CN104123542B - Device and method for positioning hub workpiece - Google Patents

Abstract

The invention discloses a kind of devices and methods therefor of hub workpiece positioning, described device includes image capture module, wheel hub Template Information extraction module, hubless feature to be detected point extraction module, Feature Points Matching module and wheel hub locating module；Four points on SIFT feature, the position in the center of circle and valve and wheel hub outward flange circumference that wheel hub Template Information extraction module is used to extract on wheel hub template image.The problems such as present invention is in view of the illumination effect and translation, rotation, dimensional variation run into during wheel hub images match, matched using Scale invariant features transform characteristic point matching method template image and altimetric image to be checked it is hollow between corresponding point it is right, then these spatial correspondences to hub area image in judge templet image and altimetric image to be checked are passed through, known calibration point in template image is finally calculated into the corresponding point of hub area in altimetric image to be checked by the spatial correspondence of the two, so as to reach the purpose of wheel hub positioning.

Description

Device and method for positioning hub workpiece

technical field

The invention relates to a workpiece positioning technology, in particular to a hub workpiece positioning device and method.

Background technique

On the automatic assembly line of automobiles, workpiece positioning is a common operation requirement. Applying industrial robots with computer vision functions to automatic assembly of automobiles can effectively reduce the interference of human factors, significantly improve production efficiency and product quality, and reduce production costs. When the industrial robot performs automatic processing of the automobile hub workpiece, it is necessary to use computer vision technology to analyze the actual workpiece image collected by the industrial camera, identify the hub from the image, calculate the geometric position information, and determine the grasping posture and position of the robot accordingly. Motion trajectory, real-time control of industrial robots to grab and handle the wheel hub.

Usually, the hub workpiece is a casting workpiece, and many casting lines will be left on the side of the hub after rough machining, and the surface of the hub is rough; in addition, in the actual working environment, there are often other disturbances around the hub, and the hub has translation and Rotation, incomplete shooting of hub workpieces, location of hubs, and complex image backgrounds, etc. In this case, the image of the hub workpiece is complex, which makes it difficult to locate the hub and its valve.

Now the prior art relevant to the present invention is introduced as follows:

1. The technical solution of prior art 1

Zhao Yuliang, Liu Weijun, and Liu Yongxian studied the automatic identification and classification of various automobile wheels in the state of random mixed flow on the conveyor belt in the literature "Research on the Online Identification System of Automobile Wheels. Mechanical Design and Manufacturing, 2007 (10): 164-166". The basic steps of the method are: image acquisition, image preprocessing, feature extraction and recognition classification. The key step of hub positioning and classification is to extract five types of features of the hub image: whether there is a hole in the center of the hub; the diameter of the hub; the number of holes in the peripheral area of the hub; the area occupied by the entire hub; The gray value with the largest pixel count.

Existing technology one uses image region segmentation, Rober operator edge detection and least square method to fit the hub contour circle. However, in the actual processing environment, there may be a complex background for placing the hub or when the background of the hub image is similar to the target color of the hub, and when the edge detection result is not good, it may cause missed and wrong judgments of image features. In addition, when the part of the hub workpiece is not fully photographed, the method cannot realize the calculation of the hub shape and position.

2. The technical solution of the second prior art

Le Ying, Xu Xinmin and Wu Xiaobo proposed a method based on area array CCD imaging and computer image processing High-precision shape and position parameter detection method of technology. The basic steps of this method are as follows: take a wheel hub image with a calibration template, and perform image edge detection through image grayscale conversion and region segmentation, and perform geometric distortion correction on the hub boundary with an optical distortion correction model; then use sub-pixel interpolation The algorithm makes the edge detection result more accurate; finally, the shape and position parameters of the hub are fitted according to the position of the mounting hole.

The positioning accuracy of prior art 2 depends on image region segmentation and edge detection results, especially the shape detail segmentation results inside the hub region. However, in the actual processing environment, when the background of the hub is more complex or the background of the hub image is similar to the target color of the hub, the image segmentation algorithm cannot well highlight the detailed shape inside the hub area, resulting in the failure of subsequent positioning steps; In addition, when the part of the hub workpiece is not fully photographed, the method cannot realize the calculation of the hub shape and position.

