12 Classification of Hand Motion Using Surface EMG Signals

The human hand has multiple degrees of freedom (DOFs) to achieve high dexterity. Identifying the five-finger movements using surface elec-tromyography (sEMG) is challenging Moreover, the success rate of identifying the hand movements is sensitive to many aspects; for example, the sEMG electrode placements, variant movement forces, or movement speeds In this chapter, a robust sEMG system for identifying the hand movements is developed First, a multichannel sEMG sensor ring is designed, which is easy to wear on the human forearm even without knowledge of the exact location of the corresponding muscles However, a new problem of using the multichannel sensor ring is followed and unsolved in the current research The problem is how to relocate the sEMG electrodes with the same sequence as the last trial. This chapter introduces the concordance correlation coefficient to investigate the relationships of all channels, and autorelocation of the sEMG electrodes is possible. The process of the successful hand motion classification can be divided into collecting original sEMG signals, calculating features basd on original signals, and classifying motions based on features If the calculated features are robust to some variances in the movement forces and speed, the motion classification results also have robustness. Thus, to make classification of the hand movements robust to some variances in the movement forces and speed, a new ratio measure of the multiple channels is defined as the feature, which is based on the results of the temporal square integral values of each channel signal Finally, real-time classification of the hand movements is possible, using the statistical classifier based on Mahalanobis distance. In addition to classification of the hand movement types, knowing the movement force and the movement speed are important In this chapter, the levels of the movement forces are described using spectral moments based on the short-time Fourier transform (STFT) results The levels of the movement speeds, which are seldom studied in the current research, are also identified. Based on the results of STFT, the spectral flatness feature is firstly introduced for the sEMG signal to describe the different speeds of the hand movements

12.1 Introduction

The surface electromyography (sEMG) signal is generated by the electrical activity of muscle fibers during a contraction and is noninvasively recorded by electrodes attached to the skin (Merletti and Parker 2004) Its application to control of artificial limbs or duplication of human movements using a remote mechanism is challenging. In the rehabilitation field, sEMG signals have been applied to control prosthetic legs (Jin et al. 2000) and prosthetic arms (Doringer and Hogan 1995; Ito et al. 1992; Saridis and Gootee 1982). Identification of human hand movements is relatively difficult, because the hand possesses more degrees of freedom (DOFs) than the legs and arms. Due to identification difficulties, the dexterity of some sEMG prosthetic hands in the market is far less than that of the human hand, achieving only a limited number of movements; that is, hand open and hand close Many researchers focus on dexterity improvement of sEMG prosthetic hands (Farry, Walker, and Barabiuk 1996; Fukuda, Tsuji, and Kaneko 1997; Hudgins and Parker 1993; Kuribayashi, Okimura, and Taniguchi 1993), and discrimination of two to six patterns can be achieved In this chapter the aim is to further increase the number of identified hand movements, and classification of seven hand movements is achieved

The placement of sEMG electrodes is a critical issue for successful identification of hand movements Most of the current research is based on the idea that the distribution of the corresponding muscles for hand movements is known. The classification success rates are dominantly determined by the placement of the sEMG electrodes However, in many applications, for example, commercial products, most users lack knowledge of the muscle distribution, and thus there is a risk of classification failure due to misalignment of the sEMG electrodes To solve this problem, in recent research, multichannel sensor rings have been designed (Y.-C. Du et al. 2010; Saponas et al. 2008). Since multichannel sensor rings envelop the whole or half circumference of the forearm, sensor rings can capture all signals from the extensor or flexor muscles of the forearm In this chapter, the sEMG sensor is designed as a half wristband, and the user can easily wear the sensor ring on the wrist as if wearing a watch

