CN103892829A - Eye movement signal identification system and method based on common spatial pattern - Google Patents

Abstract

The invention discloses an eye movement signal identification system and method based on a common spatial pattern. The eye movement signal identification system comprises an eye movement signal preprocessing module, a spatial filter training module and an eye movement signal identification module. The eye movement signal identification method comprises the steps of collecting eye movement data based on an electro-oculogram and preprocessing the eye movement data; dividing all the preprocessed data into training data and testing data; adopting a CSP algorithm to conduct spatial filtering on the training data, and using the result obtained after filtering as feature parameters which are input into an SVM classifier for SVM model training; using the CSP algorithm to conduct feature extraction on the testing data, feeding the result obtained after the feature extraction into the trained SVM classifier for identification, and finally obtaining the identification result of eye movement. The eye movement signal identification system and method have the advantages that the accuracy of eye movement signal identification is higher, the eye movement signal spreading and classification capability is higher, and the application potential is large.

Description

A kind of eye movement signal recognition system and recognition methods thereof based on common space pattern

Technical field

The present invention relates to a kind of eye movement signal recognition system and recognition methods thereof based on common space pattern (Common Spatial Pattern, CSP).

Background technology

Based on bioelectric man-machine interaction (Human-Computer Interaction, HCI) technology is supplemented as the one of conventional man-machine interaction method, under some special application scenarios, such as: disabled colony and external environment mutual, the monitoring of clinical patient, communication under special environment, the aspects such as driver's fatigue detecting all have stronger actual application value.Wherein, due to electro-ocular signal, to have amplitude strong, is easy to the features such as detection, therefore, based on EOG(electro-oculogram, electro-oculogram) human-computer interaction technology progressively become people's study hotspot.

So-called electro-ocular signal refers to the potential difference existing between electromotive force that people produces due to the motion of eyes and human body skin surface electrode electromotive force.Further medical research shows, this electromotive force relation is to be caused by the electric potential difference between the cornea of eyes and retina (as shown in Fig. 1 left side).This electromotive force is initiated by retinal pigment epithelium and photoreceptor cell, its positive pole is positioned at photoreceptor end, negative pole is positioned at retinal pigment epithelium end, the electric current producing has flowed to cornea end from retina end, thereby form an amplitude and be about 0.4mV～10mV, cornea for anodal, retina is the electromotive force of negative pole, and we claim that this electromotive force is electro-ocular signal.In the time of people's ocular movement, can there is along with the motion of eyeball continuous variation in the amplitude of electro-ocular signal, we are plotted to the electromotive force of this variation the curve that can form on time shaft, and this curve is just referred to as electro-oculogram, and Fig. 2 has shown one section of EOG collection of illustrative plates that comprises eye movement signal.

In the detection of eye movement signal, signal of blinking is relatively short owing to having stronger periodicity and persistent period, therefore the detection of signal of blinking is also easier to.And for eyeball pan signal, between different experimenters or in same subject different time sections, all can exist larger difference at aspects such as signal amplitude, persistent period length and pan angles, be unfavorable for the identification of eye movement signal.In order to address this problem, the method of dynamic time warping was once used to carry out the identification of eye movement signal, but due to the higher similarity of existence between upwards sweeping and sweep left, sweep downwards and sweeping to the right, therefore, recognition result is not desirable especially, is difficult to the stage that reaches practical.

Summary of the invention

The present invention is the weak point existing in above-mentioned prior art for avoiding, and the eye movement signal recognition method based on common space pattern that a kind of discriminator accuracy is high, extended capability is strong, application potential is large is provided.

The present invention be technical solution problem by the following technical solutions.