3. The technical solution of prior art 3

Hu Chao, Cui Jialin, Qiu Jun, etc. in the patent "Automatic Wheel Identification Device and Method: China, 103090790.A[P]. Apparatus and method for offset parameters to the outer peripheral plane of the bottom of the hub. This method automatically obtains the offset distance parameters from the assembly surface of the wheel hub to the outer peripheral plane of the bottom of the hub through a non-contact distance measuring instrument in advance, which fully covers various technical parameters of the hub, creates a hub information database, and then uses two image acquisition devices to collect the The images above and below the hub, the image above the hub is the top view of the hub, the shape parameters of the hub are obtained through the top view of the hub, and the shape parameters of the center hole of the hub and the assembly surface of the hub are obtained through the image below the hub, that is, the bottom view of the hub, for example The size, position, shape, etc. of the mounting holes on the mounting surface.

The third prior art needs to pre-store a large amount of wheel image information, and in the actual recognition process, it is necessary to obtain the frontal images of the upper and lower sides of the wheel, and the device and recognition process are relatively complicated. In addition, the purpose of the present invention is to locate a specific part of the hub, and the relative position between the camera and the hub is not fixed, so this method is not applicable to this scenario.

4. The technical solution of prior art 4

Huang Qian, Wu Yuan, and Tang Dajun proposed an automatic wheel model identification system in the patent "A detection system and detection method for identifying wheel models: China, 103425969.A[P].2013", which includes a host computer and the CCD image sensor, the host computer is connected with the CCD image sensor in turn. The patent also provides a method implemented by the above system, including the following steps: initializing settings; obtaining a hub-free image in the hub model identification area; creating a hub model database record; identifying and determining the hub model. This patent can realize automatic model identification for the hubs entering the hub identification area during the working process of the system according to the hub model database created in advance.

Prior Art 4 needs to pre-store the hub image database, and when the hub workpiece part is not fully photographed, this method cannot realize the calculation of the hub shape and position.

To sum up, the existing hub workpiece positioning technology has the following problems: (1) During the hub matching process, when the image encounters problems such as illumination, translation, rotation, and scale changes, the hub workpiece positioning deviation is relatively large; (2) ) It is difficult to locate the hub when the viewing angle of the hub changes and partial occlusion occurs.

The terms to be used in the present invention are abbreviated as follows:

SIFT: Scale-Invariant Feature Transform, scale invariant feature transformation;

DoG: Difference of Gaussian, Gaussian difference;

BBF: Best Bin First, the optimal node is preferred;

RANSAC: Random Sample Consensus, random sampling consistency.

Contents of the invention

In order to solve the above-mentioned problems in the prior art, the present invention will design a device and method for hub workpiece positioning to achieve the following two purposes:

(1) During the hub matching process, reduce the deviation of hub workpiece positioning;

(2) It is easy to locate the hub when the viewing angle of the hub changes and partial occlusion occurs.

In order to achieve the above object, the technical solution of the present invention is as follows: a device for hub workpiece positioning, including an image acquisition module, a hub template information extraction module, a hub feature point extraction module to be detected, a feature point matching module and a hub positioning module; The image acquisition module is used to collect the grayscale image of the hub workpiece; the described hub template information extraction module is used to extract the SIFT feature point on the hub template image, the position of the center of circle and the gas nozzle, and four points on the circumference of the outer edge of the hub; The detection wheel feature point extraction module is used to extract the SIFT feature point information on the wheel hub image to be detected; the feature point matching module is used to find the feature point pairs that match the wheel hub image to be detected and the wheel hub template image, and calculate the distance between the wheel hub to be detected and the template wheel hub Spatial mapping relationship; the hub positioning module is used to locate the center of the corresponding hub in the image to be detected, the position of the valve and four points on the circumference of the outer edge of the hub, and calculate the radius length of the hub in the image to be detected;

The output ends of the image acquisition module are respectively connected with the wheel hub template information extraction module and the wheel hub feature point extraction module to be detected, and the input ends of the feature point matching module are respectively connected with the wheel hub template information extraction module and the wheel hub feature point extraction module to be detected The modules are connected, and the output end of the feature point matching module is connected with the hub positioning module.

