CN110399846A - A kind of gesture identification method based on multichannel electromyography signal correlation - Google Patents
A kind of gesture identification method based on multichannel electromyography signal correlation Download PDFInfo
- Publication number
- CN110399846A CN110399846A CN201910689483.3A CN201910689483A CN110399846A CN 110399846 A CN110399846 A CN 110399846A CN 201910689483 A CN201910689483 A CN 201910689483A CN 110399846 A CN110399846 A CN 110399846A
- Authority
- CN
- China
- Prior art keywords
- signal
- classifier
- electromyography signal
- imf
- electromyography
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F3/00—Input arrangements for transferring data to be processed into a form capable of being handled by the computer; Output arrangements for transferring data from processing unit to output unit, e.g. interface arrangements
- G06F3/01—Input arrangements or combined input and output arrangements for interaction between user and computer
- G06F3/011—Arrangements for interaction with the human body, e.g. for user immersion in virtual reality
- G06F3/015—Input arrangements based on nervous system activity detection, e.g. brain waves [EEG] detection, electromyograms [EMG] detection, electrodermal response detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2203/00—Indexing scheme relating to G06F3/00 - G06F3/048
- G06F2203/01—Indexing scheme relating to G06F3/01
- G06F2203/011—Emotion or mood input determined on the basis of sensed human body parameters such as pulse, heart rate or beat, temperature of skin, facial expressions, iris, voice pitch, brain activity patterns
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
- G06F2218/06—Denoising by applying a scale-space analysis, e.g. using wavelet analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Landscapes
- Engineering & Computer Science (AREA)
- General Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Neurosurgery (AREA)
- General Health & Medical Sciences (AREA)
- Neurology (AREA)
- Health & Medical Sciences (AREA)
- Dermatology (AREA)
- Human Computer Interaction (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Biomedical Technology (AREA)
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
Abstract
The present invention proposes a kind of gesture identification method based on multichannel electromyography signal correlation, denoising is carried out to the electromyography signal of each channel acquisition first, active segment is being detected according to signal amplitude power, then structuring processing is carried out to movable segment signal, by being superimposed several continuous time windows signal processing into the format with temporal correlation, to further improve recognition accuracy, finally, based on CNN+RNN neural fusion hybrid neural networks CRNet, establish the classifier of gesture identification, the input of classifier is the signal of structuring processing, the output of classifier is gesture classification probability, gesture identification is carried out using trained classifier.The present invention only uses several myoelectric sensor acquisition original signals, does not need additional complex device, easy to operate and good environmental adaptability.The present invention can effectively remove the noise in signal, and classifier used reduces computing resource and improves recognition efficiency, be more suitable for engineer application.
Description
Technical field
The present invention relates to a kind of gestures detections and recognition methods based on multichannel myoelectric sensor signal, belong to sensor
Signal processing and identification technology field.
Background technique
In recent years, the development for promoting human-computer interaction technology is widely applied in embedded intelligent equipment.Wherein, gesture identification exists
It plays an important role in human-computer interaction, is increasingly drawn by gesture control smart machine (such as unmanned plane, robot) work
Play the concern of correlative study person.Gesture can be with the information of expressed in abundance, up to 30000 kinds of the gesture of various countries worldwide, this
It exchanges and brings great convenience between person to person and people and machine.Especially for hearing-impaired people and formula interaction of keeping silent
(such as tactics gesture instruction), protrusion embody the unique advantage that gesture has.
The development of Gesture Recognition is to realize the basis based on gesture interaction, at present according to identification method difference, substantially
Tradition and non-traditional two class can be divided into.Wherein traditional method mainly uses machine vision method to identify gesture, uses list
Mesh or binocular camera obtain images of gestures, extract gesture feature by a series of images processing technique, and then identify gesture letter
Breath.This mode is more intuitive, and meets the visual custom of people.But since image information data amount is big, to machine vision
Algorithm performance requirement is relatively high, is extremely difficult to real-time.In addition, light is affected to picture quality, it is higher to environmental requirement.
In a dark environment or under obstruction conditions, this mode for obtaining images of gestures based on camera will be unable to realize.This is just needed
A kind of new approach completion gesture identification is explored, realizes information exchange.Inertial sensor is mainly used for unorthodox method
Identify gesture motion.Research shows that can identify the motion profile of gesture based on inertial sensor, this is solved to a certain extent
The problem of conventional method.Although however this kind of sensor can characterize movement of the gesture in space, can not accurately divide
The change of finger movement is precipitated, it cannot the corresponding content information of expression gesture abundant.
