阅读说明：本技术 一种基于表面肌电信号的手势动作在线识别方法及系统 (gesture action online recognition method and system based on surface electromyographic signals ) 是由 刘聪 周淑旺 费炜 胡胜 于 2019-09-11 设计创作，主要内容包括：本发明属于人工智能技术领域,公开了一种基于表面肌电信号的手势动作在线识别方法及系统,实时采集肱三头肌、肘肌、肱二头肌、肱桡肌四个通道的表面肌电信号；然后进行小波阈值去噪处理；提取处理后信号的均方根(RMS)特征和前4阶AR模型参数特征,最后用SVM投票法和聚类思想联合判别实时动作。每间隔0.05s识别一次。每种动作的训练样本个数只采集10个,样本数和特征种类数少,方便实时识别；利用SVM投票法和聚类思想联合判别,可以提高识别正确率；聚类思想的应用在一定程度上拒绝了对异常动作的识别。(The invention belongs to the technical field of artificial intelligence, and discloses a gesture action online recognition method and system based on surface electromyographic signals, which are used for collecting the surface electromyographic signals of four channels of brachial triceps, elbow, biceps brachii and brachial radius in real time; then, carrying out wavelet threshold denoising treatment; and extracting Root Mean Square (RMS) characteristics of the processed signals and parameters of the first 4-order AR model, and finally, jointly judging real-time actions by using an SVM voting method and a clustering idea. The identification is carried out every 0.05 s. The number of training samples of each action is only 10, the number of samples and the number of characteristic types are few, and real-time identification is convenient; the SVM voting method and the clustering idea are combined for discrimination, so that the recognition accuracy can be improved; the application of the clustering idea to a certain extent rejects the identification of abnormal actions.)

1. a gesture action online recognition method based on a surface electromyogram signal is characterized by comprising the following steps:

Collecting surface electromyographic signals of a plurality of channels and carrying out wavelet threshold denoising treatment;

Extracting the parameter characteristics and RMS characteristics of the front 4-order AR model of the processed surface electromyographic signals as the characteristics used for classifying and identifying each action;

And (4) jointly judging the real-time action by utilizing an SVM voting method and a clustering idea.

2. the method for recognizing gesture actions based on surface electromyography signals on line according to claim 1, comprising the steps of:

Selecting a plurality of static gesture actions to be recognized;

Acquiring training samples, acquiring 10 samples for each action, wherein the time period for acquiring each action is 1s, and extracting an original surface electromyographic signal by using hardware equipment; and carrying out wavelet threshold denoising treatment.

Extracting characteristics, namely extracting AR model parameter characteristics and RMS characteristics from surface electromyographic signals of four channels of the brachial triceps, the elbow, the biceps brachii and the brachial radius of each action sample; extracting the first 4-order parameters of the AR model parameter features;

And fourthly, recognizing the gesture in real time, designing the SVM between any two types of motion samples, designing k (k-1)/2 SVM for the k types of samples, and simultaneously calculating the maximum distance Dmax between the average characteristic vector of each type of sample and the characteristic vector of the type of sample.

3. The method for on-line recognition of gesture actions based on surface electromyography signals as claimed in claim 2, wherein in step three, the time series root mean square RMS characteristics calculation formula is:

In the formula: RMS is the root mean square myoelectric value; n represents the number of electromyographic signal samples; x (t) is a myoelectric signal value.

4. The method for on-line recognition of gesture actions based on surface electromyography signals as claimed in claim 2, wherein in step three, the AR model is accumulated for white noise input and past p signal values, specifically:

after the model coefficient is determined, determining a system model excited by white noise; and (3) iteratively solving each order coefficient of the AR model by using a Berger method, and determining the order p before solving each order coefficient of the AR model.

5. The method for on-line recognition of gesture actions based on surface electromyography signals as claimed in claim 4, wherein the model order p, the AR model is ordered by using the final prediction error FPE criterion, the final prediction error is:

In the formula, N is the length of the time sequence, p is the order of the AR model, and sigma ² is the variance of the model.

6. The method for on-line recognition of gesture actions based on surface electromyography signals according to claim 2, wherein the SVM in step four adopts a gaussian kernel function, and the expression is:

In the formula: x and z are two different eigenvectors respectively; sigma is a standard deviation parameter;

The distance calculation adopts an Euclidean distance formula, and the expression is as follows:

where d ₁₂ is the distance between two eigenvectors, n is the dimension of the eigenvector, and x _1k and x _2k are two different eigenvectors, respectively.