An eye movement signal recognition system based on common space pattern, its construction features is to comprise eye movement signal pre-processing module, spatial filter training module and eye movement signal identification module;

Described eye movement signal pre-processing module, is used for all training and test data carry out bandpass filtering and remove equal Value Operations, in order to reduce the interference of different noise signals to original multi-lead eye movement signal, thereby improves recognition correct rate;

Described spatial filter training module, builds the many classification eye movement signal space wave filter based on common space pattern for the method by Joint Diagonalization of Matrix;

Described eye movement signal identification module, the result that is used for original multi-lead eye movement signal to obtain after above-mentioned spatial filter filtering is as characteristic parameter, and use support vector machine (Support Vector Machine, SVM) to carry out the identification of eye movement signal.

A kind of eye movement signal recognition system based on common space pattern of the present invention and the construction features of recognition methods thereof are also:

The recognition methods of described a kind of eye movement signal recognition system based on common space pattern, is characterized in comprising the steps:

Step 1: gather the eye movement data based on electro-oculogram and eye movement data is carried out to pretreatment;

Step 2: pretreated all data are divided into training data and two parts of test data; Described training data is original multi-lead eye movement signal, adopts CSP algorithm to carry out airspace filter to original multi-lead eye movement signal, calculates the spatial filter W of left pan, right pan, upper pan and lower pan four classes _l, W _r, W _uand W _d;

Step 3: use the W in above-mentioned steps 2 _l, W _r, W _uand W _dfour spatial filters carry out airspace filter to training data, and are input in svm classifier device and train result after filtering as characteristic parameter;

Step 4: to test data, use equally the W in above-mentioned steps 2 _l, W _r, W _uand W _dfour spatial filters carry out airspace filter, feature extraction, are then sent in the svm classifier device having trained and identify, and finally obtain the recognition result of eye movement signal.

In described step 1, eye movement signal being carried out to preprocessing process comprises bandpass filtering step and goes average step.

The cut-off frequency that is used for the band filter that carries out bandpass filtering step is 0.3Hz～12.5Hz.

While adopting CSP algorithm to carry out airspace filter to original multi-lead eye movement signal in described step 2, spatial filter W _l, W _r, W _uand W _dcomputational process be:

Suppose X _l, X _r, X _uwith X _drepresent respectively the original multi-lead eye movement signal under left pan, right pan, upper pan and 4 kinds of tasks of lower pan, X _l, X _r, X _uwith X _dspace covariance matrix after normalization is

$\begin{array}{cc}{C}_l=\frac{X_l{X}_l^T}{\mathrm{trace}\left(X_l{X}_l^T\right)}& C_r=\frac{X_r{X}_r^T}{\mathrm{trace}\left(X_r{X}_r^T\right)}\\ C_u=\frac{X_u{X}_u^T}{\mathrm{trace}\left(X_u{X}_u^T\right)}& C_d=\frac{X_d{X}_d^T}{\mathrm{trace}\left(X_d{X}_d^T\right)}\end{array}---\left(1\right)$

The covariance matrix of the many groups training data under same eye movement mode is averaged, obtain average covariance matrix

, with

, calculate the average covariance matrix sum of all eye movement signals

then C is carried out to Eigenvalues Decomposition,

$C={\mathrm{<figure-callout\; id="0"\; label="U"\; filenames="HDA0000492678030000012.png,HDA0000492678030000051.png"\; state="\{\{state\}\}">U</figure-callout>}}_0\mathrm{\&Sigma;}{U}_0^T---\left(2\right)$

In formula (2), U ₀representation feature vector, representing matrix U ₀transposed matrix, the diagonal matrix that Σ is eigenvalue, prewhitening transformation matrix P can be expressed as

$P=\sqrt{{\mathrm{\&Sigma;}}^{-1}}{U}_0^T---\left(3\right)$

Utilize prewhitening transformation matrix P to the average covariance matrix of above-mentioned 4 classification

with

convert, can obtain

$\begin{array}{cc}{S}_l=P\stackrel{\_}{C_l}{P}^T& S_u=P\stackrel{\_}{C_u}{P}^T\\ S_r=P\stackrel{\_}{C_r}{P}^T& S_d=P\stackrel{\_}{C_d}{P}^T\end{array}---\left(4\right)$

