CN103279751A - Eye movement tracking method on the basis of accurate iris positioning - Google Patents

Abstract

The invention discloses an eye movement tracking method on the basis of accurate iris positioning. The method mainly comprises a human face detection module, a human eye detection module, an eye movement tracking module and the like. A human face area is detected according to a detector based on the AdaBoost algorithm. A human eye is detected on the basis of human face area detection. An iris area is accurately located. A tracking template based on combined characteristics of target colors and edges is built up. The influence of saturability components is considered when the color characteristic template is built up. The condition that the iris can deform due to the fact that an eye is closed is considered in the tracking stage. Self-defined standards are used for judging, so that the tracking template is adjusted in a self-adapting mode. Consequently, eye movement tracking is finished. Compared with other methods, the method is higher in tracking speed, higher in tracking accuracy, and wide in application range, meets requirements for real-timeness, accuracy, and robustness, and has good application prospects in handheld mobile equipment.

Description

A kind of eye movement tracking based on accurate location iris

Technical field

The present invention relates to a kind of eye movement tracking of the non-intrusion type based on accurate location iris.Relate in particular to a kind of trace template based on iris region color and edge associating feature, finish the method that eye movement is followed the tracks of according to self-defining criterion.Belong to the computer vision message area.

Background technology

Eyes are windows of soul, and we can probe into the rule of many human psychology activities to see through this window.Human information processing depends on vision to a great extent, therefore, never is interrupted for the scientific research of " how the people sees things ".About this point, be considered to the most effective means of visual information processing for the research of eye movement.Eye movement follow the tracks of (Eye Tracking) when mainly referring to utilize eye movement with geostationary some feature as object of reference, thereby information such as the motion track of opening the state of closing, direction of gaze, time and blinkpunkt of acquisition eyes and translational speed have good application prospects in fields such as portable terminal, medical science, interface design and assessment, product test, scene research, performance analysis, Aero-Space, man-machine interaction, virtual reality and recreation.

At present method commonly used has based on motion, based on the algorithm of feature and template, based on the eye movement instrument with based on the method for nonlinear filtering theory etc.Based drive eye movement track algorithm belongs to the intrusive mood eye movement to be followed the tracks of, and often need comprise electroculogram method and solenoid method by extraneous utility appliance.This method tends to cause tester's ophthalmic uncomfortable, therefore is unsuitable for conventional eye movement and follows the tracks of.Eye movement tracking based on feature and template refers to set up template according to biology and geometric properties after accurately locating human eye, realizes following the tracks of, and comprises template matching method, optimum linear approximatioss, median algorithm etc.But this method is subjected to the external interference factor bigger, and accuracy and robustness are still waiting to improve.Eye movement tracking based on the eye movement instrument need adopt complexity and expensive large-scale precision instrument, just begun development abroad in 20 beginnings of the century, along with fast development of computer technology, till now, a lot of companies have all developed Related product, such as Australian Seeing Machines, U.S. ASL, Japanese ISCN, German SMI, Canadian SR company etc.Domesticly start late comparatively speaking, professor Fu Rui of Chang An University and Xian Electronics Science and Technology University have carried out the research of related fields, yet it is further perfect to still need in precision, hommization and miniaturization.Based on the develop rapidly along with control theory of the eye movement track algorithm of nonlinear filtering theory, become research focus in recent years.Theoretical method commonly used comprises Kalman wave filter, EKF wave filter, UKF wave filter etc.But sampling filter also needs more deep research for the precision, real-time and the robustness that how to improve in eye movement is followed the tracks of, and the tracking analysis and evaluation is as shown in table 1.

Table 1 eye movement tracking is analyzed

Therefore, proposing a kind of rapid eye movement tracking based on non-intrusion type is problem demanding prompt solution.Fig. 1 shows typical eye movement trace flow figure based on non-intrusion type, and core is human eye detection and tracking, and wherein human eye detection is important link, is related to the accuracy of target following, is the foundation stone of whole eye movement tracker.At the eye movement tracking phase, in order to reach accuracy, robustness, real-time requirement, many problems that need solution are arranged.

