CN103824056A - Pig posture recognition method based on Zernike moment and support vector machine - Google Patents

Abstract

The invention provides a pig posture recognition method based on a Zernike moment and a support vector machine. The method mainly includes the steps of conducting side view scene video image collection on behavior states of a pig in a pig house through a digital camera, preprocessing video images to extract a two-value skeleton map of the pig, conducting transverse moving normalization and dimension normalization on the two-value skeleton map of the pig with the method of regular moments, conducting feature extraction on the normalized images through the Zernike moment to obtain feature vectors, conducting design of a classifier for the behavior states of the pig based on the support vector machine, and conducting classification identification on the behavior states of the pig through the support vector machine for the feature vectors. Identification of the normal walking posture, the head-lowered walking posture, the head-raised walking posture and the lie-down posture of the pig can be identified through the machine vision technology and the support vector machine technology.

Description

The gesture recognition method of a kind of pig based on Zernike square and support vector machine

Technical field

The present invention relates to machine vision technique field, be specifically related to the gesture recognition method of a kind of pig based on Zernike square and support vector machine.

Background technology

Animal behavior analysis is always subject to faunist's attention, and it can be used for setting up the behavior spectrum of animal, is used for zoologizeing the various actions within a period of time and the time of occupying thereof.The research of animal behavior spectrum is understood to be again " behavioral formation " sometimes.

The behavior spectrum of setting up animal requires anthropologist to get along with animal for prolonged period of time conventionally, and in the situation that not disturbing their various daily routines, correctly and at length records the various actions type of zoologizeing.Because the behavior of animal has the specificity of planting as the form of animal, behavior and form are all the products of long-term evolution.In fact, each animal has factum spectrum, but anthropologist also just has than more comprehensive understanding the behavior spectrum of few animals so far.This is that this method wastes time and energy because traditional animal behavior analysis is by manual observation and records the behavior state of animal, and efficiency is low, human cost is large, and animal-breeding environmental baseline is poor, and staff's physical and mental health can be brought and be had a strong impact on.

Along with the development of computer technology and Digital Image Processing theory, machine vision is penetrated into agriculture every field gradually.At present, also have utilizing in varying degrees machine vision technique to measure the body weight of pig both at home and abroad, carry out simple animal tracking and identification.For example, in 2009, the people such as the J.M.Navarro-Jover of Valencia Polytechnics of Spain, studied auto color algorithm based on computer vision and followed the trail of the position of piglet, respectively piglet below and side, describe each RGB image space with the different colors mark that sprays paint.The method of this utilization is color characteristic, catches the piglet that uses different colours mark in image.But this method different color used is sprayed paint, mark can fade along with the elongated of time, and especially in pig house, environment is poor, and pig can be stained with the dirts such as spot with it, can change these colors mark that sprays paint, and causes the method not implement.Therefore this can only be experiment ideally, and practicality is very restricted.

Summary of the invention

The deficiency waste time and energy in order to overcome prior art, efficiency being low, human cost is large, the invention provides the gesture recognition method of a kind of pig based on Zernike square and support vector machine, utilize that machine vision and support vector machine technology are realized the normal walking to pig, the walking of bowing, the four kinds of attitudes such as walk, couch that come back are identified.

The present invention solves the scheme that its technical matters adopts: the gesture recognition method of a kind of pig based on Zernike square and support vector machine, comprise the steps,

(1) adopt digital camera to carry out the collection of side-looking scene video image A to the behavior state of pig in pig house;

(2) video image A is carried out to pre-service, extract the two-value profile diagram B of pig;

(3) method of employing standard square, carries out the normalization of translation and yardstick to the two-value profile diagram B of pig, obtains image C;

(4) to image C, adopt Zernike square to carry out feature extraction, obtain proper vector D;

(5) design of the behavior attitude sorter of the pig based on support vector machine;

(6), to described proper vector D, adopt support vector machine to carry out the behavior attitude of Classification and Identification pig.