The features extracted from the raw sEMG signals are another critical issue for successful identification. For research in which sEMG electrodes are pasted exactly above the corresponding muscles, methods include the temporal features (Zecca et al 2002) for noncomplex and low-speed movements and temporal-spectral features; for example, short-time Fourier transform (STFT) and short-time Thompson transform (STTT), which can provide more transient information for complex and high-speed movements (S J Du and Vuskovic 2004; Farry, Walker, and Barabiuk 1996; Hannaford and Lehman 1986) For a multichannel sensor ring, the methods of feature extraction include the ratios of the temporal and spectral features among the different channels (Saponas et al 2008) and six temporal features directly used for the motion classifier (Y.-C. Du et al. 2010). This chapter further improves the feature extraction method for the multichannel sensor ring by increasing classification robustness to some variances of the movement forces and speed. A new ratio measure of the multiple channels is defined as the fea-ture, which is based on the results of the temporal square integral values of each channel signal

For a multichannel sensor ring, a new problem has arisen, which is how to recognize the same channel sequence as the last trial after the user arbitrarily wears the sensor ring Currently, no studies have been conducted to solve this problem This chapter is inspired by research using the cross-correlation coefficient to investigate the cross-talk among the different channels (Mogk and Keir 2003). The concordance correlation coefficient is introduced to study the relationship between multiple channels and check the feasibility of channel sequence recognition

The speed of hand movements is also critical for movement description Currently, there are few studies concentrated on this topic In this chapter, the levels of the movement speeds are identified. Based on the results of STFT, the spectral flatness feature is firstly introduced for the sEMG signal to describe the different speeds of hand movements

12.2 System Configuration

The sEMG-based sensing system is configured as shown in Figure 12.1. The potential user wears a small, lightweight sEMG sensor ring on the wrist. The sEMG sensor ring has six-channel electrodes, integrated with analog circuits, an A/D converter, and a wireless module. The sEMG signals, accumulated on the skin covering the wrist, are amplified, filtered, and converted digitally and then transferred to a computer via the wireless module. The computer processes the hand movements for identification.

12.2.1 Multichannel sEMG Sensor Ring

The sensor ring has six pairs of sEMG sensor electrodes, and each pair is for one channel, as shown in Figure 12.2. The six-channel sensor ring covers the overall posterior side of the forearm; that is, it includes all of the extensor muscles for the finger movements. Each sEMG electrode has a diameter of (|10 mm, and one channel is composed of two electrodes. Each pair of channels is placed 15 mm apart. Before wearing the sEMG electrodes, the forearm should be washed to ensure good conductivity between the skin surface and the electrodes

images

FIGURE 12.1

Configuration of the sEMG-based sensing system.

images

FIGURE 12.2

Multichannel sEMG sensor ring: (a) sensor ring on the forearm and (b) six-channel sEMG electrodes.

12.2.2 sEMG Signal Preprocessing

The amplitudes of the raw sEMG signals are miniature, at the scale from several microvolts to millivolts. Moreover, several types of noise are mixed with the useful sEMG signals; that is, low-frequency cross-talk, 50-Hz AC frequency, and high-frequency noises. Therefore, the differential amplifier and filters are designed to strengthen and clarify the useful signals. The sEMG signal is amplified to the scale of V by the differential amplifier. The high-pass (>20 Hz), notch (50 Hz), and low-pass (<500 Hz) filters eliminate the three types of noise mentioned above. The amplifier and filters are miniature sized and are integrated with the sEMG sensor ring to be worn on the human forearm The advantage of integration of the sEMG sensor and the analog circuits is that there is no extra noise introduced during long-distance wire transfer

After amplification and filtration, the analog voltages of the sEMG signal are converted digitally by a 10-bit A/D converter. The digital signals are transferred to the computer via wireless communication using standard radio frequency (RF) technology.

12.3 Classification of Hand Movements

This section addresses the classification of hand movements using the six-channel sensor ring. We defined seven types of hand movements, as shown in Figure 12.3.