A positioning method of a positioning device for a hub workpiece, comprising the following steps:

A. Offline processing

In the offline processing stage, the image of the hub workpiece is collected, the SIFT feature point information of the hub image is extracted and stored, and the position of the center of the circle and the valve is pre-marked on the template image; specifically, the following steps are included:

A1. Collect the image of the hub workpiece

The grayscale image of the wheel hub workpiece is collected by the image acquisition module. When taking it, it is necessary to obtain an ideal wheel hub template image in an environment with good lighting conditions and low noise; the background color of the template image is required to be uniform, and there are only wheels in the image without other interference;

A2. Wheel template information extraction

The SIFT feature points on the hub template image are extracted by the hub template information extraction module, the center of circle and the position of the valve of the hub template are calibrated, and the radius of the hub template is measured; the specific steps are as follows:

A21. Analyze the input hub workpiece template image, search for pixels in the hub workpiece area that meet the characteristics of SIFT feature points, count and store the description information of SIFT feature points, that is, obtain the SIFT feature point template information of the hub according to steps 1) to 5). :

A211. Construct image pyramid T

The input image is defined as f(x, y), and f(x, y) is down-sampled I times to obtain the image pyramid T of the (I+1) layer, where I=log ₂ [min(M,N)]- 3. M and N are the number of rows and columns of f(x,y) respectively. The down-sampling refers to taking the average value of four adjacent pixels as the down-sampled pixel.

Define the image of the 0th layer in the image pyramid model T as T ₀ (x, y), that is, the original image f(x, y); the image of the i-th layer is defined as T _i (x, y), that is, the original image The image obtained after f(x, y) is down-sampled I times, i=0, 1, 2,..., I.

A212. Constructing a Gaussian Pyramid L

Use the Gaussian convolution kernel function G(x,y,σ) to convolve T _i (x,y), and continuously change the scale space factor σ to obtain the scale space as L _i :

L _i (x,y,σ)=G(x,y,σ)*T _i (x,y) (1)

Among them, the symbol '*' represents the convolution operator, σ is the scale space factor, i=0, 1, 2,..., I.

Do the same operation on (I+1) images in T to get L.

A213. Construct DoG pyramid D

Take the difference between every two adjacent images in L _i to get the DoG space D _i , namely

D _i (x,y,σ)=[G(x,y,kσ)-G(x,y,σ)]*T _i (x,y)=L _i (x,y,kσ)-L _i (x,y,σ) (2)

Among them, the symbol '*' represents a convolution operator, k is a constant that is a multiple of two adjacent scale spaces, and i=0, 1, 2,..., I.

Do the same operation on the (I+1) group of images in L to get D.

A214. Detect the spatial local extremum points in D

Using the Taylor expansion of the DoG function, in the case where the derivative of the function is zero

where X = (x, y, σ) ^T .

When the derivative of D(X) is set to zero, an extreme point with sub-pixel precision is obtained which is

A215. Screen unstable extreme points to obtain SIFT feature point set

First remove the points with low contrast in the image, that is, satisfy extreme point of Then use the Hessian matrix to remove the edge extreme points.

The second derivative of an image of a certain scale in the DoG pyramid in the x direction is defined as D _xx , then the Hessian matrix is expressed as:

The two eigenvalues of H are defined as λ ₁ and λ ₂ , where λ ₁ ≥ λ ₂ and λ ₁ /λ ₂ = r, where λ ₁ and λ ₂ correspond to the principal curvatures of the image pair in the x and y directions, respectively value, when r is greater than the threshold of 10, it is judged that the extreme point is located at the edge of the DoG curve.

Define Tr(H) as the trace of H, and Det(H) as the determinant of H, then

That is, by calculating Tr(H) and Det(H) to avoid directly calculating the eigenvalues, thereby reducing the amount of calculation.

A216, calculate SIFT feature point descriptor

Taking the image area of any size around the key point as the statistical range, the image area is divided into several blocks; the gradient histogram of each point in each block is counted, and the vector representing the image information of the area is calculated.

The modulus value of the defined gradient is m(x,y), and the direction is θ(x,y), then

First calculate the image area required by the descriptor, divide the neighborhood near the feature point into 4×4 sub-areas, and the size of each sub-area is 3σ, where σ is the scale space factor; then, count the gradient direction histogram of each sub-area Figure: Take the direction of the feature point as the reference direction, then calculate the angle of the gradient direction of each pixel point in each sub-region relative to the reference direction, and project it to 8 directions in the 0-2π interval with π/4 as the interval, and The accumulation of gradient values in each direction is counted, and an 8-dimensional vector descriptor is generated after the normalization operation; finally, the 8-dimensional vectors of each sub-region are collected to form a 4×4×8=126-dimensional feature point descriptor.

A22. The hub template information extraction module calibrates and stores the position information of six pixels in the hub template image in units of pixels: the center of the hub workpiece O(x ₀ ,y ₀ ), the center of the valve nozzle of the hub workpiece O _gas (x _gas ,y _gas ) and four points O ₁ (x ₁ ,y ₁ ),O ₂ (x ₂ ,y ₂ ),O ₃ (x ₃ ,y ₃ ),O ₄ (x ₄ , y ₄ ).

