CN108836322A - A kind of naked eye 3D display vision induction motion sickness detection method - Google Patents

Abstract

The invention discloses a kind of, and the naked eye 3D display vision based on EEG induces motion sickness detection method.EEG data is pre-processed first；Then feature extraction is carried out to pretreated data, and is obtained and VIMS rank (VIMS Level, VIMSL) the high character subset of data dependence by feature selecting；Finally, classified by state of the structural classification device to VIMS, vision induction motion sickness symptom is judged whether there is to generate, to realize the detection to vision induction motion sickness (VIMS), vision induction motion sickness symptom is judged whether there is to generate, to realize the detection to vision induction motion sickness (VIMS), so that the improvement for naked eye 3D display technology provides support judgment mechanism.

Description

A kind of naked eye 3D display vision induction motion sickness detection method

Technical field

The present invention relates to naked eye 3D field of video displaying, more specifically more particularly to a kind of naked eye 3D display vision Induce motion sickness detection method.

Background technique

Naked eye 3D display is due to without passing through column in the case where wearing any ancillary equipment (such as 3D glasses, the helmet) The advanced optical technology such as lens and signal processing algorithm can watch the stereoscopic visual effect with impact force, represent 3D The developing direction in display technology future can be widely used for the fields such as media advertising, display and demonstration, scientific research and education, video display amusement, Completely new experience and visual enjoyment can be brought to the production of people, life various aspects.In recent years, as 4K and 8K shows skill The fast development of art, naked eye 3D technology is increasingly mature, software and hardware constantly improve, and naked eye 3D viewing effect has than in the past at present Very big raising.With the further development of 3D technology, the reduction of cost, content gradual perfection, naked eye 3D display will become Newest display and ultimate video product spread to family and each application field.

In new industry development plan, naked eye 3D display has been included in novel display special column by emphasis, and explicitly pointing out need to Grasp the integrated technologies such as its relevant program source, transmitting, transmission, reception, display.Market prediction mechanism IDC is in a research number It is pointed out in, naked eye 3D display technology graduallys mature, it is contemplated that be up to 18.4% to the year two thousand twenty shipment specific gravity.At this stage, large-size screen monitors Naked eye 3D display is applied to household application market, and there are also some technologies and content bottlenecks, are only with naked eye 3D mobile phone, naked eye 3D PAD Example, the industry have the market scale of more than one hundred billion.

However, people long-time using naked eye 3D display equipment viewing 3D video during, often generate dizziness, The symptoms such as nausea even vomiting, these symptoms are referred to as vision induction motion sickness (Visually Induced by we Motion Sickness, VIMS), these serious symptoms affect the user experience of naked eye 3D display product, greatly restrict The popularization and application of naked eye 3D product.Therefore, carry out the detection research to VIMS to have great importance.

About the detection of VIMS, there is scholar to propose in a kind of dynamic driving environment based on brain wave in the past The online cinetosis assessment system of (Electroencephalogram, EEG).The system need more complicated brain wave acquisition equipment with The support of the EEG data of more port numbers is obtained, then carries out directly carrying out frequency band selection after channel selecting and time-frequency convert Two steps have directly abandoned the information compared with multichannel and frequency band；And after the process flow of time frequency analysis, the system There is no further make full use of the feature of frequency domain.And VIMS detection algorithm proposed by the present invention then avoids Problem is stated, can remain to reach using relatively simple brain wave acquisition equipment in the case where the EEG data in only 4 channels Higher Detection accuracy, the building suitable for VIMS detection system under the conditions of the wearable wireless device in less channel.

Summary of the invention

The vision induction motion sickness inspection based on EEG that it is an object of the invention to overcome the deficiencies of the prior art and provide a kind of Method of determining and calculating can effectively detect the generation of VIMS in the case where less channel data EEG signal, enrich people to VIMS Research, so that the improvement and development for naked eye 3D display technology provide reference.