As biology information technology develops, electromyography signal more and more attention has been paid to.Myoelectricity is by flesh during limb motion
The superposition of the flexible current potential generated of meat fiber in time, the movement of upper limb and finger can all generate myoelectricity, and difference movement generates
Signal there is unique signal rule, therefore carry out gesture identification with electromyography signal feature and provide one in order to solve the above problem
Kind may.However a kind of faint time series signal of the electromyography signal as nonlinear and nonstationary, it is past in signal acquisition process
Toward the interference for adulterating the signals such as power frequency error, low frequency movement artifact, other biological electrical interference and ambient noise, seriously affect
The accuracy of signal characteristic identification.In addition, different gesture motions is controlled by the different muscle of arm, it is past for complicated gesture
Toward the electromyography signal for needing to acquire the generation of muti-piece muscle.Therefore the correlation between research multichannel electromyography signal is to accurately identifying hand
Gesture plays an important role.Presently relevant research is realized using the Feature fusion of multichannel electromyography signal, is being identified
The correlation between multi channel signals is had ignored in journey, affects the precision of signal identification.
The development of deep learning is so that become possible in myoelectricity gesture identification for Application of Neural Network.And use convolution pair
Electromyography signal carries out feature extraction and preferably resolves the problem of conventional method loses useful information when extracting feature, also can be right
Correlative character is merged between multichannel.
In view of electromyography signal and a kind of time series, convolutional neural networks not can solve what it was relied on for a long time
Problem.Recognition with Recurrent Neural Network is good at processing sequence signal, and the historical information of sequence can be stored in multitiered network by it.But
For longer hand signal, Recognition with Recurrent Neural Network poor effect, it is difficult to which the signal before connection for a long time carries out gesture identification.Cause
This, needs a kind of method that can accurately identify longer hand signal.
Summary of the invention
The present invention provides a kind of base aiming at the problem that current method cannot accurately identify longer hand signal
In the gesture identification method of multichannel electromyography signal correlation, predefined gesture can be detected and be identified.
Gesture identification method based on multichannel electromyography signal correlation of the invention mainly includes using myoelectric sensor
Following steps:
Step 1: electromyography signal denoising, use experience mode decomposition EMD is by collected signal decomposition at several
Mode function is levied, noise and useful signal are distinguished by the statistical property of signal and noise autocorrelation function, then rebuild
Electromyography signal after being denoised.
Step 2: active segment segmentation detects active segment according to signal amplitude power and is further processed.
Step 3: structuring processing being carried out to movable segment signal, by being superimposed several continuous time windows signal processing
At the format with temporal correlation, recognition accuracy is further improved.
Step 4: establish the classifier of gesture identification, the input of classifier is the signal of structuring processing, classifier it is defeated
It is out gesture classification result；Gesture identification is carried out using trained classifier.
In the step 1, empirical mode decomposition is carried out to the electromyography signal of acquisition, obtains n intrinsic mode letters if decomposing
Number IMF1(t)~IMFn(t), t indicates the time；If the auto-correlation function to the building of n intrinsic mode function is respectively τ is the time difference of different moments；It is as follows that threshold value T is set:
Wherein, σ () is to seek standard deviation；
Standard deviation is asked to the auto-correlation function of each rank IMF component
The standard deviation of auto-correlation function is less than to the IMF of threshold value T1(t)~IMFk(t), as the mode of noise dominant point
Amount, and noise reduction process is carried out, if obtaining IMF1' (t)~IMFk′(t)；Threshold value T's is more than or equal to the standard deviation of auto-correlation function
Intrinsic mode function is without processing.
In the step 2, the mean value of the electromyography signal after solving denoising believes the myoelectricity of each moment t after denoising
Number, think that the signal is movable segment signal if its absolute value is greater than mean value, otherwise it is assumed that being non-movable segment signal.
In the step 3, the structuring processing is: (1) on the basis of selecting a time window, next continuous
C time window be added to current base window, (2) as the time is slided to the right, repeat time window (1)；C is
Integer greater than 1.Specifically, present invention preferably uses the time windows that continuous 5 sizes are 15 sampled points to be overlapped, will
Input of the superimposed signal as classifier.
In the step 4, the classifier is established using hybrid neural networks CRNet, hybrid neural networks CRNet's
Structure includes two convolutional layers, and first convolutional layer is used to excavate the characteristic spectrum of institute's input signal, and second convolution is for subtracting
The quantity of few characteristic spectrum；It is a maximum pond layer after first convolutional layer, is one after second convolutional layer
Average pond layer, then connects a LSTM network；The feature vector that average pond layer exports is converted into one-dimensional vector, is inputted
LSTM network.
Compared with prior art, the present invention having the advantage that
(1) present invention only uses several myoelectric sensor acquisition original signals, does not need additional complex device, operation side
Just and good environmental adaptability.
It (2) is a kind of adaptive noise-reduction method the present invention is based on the empirical mode decomposition Method of Noise of auto-correlation function, no
Specified cutoff frequency is needed, decomposition that can be adaptive obtains component and the further progress denoising of noise dominant, can
Effectively remove the noise in signal.