Further, step is carried out pre-service to video image A in (2), and the step that extracts the two-value profile diagram B of pig is:

A. adopt geometric transformation, method level and smooth, that strengthen to carry out pre-service to video image A, obtain image A 1;

B. select Otsu adaptive thresholding to cut apart image A 1 is carried out to binary conversion treatment, obtain image A 2;

C. image A 2 is carried out to medium filtering, mathematical morphology opening and closing operation, obtain image A 3;

D. use Canny operator to make rim detection to image A 3, extract the two-value profile diagram B of pig.

Further, in step (3), normalized step is:

A. defining standard square is:

in formula, exponent number p, q=0,1,2 ..., m _pqfor standard square, f (x, y) is target image, and x and y represent the coordinate position of pixel, and f is the numerical value of represent pixel;

B. obtain the centre of form coordinate of target image by standard square wherein

image origin is placed in the target centre of form, to solve translation problem;

C. define a scale factor α, wherein

in formula, β is predefined constant, m ₀₀it is the area of target;

D. obtain the image g (x, y) after translation, yardstick normalization by coordinate transform, its computing formula is: $g(x,y)=f(\frac{x}{\mathrm{\α}}+\stackrel{\‾}{x},\frac{y}{\mathrm{\α}}+\stackrel{\‾}{y}).$

Further, the step that step (4) adopts Zernike square to carry out feature extraction is:

A. design Zernike orthogonal polynomial;

B. utilize the Zernike polynomial construction that orthogonal p rank q is heavy to obtain multistage Zernike square, extract multistage Zernike moment characteristics value, as the input quantity of the behavior gesture recognition of pig.

Further, the design of the behavior attitude sorter of the pig of step (5) based on support vector machine, the steps include:

A. the behavior posture feature of pig is transformed into the feature space of higher-dimension by nonlinear transformation, in higher dimensional space, constructs linear decision function and realize the non-linear decision-making in former space;

B. by adopting the parameter selection method of cross validation and web search to construct a kind of study mechanism of optimization, design support vector machine classifier;

C. adopt man-to-man temporal voting strategy, two sorters are formed and meet normal walking to pig, the walking of bowing, new line four kinds of sorters that attitude is identified such as walk, couch.

Further, the support vector machine of step (5) is selected radial basis kernel function type support vector machine, and its definition is K (x _i, x _j)=exp (γ || x _i-x _j||), γ >0, wherein γ is radial basis kernel functional parameter; Exp() be that e is the exponential function at the end, any point in Xi representation space, Xj is kernel function center.

The invention has the beneficial effects as follows and utilize machine vision and support vector machine technology, four kinds of attitudes are identified to realize that normal walking to pig, the walking of bowing, new line are walked, couched etc., for building the behavior spectrum of pig, the impact of the behavior state on pig such as research breeding environment, nutritional condition, the pig-breeding pattern that forms efficient and healthful provides technical method.

Accompanying drawing explanation

Below in conjunction with the drawings and specific embodiments, the present invention is described in further detail.

Fig. 1 is the inventive method process flow diagram.

In figure: A. video image, the two-value profile diagram B of B. pig, image after C. normalization, D. proper vector.

Embodiment

As shown in Figure 1, the gesture recognition method of a kind of pig based on Zernike square and support vector machine, comprises the steps:

(1) adopt digital camera to carry out the collection of side-looking scene video image A to the behavior state of pig in pig house;

(2) video image A is carried out to pre-service, extract the two-value profile diagram B of pig;

(3) method of employing standard square, carries out the normalization of translation and yardstick to the two-value profile diagram B of pig, obtains image C;

(4) to image C, adopt Zernike square to carry out feature extraction, obtain proper vector D;

(5) design of the behavior attitude sorter of the pig based on support vector machine;

(6), to described proper vector D, adopt support vector machine to carry out the behavior attitude of Classification and Identification pig.

Embodiment explanation:

(1) adopt digital camera to carry out the collection of side-looking scene video image A to the behavior state of pig in pig house.

Adopt digital camera, adopt side to take obtain side-looking scene video image to the pig in pig house, resolution is 659*454, and frame per second is 29fps, and the image of acquisition is colored.

