CN102521563A - Method for indentifying pig walking postures based on ellipse fitting - Google Patents

Abstract

The invention provides a method for indentifying pig walking postures based on ellipse fitting, which is characterized by comprising the following steps of: step (1), utilizing a background difference method to obtain a clear outline of a target pig; step (2), utilizing the ellipse fitting to determine characteristic parameters of all parts of the outline of the pig; and step (3), realizing the classification of different postures of normally standing, standing with a lower head, laying down and the like based on a pig walking posture classifier of a support vector machine.

Description

Pig walking posture recognition methods based on ellipse fitting

Technical field

The invention belongs to image procossing and machine vision technique.It is specific to further relate to ellipse fitting and animal behavior identification and sorting technique.

Background technology

Video and image are that objective things are vividly intuitively described.Into after 21st century, the paces of digitlization and networking are progressively accelerated, video monitoring system is gained great popularity with the advantage of intuitive and real-time, and it is applied to many aspects, including government bodies, electric power telecommunications, prison, army, bank, national treasury, supermarket, market, hotel, cell, school, office building etc..

With the improvement of living standards, people increase to the focus of food requirement from quantity has been increasingly turned to quality and quantity simultaneous growth.Pork based food has been a kind of indispensable food in people live, therefore, and the Quality advance of Pork product increases no less important with yield, that is to say, that not only to pay close attention to production yields, and to pay close attention to the whole growth course of product.Therefore, the research of the behavior of pig increasingly attracts attention.

The behavior of pig generally uses manual observation and hand-kept.Using manual observation and hand-kept, one side observer labor intensity is big, and poor working environment, observes the health by staff is had a strong impact in pig house for a long time；On the other hand due to only manually observing, because people is hand-tight, information can be caused inaccurate for the factor such as the fatigue and the degree of awareness of high cost and people and the phenomenon such as careless omission occurs, so, find more convenient, accurate and reliable automatic testing method into the urgent need to.

Pig body is an organic whole, the external presentation that its shape and appearance is not only body structure be also its interior tissue organ and grow, the embodiment of physiological function, behavioral aspect, it therefore, it can be used for understanding by the Geometrical Parameter information for obtaining pig body behavior and the physiological status of pig.

The present invention builds pig body Model using ellipse, using ellipse fitting algorithm, realizes the identification to pig behavior attitude, and this is to understanding and improves the growing environment and condition of pig, improves the welfare of pig, improves swine product quality significant.

This method extracts moving target (pig) from continuous video image, and the pig extracted is tracked, and its behavior is described and understood.Specifically, according to behavioural characteristics such as the relative position of pig corporal parts and its motions, using morphology, computer vision technique and quantitative analysis method, figure's parameter of description pig behavior is calculated, corresponding model is built, realizes and the walking posture of pig is recognized.

The content of the invention

The invention mainly comprises geometric figure modeling and ellipse fitting algorithm.Rim detection is carried out to the target pig in pigsty using background subtraction first, obtain the profile of pig, the contour images of pig are pre-processed using morphological method, the noise in profile is removed, ellipse fitting then is carried out to the point on pig body profile using improved arbitrary elliptical detection algorithm.Four ellipses are fitted to according to by positions such as incidence, body and four limbs, using the maximum oval center of circle of the area representated by pig body trunk as origin, the oval major axis is exactly the transverse axis of coordinate system, sets up a two-dimensional coordinate system O '.Then geometrical model is set up by each oval relative position distribution feature in the coordinate system and determines corresponding attitude parameter.These parameters are finally inputted into support vector machine classifier, to the normal stand of pig, standing is bowed and the different attitudes such as couch is identified and classified.

Accompanying drawing table explanation

Fig. 1 is the design flow chart of the present invention.

Fig. 2 is the training process of pig behavior attitude support vector machine classifier.

Fig. 3 is background image

Fig. 4 is the target image of collection

Fig. 5 is edge detection results figure

Fig. 6 is coordinate foundation figure

Fig. 7 is pig behavior posture feature parameter list diagram

Fig. 8 is the coordinate diagram and fitted figure of the different gestures of pig

Embodiment

The design flow chart of the present invention is as shown in Figure 1.Comprise the following steps that.

1. IMAQ is carried out to the pig in pigsty by image capturing system.

2. detecting the target pig in image using background subtraction, the profile of target pig is extracted；The profile extracted is handled using Morphology Algorithm, the cavity in the bur and blank map picture in pig body is removed respectively, a complete pig body profile is obtained.

3. ellipse fitting is carried out to the pig body profile in image using improved arbitrary elliptical detection algorithm and coordinate system is set up.

