CN105640500A - Scanning signal feature extraction method based on independent component analysis and recognition method - Google Patents

Abstract

The invention discloses a scanning signal feature extraction method based on independent component analysis. The method comprises the steps that six-lead scanning eye electric signals are collected and subjected to band-pass filtering treatment, an airspace filter set corresponding to different scanning task backgrounds is built through an ICA method for filtered data, linear protection is carried out, and airspace feature parameters of scanning signals are obtained. The invention further discloses a recognition method of the scanning signal feature extraction method based on independent component analysis. An ICA airspace filter set is built for each experiment sample in an eye movement database, feature parameters are extracted, cross validation is carried out through a support vector machine, and the optimal ICA filter set and SVM model parameters are determined; the optimal ICA airspace filter set is used for filtering, and then the result is fed into an SVM classifier to be recognized. The scanning signal feature extraction method based on independent component analysis and the recognition method have the advantages of being higher in recognition accuracy rate, higher in expansibility, good in application prospect and the like.

Description

Based on pan signal characteristic extracting methods and the recognition methods of independent component analysis

Technical field

The present invention relates to a kind of pan signal characteristic extracting methods based on independent component analysis and recognition methods.

Background technology

The eye movement mode that people causes when carrying out specific activities can disclose its behavior state to a great extent, as: reading, writing, rest etc., and this kind of eye movement mode can by obtaining the tracking of eye movement situation, therefore new research focus has been become based on the design and implimentation of Human bodys' response (HumanActivityRecognition, the HAR) algorithm of the dynamic information of eye.

Present stage, the dynamic signal of the means record eye of video is adopted to be widely used, but the eye-tracking system based on video, especially wearable eye-tracking system price is expensive, volume is bigger and heavy, simultaneously also not fully up to expectations in the real-time analysis of system power dissipation and result. Electrooculogram (Electro-oculogram, EOG) as the dynamic signal measurement technique of eye of a kind of low cost, conventional video of comparing means, not only measure more accurate, simultaneously its collection equipment also have weight light, be convenient to long-time record, more easily realize the advantages such as wearable design. Therefore, it may also be useful to EOG substitutes traditional video method to carry out HAR system design and have important researching value.

EOG-HAR system refers to using EOG signal as object being observed, by it being analyzed and identify, obtains the information such as the action type of object being observed, behavior pattern. In systematic realizing program, the examination and analysb of pan EOG signal is a most key step, and for this reason, investigators make big quantifier elimination. Wherein, Clement proposes to utilize the visible angle of original EOG signal to carry out end-point detection and the identification of the dynamic signal of eye; When the people such as Aungsakun and Soltani utilize Rotation of eyeball, corresponding EOG signal changes the characteristic parameter of the signal of feature extraction pan faster; In addition, Vidal and Bulling also refer to and use the statistics sweeping signal and temporal signatures to carry out the thinking identified.Although above-mentioned detection method achieves certain success, but what this kind of method was mainly paid close attention to is the analysis of the single dynamic signal of the eye that leads, the related information that its analysis process only considers the change of single lead signals and ignores between leading, it is difficult to ensure the recognition correct rate of pan signal; In addition, when the type of the dynamic signal of eye increases, aforesaid method is difficult to effective identification, the situation of None-identified even occurs.

Summary of the invention

Technical problem to be solved by this invention is to provide a kind of pan signal characteristic extracting methods based on independent component analysis and recognition methods, and recognition correct rate is higher, extendability is stronger, application prospect is good.

In order to solve the problems of the technologies described above, the present invention adopts following technical scheme: a kind of pan signal characteristic extracting methods based on independent component analysis, it is characterised in that, comprise the steps:

Step 1, multi-lead pan signals collecting and pre-treatment: lead electro-ocular signal with the 6 of data label when using 8 bioelectrodes to obtain the pan of experimenter upper and lower, left and right; And use bandpass filter to carry out filtering original multi-lead electro-ocular signal, to remove noise jamming;

Step 2, ICA spatial filter design: use single experiment data y_i(i=1 ..., N) and carry out ICA analysis, and according to the mapping pattern of independent component at acquisition electrode, automatically select the relevant isolated component of pan and corresponding ICA wave filter, set up the ICA spatial filter group { D under corresponding different pan task context_li, D_ri, D_ui, D_di(i=1 ..., N);

Step 3, feature generate: refer to the ICA spatial filter group { D using step (2) to generate_li, D_ri, D_ui, D_diThe pan signal that leads to original 6 linearly projects, with the pan signal spatial feature parameter generated under corresponding task context.

