CN113312994A

CN113312994A - Gesture classification recognition method and application thereof

Info

Publication number: CN113312994A
Application number: CN202110542132.7A
Authority: CN
Inventors: 郭伟钰; 杨永魁; 陈瑞; 陈超; 辛锦瀚; 王峥
Original assignee: Shenzhen Institute of Advanced Technology of CAS
Current assignee: Shenzhen Institute of Advanced Technology of CAS
Priority date: 2021-05-18
Filing date: 2021-05-18
Publication date: 2021-08-27
Also published as: WO2022242133A1

Abstract

The application belongs to the technical field of data classification, and particularly relates to a gesture classification recognition method and application thereof. However, some current gesture classification recognition algorithms related to sEMG signals are low in recognition accuracy, and overfitting and under-fitting, gradient disappearance, poor robustness and long training time also exist in the model training process. The application provides a gesture classification recognition method, which comprises the following steps: acquiring a surface electromyographic signal; extracting the characteristics of the surface electromyographic signals to obtain a gesture characteristic sequence and a gesture type; and inputting the gesture feature sequence and the gesture type into a neural network of a circulating gate circuit for training to obtain a classification model, and realizing gesture classification and recognition by adopting the classification model. The accuracy of prediction classification is improved.

Description

Gesture classification recognition method and application thereof

Technical Field

The application belongs to the technical field of data classification, and particularly relates to a gesture classification recognition method and application thereof.

Background

With the development of science and technology, new human-computer interaction methods are more and more concerned by researchers. Gestures are a natural and intuitive interaction means in life and are very important perception channels in man-machine interaction. In human-computer interaction, the gesture type capable of interacting with a computer is identified, and the method has very important significance for the current control research on the artificial limb.

Compared with a mode of realizing human-computer interaction through computer vision, the current research uses surface electromyography (sEMG) to effectively avoid the influence caused by physical factors such as illumination. sEMG signals are important bioelectric signals generated as a person's muscle moves, and changes in signals at the muscle surface are recorded by electrodes. Different gesture motions correspond to different muscle motions, the generated sEMG signals are different, and the sEMG signals have great application value in gesture motion classification. Gesture classification recognition based on sEMG signals has become one of research hotspots in the fields of prosthesis control and rehabilitation training, and the recognition of future gesture actions will also be widely applied in the fields of sports medicine, clinical muscle diagnosis and the like.

The sEMG signal is a time series signal, and in the problem of processing the series signal, a Recurrent Neural Network (RNN) is a good tool for processing the series problem by memorizing and learning sequence data of the signal through the RNN, compared with a machine learning network and a convolutional neural network. However, because the cycle of RNN is only a simple linear relationship, there is a long-term dependence problem, and in multiple iterations of the calculation process, the coefficient multiplication becomes smaller and smaller, which indirectly results in the loss of data information at a longer distance.

However, some current gesture classification recognition algorithms related to sEMG signals are low in recognition accuracy, and overfitting and under-fitting, gradient disappearance, poor robustness and long training time also exist in the model training process.

Disclosure of Invention

1. Technical problem to be solved

Based on the current gesture classification recognition algorithms about sEMG signals, the recognition accuracy is low, and the problems of over-fitting and under-fitting, gradient disappearance, poor robustness and long training time can also exist in the model training process.

2. Technical scheme

In order to achieve the above object, the present application provides a gesture classification recognition method, including the following steps: acquiring a surface electromyographic signal; extracting the characteristics of the surface electromyographic signals to obtain a gesture characteristic sequence and a gesture type; and inputting the gesture feature sequence and the gesture type into a neural network of a circulating gate circuit for training to obtain a classification model, and realizing gesture classification and recognition by adopting the classification model.

Another embodiment provided by the present application is: acquiring the surface electromyographic signals comprises wiping experimental equipment and the skin surface of a testee by adopting alcohol, placing electrodes for acquiring the electromyographic signals on the skin surface of the testee, and acquiring the muscle signal changes brought by the gesture actions of different arms in real time.

Another embodiment provided by the present application is: the number of the electrodes is 8, and the 8 electrodes are distributed on the forearm at equal intervals.

Another embodiment provided by the present application is: and the surface electromyographic signals are transmitted to an intelligent terminal through a wireless device for data processing.

Another embodiment provided by the present application is: and processing the surface electromyogram signal comprises converting the surface electromyogram signal into a digital signal and extracting a characteristic value of the digital signal.

Another embodiment provided by the present application is: and the surface electromyographic signals are input to the intelligent terminal by adopting a sliding windowing method.

Another embodiment provided by the present application is: the frequency of the collected surface electromyographic signals is 2kHZ, the width of the sliding window is 100ms, and the sliding step length is 0.5ms.