7. the method for on-line recognition of gesture actions based on surface electromyographic signals according to claim 1, wherein in the real-time actions are jointly discriminated by using an SVM voting method and a clustering idea, vectors corresponding to the actions are selected as a training set during action training to obtain corresponding SVM;

During real-time classification, the corresponding vectors are respectively used for testing results; and (5) voting is adopted, and a group of results are obtained finally.

8. the method for on-line recognition of gesture actions based on surface electromyography signals according to claim 1, wherein the action recognition by combining SVM voting and clustering specifically comprises:

and if the distance between the motion characteristic vector to be classified and the average characteristic vector of the class sample with the most votes is less than or equal to Dmax, judging that the motion to be classified is the class motion, and if not, abandoning the classification and identifying once every 0.05 s.

9. a gesture online recognition control system based on the surface electromyography signals, which implements the gesture action online recognition method based on the surface electromyography signals of any one of claims 1 to 8.

10. an online gesture recognition device based on surface electromyography signals, implementing the online gesture recognition method based on surface electromyography signals of any one of claims 1 to 8.

Technical Field

The invention belongs to the technical field of artificial intelligence, and particularly relates to a gesture action online recognition method and system based on surface electromyographic signals.

Background

Currently, the closest prior art:

with the development of science and technology, the research on gesture recognition technology has become a popular subject, and the application based on gesture recognition technology also starts to penetrate into the aspects of people's life, which is a sign that the technology goes toward being popular. The application field of the surface electromyographic signals is mainly in man-machine interaction and artificial limb control.

in the traditional gesture recognition based on the surface electromyogram signal, a structure similar to voice recognition is adopted, namely signal acquisition, active segment detection, feature extraction and classification. This approach requires detection and identification of the active segment.

For the recognition of gestures based on surface myoelectric signals, most of the existing methods can achieve high recognition accuracy, but the real-time control is inconvenient due to excessive use characteristics; the accuracy can be greatly reduced by singly using the time domain characteristics or the frequency domain characteristics; in addition, the complexity of the recognition algorithm also puts corresponding requirements on hardware, otherwise, the real-time performance is difficult to guarantee.

in addition, many recognition algorithms are not robust and only have good recognition effects on several fixed gestures.

At present, a gesture online recognition method which is strong in operability and easy to implement in algorithm is needed, and both accuracy and real-time performance can be considered.

In summary, the problems of the prior art are as follows:

(1) active segment detection is required before each recognition.

(1) the use characteristics are too many, so that real-time identification is inconvenient; using either the time domain features or the frequency domain features singularly can significantly reduce the accuracy.

(2) The complexity of the recognition algorithm also puts corresponding requirements on hardware, otherwise the real-time performance is difficult to guarantee.

(3) Many recognition algorithms are not robust and only have a good recognition effect on several fixed gestures.

The difficulty of solving the technical problems is as follows:

For a plurality of fixed gestures, although the accuracy can be improved by using too many features and algorithms, the real-time performance is difficult to guarantee, and in addition, the complexity of the recognition algorithm also puts corresponding requirements on hardware. The accuracy and the real-time performance are difficult to be considered at the same time.

In order to improve the robustness of the algorithm, the characteristics with strong universality and representativeness and the muscle positions of collected signals need to be determined, so that the algorithm can be suitable for more gestures.

The significance of solving the technical problems is as follows:

Firstly, the difficulty and complexity of hardware design can be reduced, so that the software and hardware of the whole system are easy to realize; secondly, under the condition of not reducing the accuracy too much, the system performance is better due to higher real-time performance; finally, the improvement of the robustness of the algorithm helps the application range to become wider.

disclosure of Invention

The invention provides a gesture action online recognition method and system based on surface electromyographic signals, and aims to solve the problems that the traditional gesture online recognition algorithm is not strong in algorithm robustness and cannot give consideration to real-time performance and accuracy.

The invention is realized in such a way that a gesture action on-line recognition method based on surface electromyographic signals comprises the following steps: surface electromyogram signals of four channels of brachial triceps, the elbow muscle, the biceps brachii and the brachial radius are collected.

And extracting the front 4-order AR model parameter characteristics and RMS characteristics of the surface electromyogram signals after noise reduction processing as the characteristics used for each action classification identification.