As the further optimization of such scheme, the installation position of described multi-lead pan electrode signal acquisition is set to:

(A1), using 2 electrode collections vertically to sweep signal, an electrode is placed in 2.3-2.7cm place above experimenter's left eye eyeball center, and another electrode is placed in 2.3-2.7cm place below experimenter's left eye eyeball center; Or an electrode is placed in 2.3-2.7cm place above experimenter's right eye eyeball center, and another electrode is placed in 2.3-2.7cm place below experimenter's right eye eyeball center;

(A2), using 2 electrodes to gather horizontal saccade signal, an electrode is placed in the left eye eyeball horizontal center point left side 3.3-3.7cm place of experimenter; An electrode is placed in the 3.3-3.7cm place, right eye eyeball horizontal center point right of experimenter;

(A3), using 2 supporting electrodes, a supporting electrode is placed in 4.8-5.2cm place directly on the right side of the left eye eyeball of experimenter; A supporting electrode is placed in 4.8-5.2cm place directly on the left of the right eye eyeball of experimenter;

(A4), using 1 reference electrode, reference electrode is newborn convex place after being placed in the left ear of experimenter;

(A5), using 1 ground-electrode, ground-electrode is newborn convex place after being placed in the right ear of experimenter.

As the further optimization of such scheme, in data prediction process, it is 0.5-8.5Hz for carrying out the limiting frequency of the bandpass filter of bandpass filtering step.

As the further optimization of such scheme, described design ICA spatial filter comprises the steps:

(B1), random selection one group of list time pan data y from EOG database_i(i=1 ..., N) and carry out ICA analysis, obtain mixing matrix M and the separation matrix D of 6 �� 6;

(B2), according to the mapping pattern of independent component at acquisition electrode, automatically select the relevant isolated component of pan and corresponding ICA wave filter, obtain corresponding respectively to the ICA spatial filter group { D under pan task context left and right, upper and lower_li, D_ri, D_ui, D_di}��

As the further optimization of such scheme, the learning method of separation matrix D in step (B1) is comprised the steps:

(C1), taking information maximum criterion as independence measurement foundation, it may also be useful to natural water surface coatings, separation matrix D being carried out iterative processing, iterative formula is see formula (1):

��D^T��{I-E[s]}D^T(1)

In formula (1), I is unit matrix, and E [] is average computing, and s is the source signal of estimated pan signalStatistic, statistic s with pan signal source signalPass be:

$s=T\·\tanh \left(\hat{x}\right(t\left)\right)\hat{x}{\left(t\right)}^T+\hat{x}\left(t\right)\hat{x}{\left(t\right)}^T---\left(2\right)$

In formula (2), T represents probability model switching matrix, and on its diagonal lines, the value of element comes from the source signal to pan signalThe dynamic estimation of kurtosis symbol,For the source signal of estimated pan signal;

(C2), to the source signal of pan signalCarry out normalized square mean process, such as formula (3):

$\hat{x}\left(t\right)\←\hat{x}\left(t\right)/diag\[\stackrel{\‾}{\hat{x}\left(t\right)}\]---\left(3\right)$

(C3), on the basis of formula (3), mixing matrix M and separation matrix D coefficient are adjusted, such as formula (4):

$M\←M\×diag\[\stackrel{\‾}{\hat{x}\left(t\right)}\]---\left(4\right)$

D^T=M^-1

In formula (3), (4),ForStandard deviation, diag () represent computing is converted into diagonal matrix.

As the further optimization of such scheme, the relevant isolated component of described automatic selection pan, comprises the steps:

(D1), to the mixing matrix M in step B1 take absolute value, i.e. | M |, and the maximum value of element in searching in | M | every row column vector by column, record the subscript of its column and corresponding electrode number;

(D2) select 4 column vectors at 2 vertical pan electrodes and position of leading, 24, horizontal saccade electrode places with maximum value element, respectively, record the row sequence number of its correspondence;

(D3) if comprising 4 described in step D2 column vector when matrix | M | is different, then abandon based on this list time ICA filter designs, otherwise, proceed to step D4;

(D4), according to gained row sequence number, separation matrix D finds corresponding row respectively, form 4 classes corresponding to the ICA spatial filter group under pan task context left and right, upper and lower: { D_li, D_ri, D_ui, D_di, (i=1 ..., N).

As the further optimization of such scheme, in step 3, spatial domain filtering method is as follows:

Use ICA spatial filter group { D_li, D_ri, D_ui, D_di(i=1 ..., N) and to all original pan eyes electricity data y_j(j=1 ..., N) and carry out airspace filter, such as formula (7):

$x_{lj}=D_{li}^T*y_j,x_{rj}=D_{ri}^T*y_j,x_{uj}=D_{ui}^T*y_j,x_{dj}=D_{di}^T*y_j---\left(7\right)$

In formula (7), x_lj,x_rj,x_uj,x_djRepresent this list time pan eye electricity data y respectively_jResult after airspace filter, the pan signal characteristic parameter namely extracted.