B. Online processing

In the online processing stage, first extract the SIFT feature points on the image to be detected; then use the Best-Bin-First search algorithm to search for feature points that match the hub template; then use the RANSAC algorithm to eliminate mismatching points, and calculate the The spatial mapping relationship between the hub and the template image; finally, according to the marked points of the template image, calculate the center of the hub and the position of the valve in the image to be detected; specifically include the following steps:

B1. Extract SIFT feature points on the image to be detected

The SIFT feature points on the image to be detected are obtained by the feature point extraction module of the wheel hub to be detected according to step A21, that is, the input wheel hub image to be detected is analyzed, and the pixels satisfying the characteristics of the SIFT feature point in the image are searched for, counted and stored

B2. Match feature points

The feature point matching module finds the matching feature points between the wheel hub image to be detected and the hub template image, and eliminates the wrong matching points, and calculates the spatial mapping relationship between the wheel hub to be detected and the template wheel hub. Specific steps are as follows:

B21. Initially match the reference image and the image to be matched using the nearest neighbor/second nearest neighbor algorithm

Use the BBF algorithm to search for the nearest neighbor feature point p _min (feature vector v _min ) and the second neighbor feature point p _min2 (feature vector v _min2 ) with the closest Euclidean distance to the feature point p (feature vector v _i ) to be matched, then Point pairs satisfying the following conditions are matched feature points:

Among them, Dist(v _i , v _min ) represents the Mahalanobis distance between v _i and v _min , and Dist(v _i , v _min2 ) represents the Mahalanobis distance between v _i and v _min2 , namely

where the superscript T denotes the matrix transpose notation.

B22. Use the RANSAC algorithm to eliminate false matching points, and calculate the spatial correspondence between the target area and the template image.

Set point sets A and B as the initial matching point sets obtained on the template image and the detection image respectively, then the specific steps of the RANSAC algorithm are as follows:

B221. Randomly select 4 pairs of matching point pairs in point pair sets A and B, and calculate the projection transformation matrix H of these four pairs of point pairs:

For a point p(x,y) in the image, the point is transformed to a point p'(x',y') through the matrix H, that is

in,

That is, H can be obtained by matching point pairs p(x, y) and p'(x', y'). A projection transformation matrix can be calculated for every 4 pairs of matching points.

B222. Using the projective transformation matrix H calculated in step B221, perform spatial transformation on all the feature points in point set A to obtain point set B′;

Calculate the coordinate error of all corresponding points in point sets B and B′, that is, e=||B-B′||; set the error threshold σ, if e<σ, the point pair is considered as an inner point pair, otherwise it is an outer point pair Point right;

B223, repeat steps B221 and B222, find a transformation with the largest number of interior point pairs, and use the interior point pair sets obtained by this transformation as new point sets A and B, and perform a new round of iteration;

B224 Iteration Termination Judgment: When the number of internal point pairs obtained by iteration is consistent with the number of point pairs in point sets A and B before this iteration, the iteration terminates;

B225. Iterative results: A and B of the last iteration are the matching point sets after eliminating the mismatched feature point pairs, and the corresponding projection transformation matrix H represents the space transformation between the original image and the image to be detected that we require relation.

B3. Locate the center of the hub and the position of the valve in the image to be detected

The hub positioning module locates the center of the hub in the image to be detected and the position of the valve, and calculates the radius length of the hub in the image to be detected. Specific steps are as follows:

B31, according to the space transformation matrix H obtained in step B222, calculate six pixel points corresponding to the calibration points in the image to be detected: the center of circle O'(x' ₀ , y' ₀ ) of the hub workpiece, the center of the valve of the hub workpiece O′ _gas (x′ _gas ,y′ _gas ) and four points O′ ₁ (x′ ₁ ,y′ ₁ ), O′ ₂ (x′ ₂ ,y′ ₂ ), O′ ₃ (x′ ₃ , y′ ₃ ), O′ ₄ (x′ ₄ , y′ ₄ ).