12.3.1 Automatic Relocation of sEMG Electrodes

The advantages of using the multichannel sensor ring include convenience for the user and the low possibility of classification failure due to misalignment of the sEMG electrodes. However, especially for the whole sensor ring, some channels are redundant. The multichannel sensor ring always has redundant channels, and it is important to determine which channels are meaningful. To ensure that the extracted features are uniform, especially when the user arbitrarily wears the sensor ring, it is important to determine where the first-channel electrodes among the selected meaningful channels are located for every trial

images

FIGURE 12.3

Seven classes of hand movements: (a) thumb, (b) index, (c) V sign, (d) OK gesture, (e) all fingers extended, (f) four fingers extended, and (g) grasp.

For multichannel sEMG sensors, the definition of the cross-correlation coefficient, also called the Pearson's product-moment coefficient, has been used to investigate cross-talk among the channels (Mogk and Keir 2003). However, the cross-correlation only measures the extent of the linear relationship between two variables. When two variables have a nonlinear relationship, the value of the cross-correlation coefficient is zero. Thus, the cross-correlation coefficient has the risk during evaluating the relationship of two variables. Here, the concordance coefficient is introduced in this chapter to evaluate the agreement of each two-channel sEMG signal The concordance correlation coefficient, defined by Lin (1989), measures the agreement between two variables and has been widely used in studies on data reproducibility (Lin 1989) and image comparison analysis (Lange et al. 1999). In this chapter, we use the concordance coefficient to investigate the agreement between sEMG signals for each two-channel signal

The concordance correlation coefficient of the N-length variables of x and y is defined as

$p = \frac{2 0_{xy}}{o_{χ}^{2} + 0_{2}^{2} + (μ_{χ} - μ_{y})^{2}}$ $p = \frac{2 0_{xy}}{o_{χ}^{2} + 0_{2}^{2} + (μ_{χ} - μ_{y})^{2}}$ (12.1)

i_x and u_y are the mean of the two variables, and u_y has the same formula as

$μ_{χ} = \frac{1}{N} \sum_{i = 1}^{N} x_{i}$ $μ_{χ} = \frac{1}{N} \sum_{i = 1}^{N} x_{i}$ (12.2)

c_x and c_y are the variances of the two variables, and c_y has the same formula as c_x:

$o_{χ}^{2} = \frac{1}{N} \sum_{i = 1}^{N} (x_{i} - μ_{χ})^{2}$ $o_{χ}^{2} = \frac{1}{N} \sum_{i = 1}^{N} (x_{i} - μ_{χ})^{2}$ (12.3)

c_xy is the covariance of x and y:

$o_{xy} = \frac{1}{N} \sum_{i = 1}^{N} (x_{j} - μ_{χ}) (y_{j} - μ_{y})$ $o_{xy} = \frac{1}{N} \sum_{i = 1}^{N} (x_{j} - μ_{χ}) (y_{j} - μ_{y})$ (12.4)

For the generalized formulation, it is assumed that there are a total of M pairs of sEMG electrodes in the multichannel sensor ring. The M-channel sEMG signals are represented by an NxM matrix of X = [X_l7 X¡, X_M ]. Each column X_{ of an N-length vector is the time-series sEMG signal of channel i . The concordance correlation coefficient of any two channels channel i and channel j is defined as

$R_{X_{i} X_{i}} = p_{X_{i} X_{j}}, i = 1, ... M - 1, j = i + 1, ... M$ $R_{X_{i} X_{i}} = p_{X_{i} X_{j}}, i = 1, ... M - 1, j = i + 1, ... M$ (12.5)

For each hand movement, we can obtain a 1 x y (M - i) vector R of the concordance correlation coefficients as ⁱ ¹

$R = [R_{1}, R_{i'} R_{M - 1}]$ $R = [R_{1}, R_{i'} R_{M - 1}]$ (12.6)

where

$R_{1} = [R_{xX} 12'$ $R_{1} = [R_{xX} 12'$ $R_{xX} 13'$ $R_{xX} 13'$ $R_{X_{1} X_{M}}]$ $R_{X_{1} X_{M}}]$

$R_{i} = [R_{x_{i} X_{i + 1'}}$ $R_{i} = [R_{x_{i} X_{i + 1'}}$ $R_{x_{i} X_{i + 2'}}$ $R_{x_{i} X_{i + 2'}}$ $R_{X_{i} X_{M}}],$ $R_{X_{i} X_{M}}],$