Therefore, improve the eye movement tracking velocity, the requirement that really reaches accuracy, robustness, real-time becomes the problem that this class eye movement tracking presses for solution.

Summary of the invention

Purpose of the present invention is set up the target following template with regard to being that providing a kind of in order to address the above problem accurately locatees iris in the human eye detection stage, adjusts search box size by self-defined standard self-adaptation, thereby realizes the method that eye movement is followed the tracks of.

In order to achieve the above object, the present invention has adopted following technical scheme:

A kind of eye movement tracking based on accurate location iris of the present invention is by accurate location iris, sets up the trace template based on iris region color and edge associating feature, finishes eye movement according to self-defining criterion and follows the tracks of.This method specifically may further comprise the steps:

1, background video or the real-time video that provides by the user accurately located iris region;

Accurately the location iris region further comprises following idiographic flow:

1), load the eye movement video record in advance, handle every frame video;

2), analyze a two field picture, load the human-face detector location human face region based on the AdaBoost algorithm;

3), load human eye detection device coarse positioning human eye rectangular area at the human face region that extracts;

4), in the human eye area of extracting by eliminating noise, iris is accurately located in operations such as smoothing processing.

By the elimination noise, operations such as smoothing processing, the concrete steps of accurately locating iris are as follows:

1., the human eye area of extracting is carried out binary conversion treatment;

2., introduce the susan operator and eliminate most of interference such as homochromy spot noises of class eyeball such as eyebrows;

3., the bianry image after scanning and the analyzing and processing obtains connected domain and divides, eliminate the less agglomerate of unconnected region area, thereby further eliminate interference of noise;

4., use medium filtering that the iris region that obtains is carried out smoothing processing, accurately locate iris.

2, finish the eye movement tracking according to accurately locating iris.

The eye movement tracing process comprises following flow process:

1), initialization tracking target center, set up iris color and edge the associating feature templates;

Concrete steps are as follows:

1., (x y), adjust to follow the tracks of edge circle, and template size, target's center position are shown in formula (1), (2), (3) in initialization iris center.

Formula (1)

White point, the black that white_num wherein, black_num are respectively the bianry image target area x that counts _Oright, x _Oleft, y _Oup, y _OdownBe target area and judgement of color criterion in the statistics null value image array

Adjusting template size and target's center's coordinate is:

$\left\{\begin{array}{c}{\mathrm{Model}}_h=y_{\mathrm{nup}}-y_{\mathrm{ndowm}}\\ {\mathrm{Model}}_w=x_{\mathrm{nright}}-y_{\mathrm{nleft}}\end{array}\right.$ Formula (2)

$\left\{\begin{array}{c}x=x+(x_{\mathrm{nright}}+x_{\mathrm{nleft}})/2\\ y=y+(y_{\mathrm{ndown}}+y_{\mathrm{nup}})/2\end{array}\right.$ Formula (3)

2., be different from the traditional C amShift algorithm that reckons without noise effect when setting up the color characteristic template, the present invention considers noise effect, set up image HSV space S component variation curve map, it is lower to obtain the noise saturation degree by analysis, set up the relation of s min and search window s, as shown in Equation 4.

$\left\{\begin{array}{c}\mathrm{\η}=s_k-1/s_k\\ s{\min }_{k+1}=s{\min }_k*\mathrm{\η}\end{array}\right.$ Formula (4)

3., statistics region of search grey level histogram distribution hisArray (being divided into the N level), gray probability distribution hisPro calculates region of search gray average μ, record statistics with histogram globalHis sets up iris color and edge and unites feature templates;

2), prediction next frame target's center position, finish eye movement by self-defining edge feature standard and follow the tracks of.