(1) improved arbitrary elliptical detection algorithm

There is the defect of bigger error in existing arbitrary elliptical detection (Randomized Ellipse Detection, RED) algorithm, done following improvement to RED algorithms herein in pig gesture recognition：1) from adaptive threshold T_dParameter is avoided to choose the problem of oval and true oval difference that is improper and causing fitting is excessive, ensure that the adjacent data point randomly selected is controlled in certain range intervals, 2) increase iterative cycles, the point not on original hypothesis ellipse is collected and is fitted again.Used in the algorithm to parameter be shown in Table 1.

The algorithm parameter table of table 1

(problem：(1) whether that row " the 4th point " of Td is misplayedWhether " data point "It is " data point "

(2) np, f, that row whether should np represent the quantity of marginal point in set U) be

Specific algorithm after improvement is as follows：

1) initialization counter f, makes f=0.

2) four data point P on pig body contour edge are taken_i, i.e. P_i∈ U, i=(1,2,3,4), make P_i∈ U meet 2 difference P of any of which_iGeometric distance can not be less than threshold value T_a；Remove taken data in set U, make U=U- { P_i}；Work as f=T_fOr n_p＜ T_emWhen, interrupted, provide result.

3) according to P_iFit the ellipse of hypothesis：Judge whether 4 data points are eligible, i.e., whether any one data point is more than threshold value T to the distance for assuming ellipse_d, if greater than equal to threshold value T_d, then this 4 data points are returned in set U, and counter plus 1, step 2 is gone to, next step is otherwise directly entered.

4) the hypothesis ellipse fitted in step 3 is set as E_ijk, initializing n makes n=0.Judge the data point P in set U_mTo E_ijkApart from L and threshold value T_dRelation, if L ＜ T_d, then make n=n+1, and set set V to be used for collecting not assuming that data point on ellipse, makes V=U-P_m。

5) all data points in set U are traveled through using step 4, you can be met threshold value T_dCounter Value n_e=n.

6) E is obtained_ijkGirth C_ijk, judge whether to meet n_e≥T_rC_ijk, if it is satisfied, then step 7 is jumped to, otherwise it is assumed that E_ijkIt is not present, then by the n in step 4_eIndividual data point is returned in U, and counter adds 1, i.e. f=f+1, then branches to step 2.

7) maximum for setting iterations T=0 and T is T_max, minimum rate of change T_n, and make N_old=n_e。

8) data point in the set V in step 4 is fitted again, and iterations plus 1, i.e. T=T+1.

9) by assuming that the oval E existed_ijkTo solve T_dValue；Travel through the data point in V, find it is selected to may be oval apart from D ＜ T_dData point, update n_eWith possible ellipse border point set V_e, and by n_eValue be assigned to N_new；If | N_new-N_old|/N_old＞ T_nAnd T ＜ T_t, then step 2 is jumped to；Otherwise algorithm is terminated, V=V-V_e；Confirm oval expression formula.

10) ellipse E is confirmed_ijkIt is real.By counter O reset, rebound step 2 is detected to other ellipses again.

By above-mentioned fit procedure it is known that inserting the process of an iteration in former RED algorithms, to that may be present and not assuming that the data point on ellipse re-starts collection, make the process of ellipse fitting more accurate.

(2) geometrical model of pig body is set up

Due to pig profile non-rigid characteristic and the difference of body parts size, the pig body after fitting is made up of many ellipses.The ellipse that the incidence of pig body, trunk, 4 parts of forelimb and hind leg are fitted is determined by oval size respectively, and four oval size relations are limbs before and after trunk ＞ incidences ＞.In all ellipses of fitting, the ellipse representated by pig body trunk is the ellipse of area maximum in all ellipses, a two-dimensional coordinate system O ' is set up by origin of this oval center of circle, the oval major axis is exactly the transverse axis of coordinate system.Then geometrical model is set up by oval relative position parameter in a coordinate system.

If the barycenter of the oval I (x, y) of any one in pig body is

Wherein

$\stackrel{\‾}{x}=\frac{1}{N}\underset{x,y}{\mathrm{\Σ}}I(x,y)x---\left(1\right)$

$\stackrel{\‾}{y}=\frac{1}{N}\underset{x,y}{\mathrm{\Σ}}I(x,y)y---\left(2\right)$

Here N is the elliptic region all pixels, and expression formula is

$N=\underset{x,y}{\mathrm{\Σ}}I(x,y)---\left(3\right)$

The covariance matrix of the elliptic region is

$\left[\begin{array}{cc}a& d\\ c& b\end{array}\right]=\frac{1}{N}\underset{x,y}{\mathrm{\Σ}}I(x,y)\left[\begin{array}{cc}{(x-\stackrel{\‾}{x})}^2& (x-\stackrel{\‾}{x})(y-\stackrel{\‾}{y})\\ (x-\stackrel{\‾}{x})(y-\stackrel{\‾}{y})& {(y-\stackrel{\‾}{y})}^2\end{array}\right]---\left(4\right)$

The eigenvalue λ of the covariance matrix₁、λ₂With corresponding characteristic vector v₁、v₂Represent length and the direction of the oval major and minor axis.