$R_{M - 1} = [R_{X_{M - 1} X_{M}}]$ $R_{M - 1} = [R_{X_{M - 1} X_{M}}]$

In our case, there are a total of six channels for the sEMG electrodes. Before the derivation of the concordance correlation coefficients, we explain why six channels for the sEMG electrodes are chosen It is well known that most muscles responsible for hand movements are clustered on the posterior side of the forearm The extensor digitorum is responsible for the movements of the index, middle, ring, and little fingers; the extensor pollicis longus and brevis control the thumb; the extensor indicis controls the index finger; and the extensor digiti minimi controls the little finger (Miller 2008). The muscles on the posterior of the forearm must be included in the area enveloped by the sensor ring. Six channels are enough to cover the circumstance of the posterior side The six-channel sEMG signals of each hand movement are represented by an Nx6 matrix of X = \_X_1r X₂, X₃, X₄, X₅, X₆ ]. For each hand movement, we can obtain a 2x25 vector R of the concordance correlation coefficients as

$R = [R_{1},$ $R = [R_{1},$ $R_{2}, R_{3}, R_{4}, R_{5}]$ $R_{2}, R_{3}, R_{4}, R_{5}]$ (12.7)

where

$R_{1} = [R_{X_{1} X_{2}},$ $R_{1} = [R_{X_{1} X_{2}},$ $R_{X_{1} X_{3}},$ $R_{X_{1} X_{3}},$ $R_{X_{1} X_{4}},$ $R_{X_{1} X_{4}},$ $R_{X_{1} X_{5}},$ $R_{X_{1} X_{5}},$ $R_{X_{1} X_{6}}]$ $R_{X_{1} X_{6}}]$

$R_{2} = [R_{X_{2} X_{3'}}$ $R_{2} = [R_{X_{2} X_{3'}}$ $R_{X_{2} X_{4'}}$ $R_{X_{2} X_{4'}}$ $R_{X_{2} X_{5'}}$ $R_{X_{2} X_{5'}}$ $R_{X_{2} X_{6}}]$ $R_{X_{2} X_{6}}]$

$R_{3} = [R_{X_{3} X_{4'}}$ $R_{3} = [R_{X_{3} X_{4'}}$ $R_{X_{3} X_{5'}}$ $R_{X_{3} X_{5'}}$ $R_{X_{3} X_{6}}]$ $R_{X_{3} X_{6}}]$

$R_{4} = [R_{X_{4} X_{5'}}$ $R_{4} = [R_{X_{4} X_{5'}}$ $R_{X_{4} X_{6}}]$ $R_{X_{4} X_{6}}]$

$R_{5} = [R_{X_{5} X_{6}}]$ $R_{5} = [R_{X_{5} X_{6}}]$

The concordance correlation coefficients of three kinds of gestures were investigated, including thumb (Figure 12.3a), index (Figure 12.3b), and the OK configuration (Figure 12.3d). For each kind of gesture, thirty trials were sampled. In the first step, the same onset point of six-channel signals for each movement was found using the Bonato method (Staude et al. 2001). In the second step, 500-ms signals were selected. Finally, the concordance correlation coefficients were computed. In Figure 12.4, the results show that for each kind of hand movement, the concordance correlation coefficients of thirty trials have similar and stable distributions, which fall in the narrow bands with the maximum and minimum values as the boundaries

Here we apply the results in Figure 12.4 to the generalized case in which the sEMG sensor ring has M channels (M > 6) for the electrodes. Before each use, the user only needs to perform these three gestures several times as calibration For each hand movement, a

$1 \times \sum_{i = 1}^{M - 1} (M - i)$ $1 \times \sum_{i = 1}^{M - 1} (M - i)$

vector can be obtained From this vector, six continuous channels, of which concordance correlation coefficients are all located within the boundaries shown in Figure 12.4, can be selected. If, for each calibration, the concordance correlation coefficients of the same six continuous channels fall within the boundaries, it can be concluded that the channel at the beginning of these six channels is the first channel.

