct matching, a correlation assumption is made between the probability and the evaluation function, as shown in formula (7).

[0082]

[0083] The set of data points with the largest correlation function value is recorded as U n , and sample m data points from it, recorded as set M. According to the sample quality T N From the dataset U N The sampled sequence is recorded as If sequence The result after sorting by the evaluation function is as shown in formula (8).

[0084]

[0085] After descending order according to the matching quality of the feature points, every 4 feature points are grouped together, the sum of the quality of each group is calculated and sorted, and the top 4 groups of matching points are taken to calculate their homography matrices. These four groups of points are then eliminated, and the remaining points are used to calculate the corresponding projection points according to the homography matrix, and the projection errors between the projection points are calculated.

[0086] Step 5: Extract samples from step 4 to solve model parameters, and then optimize the matching of the feature matching set that is assumed to contain outliers to obtain the best match;

[0087] The feature points are compared with the set inlier threshold. If the value is less than the threshold, it is an inlier, otherwise it is an outlier. The number of inliers obtained is compared with the set inlier threshold. If it is greater than the threshold, the number of inliers is updated to the current value. Otherwise, it is iterated until the best match is obtained.

[0088] Based on the above specific implementation, the effect of the present invention is verified in combination with specific experiments below:

[0089] After extracting features, the extracted feature points are matched at multiple levels, and all matching items are marked with straight lines in each pair of images, such as Figure 2 The KTGP-ORB model first uses the KNN algorithm to quickly establish a one-to-two data association between the feature points in the two images, as shown in Figure 2As shown in (a), in order to convert a one-to-two data association into a one-to-one data association, the TF technology is used to find the best match in the one-to-two data association to form a one-to-one matching relationship, such as Figure 2 as shown in (b). As Figure 2 shown in (c), in order to remove the low-quality matches and incorrect matches after TF, the GMS algorithm is used to further eliminate the incorrect matches. The GMS algorithm can quickly distinguish correct matches from incorrect matches by considering the number of matches within the neighborhood range of the matching points and converting a higher number of feature point matches into higher-quality matches. Finally, in order to obtain the optimal match, PROSAC is used for the last screening of the KTGP-ORB multi-level refinement matching model, and the final result is as Figure 2 shown in (d). The experimental results intuitively verify the effectiveness of the KNN-TF-GMS-PROSAC (KTGP) multi-level fine matching model.

[0090] In the Leuven dataset and Bikes dataset of the scenario, the average values of two evaluation metrics, ME (Mean Error) and RMSE (Root Mean Square Error), were calculated. The Mean Error (ME) refers to the average of the distance errors between all correctly matched feature point pairs. The smaller the Mean Error, the higher the accuracy of the algorithm. The Root Mean Square Error (RMSE) refers to the root mean square of the distance errors between all correctly matched feature point pairs. The smaller the Root Mean Square Error, the higher the accuracy of the algorithm.

[0091] The average ME and RMSE values of the KTGP-ORB model of the present invention on the Leuven dataset and Bikes dataset are 23.73106, 25.23356 and 23.99844, 25.56006 respectively. The average ME and RMSE values of the ORB algorithm on the Leuven dataset and Bikes dataset are 33.90576, 37.02068 and 34.1985, 36.83376 respectively. It can be calculated therefrom that the KTGP-ORB model reduces the error by an average of 29.92% compared with the ORB algorithm in scenarios with illumination changes and blur.

[0092] This embodiment is completed using VS2019 and OpenCV under the Linux operating system. The hardware environment is a laptop with a 3.20 GHz i7 processor and 8 GB of running memory, and the experimental process is relatively stable.

[0093] Embodiment 3:

[0094] This embodiment provides a feature matching system based on a multi-level refinement strategy, including:

[0095] An image acquisition module, configured to acquire an image to be matched;

[0096] A feature matching module that performs feature matching on the image acquired by the image acquisition module by using the feature matching method based on a multi-level refinement strategy described in Embodiment 1 or Embodiment 2;

[0097] An output display module for outputting the matching result of the feature matching module.

[0098] Some steps in the embodiments of the present invention can be implemented by software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk, etc.

[0099] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A feature matching method based on a multi-level refinement strategy, characterized in that The method includes: Step 1: Obtain two images to be matched, extract the feature points of the two images respectively through the ORB algorithm, and then use the KNN algorithm to establish the data association of two given feature sets in the two images, generating a large number of one-to-two associated feature pairs; Step 2: Through threshold filtering, convert the one-to-two feature association pairs obtained in Step 1 into one-to-one associated feature pairs by calculating the distance ratio; Step 3: Adopt the GMS algorithm to remove the low-quality matches and false matches after the threshold filtering operation in the solution process of Step 2; Step 4: Use the result of Step 3 as the initial match input for the resampling-based method, and then use the progressive consistent sampling algorithm to sort the results obtained by GMS in descending order; Step 5: Extract samples from Step 4 to solve the model parameters, and then optimize the matching of the feature matching set assumed to contain outliers to obtain the optimal match.