For achieving the above object, the present invention proposes a kind of naked eye 3D display vision induction motion sickness inspection based on EEG Method of determining and calculating, which is characterized in that including following three big modules：Pretreatment, feature extraction and selection and VIMS state classification, wherein Each module divides several steps again, introduces more detail below：

S1：Preprocessing module：

S101：Data deduplication：

We have found that in the raw EEG data of acquisition, there are minimal amount of repeated data (4 channels in adjacent record Floating number identical be considered as repetition).It is duplicate for being averaged in 400,000 records of each subject there are about 500 records.This Partial data may be due to the quantization digit of Muse is less and sample frequency it is lower caused by, wherein Muse equipment is with being EEG data is acquired, completes the acquisition of basic physiological data.Repeated data can not the instant brain state of accurate characterization, Therefore it to be removed.

S102：Normalization：

Since the numerical fluctuations of raw EEG data are larger, if analyzed using original numerical value, numerical value will be exaggerated The effect of higher data.In addition, have department pattern after each dimension unevenly stretch in machine learning, it is optimal Solution and original non-equivalence, such as support vector machines (Support Vector Machine, SVM) and artificial neural network (Artificial Neural network, ANN) etc..In order to enable different numerical value is comparable, and do not change original The distribution of data, it would be desirable to which raw EEG data is normalized.

S103：Data filtering：

Scientific investigations showed that effective EEG signals frequency of normal person covers 0.5Hz to 50Hz, and in eeg signal acquisition During again inevitably influenced by high-power power frequency component.Experimental data of the present invention is cured in Harvard of the U.S. What institute collected, it is contemplated that Unite States Standard electric voltage frequency is 60Hz, therefore the present invention designs the low pass for choosing that the upper limit is 50Hz Filter removes the Hz noise in EEG signals.Simultaneously, it is contemplated that band logical area (50Hz or less) EEG signal should be kept Superperformance, and wish to decay as far as possible Hz noise in band resistance area (50Hz or more), therefore selected band logical area most It is flat, it is slowed down under stopband, but can finally decay to 0 Butterworth filter.Finally, the present invention select SC service ceiling for Butterworth (Butterworth) low-pass filter of 50Hz removes the Hz noise in EEG signal.Wherein, Butterworth filter is a kind of electronic filter, and its feature is that the frequency response curve in passband is flat, is not had Waviness variation, and in suppressed frequency band then monotonic decreasing until be zero.

S2：Feature extraction and selection module：

S201：Feature extraction：

In feature extraction phases of the invention, we used two kinds of distinct methods, and feature is extracted from raw EEG signal, Wavelet Entropy, each wavelet sub-band Coefficient Mean and the calculated maximum of statistical property obtained including the Wavelet Properties based on EEG Value, minimum value, mean value and standard deviation.Then correlation analysis is made to all 44 obtained candidate features and VIMSL column again, Selection has 22 features of significant difference as final detection feature.This stage mainly extracts candidate feature, is feature The first step extracted and selected.

The EEG data for acquiring 4 channels (TP9, FP1, FP2, TP10) in the present invention altogether, each sliding to each channel Data in window carry out 6 rank wavelet transformations respectively and acquire wavelet coefficient and Wavelet Entropy, because the frequency of normal person's EEG signal is low In 50Hz, so eliminating the subband of 55Hz~110Hz in the present invention when handling Wavelet Transform Feature.Cause This can be obtained by 1 Wavelet Entropy, 6 subband Coefficient Means and maximum value, most to each sliding window data in each channel 11 candidate features such as small value, mean value, standard deviation, 4 channels then one share 44 (11*4) a candidate features.

S202：Feature selecting：

Feature selecting is also referred to as " feature subset selection " or " Attributions selection ", generally refers to from existing candidate feature One most correlated characteristic subset of selection is concentrated to be trained so that specific objective function optimization, is to improve learning model performance Important method, even more data handling procedure particularly critical in mode identification procedure.In a sense, using more Exquisite learning sample can obtain more preferably experimental result.

In the present invention, we carry out spy using the method for calculating each feature and the correlation of response variable Sign selection.In the characteristic extraction procedure of step S201, algorithm synthesis wavelet character and statistical nature totally 44 attribute conducts Candidate feature, but there have some features and the response variable of experiment to do correlation analysis discovery in these candidate features to be not significant Sex differernce generates, and can have bad influence on training result if retaining these candidate features, therefore these should not be shown The feature for writing sex differernce is removed, boosting algorithm accuracy.