12.3.2 Feature Extraction from Multiple Channels

The spectral method of square integral feature defined in Equation (12.8) has been widely used as the calculated feature for motion classification.

$E_{j} = \sum_{i = 1}^{N} X_{i}^{2} (t)$ $E_{j} = \sum_{i = 1}^{N} X_{i}^{2} (t)$ (12.8)

images

FIGURE 12.4

Concordance correlation coefficients for three gestures: (a) thumb, (b) index, and (c) OK gesture.

where i represents the ith channel, X¡ (t) is the time-series sEMG signal of the ith channel, and N is the data number of the time-series sEGM signal from one channel

However, there is a limitation when using the spectral integral of the temporal sEMG signals as the feature. For the same type of the movement, the square integral values of the temporal signal vary with the movement force and movement speed. The changes in the feature's values can affect the classification results. The ideal case is that classification of the types of hand movements is not affected by some variances such as movement forces and speed. To make the classification result robust to these variances, the ratios of the square integral values of the multiple channels are defined as the feature. The ratio of the ith channel to first-channel signals is defined as

$R E_{i 1} = \frac{E_{i}}{E_{1}}, i = 2 M$ $R E_{i 1} = \frac{E_{i}}{E_{1}}, i = 2 M$ (12.9)

All of the ratios of single-channel to first-channel signals are defined as

$R E_{1} = [R E_{21}, R E_{M 1}]$ $R E_{1} = [R E_{21}, R E_{M 1}]$ (12.10)

The ratio of the ith channel to the jth channel signal is represented as

$R E_{i}^{*} i = \frac{E_{i}}{E_{i}}, i = 2, ... / M - 1, j = i + 1, ... M /$ $R E_{i}^{*} i = \frac{E_{i}}{E_{i}}, i = 2, ... / M - 1, j = i + 1, ... M /$ (12.11)

Normalize RE* with reference to first-channel signal:

$R E_{ι' i} = \frac{E_{l} / E_{i}}{E_{i} / E_{1}} = \frac{E_{ι'} \times E_{1}}{E_{i}^{2}}$ $R E_{ι' i} = \frac{E_{l} / E_{i}}{E_{i} / E_{1}} = \frac{E_{ι'} \times E_{1}}{E_{i}^{2}}$ (12.12)

All of the ratios of the ith channel to jth channel signals with reference to the first-channel signal are represented as

$R E_{i} = [R E_{(i + 1) i}, R E_{Mi}], i = 2, ... / M - 1$ $R E_{i} = [R E_{(i + 1) i}, R E_{Mi}], i = 2, ... / M - 1$ (12.13)

Combining Equations (12.10) and (12.13), we can obtain the newly defined feature of channel ratio, a vector formulated as

$RE = [R E_{1},, R E_{i},, R E_{M - 1}]$ $RE = [R E_{1},, R E_{i},, R E_{M - 1}]$ (12.14)

where M is the channel number. RE, is a 1 x (M-1) vector, and RE_i is a 1 x (M-i) vector. So RE is a row vector with N (M - i) columns.

In our case, there is a total of six channels, and thus RE is a 1 x 15 vector of:

$RE = [R E_{21}, R E_{61}, R E_{32}, R E_{62}, R E_{43},, R E_{63}, R E_{54}, R E_{64}, R E_{65}]$ $RE = [R E_{21}, R E_{61}, R E_{32}, R E_{62}, R E_{43},, R E_{63}, R E_{54}, R E_{64}, R E_{65}]$

Using the above definition and the six-channel sensor ring shown in Figure 12.2, the seven types of hand movements determined as shown in Figure 12.3 can be distinguished. For each type of hand movement, thirty trials were sampled. First, the Bonato method was used to find the onset points, and 500-ms signals from the onset points were selected, as shown in Figure 12.5. Then, the feature measures defined in Equation (12.14) were calculated, and the results are shown in Figure 12.6.

images

FIGURE 12.5

Six-channel raw sEMG signals for seven types of hand movements.

images

FIGURE 12.6

Channel ratio features of seven types of hand movements.

