Elimination method of mismatched point pairs based on agglomerative clustering algorithm

Abstract

The application discloses a method for removing mismatched point pairs based on a condensed clustering algorithm, and mainly solves the problems of large operation amount and low precision of existing methods for removing mismatched point pairs. The implementation scheme is as follows: extracting accelerated segment test feature key points of two original images A and B; calculating binary robust independent basic feature descriptors of each accelerated segment test feature key point, and matching the feature descriptors; calculating the main direction of each accelerated segment test feature key point by using a gray centroid method; calculating the difference value of the main direction of the matched point pairs; clustering the set of the main direction difference value of the matched point pairs by using a condensed clustering algorithm; selecting the clustering cluster with the largest number of points from the clustering result, and taking the matched points in the clustering cluster as correct matched points, and taking the other matched points as mismatched points to be deleted. The application reduces the operation amount in the process of removing mismatched point pairs, improves the precision of a positioning system, and can be used for indoor target vehicle positioning.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of visual positioning, and further relates to a method for removing mismatched points, which can be used for estimating the position of an indoor unmanned target vehicle. BACKGROUND

[0002] The indoor unmanned target vehicle is a movable device used by personnel when designing and training in an indoor target range. With the vigorous development of visual technology, intelligent modules are installed on the target vehicle. In order to facilitate the management of the indoor intelligent unmanned target vehicle, the position of the indoor unmanned target vehicle must be known, that is, the indoor positioning of the unmanned target vehicle must be performed. There are many methods applied to indoor positioning, among which the indoor positioning method based on vision has become a new algorithm due to the low price of cameras and the development of visual algorithms. The key point-based visual positioning method belongs to the visual indoor positioning method, which matches the key points between the continuous video frames extracted by the camera, removes the mismatched point pairs by using the random sample consensus method, and finally calculates the pose of the target motion by using the correctly matched key points. However, the method for removing mismatched points by using the random sample consensus method has the defects of large amount of calculation and low accuracy.

[0003] The patent application with the application number 2022103829904 and the name "monocular visual odometry target pose detection method" discloses a method for removing mismatched point pairs by using the random sample consensus algorithm to further improve the overall positioning accuracy of the system. The method randomly selects 4 samples from the matching data set, calculates a homography matrix as a model, and then tests all data in the matching data set by using the model and calculates the number of data points and the cost function that satisfy the model. If the model is the optimal model, the corresponding cost function is the minimum. However, the disadvantage of this method is that the iteration number of the random sample consensus algorithm is unstable, the amount of calculation is too high, and in essence, the mismatched point pairs are not removed in the iteration process, but are instead introduced into the iteration process, thereby leading to inaccurate calculation results and thus reducing the overall accuracy of the positioning system. SUMMARY

[0004] The purpose of the present application is to overcome the defects of the prior art and provide a mismatched point pair removal method based on a condensed clustering algorithm to reduce the amount of calculation and improve the efficiency and accuracy of visual positioning.

[0005] The technical solution for achieving the purpose of the present application includes the following steps:

[0006] (1) Obtain two original images A and B of the same size, and extract the accelerated segment test feature FAST key points of the two images by using the accelerated segment test feature FAST algorithm;

[0007] (2) define two empty sets F A and F B , all the FAST key points of the acceleration section test features extracted from the image A are stored in the first set F A , and all the FAST key points of the acceleration section test features extracted from the image B are stored in the second set F B ;

[0008] (3) the binary robust independent basic features (BRIEF) descriptors of the FAST key points of the acceleration section test features extracted in the first set F A and the second set F B are respectively calculated and stored in the third set sbDes1 and the fourth set sbDes2;

[0009] (4) the principal directions of each FAST key point of the acceleration section test features in the first set F A and the second set F B are respectively calculated by using the gray centroid method;

[0010] (5) the Hamming distances between each feature descriptor in the third set sbDes1 and each feature descriptor in the fourth set sbDes2 are calculated in sequence, and the feature descriptor with the smallest Hamming distance in the fourth set sbDes2 is selected to match each corresponding feature descriptor in the third set sbDes1;

[0011] (6) an empty set is established as the fifth set data, and the difference of the principal directions of each pair of matched points is recorded;

[0012] (7) the data in the fifth set data are clustered by using the condensed clustering algorithm;

[0013] (8) the clustering clusters with the largest number of points are selected from the clustering results, the matched points in the clustering clusters are taken as the correctly matched points, and the other matched points are deleted as the mis-matched points.