In formula (4), S _l, S _r, S _uand S _dbe respectively average covariance matrix with

carry out the matrix after prewhitening conversion, to S _l, S _r, S _uand S _dcarry out after Joint Diagonalization of Matrix, can find an orthogonal matrix U and approximate diagonal battle array Σ _l, Σ _r, Σ _uwith Σ _d, and meet following relation

$\begin{array}{cc}{\mathrm{US}}_l{U}^T={\mathrm{\&Sigma;}}_l& {\mathrm{US}}_u{U}^T={\mathrm{\&Sigma;}}_u\\ {\mathrm{US}}_r{U}^T={\mathrm{\&Sigma;}}_r& {\mathrm{US}}_d{U}^T={\mathrm{\&Sigma;}}_d\end{array}---\left(5\right)$

Σ _l+Σ _r+Σ _u+Σ _d=I （6）

In formula (5) and formula (6), I is unit matrix; For many classification problems of eye movement signal, diagonal matrix Σ _l, Σ _r, Σ _uand Σ _dcan carry out eigenvalue selection according to following formula, suppose that classification number to be sorted is N, diagonal matrix is Σ,

score(Σ)=max(Σ,(1-Σ)/(1-Σ+(N-1) ²Σ)) （7）

Calculate respectively Σ _l, Σ _r, Σ _uand Σ _dthe highest eigenvalue of middle score, and take out eigenvalue corresponding row vector U in orthogonal matrix U that this score is the highest _l, U _r, U _uand U _d, then calculate corresponding spatial filter by following formula (8),

$\begin{array}{cccc}{W}_L=U_l^{T}P& W_R=U_r^{T}P& W_U=U_u^{T}P& W_D=U_d^{T}P---\left(8\right)\end{array}$

In formula (8), W _l, W _r, W _uand W _drepresent respectively the spatial filter of left pan, right pan, upper pan and lower pan four classes.

In step 3, in eye movement signal identifying, the many classification task correlated source signal (the relevant source signal of left and right, upper and lower pan four class eye movement task) obtaining after airspace filter can pass through following formula (9) and obtain

Z _l=W _lX Z _r=W _rX Z _u=W _uX Z _d=W _dX （9）

In formula (9), X represents original multi-lead eye movement signal, Z _l, Z _r, Z _u, Z _drepresent respectively the result (be many classification task correlated source signal) of left and right, upper and lower pan signal after airspace filter, and this result is sent into svm classifier device as the characteristic parameter of eye movement signal, in order to train SVM model or test data is identified.

Compared with the prior art, beneficial effect of the present invention is embodied in:

A kind of eye movement signal recognition system and recognition methods thereof based on common space pattern of the present invention, has the feature of the following aspects.

1, the present invention has higher recognition correct rate to eye movement signal.

The present invention has adopted common space pattern (CSP) algorithm, this algorithm builds the spatial filter relevant to certain particular task according to the label data having marked in advance, in extracting the component relevant to this particular task, can effectively suppress the interference of component and the noise irrelevant with eye movement signal, several experimenters are carried out to many experiments, and its result shows that Mean accurate rate of recognition of the present invention all reaches more than 98%.

2, the present invention has stronger eye movement signal extension classification capacity.

The present invention has adopted approximate matrix to combine diagonalizable method in the time realizing multiclass CSP classification, carries out approximately joint diagonalization, to obtain the orthogonal albefaction matrix of original each observation signal by multiclass covariance matrix.The method not only can be applied to the classification of the four class eye movement signals (pan left, pan to the right, upwards pan and pan downwards) of mentioning in the present invention, in the time that eye movement signal type increases, such as: the signal of blinking of different number of times, the eye movement signal of different pan angles etc., the method does not need to carry out too much change just can realize the classification to more eye movement signals, has stronger expansion classification capacity.

3, the present invention has huge application potential.