(2) video image A is carried out to pre-service, extract the two-value profile diagram B of pig.

Before feature extraction, adopt the methods such as geometric transformation, level and smooth, enhancing to carry out pre-service, select Otsu adaptive thresholding to cut apart image is carried out to binary conversion treatment, again the image after binaryzation is carried out to medium filtering, mathematical morphology opening and closing operation, finally make rim detection with Canny operator, correctly extract the contour images of pig in original image.

(3) method of employing standard square, carries out the normalization of translation and yardstick to the two-value profile diagram B of pig, obtains image C.

Because Zernike square only has rotational invariance, so need image to carry out the normalization of translation and yardstick before bianry image is carried out to feature extraction, adopt the method for standard square to be normalized, standard square is defined as:

$m_{\mathrm{pq}}=\mathrm{\Σ\Σ}{x}^p{y}^{q}f(x,y)$

(1) in formula, exponent number p, q=0,1,2 ..., the shape function of image can be thought a conversion coefficient (standard square) m _pqthe unlimited set of composition, f (x, y) is target image, and wherein, x and y represent the coordinate position of pixel, and f is the numerical value of represent pixel.

Obtain the centre of form coordinate of target image (pig) by standard square

see formula (2).

$\stackrel{\‾}{x}=m_{10}/m_{00},\stackrel{\‾}{y}=m_{01}/m_{00}---\left(2\right)$

Image origin is placed in the target centre of form and can solves translation problem, then define a scale factor

$\mathrm{\α}=\sqrt{\mathrm{\β}/m_{00}}---\left(3\right)$

With solving scale problem, wherein β is predefined constant, and β gets 10000 herein,, if target pixel value is 1 on bianry image, background is pixel value 0, m ₀₀be the area of target, the object of doing is like this exactly that to make the area of target be a fixing size, has just solved the problem of yardstick.Obtain the image g (x, y) after translation, yardstick normalization by coordinate transform below, computing formula is shown in formula (4).

$g(x,y)=f(\frac{x}{\mathrm{\α}}+\stackrel{\‾}{x},\frac{y}{\mathrm{\α}}+\stackrel{\‾}{y})---\left(4\right)$

Carry out the feature extraction of Zernike square by the image after the normalization of standard square, just can obtain having the feature of translation, yardstick and invariable rotary.

(4) to image C, adopt Zernike square to carry out feature extraction, obtain proper vector D.

The step that employing Zernike square carries out feature extraction is as follows:

A. design Zernike orthogonal polynomial.

Zernike polynomial expression { V _pq(x, y) } be to be defined in unit circle (x ²+ y ²≤ 1) complex value Orthogonal Function Set on, that there is completeness.{ V _pq(x, y) } completeness and orthogonality make can represent to be defined in any quadractically integrable function on unit circle, its representation:

V _pq(x,y)=V _pq(r,θ)=R _pq(r)e ^ipθ （5）

In formula (5), r represents the vector length of former point-to-point (x, y), and r≤1, and θ represents the angle of vector r and x axle, and Rpq (r) is the radial polynomial of a real number value, is provided by formula (6):

$R_{\mathrm{pq}}\left(r\right)=\underset{s=0}{\overset{(p-|q\left|\right)/2}{\mathrm{\Σ}}}{(-1)}^s\frac{(q-s)!}{s!((q-2s+|p\left|\right)/2)!((q-2s-|p\left|\right)/2)!}{r}^{q-2s}---\left(6\right)$

These polynomial expressions are orthogonal and satisfied:

$\underset{x^2+y^2\≤1}{\∫\∫}{\left\{V_{\mathrm{pq}}(x,y)\right\}}^{*}\left\{V_{\mathrm{nm}}(x,y)\right\}{d}_x{d}_y=\frac{\mathrm{\π}}{n+1}{\mathrm{\δ}}_{\mathrm{pn}}{\mathrm{\δ}}_{\mathrm{qm}}---\left(7\right)$

(7) in formula,

{ V _pq(x, y) } ^*{ V _pq(x, y) } conjugation.