(3) by the present invention in that carrying out the time phase that integrated signal sequence is capable of in structuring processing to signal with time window
Closing property information, improves accuracy of identification.The present invention has by being superimposed several continuous time windows and electromyography signal being processed into
The format of temporal correlation, further improves recognition accuracy.Experiments have shown that the use of continuous 5 sizes being 15 sampled points
Time window is overlapped, input of the superimposed signal as classifier, and effect is more preferable.
(4) the CNN+RNN neural network that the present invention uses tensorflow to build is as classifier, and trains a variety of classification
Device model compares, after the parameter measures such as the Average Accuracy, model complexity and convergence rate of each model
Further model is screened, finally finds that hybrid neural networks CRNet of the invention does not need to calculate complicated signal spy
Sign, reduces computing resource and improves recognition efficiency, be more suitable for engineer application.
Detailed description of the invention
Fig. 1 is the flow diagram of the gesture identification method of the invention based on multichannel electromyography signal correlation；
Fig. 2 is the intrinsic mode function IMF schematic diagram that electromyography signal carries out after EMD in the present invention；
Fig. 3 is the auto-correlation function schematic diagram of each intrinsic mode function in the present invention；
Fig. 4 is that active segment divides schematic diagram in the present invention；
Fig. 5 is that influence comparison diagram of the window parameter to recognition accuracy is superimposed in the present invention；
Fig. 6 is the schematic diagram of 10 kinds of digital gesture motions in the present invention；
Fig. 7 is 6 kinds of grasping movement schematic diagrames in the present invention.
Fig. 8 is the result schematic diagram for passing through parameter measure classifier in the present invention.
Specific embodiment
The present invention is understood and implemented for the ease of those of ordinary skill in the art, and the present invention is made into one with reference to the accompanying drawing
The detailed and deep description of step.
A kind of gesture identification method based on multichannel electromyography signal correlation proposed by the present invention is mainly used for predetermined
The gesture of justice is detected and is identified.The present invention uses LSTM (Long Short-Term Memory, shot and long term memory network)
Unit explains the non-linear relation between electromyography signal and gesture, and experiment shows that the model can accurately identify hand exercise
Posture.This method specifically includes that (1) electromyography signal denoising: use experience mode decomposition is by collected electromyography signal
Several intrinsic mode functions are resolved into, noise and useful signal are distinguished by the statistical property of signal and noise autocorrelation function；
(2) active segment is divided: detecting active segment according to signal amplitude power and is further processed；(3) signal structureization is handled: by folded
Add several continuous time windows that signal processing at the format with temporal correlation, is further improved recognition accuracy；
(4) gesture classification identifies: by the comparison to different classifier algorithms, finally selecting a kind of hybrid production style opponent
Gesture signal is classified and is identified.The present invention can be decomposed and be denoised to the electromyography signal of nonlinear and nonstationary, can be in noise
Good denoising effect is obtained in the case where relatively low；The present invention can recognize that different Pre-defined gestures and have higher standard
True rate.
As shown in Figure 1, the embodiment of the present invention acquires electromyography signal, myoelectric sensor built in armlet using armlet (Myo)
And sensor signal transmission is carried out by bluetooth.The embodiment of the present invention receives and processes the electromyography signal of acquisition with mobile phone, complete
At gesture identification.The myoelectric sensor of armlet (Myo) containing 8 channels in the embodiment of the present invention, user wear when in use
Good armlet, does not need that special consideration should be given to wearing positions.Intelligent end executes following four step to the original electromyography signal of acquisition.
Step 1 carries out denoising to collected electromyography signal.
During actual electromyographic signal collection, electromyography signal is very faint and makes an uproar containing sampler is intrinsic
The noise etc. that sound, movement generate, therefore be difficult to find the frequency range of noise in the lower situation of signal-to-noise ratio.The invention proposes one
Empirical mode decomposition noise-reduction method (XB-EMD) of the kind based on auto-correlation function carries out denoising to electromyography signal.Traditional warp
It tests mode decomposition and obtains several modal components later, since low order components contain much noise, typically directly abandon.But this side
Method cannot find the boundary for distinguishing height order component well, often will cause the loss of useful signal.Therefore, the present invention utilizes
The random noise feature different with useful signal correlation, should using the method judgement of auto-correlation function to each modal components
Modal components whether be noise dominant component, noise is further removed for the component of this kind of noisy sound pitch.XB- of the invention
EMD method does not need specified cutoff frequency, can be effectively solved Denoising Problems.