Based on a recognition methods for the pan signal characteristic extracting methods of independent component analysis, comprise the steps:

Step S1: data gathering: lead label E OG data in 6 when gathering experimenter's pan left and right, upper and lower respectively, and it is carried out bandpass filtering treatment;

Step S2: optimum ICA spatial filter group design: comprise the following steps:

(E1), single experiment data y is used_i(i=1,, N) and (i-th single experiment data) design linear ICA spatial filter, according to the mapping pattern of isolated component on acquisition electrode, automatically select the relevant isolated component of pan, obtain the ICA bank of filters { D corresponding to pan action left and right, upper and lower_li, D_ri, D_ui, D_di(i=1 ..., N);

(E2), at random all sampled datas are divided into 5 groups, select arbitrarily wherein one group as test sample book collection, remaining 4 groups then as training sample set, all learning sample are used the ICA bank of filters { D obtained in step e 1_li, D_ri, D_ui, D_diCarry out airspace filter, the result after linearly projection is sent in SVMs (SVM) as its characteristic parameter and trains;To test sample book, the above-mentioned ICA bank of filters { D of same use_li, D_ri, D_ui, D_diCarry out airspace filter, and result after projection is sent in the SVM classifier trained as characteristic parameter and identifies; Above-mentioned steps is repeated 10 times, and each experimental result is averaged, finally obtain at this ICA bank of filters { D_li, D_ri, D_ui, D_diUnder the Mean accurate rate of recognition of different pan signal;

(E3), to all data sample repeating step E1 and step e 2 in pan database, obtain N number of ICA bank of filters and corresponding Mean accurate rate of recognition, pick out the ICA bank of filters { D corresponding to the highest average recognition rate_l, D_r, D_u, D_dAs optimum spatial filter;

Step S3: the identification of pan signal: to experimental data to be identified, after pre-treatment, by optimum spatial filter group { D_l, D_r, D_u, D_dCarry out airspace filter to obtain characteristic parameter, and this characteristic parameter is sent in step e 2 the SVM model trained and identifies, move type to obtain the eye of this experiment.

Compared with the prior art, a kind of pan signal characteristic extracting methods based on independent component analysis of the present invention and recognition methods, have the advantages such as recognition correct rate is higher, extendability is stronger, application prospect is good. The useful effect of the present invention is embodied in the feature of the following aspects.

1, pan signal is had higher recognition correct rate by the present invention: the present invention can from leading the relevant independent component of the pan isolating multiple " truly " EOG signal more, therefore the truth that eye moves relevant independent source can be described more accurately, the component that simultaneously can effectively suppress signal dynamic with eye unrelated and the interference of external noise, obtain higher recognition correct rate.

2, the present invention has stronger extended capability in the identification of the dynamic type of eye: although the present invention only gives the Feature extraction and recognition method of four class pan signals, but ICA airspace filter method is to the number not restriction of leading of input signal, therefore, institute of the present invention extracting method has stronger classification extended capability, the Feature extraction and recognition of more dynamic types can be carried out, effectively improve the actual application value of algorithm.

3, the present invention has a good application prospect: the present invention, to improve EOG-HAR system performance argument mark, mainly solves in systematic realizing program the identification problem sweeping signal. Due to EOG-HAR system can active perception user view, be therefore with a wide range of applications in fields such as intelligent video monitoring, medical diagnosis, motion analysis and man-machine interactions. In addition, the present invention, to effectively improving the conventional man-machine interactive system performance based on EOG, helps some healths have deformity and the crowd of autokinetic movement can not promote quality of life and also have important meaning.

Accompanying drawing explanation

Fig. 1 is the generative process schematic diagram of pan signal.

Fig. 2 is the algorithm flow figure of the present invention.

Fig. 3-1, Fig. 3-2 and Fig. 3-3 are distribution of electrodes figure in the dynamic signal acquisition process of eye of the present invention.

The eye that Fig. 4 is the present invention moves the relative position schematic diagram observing target and experimenter in signal acquisition process.

Fig. 5 is the single experiment normal form schematic diagram of the present invention.

Fig. 6-1 Fig. 6-8 be training an eye movement data gained different pan tasks under M and D coefficient figure.

Fig. 7 is that ICA wave filter is trained with test data all from recognition correct rate during same experimenter.

Fig. 8 is that ICA wave filter is trained from test data respectively from recognition correct rate during different experimenter.

Embodiment

Below in conjunction with accompanying drawing, the invention will be further described:

A kind of pan signal characteristic extracting methods based on independent component analysis (IndependentComponentAnalysis:ICA), comprises the steps:

Step 1, multi-lead pan signals collecting and pre-treatment: lead electro-ocular signal with the 6 of data label when using 8 bioelectrodes to obtain the pan of experimenter upper and lower, left and right; And use bandpass filter to carry out filtering original multi-lead electro-ocular signal, to remove noise jamming; Optimize, in data prediction process, it is 0.5-8.5Hz for carrying out the limiting frequency of the bandpass filter of bandpass filtering step.

Step 2, ICA spatial filter design: use single experiment data y_i(i=1 ..., N) and carry out ICA analysis, and according to the mapping pattern of independent component at acquisition electrode, automatically select the relevant isolated component of pan and corresponding ICA wave filter, set up the ICA spatial filter group { D under corresponding different pan task context_li, D_ri, D_ui, D_di(i=1 ..., N); ;

Step 3, feature generate: refer to the ICA spatial filter group { D using step (2) to generate_li, D_ri, D_ui, D_diThe pan signal that leads to original 6 linearly projects, with the pan signal spatial feature parameter generated under corresponding task context.