Judging real-time actions by combining an SVM voting method and a clustering idea; the application of the clustering idea to a certain extent rejects the identification of abnormal actions.

Further, the gesture action online recognition method based on the surface electromyogram signal comprises the following steps:

step one, selecting a plurality of actions to be identified.

And step two, collecting training samples. 10 samples are collected in each action, and original surface electromyographic signals are extracted by using hardware equipment; and carrying out wavelet threshold denoising treatment.

and step three, feature extraction and selection. For each action sample, surface electromyographic signals of four channels of the triceps brachii, the elbow muscle, the biceps brachii and the brachioradialis are extracted to obtain AR model parameter characteristics and RMS characteristics, wherein the AR model parameter characteristics are extracted to obtain the first 4-order parameters.

And step four, recognizing the gesture in real time. An SVM is designed between any two types of motion samples, k (k-1)/2 SVM needs to be designed for k types of samples, and meanwhile, the maximum distance Dmax between the average feature vector of each type of samples and the feature vector of the type of samples is calculated. .

Further, in the third step, the Root Mean Square (RMS) characteristic of the time series may be calculated by referring to the following formula:

Where RMS is a root mean square myoelectric value, N represents the number of myoelectric signal samples, and x (t) is a myoelectric signal value.

The AR model is defined as the sum of the white noise input and the past p signal values:

As can be seen from the above equation, when the model coefficients are determined, the system model excited by white noise is determined accordingly. And (3) iteratively solving coefficients of each order of the AR model by using a burger method (burg), and determining the order p of the constructed specific model before solving the coefficients of each order of the AR model.

Further, the AR model is scaled using a Final Prediction Error (FPE) criterion. The final prediction error is defined as:

Where N is the length of the time series, p is the AR model order, and σ ² is the model variance.

Further, the SVM in the fourth step adopts a gaussian kernel function, and the expression is:

In the formula, x and z are two different eigenvectors respectively, and sigma is a standard deviation parameter.

The distance calculation adopts an Euclidean distance formula, and the expression is as follows:

Where d ₁₂ is the distance between two feature vectors, n is the dimension of the feature vector, and x _1k and x _2k are two different feature vectors, respectively.

Further, in the fourth step, when the action is trained, the vector corresponding to the action is selected as a training set to obtain a corresponding SVM; during real-time classification, the corresponding vectors are respectively used for testing results; and (5) voting is adopted, and a group of results are obtained finally.

The action recognition is carried out by combining an SVM voting method and a clustering idea, and the specific operation comprises the following steps:

And if the distance between the motion characteristic vector to be classified and the average characteristic vector of the class sample with the most votes is less than or equal to Dmax, judging that the motion to be classified is the class motion, and if not, abandoning the classification and identifying once every 0.05 s.

the invention also aims to provide a surface electromyography signal-based gesture online recognition control system for implementing the surface electromyography signal-based gesture action online recognition method.

the invention also aims to provide a surface electromyography signal-based gesture action online recognition device for implementing the surface electromyography signal-based gesture action online recognition method.

in summary, the advantages and positive effects of the invention are:

According to the method, only the root mean square characteristic and the AR model characteristic of the electromyographic signal are extracted, only 10 samples are extracted from each category, the number of extracted samples and the number of characteristic types are small, and real-time identification is facilitated; the SVM algorithm is very suitable for the classification problem of small samples, and the identification accuracy can be improved by utilizing the SVM voting method and the clustering idea joint judgment, because the clustering idea is applied to reject the identification of abnormal actions to a certain extent; in addition, the characteristics of strong universality and representativeness and the muscle positions of collected signals enable the algorithm to be suitable for more gestures, the characteristics used in the patent are root mean square and four first-order AR model coefficient characteristics, and the determined muscle positions are brachial triceps, elbow, biceps, brachioradialis; finally, identification is performed every very short time, which avoids active segment detection.

Using the method described in this patent, 10 healthy adult human arm surface myoelectric signals were collected, and each person randomly performed actions, including but not limited to the following four actions: flexion of the elbow, raising of the arm, abduction and rest. And respectively counting the accuracy of 100 identifications of each person. The four specific actions, the recognition accuracy and the hardware equipment are shown in the attached drawings.

Drawings

Fig. 1 is a flowchart of a gesture action online recognition method based on a surface electromyogram signal according to an embodiment of the present invention.

fig. 2 is a schematic diagram of an SVM algorithm provided in an embodiment of the present invention.