If collected original electromyography signal is X (t), t indicates the acquisition time of electromyography signal, first believes original myoelectricity
Number carry out empirical mode decomposition, obtain several intrinsic mode function IMF, obtained IMF1~IMF8 is as shown in Figure 2.The present invention is real
Applying myoelectric sensor frequency acquisition in example is 50HZ, and the time scale of the abscissa in Fig. 2 is 1/50HZ=20ms；Ordinate table
Show the amplitude of signal.The characteristics of according to noise signal and general signal, the intrinsic mode IMF component of low order includes high-frequency noise, high
Rank IMF component includes low frequency useful signal.It is mainly found using XB-EMD denoising and distinguishes noise IMF component and signal IMF points
The order k of amount, so that it is based on signal that the IMF component before k rank, which is IMF component based on noise, after k rank,.Therefore,
Original electromyography signal may be expressed as:
Wherein, n indicates that electromyography signal carries out the number of the IMF component generated after EMD, IMFi(t) i-th of IMF points is indicated
Amount, r (t) is residual signal components.
In order to find the order k for distinguishing the IMF component that noise and signal are dominated, the present invention constructs each rank IMF component
Auto-correlation function.It is defined according to auto-correlation function, for the i-th order component IMFi, the auto-correlation function of different moments t, s is
Wherein, E indicates expectation, and μ indicates intrinsic mode function IMFiMean value, σ2Indicate intrinsic mode function IMFiSide
Difference, IMFi(t)、IMFi(s) the i-th rank modal components IMF is respectively indicatediIn t moment, the value at s moment, τ indicates t moment and s moment
Time difference, τ=t-s.
It is found that auto-correlation function obtains maximum value at τ=0, due between signal for the IMF component that signal is dominated
Correlation, the auto-correlation function of elsewhere signal is to change over time and change there is no the cracking value for decaying to very little.
And the random noise correlation of different moments is weak, therefore the auto-correlation function of the IMF component of noise dominant obtains most at τ=0
Big value, can decay to rapidly the value of very little in its elsewhere auto-correlation function value.As shown in figure 3, low order IMF component is by noise master
It leads, high-order IMF component is dominated by signal.In Fig. 3, abscissa indicates the time difference；The phase of ordinate expression different time difference signal
Pass degree, has carried out normalized here, and range is that -1~1,0 expression is uncorrelated.
In order to find the order k for distinguishing noise and signal dominant component IMF, the present invention is used as lower threshold value T differentiates:
Wherein, standard deviation is asked to the auto-correlation function of each rank IMF componentIf standard deviation is less than threshold value T
It is considered noisy high low order IMF component；Otherwise it is considered noisy low high-order IMF component, thus judgement boundary IMF component
IMFk。
It is more than or equal to the higher order eigenmode state component IMF of T to standard deviationk+1(t)~IMFn(t) it is not processed.To noisy height
Component IMF1(t)~IMFk(t), further denoising is carried out using Threshold Filter Algorithms, obtains IMF1' (t)~IMFk′
(t).Finally reconstruct former electromyography signal, the electromyography signal X ' (t) after output denoising.In the embodiment of the present invention, to component IMF1(t)
~IMFk(t) it is denoised using Wavelet noise-eliminating method, the wavelet basis used is sym8.
Step 2, to the electromyography signal after denoising, carry out active segment segmentation.
After XB-EMD denoising, need to carry out active segment detection to the electromyography signal X ' (t) after denoising.Regard
Out when gesture, the amplitude when amplitude of the electromyography signal sequence after denoising does not make gesture has apparent fluctuation, according to this
Electromyography signal active segment after Characteristics Detection to denoising, i.e., divided according to the difference of active segment and inactive segment signal amplitude
It cuts.
Solve denoising after electromyography signal X ' (t) mean value and as threshold value, for the myoelectricity after the denoising of each moment t
Signal thinks that the electromyography signal after the denoising is movable segment signal if its absolute value is greater than threshold value；Otherwise it is assumed that being inactive
Segment signal.As shown in figure 4, black region is the active segment arrived according to threshold test in figure, that is, think the corresponding signal in the region
Value is movable segment signal.
In the embodiment of the present invention, the myoelectricity transducing signal in 8 channels is acquired by Myo armlet, to the myoelectricity in each channel
Signal carries out above-mentioned step 1 processing, when step 2 carries out active segment segmentation, as long as detecting the signal amplitude in a channel
Greater than threshold value, as movable segment signal, then it is assumed that the signal in 8 channels at this moment is movable segment signal, so later
The signal length for handling and being input to each channel in network is consistent.
Step 3, using time window, structuring processing is carried out to electromyography signal.
Because electromyography signal has temporal correlation, for long-term electromyography signal, in order in RNN
(Recognition with Recurrent Neural Network) partially relies on the signal of different windows for a long time, therefore to carry out structuring processing to window.For
Meet the input format of neural network classifier needs and electromyography signal is processed into the lattice with surrounding time correlation
Formula, it is therefore desirable to which structuring processing is carried out to the electromyography signal of the active segment after segmentation.