Concrete steps are as follows:

1., the maximum between-cluster variance segmentation threshold T in the calculating region of search _Max, as shown in Equation (5), and the record searching matrix

Edge calculation point mean distance ${\mathrm{\μ}}_d=\underset{x_{\mathrm{edge}}}{\mathrm{\Σ}}\underset{y_{\mathrm{edge}}}{\mathrm{\Σ}}\sqrt{{(x-x_{\mathrm{edge}})}^2+{(y-y_{\mathrm{edge}})}^2}/\mathrm{total},$ Be recorded in array array _DisIn;

$\left\{\begin{array}{c}{\mathrm{\&mu;}}_1=\underset{i=0}{\overset{T}{\mathrm{\&Sigma;}}}i*{\mathrm{hispro}}_i\\ d_1=\underset{i=0}{\overset{T}{\mathrm{\&Sigma;}}}{(i-{\mathrm{\&mu;}}_1)}^2*{\mathrm{hispro}}_i/{\mathrm{pros}}_1\\ {\mathrm{<figure-callout\; id="2"\; label="pros"\; filenames="HSA00000912745800021.png"\; state="\{\{state\}\}">pros</figure-callout>}}_2=1-{\mathrm{pros}}_1\\ {\mathrm{\&mu;}}_2=\mathrm{\&mu;}-{\mathrm{\&mu;}}_1\\ {\mathrm{<figure-callout\; id="2"\; label="d"\; filenames="HSA00000912745800021.png"\; state="\{\{state\}\}">d</figure-callout>}}_2=\underset{i=T+1}{\overset{N}{\mathrm{\&Sigma;}}}{(i-{\mathrm{\&mu;}}_2)}^2*{\mathrm{hispro}}_i/{\mathrm{pros}}_2\\ d_w={\mathrm{pros}}_1*d_1+{\mathrm{<figure-callout\; id="2"\; label="pros"\; filenames="HSA00000912745800021.png"\; state="\{\{state\}\}">pros</figure-callout>}}_2*d_2\\ d_b={\mathrm{pros}}_1*{\mathrm{<figure-callout\; id="2"\; label="pros"\; filenames="HSA00000912745800021.png"\; state="\{\{state\}\}">pros</figure-callout>}}_2*{({\mathrm{\&mu;}}_2-{\mathrm{\&mu;}}_1)}^2\\ \mathrm{\&theta;}=d_b/d_w\end{array}\right.$ Formula (5)

Wherein T ∈ (1, N-1), when θ obtains maximal value θ _MaxThe time be the segmentation threshold T that tries to achieve _Max, wherein hispro is the probability distribution in gray level image template zone.

2., predicting tracing target's center position (x, y), as shown in Equation (6), and record;

$\left\{\begin{array}{c}{x}_{\mathrm{off}}=\underset{x}{\mathrm{\&Sigma;}}\underset{y}{\mathrm{\&Sigma;}}x*s(x,y)/\mathrm{total}\\ y_{\mathrm{off}}=\underset{x}{\mathrm{\&Sigma;}}\underset{y}{\mathrm{\&Sigma;}}y*s(x,y)/\mathrm{total}\\ x=x+x_{\mathrm{off}}-{\mathrm{Model}}_w/2\\ y=y+y_{\mathrm{off}}-{\mathrm{Model}}_h/2\end{array}\right.$ Formula (6)

Model wherein _w, Model _hIt is wide and high to be respectively template.

3., judge whether iris deformation takes place, namely whether need to upgrade, criterion is: whether two frame statistics with histogram values change greater than 3 times of marginal point standard deviation status _EdgeAnd marginal point is to the mean distance status of central point _Dis, status _Edge, status _DisCalculate as shown in Equation (7), according to front LISTMAX frame target travel law forecasting target's center position, carry out as formula (8) when deformation takes place.