[0014] Compared with the prior art, the present application has the following advantages:

[0015] Firstly, in the present application, the data in the fifth set data are clustered by using the condensed clustering algorithm, and the average distance between different clustering clusters is calculated as the condition for measuring whether two clustering clusters are merged, so that the calculation amount is smaller than that in the random sample consensus algorithm in the prior art.

[0016] Secondly, in the application, the data in the fifth set data is clustered by using the condensation clustering algorithm, the cluster with the minimum average distance is selected to be merged, the similarity of the elements in each cluster is as high as possible, and the similarity of the elements between different clusters is as low as possible, the potential internal connection between the data is deeply mined, and the positioning accuracy is improved. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is the flow chart of the application;

[0018] Figure 2 is the simulation effect diagram of the application. DETAILED DESCRIPTION

[0019] The embodiments and effects of the application will be further described below with reference to the accompanying drawings.

[0020] Reference Figure 1 The implementation steps of the present example are as follows:

[0021] Step 1, extract the accelerated segment test feature FAST key points of two original images A and B.

[0022] (1.1) for each pixel p in image A and image B, let its gray value be I p , make a circle with p as the center and 3 pixel distance as the radius, and number the pixels distributed on the periphery of the circle in order from the element in the positive direction of the y axis of pixel p as 1, 2, …, 15, 16, and the corresponding gray values are also expressed in order as I1, I2, …, I 15 , I 16 ;

[0023] (1.2) set a threshold t, where t = 0.2 * I p ;

[0024] (1.3) calculate the absolute values of the differences between I1, I2, …, I 15 , I 16 and I p , respectively, and judge them:

[0025] if there are more than 9 absolute values greater than t, then pixel p is a FAST key point,

[0026] otherwise, pixel p is not a FAST key point;

[0027] (1.4) set two sets F A and F B , and store the accelerated segment test feature FAST key points extracted from image A into the first set F AThe FAST key points extracted from the image B are stored in a second set F B .

[0028] Step 2, the binary robust independent elementary features (BRIEF) descriptors of the FAST key points are calculated.

[0029] (2.1) For each FAST key point, a square region S with the key point as the geometric center and a size of 31x31 is selected;

[0030] (2.2) 256 groups of point pairs are randomly selected in the square region S;

[0031] (2.3) For each pixel p and q in each group of point pairs, a window with a size of 5x5 is determined with the center of the two pixels, and the sum of the gray values of all pixels in the window is calculated to obtain a binary code:

[0032]

[0033] wherein I(p) and I(q) represent the sum of the gray values in the window with the center of the pixel p and q, respectively;

[0034] (2.4) Step (2.3) is repeated for the 256 groups of point pairs generated in step (2.2) to obtain a 256-dimensional bit stream binary sequence, i.e., the feature descriptor;

[0035] (2.5) Two sets sbDes1 and sbDes2 are set, and the BRIEF descriptors extracted for the FAST key points in the first set F A are stored in a third set sbDes1, and the BRIEF descriptors extracted for the FAST key points in the second set F B are stored in a fourth set sbDes2.

[0036] Step 3, the main direction of each FAST key point is calculated by using the gray centroid method.

[0037] (3.1) A circle is drawn with any FAST key point P c as the center and R as the radius, and R is 3 pixel distances;

[0038] (3.2) The total gray sum M 10 in the x direction and the total gray sum M 01 in the y direction of the image are calculated:

[0039]

[0040]

[0041] where x and y represent the coordinates in x and y directions respectively, I(x, y) is the gray value of the pixel point (x, y) in the circle;

[0042] (3.3) Calculate the principal direction θ of the key point P c according to the result of (3.2):

[0043] θ = atan2 (M 01 , M 10 ).