Time window, that is, sliding window method, the present invention first pass through the letter that sliding window method obtains several continuous windows
Number value, on the basis of selecting a time window, is added to following continuously several time windows current base window, most
Afterwards, time window as the time is slided to the right；Above step is repeated until all signal datas are all handled by structural.It will
It is expanding data dimension that window, which is overlapped purpose, is processed into temporal correlation format, to facilitate subsequent neural network to do into one
Step processing.
In structuring processing, the size of window and the quantity for being superimposed window can all influence the accuracy rate of final classification, because
This present invention is determined through experimentation different windows length and is superimposed the quantity of window.As shown in figure 5, being different superposition window numbers
Amount changes the variation relation of brought average recognition accuracy with length of window.The corresponding number of difference curve is folded in Fig. 5
The quantity of the time window added, as shown in Figure 5, hybrid neural networks CRNET network model of the invention can be long in time window
When spending 15 samples, reach preferable accuracy rate using 5 superposition windows.Hybrid neural networks CRNET network model uses deep
Degree learning framework tensorflow is built, and is the network structure mixed using CNN+RNN, is illustrated in step 4 below.
Step 4, gesture classification identification, the present invention are classified and are known to hand signal using hybrid production style
Not.
The present invention establishes the classifier of gesture identification using a kind of hybrid neural networks CRNet, the letter that structuring is handled
Number, i.e., superimposed signal, the input as classifier.
Machine learning algorithm, deep learning algorithm can be used to establish in gesture identification classifier, and wherein machine learning is calculated
Method, e.g. SVM (support vector machines), KNN (K arest neighbors) and RF (random forest).The present invention propose using mixing CNN and
The neural network CRNet of RNN carries out sorter model training.By comparing Average Accuracy, mould of the classifier on test set
Type complexity and convergence rate, the hybrid neural networks CRNet sorter model for finally selecting effect best.
Hybrid neural networks CRNet of the invention is based on convolutional neural networks CNN and Recognition with Recurrent Neural Network RNN and realizes, it
The purpose of basic structure using two stacking convolutional layers as CNN, first layer convolutional layer is under excavation particular dimensions across channel
Relevant information.By repeating the available several characteristic spectrums of convolution operation, wherein containing the correlation in different channels
Feature.It is a maximum pond layer after first convolutional layer, for polymerizeing relevant information on adjacent scale.At first
After the stage of pond, characteristic spectrum size can become smaller in the case where depth remains unchanged, and reduce the quantity of parameter.Second
The effect of convolution is to reduce the quantity of characteristic spectrum to reduce network model complexity.Therefore, and available several represent
The characteristic spectrum of fuse information within the scope of different scale.It is similar with first convolutional layer, one is added to after this convolutional layer
A average pond layer is used for information fusion.Before being connected to LSTM network structure, the multidimensional characteristic vectors for exporting CNN are needed
Be converted to one-dimensional vector.The addition of Recognition with Recurrent Neural Network RNN is so that neural network has the ability of learning time sequence.RNN's
Feature is its recursive structure, it is allowed to carry out the modeling of contextual information from the sequence of Length discrepancy.The present invention this
The RNN at place uses LSTM.Movement longer for active signal is predicted then extremely important using this structure.
LSTM network structure in the model that the present invention uses is used to learn letter from the characteristic spectrum that CNN before is extracted
The contextual information of number sequence.Before inputting LSTM network structure, signal sequence is organized into the format of one-dimensional vector, immediately
Be input to in fully-connected network and single layer LSTM network, finally export the probability to each gesture classification.It is connecting entirely
Layer, the generalization ability of network is promoted using multiple neurons, and full articulamentum can integrate all channels letter of convolutional layer extraction
Number feature.LSTM layers of main function is the information that can acquire continuous several time windows, therefore the present invention both can benefit
The window of longer-term can also be relied on actual time window, the present invention carries out the training of network using 3 layers of LSTM.
In the embodiment of the present invention, the input of hybrid neural networks CRNet is the active segment of the structuring processing in all channels
Signal.In the training stage, the electromyography signals of different gestures is acquired as training dataset, after the processing of previous step 1~3,
The data that structuring is handled are input in neural network CRNet, the parameter in neural network is trained, until obtaining most
Excellent parameter, training are completed, and trained classifier is obtained.In detection-phase, electromyography signal is obtained by mobile phone, by walking above
After rapid 1~3 processing, the data that structuring is handled are input to trained classifier, classification results can be obtained.