Wherein, the installation position of multi-lead pan electrode signal acquisition is set to:

(A1), using 2 electrode collections vertically to sweep signal, an electrode is placed in 2.3-2.7cm place above experimenter's left eye eyeball center, and another electrode is placed in 2.3-2.7cm place below experimenter's left eye eyeball center; Or an electrode is placed in 2.3-2.7cm place above experimenter's right eye eyeball center, and another electrode is placed in 2.3-2.7cm place below experimenter's right eye eyeball center;

(A2), using 2 electrodes to gather horizontal saccade signal, an electrode is placed in the left eye eyeball horizontal center point left side 3.3-3.7cm place of experimenter; An electrode is placed in the 3.3-3.7cm place, right eye eyeball horizontal center point right of experimenter;

(A3), using 2 supporting electrodes, a supporting electrode is placed in 4.8-5.2cm place directly on the right side of the left eye eyeball of experimenter; A supporting electrode is placed in 4.8-5.2cm place directly on the left of the right eye eyeball of experimenter;

(A4), using 1 reference electrode, reference electrode is newborn convex place after being placed in the left ear of experimenter;

(A5), using 1 ground-electrode, ground-electrode is newborn convex place after being placed in the right ear of experimenter.

In the present invention, select when carrying out ICA computing EOG data that vertical, level and auxiliary 6 electrodes collect as analytic target, and when gathering, data are all added the label information of correspondence every time.

In step 2, design ICA spatial filter, it may also be useful to the spatial filter group under ICA method establishment correspondence pan task context, comprises the steps:

(B1), random selection one group of list time pan data y from EOG database_i(i=1 ..., N) and carry out ICA analysis, obtain mixing matrix M and the separation matrix D of 6 �� 6;

(B2), according to the mapping pattern of independent component at acquisition electrode, automatically select the relevant isolated component of pan and corresponding ICA wave filter, obtain corresponding respectively to the ICA spatial filter group { D under pan task context left and right, upper and lower_li, D_ri, D_ui, D_di}��

Mixing matrix M and separation matrix D are defined as follows:

If y (t)=[y₁(t),��,y_n(t)]^TFor n leads original EOG observation signal, this signal definition is separate implicit " source " x (t)=[x that n pan is relevant₁(t),��,x_n(t)]^TLinear instantaneous mixes, namely

Y (t)=Mx (t) (5)

In formula (5), M represents mixing matrix.

Corresponding with the mixture model of formula (5) is decomposition model, see formula (6):

$\hat{x}\left(t\right)=Dy\left(t\right)---\left(6\right)$

In formula (6), D represents separation matrix.

In the present invention, mixing matrix M and separation matrix D are carried out following process:

(C1), taking information maximum criterion as independence measurement foundation, it may also be useful to natural water surface coatings, separation matrix D being carried out iterative processing, iterative formula is see formula (1):

��D^T��{I-E[s]}D^T(1)

In formula (1), I is unit matrix, and E [] is average computing, and s is the source signal of estimated pan signalStatistic, statistic s with pan signal source signalPass be:

$s=T\·\tanh \left(\hat{x}\right(t\left)\right)\hat{x}{\left(t\right)}^T+\hat{x}\left(t\right)\hat{x}{\left(t\right)}^T---\left(2\right)$

In formula (2), T represents probability model switching matrix, and on its diagonal lines, the value of element comes from the source signal to pan signalThe dynamic estimation of kurtosis symbol,For the source signal of estimated pan signal;

(C2), to the source signal of pan signalCarry out normalized square mean process, such as formula (3):

$\hat{x}\left(t\right)\←\hat{x}\left(t\right)/diag\[\stackrel{\‾}{\hat{x}\left(t\right)}\]---\left(3\right)$

(C3), on the basis of formula (3), mixing matrix M and separation matrix D coefficient are adjusted, such as formula (4):

$M\←M\×diag\[\stackrel{\‾}{\hat{x}\left(t\right)}\]---\left(4\right)$

D^T=M^-1

In formula (3), (4),ForStandard deviation, diag () represent computing is converted into diagonal matrix.

Wherein, the automatic selecting method of the relevant isolated component of pan, comprises the steps:

(D1), to the mixing matrix M in step B1 take absolute value, i.e. | M |, and the maximum value of element in searching in | M | every row column vector by column, record the subscript of its column and corresponding electrode number;

(D2) select 4 column vectors at 2 vertical pan electrodes and position of leading, 24, horizontal saccade electrode places with maximum value element, respectively, record the row sequence number of its correspondence;

(D3) if comprising 4 described in step D2 column vector when matrix | M | is different, then abandon based on this list time ICA filter designs, otherwise, proceed to step D4;

(D4), according to gained row sequence number, separation matrix D finds corresponding row respectively, form 4 classes corresponding to the ICA spatial filter group under pan task context left and right, upper and lower: { D_li, D_ri, D_ui, D_di, (i=1 ..., N).