$\left\{\begin{array}{c}\mathrm{\&mu;}=\underset{i=0}{\overset{\mathrm{index}}{\mathrm{\&Sigma;}}}{\mathrm{array}}_{\mathrm{edge}}\left[i\right]/\mathrm{index}\\ {\mathrm{status}}_{\mathrm{edge}}=\mathrm{sqrt}(\underset{i=0}{\overset{\mathrm{index}}{\mathrm{\&Sigma;}}}{\left({\mathrm{array}}_{\mathrm{edge}}\right[i]-\mathrm{\&mu;})}^2/\mathrm{index})\\ {\mathrm{status}}_{\mathrm{dis}}={\mathrm{array}}_{\mathrm{dis}}[\mathrm{num}-1]+\left({\mathrm{array}}_{\mathrm{dis}}\right[\mathrm{num}-1]-{\mathrm{array}}_{\mathrm{dis}}[0\left]\right)/(\mathrm{num}-1)\end{array}\right.$ Formula (7)

Formula (8)

X wherein _Pd, y _PdBe respectively record and follow the tracks of the x of coordinates matrix, y component.

4. iteration is finished coupling to tracking target in the region of search, shown in formula (9), (10).

$\left\{\begin{array}{c}{p}_{\mathrm{te}}^{\left(b_{\mathrm{in}}\right)}=C_e\underset{x_i\&Element;w}{\mathrm{\&Sigma;}}k\left({\left|\right|x_i\left|\right|}^2\right)M_{x_i}\mathrm{\&delta;}({\mathrm{\&alpha;}}_{x_i}-b_{\mathrm{in}})\\ p_{\mathrm{ce}}^{\left(b_{\mathrm{in}}\right)}=C_c\underset{x_i\&Element;w}{\mathrm{\&Sigma;}}k\left({\left|\right|\frac{y_i-x_i}{h}\left|\right|}^2\right)M_{x_i}\mathrm{\&delta;}({\mathrm{\&alpha;}}_{x_i}-b_{\mathrm{in}})\end{array}\right.$ Formula (9)

$\left\{\begin{array}{c}{\mathrm{\&rho;}}_1=\mathrm{\&rho;}(p_{\mathrm{th}},p_{\mathrm{ch}})=\underset{l=1}{\overset{16}{\mathrm{\&Sigma;}}}\sqrt{p_{{\mathrm{th}}_l}{p}_{{\mathrm{ch}}_l}}\\ {\mathrm{\&rho;}}_2=\mathrm{\&rho;}(p_{\mathrm{te}},p_{\mathrm{ce}})=\underset{m=1}{\overset{16}{\mathrm{\&Sigma;}}}\sqrt{p_{{\mathrm{te}}_m}{p}_{{\mathrm{ce}}_m}}\end{array}\right.$ Formula (10)

Final similarity measure coefficient is ρ _Final=ε ρ ₁+ β ρ ₂D (y), ε+β=1 wherein, ε, β ∈ [0,1].When

Reaching a hour iteration finishes.Wherein, C is normaliztion constant, and δ is Kronecker delta function, and k (x) is the Epanechnikov kernel function.

4. 1. repeated execution of steps arrive step, finishes or the user is interrupted eye movement and followed the tracks of up to one section Video processing.

Beneficial effect of the present invention is:

The given eye movement tracking based on accurate location iris of the present invention, the eye movement tracking that belongs to non-intrusion type adopts in the human eye detection stage and accurately locatees iris, sets up the target following template, adjust search box size by self-defined standard self-adaptation, thereby realize the eye movement tracking.Method of the present invention is compared with additive method, and tracking velocity is faster, and tracking accuracy is higher, uses more conveniently, applied widely, satisfies real-time, accuracy, robustness requirement.

Description of drawings

Fig. 1 is the eye movement trace flow figure of the described non-intrusion type of background technology of the present invention

Fig. 2 is image HSV space S component variation curve map of the present invention

Fig. 3 is the main processing flow chart of embodiment of the invention system;

Embodiment

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

The experiment of following the tracks of with the eye movement of the accurate location iris under the indoor conditions is example below, and the present invention is specifically described:

In this experiment, comparatively violent or situation such as camera shake occurs for illumination variation, though be under indoor conditions, also be the exemplary that explanation is followed the tracks of based on the eye movement of accurate location iris, its concrete treatment scheme as shown in Figure 3, concrete implementation step is as follows:

Step 1: input video

The video that Shu Ru video can be for ready for also can be handled every frame video for the real-time video of camera transmission herein, changes step 2;

Step 2: add manned face cascade classifier

People's face cascade classifier herein is the good human-face detector based on the AdaBoost algorithm of precondition, analyzes a two field picture, and the human-face detector that loads based on the AdaBoost algorithm detects people's face, changes step 3;

Step 3: detect whether there is people's face

When not having people's face, obtain input one two field picture again; Otherwise change step 4;

Step 4: detect people's face, set the emerging zone of sense

In image, detect people's face region and mark, set the emerging regional ROI of sense ₁, change step 5;

Step 5: load the human eye cascade classifier

Surely feel emerging regional ROI1 at the people's face that gets access to and load the human eye detection device;

Step 6: detect human eye, set the emerging zone of sense

The emerging regional ROI of people's face sense that is obtaining ₁Middle position of human eye, coarse positioning human eye rectangular area and the mark of detecting set the emerging regional ROI of human eye sense ₂, change step 7;

Step 7: binary image and susan operator are handled

Rectangular area, gray processing human eye place, and simultaneously gray level image is carried out susan operator and binary image processing, and with the image phase or the operation that obtain, eliminate the interference of noise like points such as eyebrow, change step 8;

Step 8: connected domain analysis, smoothing denoising

Bianry image after handling is obtained connected domain divide, eliminate the less agglomerate of unconnected region area, use medium filtering that the iris region that obtains is carried out smoothing processing at last;

Step 9: accurately locate iris

The iris region that obtains is accurately located, indicate iris region and preservation;

Step 10: set up trace template

According to the iris region initialization tracking target center that obtains and adjust template size, follow the tracks of edge circle, set up color and edge associating feature templates thus.Concrete steps are as follows:

1., (x y), adjust to follow the tracks of edge circle, and template size, target's center position are shown in formula (1), (2), (3) in initialization iris center.

Formula (1)

White point, the black that white_num wherein, black_num are respectively the bianry image target area x that counts _Oright, x _Oleft, y _Oup, y _OdownBe target area and judgement of color criterion in the statistics null value image array

Adjusting template size and target's center's coordinate is:

$\left\{\begin{array}{c}{\mathrm{Model}}_h=y_{\mathrm{nup}}-y_{\mathrm{ndowm}}\\ {\mathrm{Model}}_w=x_{\mathrm{nright}}-y_{\mathrm{nleft}}\end{array}\right.$ Formula (2)

$\left\{\begin{array}{c}x=x+(x_{\mathrm{nright}}+x_{\mathrm{nleft}})/2\\ y=y+(y_{\mathrm{ndown}}+y_{\mathrm{nup}})/2\end{array}\right.$ Formula (3)

2., be different from the traditional C amShift algorithm that reckons without noise effect when setting up the color characteristic template, the present invention considers noise effect, set up image HSV space S component variation curve map, as shown in Figure 2, because The noise, saturation degree is lower, and noise is covered as between the 510-520 frame, set up the relation of s min and search window s, as shown in Equation 4.

$\left\{\begin{array}{c}\mathrm{\&eta;}=s_k-1/s_k\\ s{\min }_{k+1}=s{\min }_k*\mathrm{\&eta;}\end{array}\right.$ Formula (4)

3., statistics region of search grey level histogram distribution hisArray (being divided into the N level), gray probability distribution hisPro calculates region of search gray average μ, record statistics with histogram globalHis sets up iris color and edge and unites feature templates;

Step 11: prediction next frame tracking target center

Calculate the maximum between-cluster variance segmentation threshold T in the region of search _Max, as shown in Equation (5), and the record searching matrix Edge calculation point mean distance ${\mathrm{\&mu;}}_d=\underset{x_{\mathrm{edge}}}{\mathrm{\&Sigma;}}\underset{y_{\mathrm{edge}}}{\mathrm{\&Sigma;}}\sqrt{{(x-x_{\mathrm{edge}})}^2+{(y-y_{\mathrm{edge}})}^2}/\mathrm{total},$ Be recorded in array array _DisIn;