In order to sufficiently verify the performance of the method for the present invention, 10 kinds of typical digital gestures and 6 kinds of daily crawls are had chosen
Gesture is tested, as shown in Figure 6 and Figure 7 respectively.Data are acquired to the movement of above-mentioned two class respectively, are divided into data set 1 and data
Collection 2, and the classifier realized with other several sorting algorithms compares, and recognition accuracy is as shown in table 1:
Accuracy rate of the 1 different classifications device of table on two class data sets
| Data set | SVM | KNN | RF | CRNet |
| Data set 1 | 46% | 53% | 83% | 97.4% |
| Data set 2 | 40% | 45% | 70% | 72.55% |
As shown in table 1, the accuracy rate of Machine learning classifiers model is all lower than CRNet, therefore the present invention uses CRNet
Sorter model.Next the parameter measure network model is used.
The present invention also provides a comparison of the network model of four kinds of complexities: 1) CNN is used only；2) LSTM is used only；3) lightweight
CRNet, using less neuronal quantity, parameter scale is smaller；4) CRNet model.Model complexity and convergence rate
Comparison is as shown in Figure 8, it is seen that CRNet model accuracy rate of the invention is higher, and cracking can restrain.
Claims (5)
1. a kind of gesture identification method based on multichannel electromyography signal correlation, passes through the myoelectric sensor being arranged on arm
Acquire electromyography signal, which is characterized in that the electromyography signal of acquisition is handled as follows in the gesture identification method:
Step 1: empirical mode decomposition being carried out to the electromyography signal of acquisition, constructs the auto-correlation letter of each rank intrinsic mode function
Number, divides noise and useful signal according to threshold value, carries out noise reduction process to noise signal, then rebuilds after being denoised
Electromyography signal；
If decomposition obtains n intrinsic mode function IMF1(t)~IMFn(t), t indicates the time；If to n intrinsic mode function structure
The auto-correlation function built is respectivelyτ is the time difference of different moments；It is as follows that threshold value T is set:
Wherein, σ () is to seek standard deviation；
Standard deviation is asked to the auto-correlation function of each rank IMF component
The standard deviation of auto-correlation function is less than to the IMF of threshold value T1(t)~IMFk(t), as the modal components of noise dominant, and
Noise reduction process is carried out, if obtaining IMF1' (t)~IMFk′(t)；It is intrinsic more than or equal to threshold value T to the standard deviation of auto-correlation function
Mode function is without processing；
Step 2: the mean value of the electromyography signal after solving denoising carries out active segment to the electromyography signal after denoising and inactive section is drawn
Point；If the absolute value of the electromyography signal at each moment after denoising is greater than mean value, which is movable segment signal, otherwise right and wrong
Movable segment signal；
Step 3: movable segment signal being divided using time window, then carries out structuring processing；
The structuring processing is: (1) on the basis of selecting a time window, following continuous c time window being folded
It is added to current base window, (2) as the time is slided to the right, repeat time window (1)；C is the integer greater than 1；
Step 4: establishing the classifier of gesture identification, the input of classifier is the signal of structuring processing, and the output of classifier is
The probability of gesture classification chooses the gesture classification classification results the most of wherein maximum probability；
The classifier is established using hybrid neural networks CRNet, and the structure of hybrid neural networks CRNet includes two convolution
Layer is a maximum pond layer after first convolutional layer, is an average pond layer after second convolutional layer, then
Connect a LSTM network；First convolutional layer is used to excavate the characteristic spectrum of institute's input signal, and second convolution is for reducing
The quantity of characteristic spectrum；The multidimensional characteristic vectors that average pond layer exports are converted into one-dimensional vector input LSTM network.
2. the method according to claim 1, wherein the use of continuous 5 sizes being 15 in the step 3
The time window of sampled point is overlapped.
3. the method according to claim 1, wherein being dropped in the step 1 using Wavelet noise-eliminating method
It makes an uproar processing.
4. method according to claim 1 or 2, which is characterized in that in the step 2, to the multichannel myoelectricity of acquisition
Signal, as long as the electromyography signal in one of channel is movable segment signal, the electromyography signal in other channels is also active segment letter
Number.
5. method according to claim 1 or 2, which is characterized in that in the step 4, hybrid neural networks CRNet's
Input is the movable segment signal of the structuring processing in all channels；In the training stage, the electromyography signal conduct of different gestures is acquired
Training dataset is trained hybrid neural networks CRNet, to training data after the processing of previous step 1~3, will tie
The data of structureization processing are input in neural network CRNet, are trained to the parameter in neural network, until obtaining optimal ginseng
Number, training are completed, and trained classifier is obtained；In detection-phase, electromyography signal is obtained by mobile phone, by previous step 1
After~3 processing, the data that structuring is handled are input to trained classifier, obtain classification results.