Use ICA spatial filter group { D_li, D_ri, D_ui, D_di(i=1 ..., N) and to all original pan eyes electricity data y_j(j=1 ..., N) and carry out airspace filter, such as formula (7):

$x_{lj}=D_{li}^T*y_j,x_{rj}=D_{ri}^T*y_j,x_{uj}=D_{ui}^T*y_j,x_{dj}=D_{di}^T*y_j---\left(7\right)$

In formula (7), x_lj,x_rj,x_uj,x_djRepresent this list time pan eye electricity data y respectively_jResult after airspace filter, the pan signal characteristic parameter namely extracted.

Based on a kind of pan signal characteristic extracting methods based on independent component analysis of the present invention, the invention also discloses the recognition methods of a kind of pan signal based on independent component analysis, comprise the steps:

Step S1: data gathering: gather 6 label E OG data of leading when experimenter is left and right, upper and lower to be swept respectively, and it is carried out bandpass filtering treatment;

Step S2: optimum ICA spatial filter group design: comprise the following steps:

(E1), single experiment data y is used_i(i=1,, N) and (i-th single experiment data) design linear ICA spatial filter, according to the mapping pattern of isolated component on acquisition electrode, automatically select the relevant isolated component of pan, obtain the ICA bank of filters { D corresponding to pan action left and right, upper and lower_li, D_ri, D_ui, D_di(i=1 ..., N);

(E2), at random all sampled datas are divided into 5 groups, select arbitrarily wherein one group as test sample book collection, remaining 4 groups then as training sample set, all learning sample are used the ICA bank of filters { D obtained in step e 1_li, D_ri, D_ui, D_diCarry out airspace filter, the result after linearly projection is sent in support vector machines as its characteristic parameter and trains;To test sample book, the above-mentioned ICA bank of filters { D of same use_li, D_ri, D_ui, D_diCarry out airspace filter, and result after projection is sent in the SVM classifier trained as characteristic parameter and identifies; Above-mentioned steps is repeated 10 times, and each experimental result is averaged, finally obtain at this ICA bank of filters { D_li, D_ri, D_ui, D_diUnder the Mean accurate rate of recognition of different pan signal;

(E3), to all data sample repeating step E1 and step e 2 in pan database, obtain N number of ICA bank of filters and corresponding Mean accurate rate of recognition, pick out the ICA bank of filters { D corresponding to the highest average recognition rate_l, D_r, D_u, D_dAs optimum spatial filter;

Step S3: the identification of pan signal: to experimental data to be identified, after pre-treatment, by optimum spatial filter group { D_l, D_r, D_u, D_dCarry out airspace filter to obtain characteristic parameter, and this characteristic parameter is sent in step e 2 the SVM model trained and identifies, move type to obtain the eye of this experiment.

See the generative process schematic diagram that Fig. 1, Fig. 1 are pan signal. Describe the EOG waveform caused when reading in the present embodiment. EOG waveform during read state, the EOG waveform that left and right two Different electrodes collect, respectively see the left electrodes signal amplitude in Fig. 1 and right electrodes signal amplitude, in figure, left electrodes signal amplitude and right electrodes signal amplitude have clear and definite corresponding relation. The potential difference that electro-ocular signal refers between the cornea eye that people causes due to the motion of eyes and retina causes. This electromotive force is initiated by retinal pigment epithelium and photoreceptor cell, its positive pole is positioned at sight sensor end, and negative pole is positioned at retinal pigment epithelium end, and the electric current produced has flowed to cornea end from retina end, thus to form a cornea be positive pole, retina is the electromotive force of negative pole. When people's eye movement, the amplitude of electro-ocular signal constantly can change along with the motion of eyeball, and the electromotive force of this kind of change is plotted on time axle then to form a curve by we, and this curve is just referred to as electrooculogram. People is when carrying out certain behavior, people's eye can present different motion rules, to read, people's eyes in reading process can constantly move to obtain new information in word interested, in this process, " pan " is the highest as the frequency of occurrences, the abundantest a kind of basic eye moves type to comprise human body behavioural information, its accuracy identified has become the major decision factor of HAR system performance, and therefore the present invention is mainly in order to solve the recognition correct rate of pan EOG signal.

See the algorithm flow figure that Fig. 2, Fig. 2 are the present invention. Describe the Feature extraction and recognition block diagram sweeping signal in the present embodiment. In concrete enforcement, mainly comprise following step:

1) gather multi-lead label E OG data and carry out bandpass filtering treatment; 2) single experiment data y is used_i(i=1,, N) and (i-th single experiment data) design linear ICA spatial filter, according to the mapping pattern of isolated component on acquisition electrode, automatically select the relevant isolated component of pan, obtain the ICA bank of filters { D corresponding to pan action left and right, upper and lower_li, D_ri, D_ui, D_di(i=1 ..., N); 3) at random all sampled datas are divided into 5 groups, select arbitrarily wherein one group as test sample book collection, remaining 4 groups then as training sample set. All learning sample are used the ICA bank of filters { D obtained in step 2_li, D_ri, D_ui, D_diCarry out airspace filter, the result after linearly projection is sent in SVMs (SVM) as its characteristic parameter and trains;To test sample book, the above-mentioned ICA bank of filters { D of same use_li, D_ri, D_ui, D_diCarry out airspace filter, and result after projection is sent in the SVM classifier trained as characteristic parameter and identifies. Above-mentioned steps is repeated 10 times, and each experimental result is averaged, finally obtain at this ICA bank of filters { D_li, D_ri, D_ui, D_diUnder the Mean accurate rate of recognition of different pan signal. 4) to data sample repeating steps 2 all in pan database and step 3, it is possible to obtain N number of ICA bank of filters and corresponding Mean accurate rate of recognition, pick out the ICA bank of filters { D corresponding to the highest average recognition rate_l, D_r, D_u, D_dAs optimum spatial filter. 5) to experimental data to be identified, by optimum spatial filter group { D after pre-treatment_l, D_r, D_u, D_dCarry out airspace filter to obtain characteristic parameter, and this characteristic parameter is sent in step 3 the SVM model trained and identifies, move type to obtain the eye of this experiment.