Take the coordinate O′(x′ ₀ ,y′ ₀ ) of the center of the hub as an example:

B32. Calculate the radius R′ of the hub workpiece in the image to be detected:

in

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

1. In order to effectively locate the hub workpiece, the present invention considers the influence of light and the problems of translation, rotation, and scale change encountered in the hub image matching process, and adopts the scale-invariant feature transformation (SIFT) feature point matching method to match out The spatial corresponding point pairs in the template image and the image to be detected, and then judge the spatial correspondence between the template image and the hub area image in the image to be detected through these point pairs, and finally pass the known calibration points in the template image through the spatial correspondence between the two The relationship calculates the points corresponding to the hub area in the image to be detected, so as to achieve the purpose of hub positioning. The document "Lowe DG. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 2004, 60(2): 91-110." proves that the feature points on the image that meet the SIFT characteristics change in illumination, translation, Good SIFT characteristics can be maintained during rotation and scale changes, so the present invention has better robustness to ambient light, viewing angle changes and partial occlusion, and can locate the hub workpiece under different interference environments, with good positioning effect.

2. Before the actual positioning, the present invention obtains the feature point information of the hub template image through offline processing, and pre-calibrates the center of the hub template and the air nozzle, which reduces the amount of calculation in the actual positioning process of the hub;

3. The present invention uses the SIFT algorithm as the feature point matching method, which can overcome the problems of illumination, translation, rotation, and scale transformation encountered in the image matching process, and has better robustness to background noise, viewing angle changes, and partial occlusions. sex.

4. The present invention adopts the random RANSAC method to eliminate mismatching point pairs, thereby improving the matching accuracy.

Description of drawings

The present invention has 7 accompanying drawings, wherein:

Fig. 1 is a flow chart of a wheel hub workpiece location method based on SIFT features.

Fig. 2 is a schematic composition diagram of a hub workpiece positioning device based on SIFT features.

Fig. 3 is a schematic diagram of marking points of the hub template.

Figure 4 shows the positioning results in the case of hub rotation and translation.

Figure 5 shows the positioning results when there are disturbances around the hub.

Figure 6 is the positioning result when the background of the hub image is uneven.

Figure 7 is the positioning result when the hub part is missing.

detailed description

The present invention will be further described below in conjunction with the accompanying drawings. The composition of a hub workpiece positioning device is shown in FIG. 2 , and the specific method flow is shown in FIG. 1 .

In order to verify the effectiveness of the present invention, objective tests and subjective tests were carried out.

1. Subjective performance test (visual effect)

When the camera captures the image of the wheel hub, there will be different noise interference situations. In order to verify the effectiveness of the method of the present invention, several images under different interference conditions were collected for experiments.

In the experiment, the size of the template image is 690×691 pixels, and the size of the detection image is 1280×960 pixels. The wheel hub template and its marking points are shown in Figure 3. In Fig. 3, the center of the hub template and the center of the air nozzle are marked with a cross, and the circumference of the outer edge of the hub template is marked with a solid circle. In the process of hub positioning, it is necessary to first select the relevant coordinate values in the template image. In this experiment, the coordinates of the hub circle center in the template image are (353,351), the coordinates of the valve center point are (127,246), and the hub radius is 339 pixels.

Considering the space limit, one image is selected under different interference conditions of the hub, and the positioning results are shown in Fig. The obtained outer edge circumference of the hub is marked with a cross to mark the center point of the wheel hub and the central position of the air nozzle detected by the present invention, and the detected outer edge circumference of the hub is marked with a dotted circle. In addition, on the right side of each wheel hub diagram, the center of the circle and the valve position are enlarged to show the detection results more clearly.

2. Objective performance standards

In order to objectively evaluate the positioning accuracy, the present invention counts the absolute difference between the radius value, the center of the circle and the center coordinates of the gas nozzle and the actual value obtained under each interference situation, that is, the absolute value of the deviation between the positioning result and the actual value, and Computes the mean absolute difference. Under different interference conditions, the average absolute difference between the hub circle center, valve center, and radius positioning results and the actual value is shown in Table 1, and the units of the data in the table are pixels.

Table 1 The wheel hub alignment result of the inventive method and the mean absolute difference of the actual value

It can be seen from Table 1 that although there are rotation-translation transformations in the hub area in the image, uneven background or interference objects, and even partial missing of the hub, the hub positioning based on SIFT feature point matching can obtain good positioning results, almost no affected by interfering factors.

3. For the technical solution of the present invention, the following alternatives can also accomplish the purpose of the invention

(1) For a more ideal working environment, such as sufficient lighting and no occlusion, other feature point matching algorithms with less robustness but less calculation, such as the SURF algorithm, can be used for matching;

(2) The SIFT algorithm is a method based on feature point matching. For a regular rotationally symmetrical shape workpiece such as a wheel hub, line features can also be used instead of point features for matching.