$\left\{\begin{array}{c}{\mathrm{\&mu;}}_1=\underset{i=0}{\overset{T}{\mathrm{\&Sigma;}}}i*{\mathrm{hispro}}_i\\ d_1=\underset{i=0}{\overset{T}{\mathrm{\&Sigma;}}}{(i-{\mathrm{\&mu;}}_1)}^2*{\mathrm{hispro}}_i/{\mathrm{pros}}_1\\ {\mathrm{<figure-callout\; id="2"\; label="pros"\; filenames="HSA00000912745800021.png"\; state="\{\{state\}\}">pros</figure-callout>}}_2=1-{\mathrm{pros}}_1\\ {\mathrm{\&mu;}}_2=\mathrm{\&mu;}-{\mathrm{\&mu;}}_1\\ {\mathrm{<figure-callout\; id="2"\; label="d"\; filenames="HSA00000912745800021.png"\; state="\{\{state\}\}">d</figure-callout>}}_2=\underset{i=T+1}{\overset{N}{\mathrm{\&Sigma;}}}{(i-{\mathrm{\&mu;}}_2)}^2*{\mathrm{hispro}}_i/{\mathrm{pros}}_2\\ d_w={\mathrm{pros}}_1*d_1+{\mathrm{<figure-callout\; id="2"\; label="pros"\; filenames="HSA00000912745800021.png"\; state="\{\{state\}\}">pros</figure-callout>}}_2*d_2\\ d_b={\mathrm{pros}}_1*{\mathrm{<figure-callout\; id="2"\; label="pros"\; filenames="HSA00000912745800021.png"\; state="\{\{state\}\}">pros</figure-callout>}}_2*{({\mathrm{\&mu;}}_2-{\mathrm{\&mu;}}_1)}^2\\ \mathrm{\&theta;}=d_b/d_w\end{array}\right.$ Formula (5)

Wherein T ∈ (1, N-1), when θ obtains maximal value θ _MaxThe time be the segmentation threshold T that tries to achieve _Max, wherein hispro is the probability distribution in gray level image template zone.

According to present frame iris characteristics of motion prediction next frame target's center position (x, y), as shown in Equation (6), and record;

$\left\{\begin{array}{c}{x}_{\mathrm{off}}=\underset{x}{\mathrm{\&Sigma;}}\underset{y}{\mathrm{\&Sigma;}}x*s(x,y)/\mathrm{total}\\ y_{\mathrm{off}}=\underset{x}{\mathrm{\&Sigma;}}\underset{y}{\mathrm{\&Sigma;}}y*s(x,y)/\mathrm{total}\\ x=x+x_{\mathrm{off}}-{\mathrm{Model}}_w/2\\ y=y+y_{\mathrm{off}}-{\mathrm{Model}}_h/2\end{array}\right.$ Formula (6)

Model wherein _w, Model _hIt is wide and high to be respectively template.

Step 12: whether deformation takes place

Consider to have situation such as close one's eyes, judge according to self-defining target deformation standard whether iris changes, namely whether need renewal, criterion is: whether two frame statistics with histogram values change greater than 3 times of marginal point standard deviation status _EdgeAnd marginal point is to the mean distance status of central point _Dis, status _Edge, status _DisCalculate as shown in Equation (7):

$\left\{\begin{array}{c}\mathrm{\&mu;}=\underset{i=0}{\overset{\mathrm{index}}{\mathrm{\&Sigma;}}}{\mathrm{array}}_{\mathrm{edge}}\left[i\right]/\mathrm{index}\\ {\mathrm{status}}_{\mathrm{edge}}=\mathrm{sqrt}(\underset{i=0}{\overset{\mathrm{index}}{\mathrm{\&Sigma;}}}{\left({\mathrm{array}}_{\mathrm{edge}}\right[i]-\mathrm{\&mu;})}^2/\mathrm{index})\\ {\mathrm{status}}_{\mathrm{dis}}={\mathrm{array}}_{\mathrm{dis}}[\mathrm{num}-1]+\left({\mathrm{array}}_{\mathrm{dis}}\right[\mathrm{num}-1]-{\mathrm{array}}_{\mathrm{dis}}[0\left]\right)/(\mathrm{num}-1)\end{array}\right.$ Formula (7)