See Fig. 3, the eye that Fig. 3-1, Fig. 3-2 and Fig. 3-3 are the present invention moves distribution of electrodes figure in signal acquisition process. Describe distribution of electrodes in the dynamic signal acquisition process of eye in the present embodiment. The collection of electro-ocular signal uses Ag/AgCl electrode. Information and more spatial positional information is moved in order to obtain the eye of experimenter's four direction left and right, upper and lower, the present invention employs 8 electrodes altogether, wherein, electrode V1 and electrode V2 respectively correspondence be placed in above and below experimenter's left eye eyeball center each about 2.5cm place, move signal in order to gather vertical eye; Electrode H1 and electrode H2 respectively correspondence be placed in each about 3cm place in the experimenter left eye left side of eyeball horizontal center point and right eye eyeball right, move signal in order to gather level eye; Supporting electrode A1 and A2 be placed in respectively on the right side of experimenter's left eye eyeball directly over about 5cm place, surface on the left of about 5cm and right eye eyeball; Reference electrode C1 is newborn convex place after being placed in left ear; Ground-electrode G is newborn convex place after being positioned at right ear.

It it is the relative position schematic diagram observing target and experimenter in the dynamic signal acquisition process of eye of the present invention see Fig. 4, Fig. 4. Describe the relative position observing target and experimenter in the dynamic signal acquisition process of eye in the present embodiment. In specific implementation process, experimenter is sitting on an armchair, and, about 2 meters, its front is provided with the observation target of upper and lower, left and right four direction respectively and is all 1.5 meters apart from experimenter's Optic center point (O).

See the single experiment normal form schematic diagram that Fig. 5, Fig. 5 are the present invention. Fig. 5 describes the detailed process of single experiment normal form in the present embodiment. When testing beginning, first " beginning " character occurs on screen, and with the sonic stimulation of a 20ms length, after 1 second kind, experimenter can see red arrow prompting (being respectively: upwards arrow, down arrow, to the left arrow and right-hand arrow) on screen, it is 3 seconds that arrow continues time of occurrence on screen, within this time, requirement of experiment experimenter rotates eyeball to arrow direction indication after seeing arrow, rotating back into central point after seeing point of observation, experimenter can not blink in this course. Afterwards, having the of short duration rest of 2 seconds, experimenter can blink, loosen.

It is M and D coefficient figure under the different pan tasks of a training eye movement data gained see Fig. 6-1 Fig. 6-8, Fig. 6-1 Fig. 6-8. Fig. 6-1 Fig. 6-8 describes under using aforesaid method to be collected the different pan tasks training a pan data gained in the present embodiment and mixes matrix M and separation matrix D coefficient figure.It may be seen that when experimenter carries out left pan, eye movement lead on the 1st (H2 position) affect maximum, therefore the mixing matrix coefficient of its correspondence position is maximum; With reason, during right pan, the 2nd (H1 position) coefficient that leads is maximum; During upper pan, the 5th (V1 position) coefficient that leads is maximum; During lower pan, the 6th (V2 position) coefficient that leads is maximum. The result shows that the 4 class pan signal differences that the spatial filter tried to achieve can make us be classified reach maximumization.

See Fig. 7, Fig. 7 be ICA wave filter training with test data all from recognition correct rate during same experimenter. Describe ICA wave filter in the present embodiment to train with test data all from recognition correct rate during same experimenter. Wherein, X-coordinate 1-10 is corresponding 10 different experimenters respectively, and ordinate zou represents recognition correct rate. It may be seen that under this experiment condition, the highest recognition correct rate reaches 100%, minimum is 97.5%, finds after statistics, and the total average recognition rate of all experimenters reaches 99.3%. The result shows that the present invention is carried the pan signal characteristic abstraction based on ICA and can from lead the eye of EOG signal isolating multiple " truly " move relevant independent component to recognition methods more, therefore can describe the truth that eye moves relevant independent source more accurately, obtain ideal recognition correct rate.