Eye movement signal is as a kind of principal mode of man-machine interaction's method, in our daily life, there is more and more important effect, especially some are had a bodily deformity and can not paleocinetic crowd, the man-machine interactive system based on eye electricity can effectively help them to promote quality of life; On the other hand, be inconvenient to make bimanual occasion for some, such as: under mine, in spacecraft, this system can effectively improve user work quality.In the man-machine interactive system based on eye electricity, the recognition accuracy of eye movement signal is particularly important, higher accuracy just more easily allows user accept, and main purpose of the present invention is exactly in order to realize the precise classification to eye movement signal, and the method can realize online function, therefore the present invention has huge application potential.

A kind of eye movement signal recognition system and recognition methods thereof based on common space pattern of the present invention, has the advantages such as higher to eye movement signal recognition correct rate, eye movement signal extension classification capacity is strong, application potential is large.

Accompanying drawing explanation

Fig. 1 is human eye internal anatomy.

Fig. 2 is eye movement signal generating principle figure in the present invention (electro-oculogram when pan of left and right).

Fig. 3 is algorithm basic flow sheet of the present invention.

Fig. 4 is distribution of electrodes figure in eye movement signal acquisition process of the present invention.

Fig. 5 is the relative position schematic diagram of observed object and experimenter in eye movement signal acquisition process of the present invention.

Fig. 6 is the moving single experiment normal form schematic diagram of the present invention.

Fig. 7 is the original electro-ocular signal oscillogram of passage of the present invention.

Fig. 8 is CSP spatial filter coefficient schematic diagram of the present invention.

Below pass through the specific embodiment, and the invention will be further described by reference to the accompanying drawings.

The specific embodiment

Referring to Fig. 1～8, a kind of eye movement signal recognition system based on common space pattern, it comprises eye movement signal pre-processing module, spatial filter training module and eye movement signal identification module;

Described eye movement signal pre-processing module, is used for all training and test data carry out bandpass filtering and remove equal Value Operations, in order to reduce the interference of different noise signals to original multi-lead eye movement signal, thereby improves recognition correct rate;

Described spatial filter training module, builds the many classification eye movement signal space wave filter based on common space pattern for the method by Joint Diagonalization of Matrix;

Described eye movement signal identification module, the result that is used for original multi-lead eye movement signal to obtain after above-mentioned spatial filter filtering is as characteristic parameter, and use support vector machine (Support Vector Machine, SVM) to carry out the identification of eye movement signal.

The recognition methods of described a kind of eye movement signal recognition system based on common space pattern, it comprises the steps:

Step 1: gather the eye movement data based on electro-oculogram and eye movement data is carried out to pretreatment;

Step 2: pretreated all data are divided into training data and two parts of test data; Described training data is original multi-lead eye movement signal, adopts CSP algorithm to carry out airspace filter to original multi-lead eye movement signal, calculates the spatial filter W of left pan, right pan, upper pan and lower pan four classes _l, W _r, W _uand W _d;

Step 3: use the W in above-mentioned steps 2 _l, W _r, W _uand W _dfour spatial filters carry out airspace filter to training data, and are input in svm classifier device and train result after filtering as characteristic parameter;

Step 4: to test data, use equally the W in above-mentioned steps 2 _l, W _r, W _uand W _dfour spatial filters carry out airspace filter, feature extraction, are then sent in the svm classifier device having trained and identify, and finally obtain the recognition result of eye movement signal.

In described step 1, eye movement signal being carried out to preprocessing process comprises bandpass filtering step and goes average step.

Described bandpass filtering step is with the interference that removes signals such as comprising baseline drift, myoelectricity EMG, electrocardio ECG, brain electricity EEG with the object of going average step.

The cut-off frequency that is used for the band filter that carries out bandpass filtering step is 0.3Hz～12.5Hz.