If change, then change step 13, otherwise change step 14;

Step 13: upgrade the target of prediction center

When deformation takes place, adjust the target of prediction center.According to LISTMAX frame target travel law forecasting center, front, as shown in Equation (8):

Formula (8)

X wherein _Pd, y _PdBe respectively record and follow the tracks of the x of coordinates matrix, y component.

Change step 14 after upgrading the target of prediction center;

Step 14: iteration is finished the target coupling in the region of search

Specific algorithm such as formula (9), (10):

$\left\{\begin{array}{c}{p}_{\mathrm{te}}^{\left(b_{\mathrm{in}}\right)}=C_e\underset{x_i\&Element;w}{\mathrm{\&Sigma;}}k\left({\left|\right|x_i\left|\right|}^2\right)M_{x_i}\mathrm{\&delta;}({\mathrm{\&alpha;}}_{x_i}-b_{\mathrm{in}})\\ p_{\mathrm{ce}}^{\left(b_{\mathrm{in}}\right)}=C_c\underset{x_i\&Element;w}{\mathrm{\&Sigma;}}k\left({\left|\right|\frac{y_i-x_i}{h}\left|\right|}^2\right)M_{x_i}\mathrm{\&delta;}({\mathrm{\&alpha;}}_{x_i}-b_{\mathrm{in}})\end{array}\right.$ Formula (9)

$\left\{\begin{array}{c}{\mathrm{\&rho;}}_1=\mathrm{\&rho;}(p_{\mathrm{th}},p_{\mathrm{ch}})=\underset{l=1}{\overset{16}{\mathrm{\&Sigma;}}}\sqrt{p_{{\mathrm{th}}_l}{p}_{{\mathrm{ch}}_l}}\\ {\mathrm{\&rho;}}_2=\mathrm{\&rho;}(p_{\mathrm{te}},p_{\mathrm{ce}})=\underset{m=1}{\overset{16}{\mathrm{\&Sigma;}}}\sqrt{p_{{\mathrm{te}}_m}{p}_{{\mathrm{ce}}_m}}\end{array}\right.$ Formula (10)

Final similarity measure coefficient is ρ _Final=ε ρ ₁+ β ρ ₂D (y), ε+β=1 wherein, ε, β ∈ [0,1].When

Reaching a hour iteration finishes.Wherein, C is normaliztion constant, and δ is Kronecker delta function, and k (x) is the Epanechnikov kernel function.

Step 15: whether video finishes dealing with, and the then system finishing of finishing dealing with continues the eye movement tracking otherwise change step 11.

Claims (7)

1. the eye movement tracking based on accurate location iris is characterized in that: by video, accurately locate iris, set up the trace template based on iris region color and edge associating feature, finish eye movement according to self-defining criterion and follow the tracks of.

2. described a kind of eye movement tracking based on accurate location iris according to claim 1 is characterized in that: comprise following two big steps:

(1), by background video or real-time video that the user provides, accurately locate iris region;

(2), finish the eye movement tracking according to accurately locating iris.

3. described a kind of eye movement tracking based on accurate location iris according to claim 2, it is characterized in that: described step (1)---by background video or the real-time video that the user provides, accurately locate iris region and further comprise following idiographic flow:

1), load the eye movement video record in advance, handle every frame video;

2), analyze a two field picture, load the human-face detector location human face region based on the AdaBoost algorithm;

3), load human eye detection device coarse positioning human eye rectangular area at the human face region that extracts;

4), in the human eye area of extracting by eliminating noise, iris is accurately located in operations such as smoothing processing.