In this process, Spearman rank phase is made with VIMSL column respectively to 44 characteristics that step S201 is extracted The analysis of closing property (does correlation analysis because of the data for being continuous data and ranking score type, selects Spearman rank correlation Property analysis it is more reasonable).

It is analyzed according to Spearman rank correlation as a result, we have finally chosen with significant difference (sig<0.05) 22 correlated characteristics as final training characteristics, be respectively：TP9_entro,TP9_1,TP9_max,TP9_min,TP9_ mean、 TP9_std、FP1_entro、FP1_1、FP1_max、FP1_min、FP1_mean、FP1_std、FP2_entro、FP2_ max、 FP2_min、FP2_std、TP10_entro、TP10_1、TP10_max、TP10_min、TP10_mean、TP10_std。

Pay attention to：In order to preferably keep VIMS information and characteristic that may be present in legacy data, only consider here Feature and response column whether there is significant difference, the size without considering related coefficient retains all related columns Get off, does not carry out dimensionality reduction from channel and frequency band level.

S3：VIMS state classification module：

Use pattern identifies the algorithm model of (pattern classification) to carry out cinetosis feature detection in the present invention.Pattern-recognition Mathematical technique method is exactly used for automatically processing for research characteristic mode using computer by (Pattern Recognition, PR) And interpretation.And pattern classification is that feature set is mapped to some or multiple known classifications using classification function or disaggregated model, It is the important component of pattern-recognition, directly influences the effect of its identification, pattern classification has been applied to section at present Learn the various aspects of research.

In the present invention, we mainly carry out pattern classification to the data characteristics obtained above with machine learning algorithm.? VIMS state classification detection-phase, data acquisition system is first divided into 70% training set and 30% test set by we, and upper 22 column features and the VIMSL column that the process of stating obtains, which are put into respectively in 3 kinds of machine learning models, carries out supervised learning training, packet Tri- kinds of random forest (Random Forests, RF), SVM and ANN models are included, finally model result is carried out with test set again Assessment and discussion.Then, in conjunction with naked eye 3D display feature and the concrete condition of experiment, for having detected whether that cinetosis symptom produces It is raw.Finally we are to each model respectively from accuracy rate, precision, recall rate, F1-Score and AUC (Area Under Curve, AUC) multiple evaluation indexes such as value are comprehensively considered and are analyzed, and determine optimal pattern classification model.Wherein, Random forest (RF) is the pattern classification model that detection algorithm of the present invention finally determines, support vector machines and artificial neural network Model as a comparison.

Detailed description of the invention

Fig. 1 is a kind of naked eye 3D display vision induction motion sickness detection method flow chart based on EEG.

Fig. 2 is that raw EEG signal filters front and back effect contrast figure.

Fig. 3 is the Receiver operating curve of three kinds of experimental results, wherein (a) is Random Forest model ROC curve； It (b) is SVM model ROC curve；It (c) is BP neural network model ROC curve.

Fig. 4 is P9 channel characteristics and VIMSL column correlation analysis result table.

Fig. 5 is FP1 channel characteristics and VIMSL column correlation analysis result table.

Fig. 6 is FP2 channel characteristics and VIMSL column correlation analysis result table.

Fig. 7 is TP10 channel characteristics and VIMSL column correlation analysis result table.

Fig. 8 is two classification results table of pattern-recognition.

Fig. 9 is single two classification experiments result table of subject.

Specific embodiment

The present invention is described in further detail with specific embodiment with reference to the accompanying drawing.

Fig. 1 be it is proposed by the present invention based on EEG naked eye 3D display vision induction motion sickness detection method flow chart.This Invention first pre-processes EEG data；Then feature extraction is carried out to pretreated data, and passes through feature selecting It obtains and VIMS rank (VIMS level, VIMSL) the high character subset of data dependence；Finally, passing through structural classification device pair The state of VIMS is classified, and is mainly included the following steps that：

S1：Preprocessing module：

S101：Data deduplication：

We have found that in the raw EEG data of acquisition, there are minimal amount of repeated data (4 channels in adjacent record Floating number identical be considered as repetition).It is duplicate for being averaged in 400,000 records of each subject there are about 500 records.This Partial data may be due to the quantization digit of Muse is less and sample frequency it is lower caused by, wherein Muse equipment is with being EEG data is acquired, completes the acquisition of basic physiological data.Repeated data can not the instant brain state of accurate characterization, Therefore it to be removed.