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910595783 | 2019-07-03 | ||
| CN2019105957835 | 2019-07-03 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110399846A true CN110399846A (en) | 2019-11-01 |
Family
ID=68326409
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910689483.3A Pending CN110399846A (en) | 2019-07-03 | 2019-07-29 | A kind of gesture identification method based on multichannel electromyography signal correlation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110399846A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111103976A (en) * | 2019-12-05 | 2020-05-05 | 深圳职业技术学院 | Gesture recognition method and device and electronic equipment |
| CN111184512A (en) * | 2019-12-30 | 2020-05-22 | 电子科技大学 | Method for recognizing rehabilitation training actions of upper limbs and hands of stroke patient |
| CN111209885A (en) * | 2020-01-13 | 2020-05-29 | 腾讯科技（深圳）有限公司 | Gesture information processing method and device, electronic equipment and storage medium |
| CN111476295A (en) * | 2020-04-07 | 2020-07-31 | 金陵科技学院 | GS-L SSVM-based multichannel physiological signal somatosensory gesture recognition method |
| CN111898526A (en) * | 2020-07-29 | 2020-11-06 | 南京邮电大学 | Myoelectric gesture recognition method based on multi-stream convolution neural network |
| CN112932508A (en) * | 2021-01-29 | 2021-06-11 | 电子科技大学 | Finger activity recognition system based on arm electromyography network |
| CN113688802A (en) * | 2021-10-22 | 2021-11-23 | 季华实验室 | Gesture recognition method, device and equipment based on electromyographic signals and storage medium |
| CN114081513A (en) * | 2021-12-13 | 2022-02-25 | 苏州大学 | Electromyographic signal-based abnormal driving behavior detection method and system |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105326500A (en) * | 2014-08-13 | 2016-02-17 | 华为技术有限公司 | Action recognition method and equipment based on surface electromyogram signals |
| US20170220923A1 (en) * | 2016-02-02 | 2017-08-03 | Samsung Electronics Co., Ltd. | Gesture classification apparatus and method using emg signal |
| CN108388348A (en) * | 2018-03-19 | 2018-08-10 | 浙江大学 | A kind of electromyography signal gesture identification method based on deep learning and attention mechanism |
| CN108491355A (en) * | 2018-02-05 | 2018-09-04 | 南京邮电大学 | A kind of ultrasonic signal noise-reduction method based on CEEMD and wavelet packet |
| CN108776783A (en) * | 2018-05-31 | 2018-11-09 | 陕西师范大学 | The gesture electromyography signal recognition methods of extreme learning machine-Hidden Markov Model |
| CN109145729A (en) * | 2018-07-13 | 2019-01-04 | 杭州电子科技大学 | Based on the electromyography signal denoising method for improving wavelet threshold and EEMD |
| WO2019118309A1 (en) * | 2017-12-11 | 2019-06-20 | Adaptive, Intelligent And Dynamic Brain Corporation (Aidbrain) | Method, module and system for analysis of physiological signal |
-
2019
- 2019-07-29 CN CN201910689483.3A patent/CN110399846A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105326500A (en) * | 2014-08-13 | 2016-02-17 | 华为技术有限公司 | Action recognition method and equipment based on surface electromyogram signals |
| US20170220923A1 (en) * | 2016-02-02 | 2017-08-03 | Samsung Electronics Co., Ltd. | Gesture classification apparatus and method using emg signal |
| WO2019118309A1 (en) * | 2017-12-11 | 2019-06-20 | Adaptive, Intelligent And Dynamic Brain Corporation (Aidbrain) | Method, module and system for analysis of physiological signal |
| CN108491355A (en) * | 2018-02-05 | 2018-09-04 | 南京邮电大学 | A kind of ultrasonic signal noise-reduction method based on CEEMD and wavelet packet |
| CN108388348A (en) * | 2018-03-19 | 2018-08-10 | 浙江大学 | A kind of electromyography signal gesture identification method based on deep learning and attention mechanism |
| CN108776783A (en) * | 2018-05-31 | 2018-11-09 | 陕西师范大学 | The gesture electromyography signal recognition methods of extreme learning machine-Hidden Markov Model |
| CN109145729A (en) * | 2018-07-13 | 2019-01-04 | 杭州电子科技大学 | Based on the electromyography signal denoising method for improving wavelet threshold and EEMD |
Non-Patent Citations (4)
| Title |
|---|
| NIHAL FATMA GLER 等: "Classification of emg signals using pca and fft", 《JOURNAL OF MEDICAL SYSTEMS》 * |
| YU HU等: "A novel attention-based hybrid CNN-RNN architecture for sEMG-based gesture recognition", 《PLOS ONE》 * |
| 俞毕洪: "基于表面肌电信号的穿戴式人体行为识别系统", 《中国优秀硕士学位论文全文数据库 (信息科技辑)》 * |
| 杨亚慧等: "基于卷积神经网络的表面肌电信号手势识别", 《微型机与应用》 * |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111103976B (en) * | 2019-12-05 | 2023-05-02 | 深圳职业技术学院 | Gesture recognition method and device and electronic equipment |
| CN111103976A (en) * | 2019-12-05 | 2020-05-05 | 深圳职业技术学院 | Gesture recognition method and device and electronic equipment |
| CN111184512A (en) * | 2019-12-30 | 2020-05-22 | 电子科技大学 | Method for recognizing rehabilitation training actions of upper limbs and hands of stroke patient |