[0044] Step 4, match the feature descriptors.

[0045] (4.1) Calculate the Hamming distance between each feature descriptor in the third set sbDes1 and each feature descriptor in the fourth set sbDes2 in turn, that is, perform XOR operation on the two binary feature descriptors, and take the number of 1 in the XOR operation result as the Hamming distance of the two binary feature descriptors;

[0046] (4.2) Select the feature descriptor with the smallest Hamming distance in the fourth set sbDes2 and match it with each corresponding feature descriptor in the third set sbDes1.

[0047] Step 5, calculate the difference of the principal directions of the matched point pairs.

[0048] (5.1) According to the result in step 4, calculate the difference of the principal directions of the matched point pairs;

[0049] (5.2) Set the set data, and record the difference of the matched point pairs into the fifth set data.

[0050] Step 6, use the condensed clustering algorithm to cluster the data in the fifth set data.

[0051] (6.1) Input the data set data = {x1, x2, …, x k ,...x N}, 1 ≤ k ≤ N, where x k represents the principal direction distance difference between the kth pair of matched points, and N represents the total number of matched point pairs;

[0052] (6.2) Set the number of clustering clusters of the final clustering result to K, and in this example, K = 2;

[0053] (6.3) Take each element x k in the data set data as a clustering cluster C k , and get N clustering clusters:

[0054] C1, C2,... Ck ..., C N ;

[0055] (6.4) Calculate the average distance d between the clustering clusters:

[0056]

[0057] where C i and C j are two different clustering clusters, |C i | and |C j | represent the number of elements in the clustering clusters C i and C j , respectively, and x a and x b are elements belonging to the two clustering clusters C i and C j , respectively;

[0058] (6.5) According to the calculation result of (6.4), find out the two clustering clusters with the smallest average distance, and merge the two clustering clusters into one clustering cluster;

[0059] (6.6) Determine whether the total number of clustering clusters after merging is equal to K:

[0060] If it is equal to K, the clustering is completed,

[0061] If it is not equal to K, return to step (6.4).

[0062] Step 7, select the clustering cluster with the most number of points from the clustering result, and take the matching points in these clustering clusters as the correctly matched points, and delete the other matching points as the mis-matched points.

[0063] The effects of the present application can be further illustrated by the following simulation experiment:

[0064] 1. Simulation conditions

[0065] The image used in this experiment is from the public data set TUM, and the image is two images with relative motion relationship.

[0066] This experiment is implemented in a Windows system using a 12th Gen Intel(R) Core(TM) i7-12700H, and the simulation is realized through Python language.

[0067] 2. Simulation content and result analysis

[0068] Under the above experimental simulation conditions, the correct matching points of the two images with relative motion relationship are extracted by the method of the present application, and the results are as follows Figure 2 .

[0069] From Figure 2 It can be seen that the environment reflected by the image has display screen, keyboard, potted plant, these complex indoor environment elements, each line segment represents the same three-dimensional space point in two images with relative motion relationship in the two-dimensional projection point of the line, each end point of the line segment represents the correctly matched point pair from two images with relative motion relationship, and finally 294 correctly matched point pairs are extracted.

[0070] The simulation experiment shows that: in the process of removing the mis-matched point pairs from two images with relative motion relationship, since the method uses the condensation clustering algorithm to remove the mis-matched points, the defect of large amount of calculation in the prior art is overcome. At the same time, the method clusters the matched point pairs according to the average distance, merges the two clustering clusters with the minimum distance difference, and executes the cycle until the final set number of clustering clusters is reached, the relationship between the data is deeply mined, and the simulation result Figure 2 The lines in the simulation result are parallel, which shows that the quality of the matched point pairs is high, and the positioning accuracy is improved.