S102：Normalization：

Since the numerical fluctuations of raw EEG data are larger, if analyzed using original numerical value, numerical value will be exaggerated The effect of higher data.In addition, have department pattern after each dimension unevenly stretch in machine learning, it is optimal Solution and original non-equivalence, such as support vector machines (Support Vector Machine, SVM) and artificial neural network (Artificial Neural network, ANN) etc..In order to enable different numerical value is comparable, and do not change original The distribution of data, it would be desirable to which raw EEG data is normalized.

S103：Data filtering：

Scientific investigations showed that effective EEG signals frequency of normal person covers 0.5Hz to 50Hz, and in eeg signal acquisition During again inevitably influenced by high-power power frequency component.Experimental data of the present invention is cured in Harvard of the U.S. What institute collected, it is contemplated that Unite States Standard electric voltage frequency is 60Hz, therefore the present invention designs the low pass for choosing that the upper limit is 50Hz Filter removes the Hz noise in EEG signals.Simultaneously, it is contemplated that band logical area (50Hz or less) EEG signal should be kept Superperformance, and wish to decay as far as possible Hz noise in band resistance area (50Hz or more), therefore selected band logical area most It is flat, it is slowed down under stopband, but can finally decay to 0 Butterworth filter.Finally, the present invention select SC service ceiling for Butterworth (Butterworth) low-pass filter of 50Hz removes the Hz noise in EEG signal.Wherein, Butterworth filter is a kind of electronic filter, and its feature is that the frequency response curve in passband is flat, is not had Waviness variation, and in suppressed frequency band then monotonic decreasing until be zero.

According to data flow shown in FIG. 1, inputs 4 channel (TP9, FP1, FP2, TP10) raw EEG signals and enter and be Next system is marked the receipt of subject and integrate, when by the VIMSL of each subject and EEG data according to acquisition Between be mapped, the state between two of them VIMSL is identical with previous moment.Then, step 102 duplicate removal is carried out to data Processing, the duplicate data in removal initial data the inside then carry out step S102 normalization and step S103 data filtering behaviour Make, SC service ceiling is that the Butterworth filter of 50Hz filters out Hz noise and myoelectricity interference in EEG signal.

Time domain and the frequency domain comparison diagram for filtering the channel FP1 of front and back are as shown in Figure 2.Wherein, upper left and upper right are original EEG The time-domain image of signal filtering front and back, it can be seen that the graph line on right side is obviously lighter than left hand view, and high frequency section is less；And Lower-left and bottom right are then the frequency domain image that EEG signal is filtered front and back, it can clearly be seen that low by Butterworth After bandpass filter filtering, raw EEG signal has obtained great decaying in the part that frequency is more than 50Hz, and frequency exists 50Hz and partial data below are held essentially constant.It is good that this illustrates that filter plays Hz noise and myoelectricity interference Good inhibiting effect eliminates the noise in EEG signal, while protecting the effective EEG signals of normal person to greatest extent Frequency band.

S2：Feature extraction and selection module：

S201：Feature extraction：

In feature extraction phases of the invention, we used two kinds of distinct methods, and feature is extracted from raw EEG signal, Wavelet Entropy, each wavelet sub-band Coefficient Mean and the calculated maximum of statistical property obtained including the Wavelet Properties based on EEG Value, minimum value, mean value and standard deviation.Then correlation analysis is made to all 44 obtained candidate features and VIMSL column again, Selection has 22 features of significant difference as final detection feature.This stage mainly extracts candidate feature, is feature The first step extracted and selected.