While adopting CSP algorithm to carry out airspace filter to original multi-lead eye movement signal in described step 2, spatial filter W _l, W _r, W _uand W _dcomputational process be:

Suppose X _l, X _r, X _uwith X _drepresent respectively the original multi-lead eye movement signal under left pan, right pan, upper pan and 4 kinds of tasks of lower pan, X _l, X _r, X _uwith X _dspace covariance matrix after normalization is

$\begin{array}{cc}{C}_l=\frac{X_l{X}_l^T}{\mathrm{trace}\left(X_l{X}_l^T\right)}& C_r=\frac{X_r{X}_r^T}{\mathrm{trace}\left(X_r{X}_r^T\right)}\\ C_u=\frac{X_u{X}_u^T}{\mathrm{trace}\left(X_u{X}_u^T\right)}& C_d=\frac{X_d{X}_d^T}{\mathrm{trace}\left(X_d{X}_d^T\right)}\end{array}---\left(1\right)$

The covariance matrix of the many groups training data under same eye movement mode is averaged, obtain average covariance matrix ,

with , calculate the average covariance matrix sum of all eye movement signals

then C is carried out to Eigenvalues Decomposition,

$C={\mathrm{<figure-callout\; id="0"\; label="U"\; filenames="HDA0000492678030000012.png,HDA0000492678030000051.png"\; state="\{\{state\}\}">U</figure-callout>}}_0\mathrm{\&Sigma;}{U}_0^T---\left(2\right)$

In formula (2), U ₀representation feature vector,

representing matrix U ₀transposed matrix, the diagonal matrix that Σ is eigenvalue, prewhitening transformation matrix P can be expressed as

$P=\sqrt{{\mathrm{\&Sigma;}}^{-1}}{U}_0^T---\left(3\right)$

Utilize prewhitening transformation matrix P to the average covariance matrix of above-mentioned 4 classification

with

convert, can obtain

$\begin{array}{cc}{S}_l=P\stackrel{\_}{C_l}{P}^T& S_u=P\stackrel{\_}{C_u}{P}^T\\ S_r=P\stackrel{\_}{C_r}{P}^T& S_d=P\stackrel{\_}{C_d}{P}^T\end{array}---\left(4\right)$

In formula (4), S _l, S _r, S _uand S _dbe respectively average covariance matrix with

carry out the matrix after prewhitening conversion, to S _l, S _r, S _uand S _dcarry out after Joint Diagonalization of Matrix, can find an orthogonal matrix U and approximate diagonal battle array Σ _l, Σ _r, Σ _uwith Σ _d, and meet following relation

$\begin{array}{cc}{\mathrm{US}}_l{U}^T={\mathrm{\&Sigma;}}_l& {\mathrm{US}}_u{U}^T={\mathrm{\&Sigma;}}_u\\ {\mathrm{US}}_r{U}^T={\mathrm{\&Sigma;}}_r& {\mathrm{US}}_d{U}^T={\mathrm{\&Sigma;}}_d\end{array}---\left(5\right)$

Σ _l+Σ _r+Σ _u+Σ _d=I （6）

In formula (5) and formula (6), I is unit matrix; For many classification problems of eye movement signal, diagonal matrix Σ _l, Σ _r, Σ _uand Σ _dcan carry out eigenvalue selection according to following formula, suppose that classification number to be sorted is N, diagonal matrix is Σ,

score(Σ)=max(Σ,(1-Σ)/(1-Σ+(N-1) ²Σ)) （7）

Calculate respectively Σ _l, Σ _r, Σ _uand Σ _dthe highest eigenvalue of middle score, and take out eigenvalue corresponding row vector U in orthogonal matrix U that this score is the highest _l, U _r, U _uand U _d, then calculate corresponding spatial filter by following formula (8),

$\begin{array}{cccc}{W}_L=U_l^{T}P& W_R=U_r^{T}P& W_U=U_u^{T}P& W_D=U_d^{T}P---\left(8\right)\end{array}$

In formula (8), W _l, W _r, W _uand W _drepresent respectively the spatial filter of left pan, right pan, upper pan and lower pan four classes.