B.Zernike square.

Due to the polynomial orthogonal and complete properties of Zernike, so any piece image f (x, y) in unit circle can be launched by formula (8) uniquely:

$f(x,y)=\underset{p=0}{\overset{\∞}{\mathrm{\Σ}}}\underset{q=0}{\overset{\∞}{\mathrm{\Σ}}}{A}_{\mathrm{pq}}{V}_{\mathrm{pq}}(r,\mathrm{\θ})---\left(8\right)$

(8) in formula, function V _pq(r, q) is the heavy Zernike polynomial expression of orthogonal p rank q in the unit circle of polar coordinate system.P is a nonnegative integer, and q is that the integer meeting the following conditions a: p-|q| is even number, and | q|≤p.

(8) complex coefficient value in formula:

$A_{\mathrm{pq}}=\frac{n+1}{\mathrm{\π}}\underset{x^2+y^2\≤1}{\∫\∫}f(x,y){\{V_{\mathrm{pq}}(r,\mathrm{\θ})}^{*}{d}_x{d}_y----\left(9\right)$

Integrated form can be changed into for discrete picture:

$A_{\mathrm{pq}}=\frac{n+1}{\mathrm{\π}}\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}f(x,y){\{V_{\mathrm{pq}}(r,\mathrm{\θ})}^{*}----\left(10\right)$

(10) in formula, x ²+ y ²≤ 1.

In the time calculating the Zernike square of piece image, using the center of gravity of image as initial point, and all pixels are mapped in unit circle, the point outside unit circle does not calculate, and can obviously find out from formula (9)

the amplitude of optional like this Zernike square | A _pq| as the feature that when classification needs, calculate high-order from 0 to 7 rank of Zernike totally 20 groups of moment characteristics.

(5) design of the behavior attitude sorter of the pig based on support vector machine.

After extracting the feature of image, selecting suitable sorter is the key of identification problem.The present invention classifies to four of pig kinds of daily attitudes, is a polytypic problem, need to construct support vector machine multicategory classification device.Support vector machine method is that the VC that is based upon Statistical Learning Theory ties up on theoretical and structure risk minimum principle basis, between the complicacy (i.e. the study precision to specific training sample) of model and learning ability (identifying error-free the ability of arbitrary sample), seek optimal compromise according to limited sample information, to obtaining best Generalization Ability.

A. support vector machine.

Support vector machine classification problem can be described as: be provided with classified sample set (x ₁, y ₁) ..., (x _l, y _l), x ∈ R ⁿ, y ∈ 1, and-1}, wherein l is classification samples number, and n is sample dimension, and y is two dissimilar marks, builds a decision function by these samples, makes it test data as far as possible correctly classify.Classification is exactly to seek a lineoid ω ^tx+b=0(11) make it two class samples separate completely ω ∈ R here ⁿfor planar process vector, b is amount of bias, and wherein, mutual spacing is 2/|| ω from two of maximum such lineoid spacings ||, represent the interval between two class samples.Support vector machine classification tries hard to make two class sample interval maximums, is equivalent to make || ω || and minimize, for making classification there is certain flexibility, introduce a lax ξ simultaneously _i(ξ _i>0) and corresponding wrong penalty coefficient C, finally can be summed up as and solve following quadratic programming problem:

$\min (\frac{1}{2}{\left|\right|\mathrm{\ω}\left|\right|}^2+C\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\ξ}}_i)---\left(12\right)$

y _i(ω·x _i+b)≥1-ξ _i i=1,...l （13）

Introduce Lagrange's multiplier a _i(i=1,2 ... l, a _i>=0) build Lagrangian function and solve, be converted into formula (14) dual problem:

$\max W\left(a\right)=\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{a}_i-\frac{1}{2}\underset{i=1}{\overset{l}{\mathrm{\Σ}}}\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{a}_i{a}_j{y}_i{y}_j{x}_i^T{x}_j---\left(14\right)$

In formula (14), $\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{a}_i{y}_i=0,0\≤a_i\≤C,i=1,...l.$