(3) Based on the present invention, in the case of good lighting conditions, the hub area can also be pre-segmented by image area segmentation, and then the SIFT feature point matching can be performed, which can effectively reduce the calculation amount of the positioning method.

Claims (2)

1. A device for locating a hub workpiece is characterized in that: the hub positioning device comprises an image acquisition module, a hub template information extraction module, a to-be-detected hub characteristic point extraction module, a characteristic point matching module and a hub positioning module; the image acquisition module is used for acquiring a gray level image of the hub workpiece; the hub template information extraction module is used for extracting SIFT feature points, the positions of a circle center and an air tap on a hub template image and four points on the periphery of the outer edge of the hub; the hub feature point extraction module to be detected is used for extracting SIFT feature point information on a hub image to be detected; the characteristic point matching module is used for searching a characteristic point pair matched with the hub image to be detected and the hub template image and calculating a spatial mapping relation between the hub to be detected and the template hub; the hub positioning module is used for positioning the circle center of the corresponding hub in the image to be detected, the air tap and the positions of four points on the circumference of the outer edge of the hub, and calculating the radius length of the hub in the image to be detected;

the output end of the image acquisition module is respectively connected with the hub template information extraction module and the hub characteristic point extraction module to be detected, the input end of the characteristic point matching module is respectively connected with the hub template information extraction module and the hub characteristic point extraction module to be detected, and the output end of the characteristic point matching module is connected with the hub positioning module.

2. A positioning method of a positioning device of a hub workpiece is characterized in that: the method comprises the following steps:

A. off-line processing

In the off-line processing stage, collecting a wheel hub workpiece image, extracting and storing SIFT feature point information of the wheel hub image, and marking the position of a circle center and an air nozzle on a template image in advance; the method specifically comprises the following steps:

a1, collecting the wheel hub workpiece image

The method comprises the steps that an image acquisition module acquires a gray level image of a hub workpiece, and when the image acquisition module is used, an ideal hub template image needs to be acquired in an environment with good illumination condition and low noise; the background color of the template image is required to be uniform, and the image only has a hub and no other interferents;

a2, extracting hub template information

Extracting SIFT feature points on the hub template image by a hub template information extraction module, calibrating the circle center and the air tap position of the hub template, and measuring the radius of the hub template; the method comprises the following specific steps:

a21, analyzing the input hub workpiece template image, searching pixel points which meet SIFT feature point characteristics in a hub workpiece area, counting and storing SIFT feature point description information, namely obtaining SIFT feature point template information of the hub according to the steps A211-A216:

a211, constructing an image pyramid T

Defining an input image as f (x, y), and performing I-time down-sampling on the f (x, y) to obtain an image pyramid T of an (I +1) layer, wherein I is log₂[min(M,N)]-3, M and N are the number of rows and columns, respectively, of f (x, y); the down-sampling is to take the average value of four adjacent pixels as the pixel after down-sampling;

defining the image of the 0 th layer in the image pyramid model T as T₀(x, y), i.e. the original image f (x, y); the image of the ith layer is defined as T_i(x, y), i.e. an image obtained by performing I-time down-sampling on the original image f (x, y), wherein I is 0, 1, 2., I;

a212, constructing a Gaussian pyramid L

Using a Gaussian convolution kernel G (x, y, sigma) to T_i(x, y) performing convolution, and continuously changing the scale space factor sigma to obtain a scale space L_i：

L_i(x,y,σ)＝G(x,y,σ)*T_i(x,y) (1)

Wherein, the symbol' represents the convolution operator,σ is a scale space factor, I ═ 0, 1,2,. and I;

performing the same operation on (I +1) images in the T to obtain L;

a213 constructing a DoG pyramid D

Get L_iEvery two adjacent images are subjected to difference to obtain a DoG space D_iI.e. by

D_i(x,y,σ)＝[G(x,y,kσ)-G(x,y,σ)]*T_i(x,y)＝L_i(x,y,kσ)-L_i(x,y,σ) (2)

Wherein, the symbol' denotes a convolution operator, k is a constant of two adjacent scale space multiples, I is 0, 1, 2.