The features shown in Figure 12.6 are selected as the trained data sets and then clustered by the Mahalanobis distance theory Another ten trials for each type of hand movement were sampled as the test data The success rate was around 90%. The results show that the proposed feature measure is effective. In future work, we will improve the proposed feature measure to check the feasibility of applying the same trained data set to different users

12.4 Identification of Movement Force and Speed

Force and speed are also important parameters for movement description Currently, there is no effective method for identifying movement speed Movement speed is related to the transient portion of the sEMG signals Here, the STFT results are used to obtain the distribution of the sEMG signal in both time and frequency domains. Then, the spectral flatness feature is introduced and used to describe the speed. The description of the movement speed is also based on STFT results, and then the spectral moment is calculated as the feature

12.4.1 STFT Method

The basic idea of STFT is to divide a signal into several segments in the time domain and apply discrete Fourier transform (DFT) to each time segment.

Given a finite-length time sequence x ( i), i e [ö, L -1 ], the DFT of the time sequence x(i) is

$X [m] = X [mF] = \sum_{i = 0}^{L - 1} x [i] e^{(- j 2 π (mF) (i T_{s}))}$ $X [m] = X [mF] = \sum_{i = 0}^{L - 1} x [i] e^{(- j 2 π (mF) (i T_{s}))}$ (12.15)

where T_s is the sampling frequency, and F = 1/LT_s is the frequency sampling step size. STFT for each segmented window is the sum of these DFTs with respect to T_s and F,

$P [k, m] = P [k T_{s}, mF] = \sum_{ι' = 0}^{L - 1} x [i] WW [i - t] e^{(- j 2 π mi / L)}$ $P [k, m] = P [k T_{s}, mF] = \sum_{ι' = 0}^{L - 1} x [i] WW [i - t] e^{(- j 2 π mi / L)}$ (12.16)

where W[i] is the window function. The sampling step size in the time domain is T = kT_s .

A critical issue for STFT is tradeoff between time and frequency resolution This means that a spectrum computed from a relatively long time win-dow will resolve detailed frequency features but change little if the center time of the window is shifted by a small amount. Conversely, a narrow window will have a high time resolution but show few details in the computed spectrum Thus, the acceptable resolutions of time and frequency are lower bounded. Balance can be found from the time-bandwidth uncertainty principle or Heisenberg inequality,

$Δ t \times Δ f ≙ \frac{1}{4 π}$ $Δ t \times Δ f ≙ \frac{1}{4 π}$ (12.17)

12.4.2 Features Based on STFT

Based on the results of STFT, the nth spectral moment of the frequency distribution at time t is defined as

$M_{n} (t) = \sum_{f = l_{b}}^{u_{b}} (f^{n} \hat{p} (t, f))$ $M_{n} (t) = \sum_{f = l_{b}}^{u_{b}} (f^{n} \hat{p} (t, f))$ (12.18)

where n is the order of the spectral moment, l_b and u_b are the lower and upper boundaries of the frequency, and p(t,f) is the calculated spectrum from STFT.

Based on the results of STFT, the spectral flatness feature is introduced to describe the characteristics of the sEMG signal. Spectral flatness has been widely used in the field of audio for description of the spectral power distribution of the sound. Here we introduce this definition to the sEMG signal. The spectral flatness is defined as

$SF (t) = \frac{[\prod_{f = l_{b}} u_{b^{\land}} p^{2} (t, f)]^{\frac{1}{u_{b} - l_{b} + 1}}}{\frac{1}{u_{b} - l_{b} + 1} \sum_{f = l_{b}}^{u_{b^{\land}}} p^{2} (t, f)}$ $SF (t) = \frac{[\prod_{f = l_{b}} u_{b^{\land}} p^{2} (t, f)]^{\frac{1}{u_{b} - l_{b} + 1}}}{\frac{1}{u_{b} - l_{b} + 1} \sum_{f = l_{b}}^{u_{b^{\land}}} p^{2} (t, f)}$ (12.19)

where l_b and u_b are the lower and upper boundaries of the frequency. p (t, f) is the calculated spectrum from STFT in Equation (12.16).

