CN104850851A - ORB feature point matching method with scale invariance - Google Patents

Abstract

The present invention relates to an ORB feature point matching method with scale invariance. The method is characterized by comprising the steps of a step S1 of inputting an image to be detected, performing improved SURF feature point detection on the image, and determining coordinates of the feature points; a step S2 of establishing an image pyramid for the image in the step S1; a step S3 of removing the feature points close to edges of the image; a step S4 of calculating directions of centers of mass of the remained feature points; a step S5 of calculating ORB feature point descriptors; a step S6 of adopting a K-nearest neighbor algorithm to carry out feature point matching; and a step S7 of screening feature point matching pairs and outputting the detected image. According to the method provided by the present invention, the SURF with the scale invariance and the ORB are combined, the image pyramid is introduced, and an ORB feature point matching algorithm is improved, so as to enable the method to have the scale invariance and maintain the characteristic of fastness of the ORB algorithm.

Description

A kind of ORB characteristic point matching method with scale invariability

Technical field

The invention belongs to figurepicture process field, particularly a kind of ORB characteristic point matching method with scale invariability.

Background technology

figurethe picture detection of unique point, description and coupling, be figurean extremely important branch in picture process field. figurepicture identification, video tracking, figurethe realization of many technology such as picture splicing, three-dimensional reconstruction, all will rely on figurethe picture detection of unique point, description and coupling.

ORB is a kind of outstanding feature point detection, description and matching algorithm, uses FAST Corner Detection and binary-coded Brief descriptor fast, its fast operation, illumination robustness is good, affinity is also good, but does not have scale invariability, causes ORB to have limitation in application.SURF is steadily and surely a kind of figurepicture identifies and describes algorithm, adopts integration figurepicture and box filter carry out convolution, using gaussian kernel function to build metric space, by solving quick Hessian response extraction unique point, utilizing moving window to build to the unique point extracted vertical maindirection, generates the feature descriptor of a 64bit.SURF has good scale invariability, but arithmetic speed is comparatively slow, is difficult to be applied in in the higher system of requirement of real-time.

Summary of the invention

In view of this, the object of this invention is to provide a kind of ORB characteristic point matching method with scale invariability, in conjunction with SURF, ORB is improved, introduce simultaneously figurepicture pyramid, makes it the outstanding illumination robustness of existing ORB, fast arithmetic capability, has again the scale invariability of SURF.

The present invention adopts following scheme to realize: a kind of ORB characteristic point matching method with scale invariability, and it comprises the following steps:

Step S1: input to be detected figurepicture, right figurepicture carries out the SURF feature point detection improved, and determines unique point coordinate;

Step S2: in described step S1 figurepicture is set up figurepicture pyramid;

Step S3: remove close figurethe unique point at picture edge;

Step S4: the barycenter direction calculating residue character point;

Step S5: calculate ORB unique point descriptor;

Step S6: adopt k nearest neighbor algorithm to carry out Feature Points Matching;

Step S7: screening Feature Points Matching to and output detections after figurepicture.

Further, the SURF feature point detection improved in described step S1 adopts integration figurepicture and box filter carry out convolution, build metric space solve quick Hessian response extraction unique point by gaussian kernel function, and at the 3*3*3 neighbor interpolation of described unique point, determine described unique point coordinate.

Further, described metric space is 16 layers of metric space structure of not dividing into groups.

Further, set up in described step S2 figureas pyramidal concrete grammar be: according to the size of Filtering Template in the SURF feature point detection improved, to be detected to what input figureit is down-sampled that picture carries out Gauss, and the down-sampled mesoscale value of described Gauss is wherein N is the size of Filtering Template.

Further, the square by calculating unique point in described step S4 draws the barycenter direction of unique point, and concrete grammar is: definition figurein picture, (p+q) rank square of any unique point neighborhood is: $m_{\mathrm{pq}}=\underset{x,y}{\mathrm{\Σ}}{x}^p{y}^{q}I(x,y),$ Then the center-of-mass coordinate C of its neighborhood is: $C=(C_x,C_y)=(\frac{m_{10}}{m_{00}},\frac{m_{01}}{m_{00}}),$ Can computing formula draw the barycenter direction θ of this unique point; Described formula is used to draw the barycenter direction of all residue characters point.