See Fig. 8, Fig. 8 be ICA wave filter training from test data respectively from recognition correct rate during different experimenter. Describe ICA wave filter in the present embodiment to train from test data respectively from recognition correct rate during different experimenter. Wherein, X-coordinate represents the experimenter's index for training ICA wave filter, and ordinate zou represents that ICA wave filter that this experimenter trains is to the test result of other experimenter. Finding after statistics, under this experiment condition, the total average recognition rate of all experimenters reaches 94.2%. Compare the average accuracy of Fig. 7, under this experiment condition, have dropped 5.1%. Produce this result to cause primarily of the factor of following two aspects: 1) from experimenter's own characteristic, the position that different experimenter's eye moves relevant independent component may difference to some extent, simultaneously, in data gathering process, difference can be also there is in different experimenter in saccade velocity, time length, pan angle etc., this species diversity will make the EOG signal collected different, therefore, between different experimenter can there is difference to a certain degree in ICA wave filter, thus causes recognition rate to decline; 2) from the dynamic relevant independent component distribution of eye, different experimenter installs at electrode will inevitably exist some difference on position, and these difference will make eye move, and impact that the electrode installed at same position produces by relevant independent component is also different, there is larger difference in the coefficient namely mixing matrix M between different experimenter, this also can cause recognition rate to decline to some extent.

It it is more than one embodiment of the present invention. It is noted that for the person of ordinary skill of the art, under the premise without departing from the principles of the invention, it is also possible to make some distortion and improvement.

The foregoing is only the better embodiment of the present invention, not in order to limit the present invention, all any amendment, equivalent replacement and improvement etc. done within the spirit and principles in the present invention, all should be included within protection scope of the present invention.

Being to be understood that example as herein described and enforcement mode have been only explanation, those skilled in the art can make various amendment or change according to it, when not departing from spirit of the invention, all belongs to protection scope of the present invention.

Claims (8)

1. the pan signal characteristic extracting methods based on independent component analysis, it is characterised in that, comprise the steps:

Step 1, multi-lead pan signals collecting and pre-treatment: lead electro-ocular signal with the 6 of data label when using 8 bioelectrodes to obtain the pan of experimenter upper and lower, left and right; And use bandpass filter to carry out filtering original multi-lead electro-ocular signal, to remove noise jamming;

Step 2, ICA spatial filter design: use single experiment data y_i(i=1 ..., N) and carry out ICA analysis, and according to the mapping pattern of independent component at acquisition electrode, automatically select the relevant isolated component of pan and corresponding ICA wave filter, set up the ICA spatial filter group { D under corresponding different pan task context_li, D_ri, D_ui, D_di(i=1 ..., N);

Step 3, feature generate: refer to the ICA spatial filter group { D using step (2) to generate_li, D_ri, D_ui, D_diThe pan signal that leads to original 6 linearly projects, with the pan signal spatial feature parameter generated under corresponding task context.

2. the pan signal characteristic extracting methods based on independent component analysis according to claim 1, it is characterised in that: the installation position of described multi-lead pan electrode signal acquisition is set to:

(A1), using 2 electrode collections vertically to sweep signal, an electrode is placed in 2.3-2.7cm place above experimenter's left eye eyeball center, and another electrode is placed in 2.3-2.7cm place below experimenter's left eye eyeball center; Or an electrode is placed in 2.3-2.7cm place above experimenter's right eye eyeball center, and another electrode is placed in 2.3-2.7cm place below experimenter's right eye eyeball center;

(A2), using 2 electrodes to gather horizontal saccade signal, an electrode is placed in the left eye eyeball horizontal center point left side 3.3-3.7cm place of experimenter; An electrode is placed in the 3.3-3.7cm place, right eye eyeball horizontal center point right of experimenter;

(A3), using 2 supporting electrodes, a supporting electrode is placed in 4.8-5.2cm place directly on the right side of the left eye eyeball of experimenter; A supporting electrode is placed in 4.8-5.2cm place directly on the left of the right eye eyeball of experimenter;

(A4), using 1 reference electrode, reference electrode is newborn convex place after being placed in the left ear of experimenter;

(A5), using 1 ground-electrode, ground-electrode is newborn convex place after being placed in the right ear of experimenter.

3. the pan signal characteristic extracting methods based on independent component analysis according to claim 1, it is characterised in that: in data prediction process, it is 0.5-8.5Hz for carrying out the bandpass filter limiting frequency of bandpass filtering step.

4. the pan signal characteristic extracting methods based on independent component analysis according to claim 1, it is characterised in that: described design ICA spatial filter comprises the steps:

(B1), random selection one group of list time pan data y from EOG database_i(i=1 ..., N) and carry out ICA analysis, obtain mixing matrix M and the separation matrix D of 6 �� 6;

(B2), according to the mapping pattern of independent component at acquisition electrode, automatically select the relevant isolated component of pan and corresponding ICA wave filter, obtain corresponding respectively to the ICA spatial filter group { D under pan task context left and right, upper and lower_li, D_ri, D_ui, D_di(i=1 ..., N).