A high spectral flatness indicates that the spectrum has a similar amount of power in all spectral bands. A low spectral flatness indicates that the spectral power is concentrated within a relatively small number of bands

12.4.3 Experimental Results

To test the levels of the forces and speeds, the movement is designed as ball grasping, shown in Figure 12.7. We chose the sEMG signal from a middle channel, where the signals have large amplitude First, the raw sEMG signals were processed by STFT, and the results are shown in Figures 12.8 and 12.9. According to the Heisenberg inequality in Equation (12.17), the segmented window number in the time domain was chosen as fifty, with a window overlap of 30%, and 100 points were selected over the frequency range from 0 to 500 Hz.

Based on the STFT results, the spectral moment features were calculated, as shown in Figure 12.10. It can be seen that the spectral moment features of the higher force have larger amplitudes than the lower force Thus, the force levels can be distinguished. Based on the STFT results, the spectral flatness features were calculated, as shown in Figure 12.11. The significant differences of the spectrum distribution at the different time points can be seen The spectral flatness feature showed that the flatness changes at low speed were more significant than those at high speed. Thus, the speed classification can be differentiated based on the spectral flatness features.

images

FIGURE 12.7

Ball grasping.

12.5 Summary

In this chapter, a robust sEMG sensor system was designed for identification of human hand movements. The sEMG sensor was designed as a multichannel ring. This chapter addresses the advantages of the multichannel sensor ring and also states the necessity of relocating the electrodes for each trial. Autorelocation of the electrodes is proposed using the concordance correlation coefficients for each two channels. The results show that the proposed method is effective. For classification of the move-ment types, to make the classification result robust to some variances such as the movement forces and speeds, a new feature measure based on multiple channels was proposed and applied to distinguish seven types of hand movements. The classification success rate was as high as 90% using the proposed feature measure of multiple channels and statistical Mahalanobis distance classifier. In future work, we will further improve the method of feature extraction and investigate the feasibility of using the same features for different users. Finally, in this chapter, identification of different movement speeds was achieved by introducing the spectral flatness feature to describe the spectral power distribution of sEMG signals. Different movement forces were also identified by the spectral moment feature based on the STFT results.

images

FIGURE 12.8

STFT results at different forces: (a) high force and (b) low force.

images

FIGURE 12.9

STFT results at different speeds: (a) low speed and (b) high speed.

images

FIGURE 12.10

Spectral moment features based on STFT at different forces.

images

FIGURE 12.11

Spectral flatness features based on STFT at different speeds.

References

Doringer, J.A. and Hogan, N. 1995. Performance of above elbow body-powered prostheses in visually guided unconstrained motion tasks. IEEE Transactions on Biomedical Engineering, 42: 621-631.

Du, S. J. and Vuskovic, M. 2004. Temporal vs. spectral approach to feature extraction from prehensile EMG signals. Paper read at the IEEE International Conference on Information Reuse and Integration, Las Vegas, NV, November 2009.

Du, Y.-C., Lin, C.-H., Shyu, L.-Y. and Chen, T. 2010. Portable hand motion classifier for multi-channel surface electromyography recognition using grey relational analysis. Expert Systems with Applications, 37(6): 4283-4291.

Farry, K. A., Walker, I. D., and Barabiuk, R. G. 1996. Myoelectric teleoperation of a complex robotic hand. IEEE Transactions on Robotics and Automation, 12: 775-788.