| CN111184512B (en) * | 2019-12-30 | 2021-06-01 | 电子科技大学 | Method for recognizing rehabilitation training actions of upper limbs and hands of stroke patient |
| CN111209885A (en) * | 2020-01-13 | 2020-05-29 | 腾讯科技（深圳）有限公司 | Gesture information processing method and device, electronic equipment and storage medium |
| WO2021143353A1 (en) * | 2020-01-13 | 2021-07-22 | 腾讯科技（深圳）有限公司 | Gesture information processing method and apparatus, electronic device, and storage medium |
| US11755121B2 (en) | 2020-01-13 | 2023-09-12 | Tencent Technology (Shenzhen) Company Limited | Gesture information processing method and apparatus, electronic device, and storage medium |
| CN111209885B (en) * | 2020-01-13 | 2023-05-30 | 腾讯科技（深圳）有限公司 | Gesture information processing method and device, electronic equipment and storage medium |
| CN111476295A (en) * | 2020-04-07 | 2020-07-31 | 金陵科技学院 | GS-L SSVM-based multichannel physiological signal somatosensory gesture recognition method |
| CN111898526A (en) * | 2020-07-29 | 2020-11-06 | 南京邮电大学 | Myoelectric gesture recognition method based on multi-stream convolution neural network |
| CN111898526B (en) * | 2020-07-29 | 2022-07-22 | 南京邮电大学 | Myoelectric gesture recognition method based on multi-stream convolution neural network |
| CN112932508A (en) * | 2021-01-29 | 2021-06-11 | 电子科技大学 | Finger activity recognition system based on arm electromyography network |
| CN113688802A (en) * | 2021-10-22 | 2021-11-23 | 季华实验室 | Gesture recognition method, device and equipment based on electromyographic signals and storage medium |
| CN114081513A (en) * | 2021-12-13 | 2022-02-25 | 苏州大学 | Electromyographic signal-based abnormal driving behavior detection method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108491077B (en) | Surface electromyographic signal gesture recognition method based on multi-stream divide-and-conquer convolutional neural network | |
| CN110399846A (en) | A kind of gesture identification method based on multichannel electromyography signal correlation | |
| CN109409222B (en) | Multi-view facial expression recognition method based on mobile terminal | |
| CN111062250B (en) | Multi-subject motor imagery electroencephalogram signal identification method based on deep feature learning | |
| CN109726751B (en) | Method for recognizing electroencephalogram based on deep convolutional neural network | |
| Shen et al. | Movements classification of multi-channel sEMG based on CNN and stacking ensemble learning | |
| CN105242779B (en) | A kind of method and mobile intelligent terminal of identification user action | |
| CN109598222B (en) | EEMD data enhancement-based wavelet neural network motor imagery electroencephalogram classification method | |
| CN113505822B (en) | Multi-scale information fusion upper limb action classification method based on surface electromyographic signals | |
| CN110658915A (en) | Electromyographic signal gesture recognition method based on double-current network | |
| CN110333783B (en) | Irrelevant gesture processing method and system for robust electromyography control | |
| CN105841961A (en) | Bearing fault diagnosis method based on Morlet wavelet transformation and convolutional neural network | |
| CN111832416A (en) | Motor imagery electroencephalogram signal identification method based on enhanced convolutional neural network | |
| CN113128552B (en) | Electroencephalogram emotion recognition method based on depth separable causal graph convolution network | |
| CN112732092B (en) | Surface electromyogram signal identification method based on double-view multi-scale convolution neural network | |
| CN104751186A (en) | Iris image quality classification method based on BP (back propagation) network and wavelet transformation | |
| CN109730818A (en) | A kind of prosthetic hand control method based on deep learning | |
| CN108567418A (en) | A kind of pulse signal inferior health detection method and detecting system based on PCANet | |
| CN109685071A (en) | Brain electricity classification method based on the study of common space pattern feature width | |
| Jinliang et al. | EEG emotion recognition based on granger causality and capsnet neural network | |
| CN109009098B (en) | Electroencephalogram signal feature identification method under motor imagery state | |
| CN112784892A (en) | Electroencephalogram movement intention identification method and system | |
| CN113208593A (en) | Multi-modal physiological signal emotion classification method based on correlation dynamic fusion | |
| CN115273236A (en) | Multi-mode human gait emotion recognition method | |
| CN113627391A (en) | Cross-mode electroencephalogram signal identification method considering individual difference |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191101 |
|
| RJ01 | Rejection of invention patent application after publication |