When classified sample set is while being non-linear, can training sample be mapped to higher-dimension by nonlinear transformation φ () and even in Infinite-Dimensional Space, seek linear separability.In conjunction with the relevant theory of functional, meet the kernel function of Mercer condition corresponding to the inner product of a certain transformation space, formula (15) provides:

<x _i·x _j>→K(x _i,x _j)=φ(x _i) ^T·φ(x _j) （15）

Support vector machine is incorporated into kernel function in dual problem, obtains new quadratic programming objective function to be:

$\left(a\right)=\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{a}_i-\frac{1}{2}\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{a}_i{a}_j{y}_i{y}_{j}K(x_i,x_j)---\left(14\right)$

After solving, obtaining categorised decision function is:

$f\left(x\right)=\mathrm{sgn}(\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{a}_i^{*}{y}_{i}K(x_i,y_i)+b)---\left(17\right)$

Conventional kernel function comprises linear kernel (Linear), polynomial kernel (Polynomial), radial basis core (RBF) and S type (Sigmoid) etc., and this selects radial basis kernel function type support vector machine (SVM-RBF), and formula (18) is shown in its definition:

K(x _i,x _j)=exp(-γ||x _i-x _j||),γ>0 (18)

In formula (18), γ is radial basis kernel functional parameter, exp() be that e is the exponential function at the end, any point in Xi representation space, Xj is kernel function center.

B. the design of the behavior attitude sorter based on support vector machine.

The present invention identifies the side view of four of pig kinds of attitudes, and four kinds of attitudes are respectively normal walkings, the walking of bowing, the walking that comes back, couch.First from the side view gathering, choose the image pattern collection that quantity is enough, comprise posture feature as much as possible, and above-mentioned image pattern collection is divided into training set and test set, realize the training and testing checking of attitude sorter.In this example, choose respectively four kinds of attitude image 25 width, totally 100 width.Adopt Zernike square method to extract its eigenwert to the two-value profile diagram that obtains pig after pre-service; With selecting every kind of attitude image 15 width images as training sample, training obtains the behavior attitude sorter of support vector machine, carry out testing authentication with every kind of attitude image of other 10 width as test sample book again, finally form behavior attitude sorter, realize the Classification and Identification of behavior attitude.

Support vector machine belongs to binary classification device, and for multiclass pattern recognition problem, support vector machines can be realized by the combination of two class problems.Conventionally strategy has: the peak response strategy of one-to-many, man-to-man temporal voting strategy, man-to-man replacement policy etc.This adopts man-to-man temporal voting strategy, and for 4 class problems, temporal voting strategy is by above-mentioned four class attitude A one to one, B, C, D tetra-class sample combination of two become training set, i.e. (A, B), (A, C), (A, D), (B, C), (B, D), (C, D), obtain 6 (for n class problem, being n (n-1)/2) SVM riffles.In test, test sample book is sent into this 6 two sorters successively, take ballot form, finally obtain one group of result.

Choose and in the situation of gaussian radial basis function kernel function (RBF), always have two parameters and need to determine, be i.e. the parameter γ of RBF core self and wrong cost coefficient C.This adopts the parameter selection method based on cross validation and grid search to carry out parameter optimization, and target function value is exactly the discrimination of SVM for test set.

(6), to proper vector D, adopt support vector machine to carry out the behavior attitude of Classification and Identification pig.

Bianry image after normalization is extracted to Zernike square, can obviously find out from formula (9)

the amplitude of optional like this Zernike square } A _pq| as the feature that when classification needs, calculate high-order from 0 to 7 rank of Zernike totally 20 groups of moment characteristics, 60, training picture, 40, test picture.

Using the moment characteristics of the every class 15 width two-value contour images of 4 class attitude as training input, be input to and in support vector machine, train correlation parameter.Then to the every class 10 width images of 4 class attitude two-value profile as test pattern, extract the Zernike square of test pattern, select Gaussian radial basis function (RBF) as kernel function.Adopt parameter selection method based on cross validation and grid search parameter γ and the wrong cost coefficient C parameter optimization to RBF core self, represent respectively the classification that the normal walking of identification, the walking of bowing, new line are walked, couched with 1,2,3,4.