Fukuda, O., Tsuji, T., and Kaneko, M. 1997. An EMG controlled robotics manipulator using neural networks Paper read at the IEEE International Workshop on Robot and Human Communication, Sendai, Japan, October 1997

Hannaford, B and Lehman, S 1986 Short time fourier analysis of the electromyo-gram: fast movements and constant contraction. IEEE Transactions on Biomedical Engineering, 33: 1773-1181.

Hudgins, B. and Parker, P. 1993. A new strategy for multifunction myoelectric control. IEEE Transactions on Biomedical Engineering, 40: 82-94.

Ito, K., Tsuji, T., Kato, A., and Ito, M. 1992. EMG pattern classification for a prosthetic forearm with three degrees of freedom Paper read at the IEEE International Workshop on Robot and Human Communication, Toyko, Japan, September 1992

Jin, D. W., Zhang, R. H., Zhang, J. C., Wang, R. C., and Gruver, W. A. 2000. An intelligent above-knee prosthesis with EMG-based terrain identification. Paper read at the IEEE International Conference on Systems, Man and Cybernetics, Nashville, TN, October 2002

Kuribayashi, K., Okimura, K., and Taniguchi, T. 1993. A discrimination system using neural network for EMG-controller prostheses—Integral type of EMG signal processing. Paper read at the IEEE/RSJ International Conference on Intelligent Robots and Systems, Tokyo, Japan, July 1993

Lange, N., Strother, S C, Anderson, J R, Nielsen, F A, Holmes, A P, Kolenda, T., Savoy, R., and Hansen, L K 1999 Plurality and resemblance in fMRI data analysis. Neuroimage, 10: 282-303.

Lin, I. -K. 1989. A concordance correlation coefficient to evaluate reproducibility. Biometrics, 45(1): 255-268

Merletti, R. and Parker, P. 2004. Electromyography: Physiology, Engineering and Noninvasive Applications. Hoboken, NJ: Wiley.

Miller, C.J. 2008. Real-time feature extraction and classification of prehensile EMG signals. Master's thesis, San Diego State University.

Mogk, P M and Keir, J 2003 Crosstalk in surface electromyography of the proximal forearm during gripping tasks. Journal of Electromyography and Kinesiology, 13: 63-71.

Saponas, T S, Tan, D., Morris, D and Balakrishnan, R 2008 Demonstrating the feasibility of using forearm electromyography for muscle-computer inter-faces Paper read at the Conference on Human Factors in Computing Systems, Florence, Italy, April 2008

Saridis, G.N. and Gootee, T. P. 1982. EMG pattern analysis and classification for a prosthetic arm. IEEE Transactions on Biomedical Engineering, 29: 403^12.

Staude, G., Flachenecker, C., Daumer, M., and Wolf, W 2001 Onset detection in surface electromyographic signals: A systematic comparison of methods. EURASIP Journal on Applied Signal Processing, (1): 67-81

Zecca, M., Micera, S., Carrozza, M.C., and Dario, P. 2002. Control of multifunctional prosthetic hands by processing the electromyographic signal. Critical Review in Biomedical Engineering, 30: 459-485

..................Content has been hidden....................

You can't read the all page of ebook, please click here login for view all page.

Table of Contents for
12 Classification of Hand Motion Using Surface EMG Signals

12

Classification of Hand Motion Using Surface EMG Signals

CONTENTS

12.1 Introduction

12.2 System Configuration

12.2.1 Multichannel sEMG Sensor Ring

12.2.2 sEMG Signal Preprocessing

12.3 Classification of Hand Movements

12.3.1 Automatic Relocation of sEMG Electrodes

12.3.2 Feature Extraction from Multiple Channels

12.4 Identification of Movement Force and Speed

12.4.1 STFT Method

12.4.2 Features Based on STFT

12.4.3 Experimental Results

12.5 Summary

References

Table of Contents for 12 Classification of Hand Motion Using Surface EMG Signals

Create new playlist

Sign In

Sign Up

Classification of Hand Motion Using Surface EMG Signals

CONTENTS

12.2.2 sEMG Signal Preprocessing

Table of Contents for
12 Classification of Hand Motion Using Surface EMG Signals