The EEG data for acquiring 4 channels (TP9, FP1, FP2, TP10) in the present invention altogether, each sliding to each channel Data in window carry out 6 rank wavelet transformations respectively and acquire wavelet coefficient and Wavelet Entropy, because the frequency of normal person's EEG signal is low In 50Hz, so eliminating the subband of 55Hz~110Hz in the present invention when handling Wavelet Transform Feature.Cause This can be obtained by 1 Wavelet Entropy, 6 subband Coefficient Means and maximum value, most to each sliding window data in each channel 11 candidate features such as small value, mean value, standard deviation, 4 channels then one share 44 (11*4) a candidate features.Because data volume compared with Greatly, sliding window is not overlapped by data sectional into every section is exactly a window so taking.In present invention test, window size It is 500.

S202：Feature selecting：

Feature selecting is also referred to as " feature subset selection " or " Attributions selection ", generally refers to from existing candidate feature One most correlated characteristic subset of selection is concentrated to be trained so that specific objective function optimization, is to improve learning model performance Important method, even more data handling procedure particularly critical in mode identification procedure.In a sense, using more Exquisite learning sample can obtain more preferably experimental result.

In the present invention, we carry out spy using the method for calculating each feature and the correlation of response variable Sign selection.In the characteristic extraction procedure of step S201, algorithm synthesis wavelet character and statistical nature totally 44 attribute conducts Candidate feature, but there have some features and the response variable of experiment to do correlation analysis discovery in these candidate features to be not significant Sex differernce generates, and can have bad influence on training result if retaining these candidate features, therefore these should not be shown The feature for writing sex differernce is removed, boosting algorithm accuracy.

In this during, Spearman rank is made with VIMSL column respectively to 44 characteristics that step S201 is extracted Correlation analysis (does correlation analysis because of the data for being continuous data and ranking score type, selects Spearman rank phase The analysis of closing property is more reasonable).It briefly introduces Spearman rank correlation in experimentation and analyzes result (this point in following part Analysis result is calculated by statistics software SPSS).

It is analyzed according to Spearman rank correlation as a result, we have finally chosen with significant difference (sig<0.05) 22 correlated characteristics as final training characteristics, be respectively：TP9_entro,TP9_1,TP9_max,TP9_min,TP9_ mean、 TP9_std、FP1_entro、FP1_1、FP1_max、FP1_min、FP1_mean、FP1_std、FP2_entro、FP2_ max、 FP2_min、FP2_std、TP10_entro、TP10_1、TP10_max、TP10_min、TP10_mean、TP10_std。

Pay attention to：In order to preferably keep VIMS information and characteristic that may be present in legacy data, only consider here Feature and response column whether there is significant difference, the size without considering related coefficient retains all related columns Get off, does not carry out dimensionality reduction from channel and frequency band level.

S3：VIMS state classification module：

Use pattern identifies the algorithm model of (pattern classification) to carry out cinetosis feature detection in the present invention.Pattern-recognition Mathematical technique method is exactly used for automatically processing for research characteristic mode using computer by (Pattern Recognition, PR) And interpretation.And pattern classification is that feature set is mapped to some or multiple known classifications using classification function or disaggregated model, It is the important component of pattern-recognition, directly influences the effect of its identification, pattern classification has been applied to section at present Learn the various aspects of research.

In the present invention, we mainly carry out pattern classification to the data characteristics obtained above with machine learning algorithm.I First carry out two classification experiments, main purpose is whether detection subject has cinetosis generation.Rank is detected in VIMS state classification Section, data acquisition system is first divided into 70% training set and 30% test set by we, and 22 column that the above process is obtained are special The VIMSL that seeks peace column, which are put into respectively in 3 kinds of machine learning models, carries out supervised learning training, including random forest (Random Forests, RF), tri- kinds of models of SVM and ANN, finally model result is assessed and is discussed with test set again.Then, it ties The concrete condition of naked eye 3D display feature and experiment is closed, for having detected whether that cinetosis symptom generates.Finally we are to each mould Type is respectively from multiple evaluations such as accuracy rate, precision, recall rate, F1-Score and AUC (Area Under Curve, AUC) values Index is comprehensively considered and is analyzed, and determines optimal pattern classification model, and wherein Fig. 8 is two classification results of pattern-recognition Table.Wherein, random forest (RF) is the pattern classification model that detection algorithm of the present invention finally determines, support vector machines and artificial Neural network model as a comparison.