In step 3, in eye movement signal identifying, the many classification task correlated source signal (the relevant source signal of left and right, upper and lower pan four class eye movement task) obtaining after airspace filter can pass through following formula (9) and obtain

Z _l=W _lX Z _r=W _rX Z _u=W _uX Z _d=W _dX （9）

In formula (9), X represents original multi-lead eye movement signal, Zl, Zr, Zu, Zd represent respectively the result (be many classification task correlated source signal) of left and right, upper and lower pan signal after airspace filter, and this result is sent into svm classifier device as the characteristic parameter of eye movement signal, in order to train SVM model or test data identified.

Fig. 1 is human eye internal anatomy.

Referring to Fig. 2, the generating principle of eye movement signal in the present embodiment is described, the electro-oculogram while being left and right pan.

Referring to Fig. 3, illustrate that in the present embodiment, method is mainly made up of eye movement signal pre-processing module, spatial filter training module and eye movement signal identification module three parts.Method adopts common space pattern (CSP) method to carry out after feature extraction, utilizes support vector machine (SVM) to classify, to realize the identification to eye movement signal.Method mainly comprises following step: 1) gather EOG data and carry out pretreatment; 2) pretreated all data are divided into training data and two parts of test data.For training data, adopt CSP algorithm to carry out airspace filter, calculate the spatial filter W of left pan, right pan, upper pan and lower pan four classes _l, W _r, W _uand W _d; 3) use above-mentioned four wave filter to carry out airspace filter to training data, and be input to result after filtering as characteristic parameter and in svm classifier device, carry out SVM model training; 4) use four spatial filters to carry out filtering, feature extraction to test data is same, be then sent in the svm classifier device having trained and identify, obtain final recognition correct rate, and it is analyzed, the performance of appraisal procedure.

Referring to Fig. 4, distribution of electrodes in eye movement signal acquisition process is described in the present embodiment.The collection of electro-ocular signal is used Ag/AgCl electrode.In order to obtain, experimenter is left and right, the eye movement information of upper and lower four direction and more spatial positional information, we have used 9 electrodes altogether, wherein, electrode V1 and electrode V2 are placed in 3cm and lower 3cm place on (or right side) eyeball of experimenter left side, in order to gather vertical eye movement signal; Electrode H1 and electrode H2 are placed in respectively experimenter's left eye left side 3cm and 3cm place, right eye right side, in order to gather horizontal eye movement signal; Electrode A 1 is placed in forehead position with electrode A 2, in order to strengthen spatial information; Reference electrode C1 and C2 are positioned over respectively the protruding place of breast, the left and right sides, and ground electrode G is positioned at center, the crown.

Referring to Fig. 5, the relative position of observed object and experimenter in eye movement signal acquisition process is described in the present embodiment.In specific implementation process, experimenter is sitting on an armchair, and, approximately 2 meters, its front is provided with respectively the observed object of upper and lower, left and right four direction and is all 1.5 meters apart from experimenter's optic centre point (O).

Referring to Fig. 6, the detailed process of single experiment normal form in the present embodiment is described.In the time that experiment starts, first on screen, there is " beginning " character, and be accompanied by the long sonic stimulation of 20ms, after 1 second kind, experimenter can see red arrow prompting (being respectively: to upward arrow, arrow, arrow and arrow to the right left downwards) on screen, it is 3 seconds that arrow continues time of occurrence on screen, within this time, requirement of experiment experimenter is seeing after arrow to arrow direction indication rotation eyeball, rotate back into central point seeing after observation station, experimenter can not blink in this course.Afterwards, have 2 seconds takes a break, experimenter can blink, loosen.

Referring to Fig. 7, the original waveform figure that uses the eye movement signal that said method collects in the present embodiment is described.This waveform has comprised experimenter and has upwards swept, sweeps, swept and sweep 4 actions downwards left to the right.