Performing the same operation on the (I +1) group of images in the L to obtain D;

a214, detecting the spatial local extreme point in D

Using Taylor expansion of the DoG function, when solving the condition that the derivative of the function is zero

$D\left(X\right)=D+\frac{\&part;D^T}{\&part;X}X+\frac{1}{2}{X}^T\frac{{\&part;}^2{D}^T}{{\&part;}^{2}X}X---\left(3\right)$

Where X ═ (X, y, σ)^T；

Let the derivative of D (X) be zero, obtain the extreme point of sub-pixel level precisionNamely, it is

$\hat{X}=-\frac{{\&part;}^2{D}^{-1}}{\&part;X^2}\frac{\&part;D}{\&part;X}---\left(4\right)$

A215, screening unstable extreme points to obtain an SIFT feature point set

Firstly, the points with lower contrast in the image are removed, namely, the method satisfiesExtreme point ofThen removing edge extreme points by using a Hessian matrix;

the second derivative of the image of a certain scale in the DoG pyramid in the x direction is defined as D_xxThen the Hessian matrix is expressed as:

$H=\left[\begin{array}{cc}{D}_{xx}& D_{xy}\\ D_{xy}& D_{yy}\end{array}\right]---\left(5\right)$

two eigenvalues of H are defined as λ₁And λ₂Wherein λ is₁≥λ₂And lambda₁/λ₂R, here λ₁And λ₂Respectively corresponding to the main curvature values of the image pair in the x direction and the y direction, and judging that the extreme point is positioned at the edge position of the DoG curve when r is greater than a threshold value 10;

tr (H) is defined as the trace of H, det (H) is the determinant of H, then

$\frac{Tr{\left(H\right)}^2}{Det\left(H\right)}=\frac{{({\mathrm{\&lambda;}}_1+{\mathrm{\&lambda;}}_2)}^2}{{\mathrm{\&lambda;}}_1{\mathrm{\&lambda;}}_2}=\frac{{({\mathrm{r\&lambda;}}_2+{\mathrm{\&lambda;}}_2)}^2}{{\mathrm{r\&lambda;}}_2{\mathrm{\&lambda;}}_2}=\frac{{(r+1)}^2}{r}>\frac{{(10+1)}^2}{10}---\left(6\right)$

The direct characteristic value is avoided by calculating Tr (H) and Det (H), thereby reducing the calculation amount;

a216, calculating SIFT feature point descriptors

Dividing an image area into a plurality of blocks by taking the image area with any size around the key point as a statistical range; counting the gradient histogram of each point in each block, and calculating a vector representing the image information of the area;

the modulus defining the gradient is m (x, y) and the direction is θ (x, y), then

$m(x,y)=\sqrt{{\left(L\right(x+1,y)-L(x-1,y\left)\right)}^2+{\left(L\right(x,y+1)-L(x,y-1\left)\right)}^2}---\left(7\right)$

Firstly, calculating an image region required by a descriptor, dividing a neighborhood near a feature point into 4 multiplied by 4 sub-regions, wherein the size of each sub-region is 3 sigma, and sigma is a scale space factor; then, the histogram of gradient direction of each sub-region is counted: taking the direction of the characteristic point as a reference direction, then calculating the angle of the gradient direction of each pixel point in each sub-region relative to the reference direction, projecting the angle to 8 directions at intervals of pi/4 in a 0-2 pi interval, counting the accumulation of gradient values in each direction, and generating an 8-dimensional vector descriptor after normalization operation; finally, 8-dimensional vectors of each sub-region are collected to form a feature point descriptor with dimensions of 4 multiplied by 8 which are 126;

a22, calibrating a hub template information extraction module and storing position information of six pixel points in a hub template image by taking a pixel as a unit: center O (x) of wheel hub workpiece₀,y₀) Air tap center O of wheel hub workpiece_gas(x_gas,y_gas) And four points O on the outer edge circumference of the hub₁(x₁,y₁),O₂(x₂,y₂),O₃(x₃,y₃),O₄(x₄,y₄)；

B. On-line processing

In the online processing stage, SIFT feature points on an image to be detected are extracted; then searching characteristic points matched with the hub template by using a Best-Bin-First search algorithm; then eliminating mismatching points by using an RANSAC algorithm, and calculating a spatial mapping relation between a hub and a template image in an image to be detected; finally, calculating the circle center of the wheel hub and the position of the air tap in the image to be detected according to the mark points of the template image; the method specifically comprises the following steps:

b1, extracting SIFT feature points on the image to be detected

The hub feature point extraction module to be detected obtains SIFT feature points on the image to be detected according to the step A21, namely, analyzes the input hub image to be detected, searches pixel points which meet the SIFT feature point characteristics in the image, and counts and stores the pixel points

B2, matching feature points

Searching characteristic points matched with the hub template image by the characteristic point matching module, eliminating mismatching points, and calculating a spatial mapping relation between the hub to be detected and the template hub; the method comprises the following specific steps:

b21, carrying out initial matching on the reference image and the image to be matched by using a nearest neighbor/next nearest neighbor algorithm