Transverse length is L, then L=λ₁, minor axis length M, then M=λ₂。

Oval deflection angle θ is exactly the angle of the transverse axis in oval major axis and coordinate system, i.e.,

$\mathrm{\θ}=\∠v_1={\cos }^{-1}\frac{v_1\left(\mathrm{1,0}\right)}{\left|v_1\right|}---\left(5\right)$

Characteristic vector relative to each frame is exactlyWherein i is four oval subscripts.

By assuming that the oval E existed_ijkTo solve adaptive threshold T_d：

$\left\{\begin{array}{c}{T}_d=|{\mathrm{au}}_A^2+u_A+v_A+{\mathrm{cv}}_A^2+{\mathrm{du}}_A+{\mathrm{ev}}_A+1|\\ u_A=u_0+(L+d_{\mathrm{dif}})\cos \mathrm{\θ}\\ v_A=v_0+(L+d_{\mathrm{dif}})\sin \mathrm{\θ}\end{array}\right.---\left(6\right)$

Wherein (u₀, v₀) it is oval center point coordinate, d_difTo put the ultimate range to oval border, L be oval long axis length respectively with θ and the different ellipses of oval deflection angle 4 in 4 parameters in respective F, including barycenter

The characteristic parameter of 16 is had with long axis length L and oval deflection angle theta, the feature of pig body gesture recognition is used as using this 16 features.

4. attitude is classified based on SVMs (SVM)

(1) the behavior attitude support vector machine classifier of pig is built

The main thought of SVMs is that the data of Nonlinear separability are passed through a conversion φ：R^N→ F is mapped to a High-dimensional Linear feature space F, then by solving constraint optimization problem

s.t.y_i(w·φ(x_i)+b)≥1-ξ_i, ξ_i>=0, i=1 ..., l (8)

Construct optimal separating hyper plane

H：F (x)=w φ (x)+b (9)

Wherein, w is the coefficient vector of Optimal Separating Hyperplane in feature space, and b is the threshold value of classifying face, and ξ is the relaxation factor for considering error in classification and introducing, and C is the penalty factor to error.

Feature space F dimension is general very big, directly calculates be nearly impossible wherein, but due to

Thus SVM all computings in feature space are all dot-product operations, Kernel-Based Methods are introduced in SVM, i.e.,

K(x_i, x_j)=φ (x_i)·φ(x_j)                           (10)

The concrete form for being then not required to clearly know can just be converted into the dot-product operation in high-dimensional feature space the kernel function computing of the low-dimensional input space, dexterously solve " dimension disaster " problem that calculating is brought in higher dimensional space.The present invention uses RBF

As kernel function, the performance of grader is directly influenceed by parameter size.The performance that radial direction base nuclear parameter σ directly affects SVM classifier is good and bad.Error punishment parameter C in formula (7) is used for realizing trading off between the ratio and algorithm complex of mistake point sample, i.e. it is determined that proper subspace in regulation machine learning fiducial range and empiric risk ratio, make the generalization ability of machine learning best.The present invention uses decision-directed acyclic graph method, and its basic thought is that multiple binary classifiers are combined into multi classifier.This method is in the training stage, and when number of classifying is M, the SVM numbers that it is constructed are M (M-1)/2；But, for the decision phase, party's rule is used by the guiding circulation figure root node.This circulation figure has M (M-1)/2 internal node and M leaf node, and each internal node is a binary classifier, and the node of leaf is last classification；For a given test sample, since root node, determine that it is walked right hand path or walks left hand path using the output valve of grader as foundation, until leaf node is terminated, so as to show that sample belongs to such result.

(2) training of SVMs (SVM) attitude grader

Collection is used as the different normal stands (walking) of the pig of training, and the classification sample images such as standing (walking) and prostrate of bowing extract corresponding characteristic, and form pig attitude data storehouse using the method in step 3.SVMs (SVM) grader is trained using these training set sample datas simultaneously, its training process is as shown in Figure 2.

(3) support vector machine classifier is classified to the behavior attitude of pig

The method in step 3 is used to extract corresponding characteristic parameter data the image of the behavior attitude of tested pig, these supplemental characteristics are input in the SVM trained, the normal stand (walking) to pig is realized, standing (walking) and prostrate etc. behavior attitude of bowing are identified.