Referring to Fig. 8, the spatial filter coefficient schematic diagram calculating by CSP algorithm in the present embodiment has been described, its abscissa represents filter coefficient residing position in this wave filter, vertical coordinate represents the size of coefficient amplitude.Can find out, in the time that experimenter upwards sweeps, in all filter coefficients, the 6th value is maximum; When pan downwards, the 5th value is maximum; In like manner, while pan left, the 1st value is maximum; While pan to the right, the 2nd value is maximum.Thus, we presumably (V1), under (V2), left (H2), right (H1) 4 electrodes correspond respectively to wave filter the 6th, the coefficient of 5,1,2 positions.The above results has illustrated that this spatial filter can make the 4 class eye movement signals that we will classify reach difference maximization.

Can find out, in the time that experimenter sweeps left, in all filter coefficients, the 1st value is in all middle maximums of leading; While pan to the right, the 2nd value is maximum, and in like manner, while upwards pan, the 6th value is maximum; When pan downwards, the 5th value is maximum, thus, we presumably (V1), under (V2), left (H2), right (H1) 4 electrodes correspond respectively to wave filter the 6th, the coefficient of 5,1,2 positions.Therefore, this spatial filter can make the 4 class eye movement signals that we will classify reach difference maximization.

Claims (6)

1. the eye movement signal recognition system based on common space pattern, is characterized in that, comprises eye movement signal pre-processing module, spatial filter training module and eye movement signal identification module;

Described eye movement signal pre-processing module, is used for all training and test data carry out bandpass filtering and remove equal Value Operations, in order to reduce the interference of different noise signals to original multi-lead eye movement signal, thereby improves recognition correct rate;

Described spatial filter training module, builds the many classification eye movement signal space wave filter based on common space pattern for the method by Joint Diagonalization of Matrix;

Described eye movement signal identification module, the result that is used for original multi-lead eye movement signal to obtain after above-mentioned spatial filter filtering is as characteristic parameter, and use support vector machine (Support Vector Machine, SVM) to carry out the identification of eye movement signal.

2. the recognition methods of a kind of eye movement signal recognition system based on common space pattern according to claim 1, is characterized in that, comprises the steps:

Step 1: gather the eye movement data based on electro-oculogram and eye movement data is carried out to pretreatment;

Step 2: pretreated all data are divided into training data and two parts of test data; Described training data is original multi-lead eye movement signal, adopts CSP algorithm to carry out airspace filter to original multi-lead eye movement signal, calculates the spatial filter W of left pan, right pan, upper pan and lower pan four classes _l, W _r, W _uand W _d;

Step 3: use the W in above-mentioned steps 2 _l, W _r, W _uand W _dfour spatial filters carry out airspace filter to training data, and are input in svm classifier device and train result after filtering as characteristic parameter.

Step 4: to test data, use equally the W in above-mentioned steps 2 _l, W _r, W _uand W _dfour spatial filters carry out airspace filter, feature extraction, are then sent in the svm classifier device having trained and identify, and finally obtain the recognition result of eye movement signal.

3. the recognition methods of a kind of eye movement signal recognition system based on common space pattern according to claim 1, is characterized in that, in described step 1, eye movement signal is carried out to preprocessing process and comprises bandpass filtering step and go average step.

4. the recognition methods of a kind of eye movement signal recognition system based on common space pattern according to claim 3, is characterized in that, is 0.3Hz～12.5Hz for the cut-off frequency of the band filter that carries out bandpass filtering step.

5. the recognition methods of a kind of eye movement signal recognition system based on common space pattern according to claim 2, is characterized in that, while adopting CSP algorithm to carry out airspace filter to original multi-lead eye movement signal in described step 2, and spatial filter W _l, W _r, W _uand W _dcomputational process be:

Suppose X _l, X _r, X _uwith X _drepresent respectively the original multi-lead eye movement signal under left pan, right pan, upper pan and 4 kinds of tasks of lower pan, X _l, X _r, X _uwith X _dspace covariance matrix after normalization is

$\begin{array}{cc}{C}_l=\frac{X_l{X}_l^T}{\mathrm{trace}\left(X_l{X}_l^T\right)}& C_r=\frac{X_r{X}_r^T}{\mathrm{trace}\left(X_r{X}_r^T\right)}\\ C_u=\frac{X_u{X}_u^T}{\mathrm{trace}\left(X_u{X}_u^T\right)}& C_d=\frac{X_d{X}_d^T}{\mathrm{trace}\left(X_d{X}_d^T\right)}\end{array}---\left(1\right)$