Fig. 3 is the Receiver operating curve of three kinds of experimental results.The Random Forest model it can be seen from experimental result There are higher accuracy rate and AUC value, and precision, recall rate and F1-Score value are also higher, have absolutely proved this experiment Select Random Forest model have preferable stability to the classification of data set, VIMS Detection accuracy is preferable, generalization ability compared with By force, it can be good at handling more complicated nonlinear data, have preferable modeling energy particularly with the physiological signal of non-stationary Power can be used as the detection model of vision induction motion sickness；SVM model has preferable recall rate, but precision is lower, explanation Model is easy excessive early warning, and non-cinetosis situation is detected as cinetosis situation by meeting greater probability；The artificial neural network of shallow-layer is quasi- True rate is relatively low, although there is bigger fault-tolerant ability to training data, is also required to bigger data volume or more preferably feature Extracting method is even similar to the automatic method for carrying out feature extraction such as CNN, the data and feature that this experiment acquires at present Extracting method cannot still reach better experiment effect.Above-mentioned three kinds of learning models are compared, random forest (RF) model is to divorced Point and fluctuate violent data and have a preferable processing capacity, performance is stablized, best to the experiment effect of this paper, therefore select with Machine forest is as the VIMS detection model tested herein.

We classify the characteristic of 8 subjects below, raw to each single subject with above-mentioned 3 models respectively It manages data and carries out model verifying progress collective model assessment.Single two classification experiments result table of subject is shown in Fig. 9, at random Forest model still has higher accuracy rate and AUC value, illustrates that the model has a preferable Generalization Capability, and the feature of selection can be compared with Good detects vision induction motion sickness, which also further demonstrates the mould selected in model above evaluation process The correctness of type.Meanwhile comparison combine 8 subject data experimental result and single subject data as a result, it has been found that： Result of the trained Random Forest model in single subject data has than the result on integrated data test set obviously to be mentioned Height illustrates that each subject has identical feature to occur when VIMS is generated, but also has individual difference, and from individually by The feature of cinetosis cannot be completely extracted in examination person's data, can be learnt from the data of other individual subject to mending It fills.

Any characteristic disclosed in this specification or disclosed all methods or in the process the step of, in addition to mutually exclusive Feature and/or step except, can combine in any way.

Although the illustrative specific embodiment of the present invention is described above, in order to the technology of the art Personnel understand the present invention, it should be apparent that the present invention is not limited to the ranges of specific embodiment, to the common of the art For technical staff, if various change the attached claims limit and determine the spirit and scope of the present invention in, this A little variations are it will be apparent that all utilize the innovation and creation of present inventive concept in the column of protection.

Any feature disclosed in this specification (including any accessory claim, abstract and attached drawing), except non-specifically chatting State, can it is equivalent by other or tool similar purpose alternative features substituted.That is, unless specifically stated, each feature is An example in a series of equivalent or similar characteristics.

The invention is not limited to specific embodiments above-mentioned.The present invention, which expands to, any in the present specification to be disclosed New feature or any new combination, and disclose any new method or process the step of or any new combination.

Claims (7)

1. a kind of naked eye 3D display vision induces motion sickness detection method, which is characterized in that include the following steps：

S1：Preprocessing module：

S101：Data deduplication：We have found that having minimal amount of repeated data (adjacent record in the raw EEG data of acquisition In the floating number in 4 channels identical be considered as repetition).There are about 500 records in 400,000 records of average each subject is It is duplicate.This partial data may be due to the quantization digit of Muse is less and sample frequency it is lower caused by, wherein Muse Equipment completes the acquisition of basic physiological data with EEG data is acquired.Repeated data can not the instant brain of accurate characterization Portion's state, therefore to be removed.

S102：Normalization：It, will if analyzed using original numerical value since the numerical fluctuations of raw EEG data are larger Exaggerate the effect of the higher data of numerical value.In addition, there is department pattern unevenly to be stretched in machine learning in each dimension Afterwards, optimal solution and original non-equivalence, such as SVM, ANN etc..In order to enable different numerical value is comparable, and do not change original The distribution of beginning data, it would be desirable to which raw EEG data is normalized.