Another embodiment provided by the present application is: the characteristic value extraction is carried out by calculating a root mean square value.

Another embodiment provided by the present application is: the recurrent gate neural network includes a fully connected layer.

The application also provides a gesture classification recognition method which is applied to a human-computer interaction system.

3. Advantageous effects

Compared with the prior art, the gesture classification recognition method and the application thereof have the beneficial effects that:

the gesture classification and recognition method provided by the application is a gesture classification and recognition method of surface electromyography (sEMG) based on a GRU model.

According to the gesture classification and recognition method, the collected muscle electrode signals are subjected to data processing, and the processed gesture feature sequences and gesture types are input into a neural network (GRU) of a circulating Gate circuit for training to obtain a classification model, so that gesture classification and recognition are achieved. Compared with the prior machine learning classification algorithm, the method and the device have the advantages that the problems of over-fitting and under-fitting in the gradient calculation process are greatly solved, the robustness is improved, meanwhile, the calculation amount is greatly reduced, and the waste of resources is reduced.

The gesture classification recognition system provided by the application is a gesture classification recognition algorithm based on a Gate circuit neural network (GRU), and accuracy of prediction classification is improved.

According to the gesture classification recognition method, the GRU is an improved neural network of the recurrent neural network, the dependence problem of the RNN on data processing is effectively avoided, and the problems of gradient disappearance, gradient explosion and the like of data of a conventional long-sequence signal in a traditional machine learning training stage can be solved.

According to the gesture classification recognition method provided by the application, the structure of the GRU allows a network to capture dependent items from a large number of data sequences in an adaptive mode, and information of an early part of the sequence is not discarded. The GRU is similar to the long-short term memory neural network LSTM, but removes part of the state, and directly uses the hidden state for information transfer. Compared with a recurrent neural network, the GRU only comprises a reset gate and an update gate, and the operation amount is greatly simplified.

The gesture classification and recognition method is high in accuracy and strong in robustness.

Drawings

FIG. 1 is a schematic illustration of the experimental set-up of the present application;

FIG. 2 is a schematic flow chart of a gesture classification recognition method according to the present application;

FIG. 3 is a diagram illustrating comparison of the results of the gesture classification recognition method according to the present application.

Detailed Description

Hereinafter, specific embodiments of the present application will be described in detail with reference to the accompanying drawings, and it will be apparent to those skilled in the art from this detailed description that the present application can be practiced. Features from different embodiments may be combined to yield new embodiments, or certain features may be substituted for certain embodiments to yield yet further preferred embodiments, without departing from the principles of the present application.

Referring to fig. 1 to 3, the present application provides a gesture classification recognition method, including the following steps: acquiring a surface electromyographic signal; extracting the characteristics of the surface electromyographic signals to obtain a gesture characteristic sequence and a gesture type; and inputting the gesture feature sequence and the gesture type into a neural network of a circulating gate circuit for training to obtain a classification model, and realizing gesture classification and recognition by adopting the classification model.

Further, acquiring the surface electromyographic signals comprises wiping the experimental equipment and the skin surface of the testee by adopting alcohol, placing electrodes for acquiring the electromyographic signals on the skin surface of the testee, and acquiring muscle signal changes brought by gesture motions of different arms in real time.

Further, the number of the electrodes is 8, and the 8 electrodes are distributed on the forearm at equal intervals.

Further, the surface electromyogram signal is transmitted to an intelligent terminal through a wireless device for data processing. The intelligent terminal is a computer, a tablet computer, a mobile phone or other electronic equipment capable of realizing data processing.

Further, processing the surface electromyogram signal includes converting the surface electromyogram signal into a digital signal, and performing feature value extraction on the digital signal.

Further, the surface electromyographic signals are input to the intelligent terminal by adopting a sliding windowing method.

Furthermore, the frequency of the collected surface electromyogram signal is 2kHZ, the width of the sliding window is 100ms, and the sliding step length is 0.5ms. The data volume can be greatly improved, the overfitting problem in the training process is effectively avoided, and the data processing speed is also improved.

Further, the feature value extraction is performed by calculating a root mean square value. The amplitude change of the electromyographic signal can be better described.

Further, the circular gate neural network includes a fully connected layer.