Searching the nearest characteristic point p closest to the characteristic point p to be matched in Euclidean distance by using BBF algorithm_minAnd sub-adjacent feature point p_min2Wherein the feature vector of the feature point p to be matched is v_iNearest feature point p_minHas a feature vector of v_minNext adjacent feature point p_min2Has a feature vector of v_min2Then, the point pairs satisfying the following conditions are matched feature points:

$\frac{Dist(v_i,v_{min})}{Dist(v_i,v_{\min 2})}<0.8---\left(8\right)$

wherein Dist (v)_i,v_min) Denotes v_iAnd v_minMahalanobis distance between them, Dist (v)_i,v_min2) Denotes v_iAnd v_min2The mahalanobis distance therebetween, i.e.

$Dist(v_i,v_{min2})=\sqrt{(v_i-v_{min2}){(v_i-v_{min2})}^T}---\left(9\right)$

$Dist(v_i,v_{min})=\sqrt{(v_i-v_{min}){(v_i-v_{min})}^T}---\left(10\right)$

Wherein the superscript T represents a matrix transpose symbol;

b22, eliminating mismatching points by using a RANSAC algorithm, and calculating a spatial corresponding relation between a target area and a template image;

if the point sets a and B are initial matching point sets obtained on the template image and the detection image respectively, the RANSAC algorithm specifically comprises the following steps:

b221, randomly selecting 4 pairs of matching point pairs in the point pair sets A and B, and calculating a projection transformation matrix H of the four pairs of point pairs:

for a point p (x, y) in the image, the point is transformed by the matrix H to a point p ' (x ', y '), i.e.

$\left[\begin{array}{c}{x}^{\&prime;}\\ y^{\&prime;}\\ 1\end{array}\right]=H\left[\begin{array}{c}x\\ y\\ 1\end{array}\right]---\left(11\right)$

Wherein,

that is, H can be obtained by matching the pair of points p (x, y) and p ' (x ', y '); calculating a projective transformation matrix for every 4 pairs of matching points;

b222, performing spatial transformation on all characteristic points in the point set A by using the projection transformation matrix H obtained in the step B221 to obtain a point set B';

calculating the coordinate errors of all corresponding points in the point sets B and B ', namely e | | | B-B' | |; setting an error threshold value sigma, if e is less than sigma, considering the point pair as an inner point pair, otherwise, considering the point pair as an outer point pair;

b223, repeating the steps B221 and B222, finding out one transformation with the largest number of interior point pairs, taking the interior point pair set obtained by the transformation as a new point set A and B, and performing a new iteration;

b224 iteration end judgment: when the number of the inner point pairs obtained by iteration is consistent with the number of the point pairs in the point sets A and B before the iteration, the iteration is terminated;

b225, iteration result: finally, A and B of the iteration are the matching point sets after the mismatching characteristic point pairs are removed, and the corresponding projection transformation matrix H represents the required space transformation relation between the original image and the image to be detected;

b3, positioning the circle center of the hub and the position of the air tap in the image to be detected

Positioning the circle center of the wheel hub and the position of the air tap in the image to be detected by the wheel hub positioning module, and calculating the radius length of the wheel hub in the image to be detected; the method comprises the following specific steps:

b31, calculating six pixel points corresponding to the calibration points in the image to be detected according to the spatial transformation matrix H obtained in the step B221: circle center O ' (x ') of hub workpiece '₀,y′₀) Air nozzle center O 'of wheel hub workpiece'_gas(x′_gas,y′_gas) And four points O on the outer edge circumference of the hub₁′(x₁′,y₁′),O₂′(x′₂,y′₂),O₃′(x₃′,y₃′),O₄′(x′₄,y′₄)；

Coordinate O ' (x ') of center position of hub '₀,y′₀) For example, the following steps are carried out:

$\left\{\begin{array}{c}{x}_0^{\&prime;}=\frac{h_{00}{x}_0+h_{01}{y}_0+h_{02}}{h_{20}+h_{21}+1}\\ y_0^{\&prime;}=\frac{h_{10}{x}_0+h_{11}{y}_0+h_{12}}{h_{20}+h_{21}+1}\end{array}\right.---\left(12\right)$

b32, calculating the radius R' of the hub workpiece in the image to be detected:

$R^{\&prime;}=\frac{d_1+d_2+d_3+d_4}{4}---\left(13\right)$

whereini＝1,2,3,4。