Claims (6)

1. a gesture recognition method for the pig based on Zernike square and support vector machine, is characterized in that, comprises the steps:

(1) adopt digital camera to carry out the collection of side-looking scene video image A to the behavior state of pig in pig house;

(2) described video image A is carried out to pre-service, extract the two-value profile diagram B of pig;

(3) method of employing standard square, carries out the normalization of translation and yardstick to the two-value profile diagram B of described pig, obtains image C;

(4) to described image C, adopt Zernike square to carry out feature extraction, obtain proper vector D;

(5) design of the behavior attitude sorter of the pig based on support vector machine;

(6), to described proper vector D, adopt support vector machine to carry out the behavior attitude of Classification and Identification pig.

2. the gesture recognition method of a kind of pig based on Zernike square and support vector machine according to claim 1, is characterized in that, in described step (2), described video image A is carried out to pre-service, and the step that extracts the two-value profile diagram B of pig is:

A. adopt geometric transformation, method level and smooth, that strengthen to carry out pre-service to described video image A, obtain image A 1;

B. select Otsu adaptive thresholding to cut apart described image A 1 is carried out to binary conversion treatment, obtain image A 2;

C. described image A 2 is carried out to medium filtering, mathematical morphology opening and closing operation, obtain image A 3;

D. use Canny operator to make rim detection to described image A 3, extract the two-value profile diagram B of pig.

3. the gesture recognition method of a kind of pig based on Zernike square and support vector machine according to claim 1, is characterized in that, in described step (3), normalized step is:

A. defining standard square is: in formula, exponent number p, q=0,1,2 ..., m _pqfor standard square, f (x, y) is target image, and x and y represent the coordinate position of pixel, the numerical value of f represent pixel;

B. obtain the centre of form coordinate of target image by standard square

wherein

image origin is placed in the target centre of form, to solve translation problem;

C. define a scale factor α, wherein

in formula, β is predefined constant, m ₀₀it is the area of target;

D. obtain the image g (x, y) after translation, yardstick normalization by coordinate transform, its computing formula is: $g(x,y)=f(\frac{x}{\mathrm{\&alpha;}}+\stackrel{\&OverBar;}{x},\frac{y}{\mathrm{\&alpha;}}+\stackrel{\&OverBar;}{y}).$

4. the gesture recognition method of a kind of pig based on Zernike square and support vector machine according to claim 1, is characterized in that, the step that described step (4) adopts Zernike square to carry out feature extraction is:

A. design Zernike orthogonal polynomial;

B. utilize the Zernike polynomial construction that orthogonal p rank q is heavy to obtain multistage Zernike square, extract multistage Zernike moment characteristics value, as the input quantity of the behavior gesture recognition of pig.

5. the gesture recognition method of a kind of pig based on Zernike square and support vector machine according to claim 1, is characterized in that, the design of the behavior attitude sorter of the pig of described step (5) based on support vector machine, the steps include:

A. the behavior posture feature of pig is transformed into the feature space of higher-dimension by nonlinear transformation, in higher dimensional space, constructs linear decision function and realize the non-linear decision-making in former space;

B. by adopting the parameter selection method of cross validation and web search to construct a kind of study mechanism of optimization, design support vector machine classifier;

C. adopt man-to-man temporal voting strategy, two sorters are formed and meet normal walking to pig, the walking of bowing, new line four kinds of sorters that attitude is identified such as walk, couch.

6. the gesture recognition method of a kind of pig based on Zernike square and support vector machine according to claim 1, is characterized in that, the support vector machine of described step (5) is selected radial basis kernel function type support vector machine, and its definition is K (x _i, x _j)=exp (γ || x _i-x _j||), γ >0, wherein γ is radial basis kernel functional parameter, exp() be that e is the exponential function at the end, any point in Xi representation space, Xj is kernel function center.

Images