Further, the concrete grammar calculating ORB unique point descriptor in described step S5 is: step S51: adopt exhaust algorithm calculate n related coefficient close to 0.5 random point pair; Step S52: the random point in step S51 is rotated the barycenter direction according to unique point described in step S4, at Feature point correspondence figurein picture pyramidal layer, adopt formula $\mathrm{\&tau;}(p;x,y)=\left\{\begin{array}{ccc}1& \mathrm{if}& p\left(x\right)<p\left(y\right)\\ 0& \mathrm{otherwise}& \end{array}\right.$ Generate scale-of-two descriptor, in formula, (x, y) is postrotational random point pair, and p (x) is figurethe gray-scale value of picture block p at pixel x=(u, v) place; Show that n position binary bits string is unique point descriptor according to described formula.

Further, in described step S6, k nearest neighbor algorithm is specially: set K as 2, calculate each unique point to all unique points to be matched Hamming distance from, retain described 2 nearest points, with this unique point form mate right.

Further, screening the right concrete grammar of Feature Points Matching in described step S7 is: by the coupling of the coupling centering arest neighbors matching distance/time neighborhood matching distance > 0.6 in step S6 to rejecting, get the final match point that arest neighbors match point is this unique point.

Preferably, a coupling forms by the coordinate of two unique points and respective descriptor thereof, matching distance be Hamming distance between two descriptors from.

The invention has the advantages that and overcome the shortcoming that ORB does not have scale invariability, in conjunction with SURF, ORB is improved, and introduce figurepicture pyramid, makes the algorithm after improvement have scale invariability, retains the feature of the good illumination robustness of ORB and fast operation simultaneously.

Accompanying drawing explanation

figure1 is the overall flow of the embodiment of the present invention figure.

Embodiment

Below in conjunction with attached figureand embodiment the present invention will be further described.

The present embodiment provides a kind of ORB characteristic point matching method with scale invariability, as figureshown in 1, comprise the following steps:

Step S1: input to be detected figurepicture, right figurepicture carries out the SURF feature point detection improved, and determines unique point coordinate;

Step S2: in described step S1 figurepicture is set up figurepicture pyramid;

Step S3: remove close figurethe unique point at picture edge;

Step S4: the barycenter direction calculating residue character point;

Step S5: calculate ORB unique point descriptor;

Step S6: adopt k nearest neighbor algorithm to carry out Feature Points Matching;

Step S7: screening Feature Points Matching to and output detections after figurepicture.

In the present embodiment, the SURF feature point detection improved in described step S1 adopts integration figurepicture and box filter carry out convolution, build metric space solve quick Hessian response extraction unique point by gaussian kernel function, and at the 3*3*3 neighbor interpolation of described unique point, determine described unique point coordinate.

In the present embodiment, preferably, described metric space is 16 layers of metric space structure of not dividing into groups.

In the present embodiment, set up in described step S2 figureas pyramidal concrete grammar be: according to the size of Filtering Template in the SURF feature point detection improved, to be detected to what input figureit is down-sampled that picture carries out Gauss, and the down-sampled mesoscale value of described Gauss is wherein N is the size of Filtering Template.

In the present embodiment, the square by calculating unique point in described step S4 draws the barycenter direction of unique point, and concrete grammar is: definition figurein picture, (p+q) rank square of any unique point neighborhood is: then the center-of-mass coordinate C of its neighborhood is: $C=(C_x,C_y)=(\frac{m_{10}}{m_{00}},\frac{m_{01}}{m_{00}}),$ Can computing formula $\mathrm{\&theta;}=\arctan \left(\frac{C_y}{C_x}\right)=\arctan \left(\frac{m_{10}}{m_{01}}\right)$ Draw the barycenter direction θ of this unique point; Described formula is used to draw the barycenter direction of all residue characters point.

In the present embodiment, the concrete grammar calculating ORB unique point descriptor in described step S5 is:

Step S51: adopt exhaust algorithm calculate n related coefficient close to 0.5 random point pair;

Step S52: the random point in step S51 is rotated the barycenter direction according to unique point described in step S4, at Feature point correspondence figurein picture pyramidal layer, adopt formula $\mathrm{\&tau;}(p;x,y)=\left\{\begin{array}{ccc}1& \mathrm{if}& p\left(x\right)<p\left(y\right)\\ 0& \mathrm{otherwise}& \end{array}\right.$ Generate scale-of-two descriptor, in formula, (x, y) is postrotational random point pair, and p (x) is figurethe gray-scale value of picture block p at pixel x=(u, v) place; Show that n position binary bits string is unique point descriptor according to described formula.