Examples

The application provides a gesture classification method of a GRU model based on surface electromyogram signal acquisition, which comprises the following specific implementation steps of:

step 1: surface electromyogram signals of different gestures of a tested person are collected

The experimental device and the skin surface of a human subject are wiped by alcohol to avoid the interference of noise signals caused by skin epidermis and grease, the electrodes for acquiring the electromyographic signals are placed on the skin surface of the human subject, 8 electrodes are equidistantly distributed on the forearm, the change of muscle signals caused by the gesture motions of different arms is acquired in real time, and the experimental device is shown in figure 1. In the experimental process, the method and the device collect various gestures. In addition, in order to reduce errors caused by muscle fatigue of a testee in the acquisition process, various gestures are repeated for 6 times, the gesture action lasts for 5s every time, and the interval rest is carried out for 5s every time. The collected sEMG signals are transmitted to a computer via a wireless device.

Step 2: processing of surface electromyographic signals

(a) Converting the transmitted analog signal into a digital signal;

(b) in order to obtain more experimental data, the experiment adopts a sliding windowing method, a window contains 100ms of data, and the data in two windows before and after the window slides backwards by 0.5ms. each time has 99.5ms of overlap. Since the frequency of the acquired signal of the experimental device is 2kHZ, the width of the sliding window set by the application is 100ms, and the step length of the sliding is 0.5ms. Meanwhile, before training, data are disorganized, and the convergence speed of the model is accelerated.

(c) The extraction of the feature value is performed by calculating the root mean square value RMS.

And step 3: classification recognition of gesture actions

And inputting the gesture signals after the characteristics are extracted and the corresponding gesture categories into a designed GRU neural network, and obtaining a neural network model as a gesture classifier through training.

Compared with a traditional machine learning-based recognition and classification method, such as a Decision Tree (Decision Tree), a Random Forest (Random Forest), a Support Vector Machine (SVM), a K-nearest neighbor algorithm (KNN) and Naive Bayes (NB), the method has the advantages that when gesture classification is carried out, robustness on different postures of the arm is higher, gesture classification accuracy is higher, and the method can serve a more complex human-computer interaction system with a scene.

The feasibility of the application is verified through experiments. The experiment is realized by adopting a pyrrch frame, and the performance of the method is verified on the collected electromyographic signals of the 8 testees. P1, P2 and P3 respectively represent the arm lying on the table, the arm at an angle of 45 degrees to the table top and the arm lying in the air parallel. The training data were only the front fraction data in P1 arm posture, and the test data were the back fraction data in P1, P2, and P3 arm postures. Experimental results as shown in the following figures, the accuracy of the gesture classification method based on GRU provided by the present application is as high as 0.96, which is higher than that of other methods based on traditional machine learning. And for the arm postures of P2 and P3, the gesture classification accuracy of the method is reduced to the minimum, namely the robustness is strongest.

Although the present application has been described above with reference to specific embodiments, those skilled in the art will recognize that many changes may be made in the configuration and details of the present application within the principles and scope of the present application. The scope of protection of the application is determined by the appended claims, and all changes that come within the meaning and range of equivalency of the technical features are intended to be embraced therein.

Claims

1. a gesture classification and recognition method, is characterized in that: described method comprises the steps:

Obtaining surface EMG signals; performing feature extraction on the surface EMG signals to obtain gesture feature sequences and gesture types; inputting the gesture feature sequences and gesture types into a cyclic gate neural network training to obtain a classification model, and using the classification model The model implements gesture classification and recognition.

2. The method for classifying and recognizing gestures as claimed in claim 1, characterized in that: acquiring the surface EMG signal comprises using alcohol to wipe the surface of the experimental equipment and the subject's skin, and placing the electrode for collecting the EMG signal on the subject's skin On the surface, the muscle signal changes caused by the gestures of different arms are collected in real time.

3 . The gesture classification and recognition method according to claim 2 , wherein the number of the electrodes is 8, and the 8 electrodes are equally spaced on the forearm. 4 .

4 . The gesture classification and recognition method according to claim 2 , wherein the surface EMG signal is transmitted to a smart terminal through a wireless device for data processing. 5 .

5 . The gesture classification and recognition method according to claim 1 , wherein: processing the surface electromyography signal comprises converting the surface electromyography signal into a digital signal, and performing feature value extraction on the digital signal. 6 .

6 . The gesture classification and recognition method according to claim 4 , wherein the surface EMG signal is input to the smart terminal by a sliding window method. 7 .

7 . The gesture classification and recognition method according to claim 6 , wherein the frequency of the collected surface EMG signal is 2 kHz, the width of the sliding window is 100 ms, and the sliding step is 0.5 ms. 8 .

8 . The gesture classification and recognition method according to claim 5 , wherein the feature value extraction is performed by calculating a root mean square value. 9 .

9 . The gesture classification and recognition method according to claim 1 , wherein the recurrent gate neural network comprises a fully connected layer. 10 .

10 . An application of a gesture classification and recognition method, characterized in that: the gesture classification and recognition method according to any one of claims 1 to 9 is applied to a human-computer interaction system.