The behavior gesture recognition process of pig

Before collection pig image, the pigsty background existed without any pig is updated, background image is shown in Fig. 3.The image of pig in a width pigsty is gathered first, Fig. 4 is seen, the profile of pig in pigsty is then detected using background difference algorithm, as shown in Figure 5.Ellipse fitting is carried out to the profile of pig using improved arbitrary elliptical detection algorithm；In all ellipses of fitting, the ellipse representated by pig body trunk is the ellipse of area maximum in all ellipses, a new two-dimensional coordinate system O ' is set up by origin of this oval center of circle, the oval major axis is exactly the transverse axis of coordinate system, as shown in Figure 6.Checking area is only smaller than whether the oval incidence of trunk is oval is located at the oval left side of trunk, if incidence ellipse carries out setting up coordinate system again positioned at trunk oval right side after picture is carried out into horizontal deflection.

Pig body is modeled using 4 ellipses, 4 ellipses are respectively the oval I of incidence_headThe oval I of (x, y), body_bodyThe oval I of (x, y), forelimb_forelegThe oval I of (x, y), hind leg_hindleg(x, y).In 4 different ellipses, 4 parameters in each ellipse, including barycenter

Long axis length L and oval deflection angle theta have 16 features, and the feature of pig body gesture recognition is used as using this 16 features.As shown in fig. 7, wherein coordinate system is O ', I_head(x, y), I_body(x, y), I_foreleg(x, y), I_hindleg(x, y) is four ellipses in coordinate system respectively, and L is oval major axis, in figure (7), with I_bodyRepresented exemplified by (x, y).θ is oval I_headThe major axis of (x, y) and ellipse I_bodyThe angle of the major axis of (x, y), referred to as deflection angle, in figure (7), with I_head(x, y) is represented exemplified by the deflection angle of coordinate system.

When judging pig midstance, the position of forward and backward limb fitted ellipse can not consider, only consider ellipse I_head(x, y) and ellipse I_bodyThe relative position of (x, y).When pig normal stand, its incidence fitted ellipse is centrally located at reference frame O ' second or third quadrant, the I that such as Fig. 8 (a) shows_head(x, y), and ellipse I_headThe major axis of (x, y) and ellipse I_bodyThe angle of the major axis of (x, y), i.e. deflection angle are less than 30 °；And pig is bowed when standing, the center of its incidence fitted ellipse is then located at the third quadrant of reference frame, the I as shown in Fig. 8 (b)_head(x, y) most of third quadrant in a coordinate system, but deflection angle is more than 30 °；

When judging whether pig couches, the position of forward and backward limb fitted ellipse is mainly considered.When pig couches, forward and backward limb fitted ellipse is then located at reference frame O ' third quadrant and fourth quadrant, and ellipse I_foreleg(x, y) and I_hindlegThe major axis of (x, y) and ellipse I_bodyThe major axis of (x, y) is almost parallel, and its angle is less than 10 °, and two ellipse areas also very little, shown in such as Fig. 8 (c).

Claims (4)

1. the pig walking posture recognition methods based on ellipse fitting, it is characterised in that comprise the following steps：

Step（One）The method that the clearly profile of destination object pig is obtained using background subtraction；

Step（Two）Each genius loci parameter of pig body profile is determined using ellipse fitting；

Step（Three）Pig walking posture grader based on SVMs, so as to realize to normal stand, bows standing and the different attitude such as couch is classified.

2. the pig walking posture recognition methods according to claim 1 based on ellipse fitting, it is characterised in that the step（One）Comprise the following steps：

Step（1）The target pig in image is detected using background subtraction, the profile of target pig is extracted；

Step（2）Morphological scale-space is carried out to the profile extracted using Morphological scale-space method, the cavity in the bur and blank map picture in pig body is removed respectively；

Step（3）The edge contour of pig in image is extracted using edge detection operator.

3. the pig walking posture recognition methods according to claim 1 based on ellipse fitting, it is characterised in that the step（Two）In determine that the feature of each genius loci parametric technique of pig body profile is that ellipse fitting is carried out to the pig body profile each several part in image and coordinate system is set up using ellipse fitting, pig body after fitting is made up of many ellipses, the ellipse that 4 parts of pig body are fitted is determined by oval size respectively, then by representing head, the ellipse of trunk, forelimb and hind leg determines each genius loci parameter of pig body profile in the position of coordinate system.

4. the pig walking posture recognition methods according to claim 1 based on ellipse fitting, it is characterised in that the step（Three）In the pig walking posture grader based on SVMs feature be using SVMs (SVM) design pig walking posture grader, the input of grader, normal stand of the realization to pig are used as using each genius loci parameter of pig body profile that ellipse fitting is determined（Walking）, standing of bowing（Walking）With the identification of the attitude such as prostrate.

Images