S103：Data filtering：Scientific investigations showed that effective EEG signals frequency of normal person covers 0.5Hz to 50Hz, and in brain It is inevitably influenced again by high-power power frequency component during electrical signal collection.The experimental data of this paper is in the U.S. Harvard Medical School collects, it is contemplated that Unite States Standard electric voltage frequency is 60Hz, therefore it is the low of 50Hz that the upper limit is chosen in design herein Bandpass filter removes the Hz noise in EEG signals.Simultaneously, it is contemplated that band logical area (50Hz or less) EEG signal should be kept Superperformance, and wish to decay as far as possible Hz noise in band resistance area (50Hz or more), therefore selected band logical area most It is flat, it is slowed down under stopband, but can finally decay to 0 Butterworth filter.Finally, this algorithm experimental selects SC service ceiling The Hz noise in EEG signal is removed for the Butterworth low-pass filter of 50Hz.

S2：Feature extraction and selection module：

S201：Feature extraction：In feature extraction phases of the invention, we used two kinds of distinct methods from raw EEG signal Middle extraction feature, the Wavelet Entropy obtained including the Wavelet Properties based on EEG, each wavelet sub-band Coefficient Mean and statistical property calculate Maximum value, minimum value, mean value and standard deviation out.Then related to VIMSL column work to all 44 obtained candidate features again Property analysis, selection have 22 features of significant difference as final detection feature.This stage mainly extracts candidate feature, It is the first step of feature extraction and selection.

S202：Feature selecting：In the present invention, we are using the side for calculating each feature and the correlation of response variable Method carries out feature selecting.In the characteristic extraction procedure of step S201, algorithm synthesis wavelet character and statistical nature totally 44 A attribute has some features and the response variable of experiment to do correlation analysis discovery as candidate feature, but in these candidate features There is no significant difference generation, there can be bad influence on training result if retaining these candidate features, therefore should be this The feature of significant difference is not removed a bit, boosting algorithm accuracy.

S3：VIMS state classification module：The algorithm model of use pattern identification (pattern classification) is special to carry out cinetosis in the present invention Sign detection.Pattern-recognition (Pattern Recognition, PR) is exactly used to mathematical technique method using computer study spy Sign mode automatically process and interpretation.And pattern classification be using classification function or disaggregated model by feature set be mapped to some or Multiple known classifications, it is the important component of pattern-recognition, directly influences the effect of its identification, at present pattern classification It has been applied to the various aspects of scientific research.

2. a kind of naked eye 3D display vision according to claim 1 induces motion sickness detection method, it is characterized in that, training It during model, needs aggregation of data together, then carries out data mark according to the data in the middle position of sliding window Then the mark of label carries out the training of supervised learning.

3. a kind of naked eye 3D display vision according to claim 1 induces motion sickness detection method, it is characterized in that, it is described The Butterworth filter that the upper limit is 50Hz is carried out to labeled data in step S103, the power frequency removed in EEG data is dry It disturbs and myoelectricity interference.

4. a kind of naked eye 3D display vision according to claim 1 induces motion sickness detection method, it is characterized in that, it is described The wavelet transformation in sliding window is carried out to the data after pretreated in step S201, acquires number in each window in each channel According to wavelet band Coefficient Mean and Wavelet Entropy and window in data maxima and minima, mean value and standard deviation are made For the candidate feature of EEG data.

5. a kind of naked eye 3D display vision according to claim 1 induces motion sickness detection method, it is characterized in that, it is described For candidate feature in step S202, according to the analysis of candidate feature and the Spearman rank correlation of VIMSL as a result, selection Have the level of signifiance as final training characteristics.Wherein, the significance for only considering whether correlation, without considering phase The size of relationship number.

6. a kind of naked eye 3D display vision according to claim 1 induces motion sickness detection method, it is characterized in that, it is described Select random forest as the classifier of VIMS state classification in step S3, and comprehensive all experimental datas obtain training characteristics It is trained.

7. a kind of naked eye 3D display vision according to claim 1 induces motion sickness detection method, it is characterized in that, entirely Detection algorithm is all linked with one another, and every a one-step process all has great influence to final testing result, although each step has Oneself the characteristics of, but the flow chart of data processing of entire algorithm is also critically important.