Claims

1. A method for rejecting mis-matched point pairs based on a coagulation clustering algorithm, characterized in that, Comprising the following steps: (1) Obtain two original images of the same size and , use the accelerated segment test feature algorithm to extract the accelerated segment test features of the two images; implementation as follows: (1a) For images and images Each pixel in Set its grayscale value to ,by Draw a circle with a center of 3 pixels and a radius of 3 pixels. of Starting with the elements along the positive axis, the pixels distributed around the perimeter are numbered sequentially as 1, 2, ..., 15, 16, and their corresponding grayscale values ​​are also represented sequentially as follows: ; (1 b) setting a threshold wherein ; (1c) the absolute value of the difference between the values calculated respectively for and is determined. If there are more than 9 absolute values greater than , then the pixel is considered a key point, Otherwise, consider the pixel not a key point; (2) define two empty sets and put all the acceleration segment test features key points extracted from the image into the first set put all the acceleration segment test features key points extracted from the image into the second set ; (3) calculate the binary robust independent elementary features of the key points of the extracted acceleration segment test features in the first set and the second set respectively, and store them in the third set and the fourth set respectively; the implementation is as follows: (3a) for each acceleration segment test feature in the first or second set The key point, select the square region with the key point as the geometric center and the size of 31x31 ; (3b) randomly select 256 pairs of points within the square region ; (3c) For each pair of pixels in each group of points and a 5x5 window is determined with its center at the respective pixel and the sum of the gray values of all pixels in this window is calculated and binary coded according to the following formula: ； wherein, and respectively represent the sum of the gray values within the window centered at the pixel and ​ (3d) repeating step (3c) for the 256 sets of point pairs generated in step (3b) to obtain a 256-dimensional bit stream binary sequence, i.e. a feature descriptor; (4) using the gray-scale centroid method, the main direction of each key point of the test feature of each acceleration section in the first set and the second set is calculated respectively . (5) sequentially calculating third sets Hamming distance of each feature descriptor in the fourth set Hamming distance of each feature descriptor in the fourth set Hamming distance of each feature descriptor in the fourth set matching each corresponding feature descriptor in the third set (6) Establishing an empty set as the fifth set and record the difference of the main direction of each pair of matching points. (7) using a coagulation clustering algorithm to cluster the data in the fifth set ; (8) selecting the cluster with the largest number of points from the clustering results, and taking the matching points in these clusters as the correctly matched points, and deleting the other matching points as the mis-matched points.

2. The method of claim 1, wherein, In step (4), the gray-scale centroid method is used to calculate the first set respectively. Second set Test features for each acceleration segment The main direction of the key points is implemented as follows: (4a) take any key point in either set as the center of the circle, make a circle with a radius of take 2 pixel distance or 3 pixel distance; (4b) calculating the sum of the gray levels of the image in the direction of the and the sum of the gray levels of the image in the direction of the : ； ； wherein and respectively represent and coordinates in the direction, the gray value of the pixel point in the circle. (4c) calculating the key points from the result of (4b) main direction of : 。 3. The method of claim 1, wherein, calculating the third set in step (5) hamming distance of each feature descriptor in the fourth set The hamming distance of each feature descriptor is the number of 1s in the result of the XOR operation of two binary feature descriptors.

4. The method of claim 1, wherein, In step (7), the agglomerative clustering algorithm is used to analyze the fifth set. Clustering of the data is performed as follows: (7a) Input dataset , wherein represents the first the main direction distance difference between the matching points, denotes a total of matching points; (7b) setting the number of clusters of the clustering final result equal to wherein ; (7c) creating a dataset each element in the dataset as a cluster , obtaining clusters ; (7d) calculating the average distance d between the clusters: ； wherein, and are two clusters of clusters, and represent the number of elements in the clusters of clusters and respectively, and are elements belonging to the clusters of clusters and respectively. (7e) according to the calculation result of (7d), finding out the two clusters with the smallest average distance, and merging the two clusters into one cluster; (7f) determining whether the total number of merged cluster groups is equal to : If equal to then clustering is complete, If not equal to then return to step (7d).

Images