2. The feature matching method based on a multi-level refinement strategy according to claim 1, wherein In Step 1, the nearest neighbor matching with K = 2 is selected, specifically including: Let be the sampling set of feature points in the target image I t+1 , and be the template set of feature points in the reference image I t . Generate an index tree for the feature descriptors using the Kd-tree algorithm to achieve K-nearest neighbor search from to . For each feature on the set, find the feature points with the first and second smallest Hamming distances in the set, as shown in Equation (1): where a and b are the numbers of feature points in two images I t+1 and I t respectively, denotes the exclusive OR operation, respectively represent 256-bit binary vectors; For each feature point in the set find the nearest neighbor point and the second nearest neighbor point in the set, as shown in Equation (2): P i,r→t = {(p 1,s1 , p 1,s2 ), (p 2,s1 , p 2,s2 ),..., (p a,s1 , p a,s2 )} (2) Among them, {p 1,s1 , p 2,s1 , …, p a,s1} represents the nearest neighbor point set corresponding in 1,s2 , p 2,s2 , …, p a,s2} represents the next nearest neighbor point set corresponding in.

3. The feature matching method based on a multi-level refinement strategy according to claim 2, wherein Step 2 includes: Hypothetical point to point p 1,s1 、p 1,s2 are respectively Then the ratio of the nearest neighbor distance to the second nearest neighbor distance is the ratio of d1 and d2, and the calculation method is shown in Equation (3): When W is less than a given threshold T w then, among a pair of two - to - one associations, retain to become one - to - one associations; conversely, if the ratio is greater than the threshold T w then remove both of the two associations between (p 1,s1 , p 1,s2 ).

4. The feature matching method based on a multi-level refinement strategy according to claim 1, wherein In Step 3, the GMS algorithm is used to convert the smoothness constraint of motion into the number of matches in the neighborhood of the statistical match pair, which is used as the basis for judging true and false matches.

5. The feature matching method based on a multi-level refinement strategy according to claim 4, wherein Step 3 includes: Assume that the two input images are {I t , I t+1}, which have {N, M} feature points respectively. Denote X = {x1, x2, …, x i , …, x N} as the set of all matching pairs from image I t to image I t+1 . The region {t, t + 1} is the neighborhood of the matching pair x i in the images {I t , I t+1}, and each neighborhood has {n, m} supporting matching pairs respectively; the number of matching pairs within the neighborhood of each matching pair is also called the neighborhood support, and its calculation is shown in Equation (4) as follows: S i = |X i | - 1 (4) Among them, X i ∈ X is the matching subset between the matching pair x i and its corresponding neighborhood {t, t + 1}, and S i is the neighborhood support amount of the matching pair x i ; if the neighborhood support amount is less than the given threshold T G , then this matching pair is considered as a wrong match or a low-quality match, otherwise it is recorded as a correct match.

6. The feature matching method based on a multi-level refinement strategy according to claim 1, wherein In Step 4, the progressive consistent sampling method is used to estimate the optimal homography matrix and distinguish inliers and outliers, so as to further remove low-quality matches and redundant matches.

7. The feature matching method based on a multi-level refinement strategy according to claim 6, wherein Step 4 includes: Introduce the evaluation function q(u) to represent the probability that a data point becomes an inlier, and denote the set of all N data points as U N , then perform the sorting calculation on U N as shown in Equation (5): For each pair of feature points in the image, use the ratio β of the Euclidean distance to measure the quality of the feature point matching. The calculation of β is shown in Equation (6): where d min is the minimum Euclidean distance, and d min2 is the second minimum Euclidean distance. The smaller the β value, the greater the probability q(u) of becoming an inlier, and the better the matching quality. Subsequently, the initial matching pairs of the input are sorted in descending order according to the above evaluation function, and p{u i} is used to represent the probability that u i is correctly matched in the sorted data point subset. A correlation hypothesis is made for the correlation between this probability and the evaluation function, as shown in Equation (7): Let the set of the first n data points with the largest correlation function values be denoted as U n , and sample m data points from it, denoted as set M; according to the sample quality, the sequence sampled from the data set U N for the T N th time is denoted as If the sequence is the result sorted by the evaluation function, then as shown in formula (8):

8. The feature matching method based on a multi-level refinement strategy according to claim 1, wherein Step 5 includes: Compare the feature points with the set inlier threshold. If it is less than the threshold, it is an inlier; otherwise, it is an outlier; Compare the obtained number of inliers with the set inlier number threshold. If it is greater than the threshold, update the number of inliers to the current value; otherwise, perform iteration until the optimal match is obtained.

9. A feature matching system based on a multi-level refinement strategy, characterized in that, The system includes: An image acquisition module for acquiring images to be matched; A feature matching module that uses the feature matching method based on the multi-level refinement strategy described in any one of claims 1-8 to perform feature matching on the images acquired by the image acquisition module; An output display module for outputting the matching results of the feature matching module.

10. A computer-readable storage medium, characterized in that, A computer program is stored on the storage medium. When the computer program is executed by a processor, the feature matching method based on the multi-level refinement strategy described in any one of claims 1-8 is implemented.

Images