In the present embodiment, in described step S6, k nearest neighbor algorithm is specially: set K as 2, calculate each unique point to all unique points to be matched Hamming distance from, retain described 2 nearest points, with this unique point form mate right.

In the present embodiment, screening the right concrete grammar of Feature Points Matching in described step S7 is: by the coupling of the coupling centering arest neighbors matching distance/time neighborhood matching distance > 0.6 in step S6 to rejecting, get the final match point that arest neighbors match point is this unique point.

In the present embodiment, especially, if the unique point that existence two groups is to be matched, then first can calculate the k nearest neighbor coupling of 2 stack features points respectively, reject asymmetric coupling right, then the coupling rejecting arest neighbors matching distance/time neighborhood matching distance > 0.6 is right.The coupling stayed after screening is right, gets the final match point that its arest neighbors match point is this feature.

In the present embodiment, a coupling forms by the coordinate of two unique points and respective descriptor thereof, matching distance be Hamming distance between two descriptors from.

The foregoing is only preferred embodiment of the present invention, all equalizations done according to the present patent application the scope of the claims change and modify, and all should belong to covering scope of the present invention.

Claims (8)

1. there is an ORB characteristic point matching method for scale invariability, it is characterized in that comprising the following steps:

Step S1: input image to be detected, to the SURF feature point detection that image improves, and determines unique point coordinate;

Step S2: image pyramid is set up to the image in described step S1;

Step S3: remove the unique point near image border;

Step S4: the barycenter direction calculating residue character point;

Step S5: calculate ORB unique point descriptor;

Step S6: adopt k nearest neighbor algorithm to carry out Feature Points Matching;

Step S7: screening Feature Points Matching to and export image to be detected.

2. a kind of ORB characteristic point matching method with scale invariability according to claim 1, it is characterized in that: the SURF feature point detection improved in described step S1 adopts integral image and box filter to carry out convolution, build metric space by gaussian kernel function and solve quick Hessian response extraction unique point, and at the 3*3*3 neighbor interpolation of described unique point, determine described unique point coordinate.

3. a kind of ORB characteristic point matching method with scale invariability according to claim 2, is characterized in that: described metric space is 16 layers of metric space structure of not dividing into groups.

4. a kind of ORB characteristic point matching method with scale invariability according to claim 1, it is characterized in that: the concrete grammar setting up image pyramid in described step S2 is: according to the size of Filtering Template in the SURF feature point detection improved, carry out Gauss to the image to be detected of input down-sampled, the down-sampled mesoscale value of described Gauss is wherein N is the size of Filtering Template.

5. a kind of ORB characteristic point matching method with scale invariability according to claim 1, it is characterized in that: the square by calculating unique point in described step S4 draws the barycenter direction of unique point, and concrete grammar is: in definition image, (p+q) rank square of any unique point neighborhood is: then the center-of-mass coordinate C of its neighborhood is: can computing formula draw the barycenter direction θ of this unique point; Described formula is used to draw the barycenter direction of all residue characters point.

6. a kind of ORB characteristic point matching method with scale invariability according to claim 1, is characterized in that: the concrete grammar calculating ORB unique point descriptor in described step S5 is:

Step S51: adopt exhaust algorithm calculate n related coefficient close to 0.5 random point pair; Step S52: rotated the barycenter direction according to unique point described in step S4 by the random point in step S51, on the image pyramid layer of Feature point correspondence, adopts formula generate scale-of-two descriptor, in formula, (x, y) is postrotational random point pair, and p (x) is for image block p is at the gray-scale value at pixel x=(u, v) place; Show that n position binary bits string is unique point descriptor according to described formula.

7. a kind of ORB characteristic point matching method with scale invariability according to claim 1, it is characterized in that: in described step S6, k nearest neighbor algorithm is specially: set K as 2, calculate each unique point to all unique points to be matched Hamming distance from, retain described 2 nearest points, with this unique point form mate right.

8. a kind of ORB characteristic point matching method with scale invariability according to claim 1, it is characterized in that: screening the right concrete grammar of Feature Points Matching in described step S7 is: by the coupling of the coupling centering arest neighbors matching distance/time neighborhood matching distance > 0.6 in step S6 to rejecting, get the final match point that arest neighbors match point is this unique point.