5. the pan signal characteristic extracting methods based on independent component analysis according to claim 4, it is characterised in that: the learning method of separation matrix D in step (B1) is comprised the steps:

(C1), taking information maximum criterion as independence measurement foundation, it may also be useful to natural water surface coatings, separation matrix D being carried out iterative processing, iterative formula is see formula (1):

��D^T��{I-E[s]}D^T(1)

In formula (1), I is unit matrix, and E [] is average computing, and s is the source signal of estimated pan signalStatistic, statistic s with pan signal source signalPass be:

$s=T\·\tanh \left(\hat{x}\right(t\left)\right)\hat{x}{\left(t\right)}^T+\hat{x}\left(t\right)\hat{x}{\left(t\right)}^T---\left(2\right)$

In formula (2), T represents probability model switching matrix, and on its diagonal lines, the value of element comes from the source signal to pan signalThe dynamic estimation of kurtosis symbol,For the source signal of estimated pan signal;

(C2), to the source signal of pan signalCarry out normalized square mean process, such as formula (3):

$\hat{x}\left(t\right)\←\hat{x}\left(t\right)/diag\[\stackrel{\‾}{\hat{x}\left(t\right)}\]---\left(3\right)$

(C3), on the basis of formula (3), mixing matrix M and separation matrix D coefficient are adjusted,

Such as formula (4):

$\begin{array}{c}M\←M\×diag\[\stackrel{\‾}{\hat{x}\left(t\right)}\]\\ D^T=M^{-1}\end{array}---\left(4\right)$ In formula (3), (4),ForStandard deviation, diag () represent computing is converted into diagonal matrix.

6. the pan signal characteristic extracting methods based on independent component analysis according to claim 5, it is characterised in that: the relevant isolated component of described automatic selection pan, comprises the steps:

(D1), to the mixing matrix M in step B1 take absolute value, i.e. | M |, and the maximum value of element in searching in | M | every row column vector by column, record the subscript of its column and corresponding electrode number;

(D2) select 4 column vectors at 2 vertical pan electrodes and position of leading, 24, horizontal saccade electrode places with maximum value element, respectively, record the row sequence number of its correspondence;

(D3) if comprising 4 described in step D2 column vector when matrix | M | is different, then abandon based on this list time ICA filter designs, otherwise, proceed to step D4;

(D4), according to gained row sequence number, separation matrix D finds corresponding row respectively, form 4 classes corresponding to the ICA spatial filter group under pan task context left and right, upper and lower: { D_li, D_ri, D_ui, D_di, (i=1 ..., N).

7. the pan signal characteristic extracting methods based on independent component analysis according to claim 4, it is characterised in that, in claim 1 step 3, spatial domain filtering method is as follows:

Use ICA spatial filter group { D_li, D_ri, D_ui, D_di(i=1 ..., N) and to all original pan eyes electricity data y_j(j=1 ..., N) and carry out airspace filter, such as formula (7):

$x_{lj}=D_{li}^T*y_j,x_{rj}=D_{ri}^T*y_j,x_{uj}=D_{ui}^T*y_j,x_{dj}=D_{di}^T*y_j---\left(7\right)$

In formula (7), x_lj,x_rj,x_uj,x_djRepresent this list time pan eye electricity data y respectively_jResult after airspace filter, the pan signal characteristic parameter namely extracted.

8. based on the recognition methods of the arbitrary described pan signal characteristic extracting methods based on independent component analysis of claim 1-7, it is characterised in that, comprise the steps:

Step S1: data gathering: gather 6 EOG data of leading when experimenter is left and right, upper and lower to be swept respectively, and it is carried out bandpass filtering treatment;

Step S2: optimum ICA spatial filter group design: comprise the following steps:

(E1), single experiment data y is used_i(i=1,, N) and (i-th single experiment data) design linear ICA spatial filter, according to the mapping pattern of isolated component on acquisition electrode, automatically select the relevant isolated component of pan, obtain the ICA bank of filters { D corresponding to pan action left and right, upper and lower_li, D_ri, D_ui, D_di(i=1 ..., N);

(E2), at random all sampled datas are divided into 5 groups, select arbitrarily wherein one group as test sample book collection, remaining 4 groups then as training sample set, all learning sample are used the ICA bank of filters { D obtained in step e 1_li, D_ri, D_ui, D_diCarry out airspace filter, the result after linearly projection is sent in SVMs (SVM) as its characteristic parameter and trains; To test sample book, the above-mentioned ICA bank of filters { D of same use_li, D_ri, D_ui, D_diCarry out airspace filter, and result after projection is sent in the SVM classifier trained as characteristic parameter and identifies; Above-mentioned steps is repeated 10 times, and each experimental result is averaged, finally obtain at this ICA bank of filters { D_li, D_ri, D_ui, D_diUnder the Mean accurate rate of recognition of different pan signal;

(E3), to all data sample repeating step E1 and step e 2 in pan database, obtain N number of ICA bank of filters and corresponding Mean accurate rate of recognition, pick out the ICA bank of filters { D corresponding to the highest average recognition rate_l, D_r, D_u, D_dAs optimum spatial filter;

Step S3: the identification of pan signal: to experimental data to be identified, after pre-treatment, by optimum spatial filter group { D_l, D_r, D_u, D_dCarry out airspace filter to obtain characteristic parameter, and this characteristic parameter is sent in step e 2 the SVM model trained and identifies, move type to obtain the eye of this experiment.