4. described a kind of eye movement tracking based on accurate location iris according to claim 3, it is characterized in that: described idiographic flow 4)---pass through to eliminate noise in the human eye area of extracting, operations such as smoothing processing, accurately locate iris and comprise following concrete steps:

1., the human eye area of extracting is carried out binary conversion treatment;

2., introduce the susan operator and eliminate most of interference such as homochromy spot noises of class eyeball such as eyebrows;

3., the bianry image after scanning and the analyzing and processing obtains connected domain and divides, eliminate the less agglomerate of unconnected region area, thereby further eliminate interference of noise;

4., use medium filtering that the iris region that obtains is carried out smoothing processing, accurately locate iris.

5. described a kind of eye movement tracking based on accurate location iris according to claim 2 is characterized in that: described step (2)---finishes eye movement according to accurate location iris and follows the tracks of and further comprise following flow process:

1), initialization tracking target center, set up iris color and edge the associating feature templates;

2), prediction next frame target's center position, finish eye movement by self-defining edge feature standard and follow the tracks of.

6. described a kind of eye movement tracking based on accurate location iris according to claim 5 is characterized in that: described flow process 1)---initialization tracking target center, and set up iris color and edge associating feature templates and comprise following concrete steps:

1., (x y), adjust to follow the tracks of edge circle, and template size, target's center position are shown in formula (1), (2), (3) in initialization iris center.

Formula (1)

White point, the black that white_num wherein, black_num are respectively the bianry image target area x that counts _Oright, x _Oleft, y _Oup, y _OdownBe target area and judgement of color criterion in the statistics null value image array

Adjusting template size and target's center's coordinate is:

Formula (2)

Formula (3)

2., be different from the traditional C amShift algorithm that reckons without noise effect when setting up the color characteristic template, the present invention considers noise effect, set up image HSV space S component variation curve map, it is lower to obtain the noise saturation degree by analysis, set up the relation of s min and search window s, as shown in Equation 4.

Formula (4)

3., statistics region of search grey level histogram distribution hisArray (being divided into the N level), gray probability distribution hisPro calculates region of search gray average μ, record statistics with histogram globalHis sets up iris color and edge and unites feature templates.

7. described a kind of eye movement tracking based on accurate location iris according to claim 5, it is characterized in that: described flow process 2)---prediction next frame target's center position, finish the eye movement tracking by self-defining edge feature standard and comprise following concrete steps:

1., the maximum between-cluster variance segmentation threshold T in the calculating region of search _Max, as shown in Equation (5), and the record searching matrix

Edge calculation point mean distance

Be recorded in array array _DisIn;

Formula (5)

Wherein T ∈ (1, N-1), when θ obtains maximal value θ _MaxThe time be the segmentation threshold T that tries to achieve _Max, wherein hispro is the probability distribution in gray level image template zone.

2., predicting tracing target's center position (x, y), as shown in Equation (6), and record;

Formula (6)

Model wherein _w, Model _hIt is wide and high to be respectively template.

3., judge whether iris deformation takes place, namely whether need to upgrade, criterion is: whether two frame statistics with histogram values change greater than 3 times of marginal point standard deviation status _EdgeAnd marginal point is to the mean distance status of central point _Dis, status _Edge, status _DisCalculate as shown in Equation (7), according to front LISTMAX frame target travel law forecasting target's center position, carry out as formula (8) when deformation takes place.

Formula (7)

Formula (8)

X wherein _Pd, y _PdBe respectively record and follow the tracks of the x of coordinates matrix, y component.

4. iteration is finished coupling to tracking target in the region of search, shown in formula (9), (10).

Formula (9)

Formula (10)

Final similarity measure coefficient is ρ _Final=ε ρ ₁+ β ρ ₂D (y), ε+β=1 wherein, ε, β ∈ [0,1].When

Reaching a hour iteration finishes.Wherein, C is normaliztion constant, and δ is Kronecker delta function, and k (x) is the Epanechnikov kernel function.

4. 1. repeated execution of steps arrive step, finishes or the user is interrupted eye movement and followed the tracks of up to one section Video processing.

Images