The covariance matrix of the many groups training data under same eye movement mode is averaged, obtain average covariance matrix

,

with

, calculate the average covariance matrix sum of all eye movement signals

then C is carried out to Eigenvalues Decomposition,

$C=U_0\mathrm{\&Sigma;}{U}_0^T---\left(2\right)$

In formula (2), U ₀representation feature vector,

representing matrix U ₀transposed matrix, the diagonal matrix that Σ is eigenvalue, prewhitening transformation matrix P can be expressed as

$P=\sqrt{{\mathrm{\&Sigma;}}^{-1}}{U}_0^T---\left(3\right)$

Utilize prewhitening transformation matrix P to the average covariance matrix of above-mentioned 4 classification

with

convert, can obtain

$\begin{array}{cc}{S}_l=P\stackrel{\_}{C_l}{P}^T& S_u=P\stackrel{\_}{C_u}{P}^T\\ S_r=P\stackrel{\_}{C_r}{P}^T& S_d=P\stackrel{\_}{C_d}{P}^T\end{array}---\left(4\right)$

In formula (4), S _l, S _r, S _uand S _dbe respectively average covariance matrix

with

carry out the matrix after prewhitening conversion, to S _l, S _r, S _uand S _dcarry out after Joint Diagonalization of Matrix, can find an orthogonal matrix U and approximate diagonal battle array Σ _l, Σ _r, Σ _uwith Σ _d, and meet following relation:

$\begin{array}{cc}{\mathrm{US}}_l{U}^T={\mathrm{\&Sigma;}}_l& {\mathrm{US}}_u{U}^T={\mathrm{\&Sigma;}}_u\\ {\mathrm{US}}_r{U}^T={\mathrm{\&Sigma;}}_r& {\mathrm{US}}_d{U}^T={\mathrm{\&Sigma;}}_d\end{array}---\left(5\right)$

Σ _l+Σ _r+Σ _u+Σ _d=I （6）

In formula (5) and formula (6), I is unit matrix; For many classification problems of eye movement signal, diagonal matrix Σ _l, Σ _r, Σ _uand Σ _dcan carry out eigenvalue selection according to following formula, suppose that classification number to be sorted is N, diagonal matrix is Σ,

score(Σ)=max(Σ,(1-Σ)/(1-Σ+(N-1) ²Σ)) （7）

Calculate respectively Σ _l, Σ _r, Σ _uand Σ _dthe highest eigenvalue of middle score, and take out eigenvalue corresponding row vector U in orthogonal matrix U that this score is the highest _l, U _r, U _uand U _d, then calculate corresponding spatial filter by following formula (8),

$\begin{array}{cccc}{W}_L=U_l^{T}P& W_R=U_r^{T}P& W_U=U_u^{T}P& W_D=U_d^{T}P---\left(8\right)\end{array}$

In formula (8), W _l, W _r, W _uand W _drepresent respectively the spatial filter of left pan, right pan, upper pan and lower pan four classes.

6. the recognition methods of a kind of eye movement signal recognition system based on common space pattern according to claim 5, it is characterized in that, in step 3, in eye movement signal identifying, the many classification task correlated source signal (the relevant source signal of left and right, upper and lower pan four class eye movement task) obtaining after airspace filter can pass through following formula (9) and obtain

Z _l=W _lX Z _r=W _rX Z _u=W _uX Z _d=W _dX （9）

In formula (9), X represents original multi-lead eye movement signal, Z _l, Z _r, Z _u, Z _drepresent respectively the result (be many classification task correlated source signal) of left and right, upper and lower pan signal after airspace filter, and this result is sent into svm classifier device as the characteristic parameter of eye movement signal, in order to train SVM model or test data is identified.

Images