Fig. 3 is a schematic diagram of four actions of elbow bending, arm raising, abduction and stillness involved in the experiment of the present invention.

FIG. 4 is a scatter plot of the accuracy of the experimental identification made in the present invention.

fig. 5 is a surface electromyogram signal acquisition apparatus used in experiments performed in the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

For the recognition of simple coarse gestures, due to excessive use characteristics, real-time control is inconvenient; using either the time domain features or the frequency domain features singularly can significantly reduce the accuracy. The complexity of the recognition algorithm also puts corresponding requirements on hardware, otherwise the real-time performance is difficult to guarantee. In the prior art, the accuracy and the real-time performance are difficult to be considered in the gesture online identification process, and the robustness of the algorithm is not good enough.

In view of the problems in the prior art, the present invention provides an online gesture recognition method based on surface electromyogram signals, and the following describes the present invention in detail with reference to the accompanying drawings.

the gesture action online recognition method based on the surface electromyographic signals, provided by the embodiment of the invention, comprises the steps of collecting surface electromyographic signals of four channels of brachial triceps, elbow muscle, biceps brachii and brachial radius muscle; extracting the front 4-order AR model parameter characteristics and RMS characteristics of the preprocessed surface electromyogram signals as the characteristics used by each action classification identification; and (4) judging the real-time action by combining an SVM voting method and a clustering idea. The application of the clustering idea to a certain extent rejects the identification of abnormal actions.

As shown in fig. 1, an online gesture recognition method based on surface electromyogram signals includes the following steps:

S101: several motions to be recognized are selected, such as the motions of bending the elbow, raising the arm, etc.

S102: training samples are collected. 10 samples are collected for each action, and a hardware device is used for extracting an original surface electromyogram signal and then carrying out signal noise reduction processing.

S103: and (5) extracting and selecting the features. For each action sample, surface electromyographic signals of four channels of the triceps brachii, the elbow muscle, the biceps brachii and the brachioradialis are extracted to obtain AR model parameter characteristics and RMS characteristics, wherein the AR model parameter characteristics are extracted to obtain the first 4-order parameters.

S104: and (4) recognizing the gesture in real time. An SVM is designed between any two types of motion samples, k (k-1)/2 SVM needs to be designed for k types of samples, and meanwhile, the maximum distance Dmax between the average feature vector of each type of samples and the feature vector of the type of samples is calculated.

In the embodiment of the present invention, in step S103, the Root Mean Square (RMS) characteristic of the time series may be calculated by referring to the following formula:

Where RMS is a root mean square myoelectric value, N represents the number of myoelectric signal samples, and x (t) is a myoelectric signal value.

The AR model is defined as the sum of the white noise input and the past p signal values:

The key to model building and identification is to determine its coefficients. As can be seen from the above equation, when the model coefficients are determined, the system model that is supposed to be excited by white noise is determined. Here, the coefficients of each order of the AR model are solved iteratively using a burger method (burg), and before solving the coefficients of each order of the AR model, the order p of the constructed specific model should be determined.

In an embodiment of the invention, the AR model is scaled using a Final Prediction Error (FPE) criterion. The final prediction error is defined as:

where N is the length of the time series, p is the AR model order, and σ ² is the model variance.

In the embodiment of the present invention, in step S104, the SVM uses a gaussian kernel function, and the expression is:

in the formula, x and z are two different eigenvectors respectively, and sigma is a standard deviation parameter.

The distance calculation adopts an Euclidean distance formula, and the expression is as follows:

where d ₁₂ is the distance between two feature vectors, n is the dimension of the feature vector, and x _1k and x _2k are two different feature vectors, respectively.

In the embodiment of the present invention, in step S104, during the training of the action, a vector corresponding to the action is selected as a training set to obtain a corresponding SVM; during real-time classification, the corresponding vectors are respectively used for testing results; and (5) voting is adopted, and a group of results are obtained finally.

in the embodiment of the present invention, in step S104, the action recognition is performed by using an SVM voting method and a clustering idea in a combined manner, and the specific operations include:

and if the distance between the motion characteristic vector to be classified and the average characteristic vector of the class sample with the most votes is less than or equal to Dmax, judging that the motion to be classified is the class motion, and if not, abandoning the classification and identifying once every 0.05 s.

The present invention will be further described with reference to the following examples.