A Low-Cost, Wireless, 3-D-Printed Custom Armband for sEMG Hand Gesture Recognition

Abstract

Wearable technology can be employed to elevate the abilities of humans to perform demanding and complex tasks more efficiently. Armbands capable of surface electromyography (sEMG) are attractive and noninvasive devices from which human intent can be derived by leveraging machine learning. However, the sEMG acquisition systems currently available tend to be prohibitively costly for personal use or sacrifice wearability or signal quality to be more affordable. This work introduces the 3DC Armband designed by the Biomedical Microsystems Laboratory in Laval University; a wireless, 10-channel, 1000 sps, dry-electrode, low-cost (∼150 USD) myoelectric armband that also includes a 9-axis inertial measurement unit. The proposed system is compared with the Myo Armband by Thalmic Labs, one of the most popular sEMG acquisition systems. The comparison is made by employing a new offline dataset featuring 22 able-bodied participants performing eleven hand/wrist gestures while wearing the two armbands simultaneously. The 3DC Armband systematically and significantly ( p < 0.05 ) outperforms the Myo Armband, with three different classifiers employing three different input modalities when using ten seconds or more of training data per gesture. This new dataset, alongside the source code, Altium project and 3-D models are made readily available for download within a Github repository.

Keywords: acquisition system; gesture recognition; sEMG; surface electromyogram; wearable sensors.

Figure A1

Confusion Matrices for the Myo and the 3DC Armband employing LDA for classification and one cycle of training. Lighter is better.

Figure A2

Confusion Matrices for the Myo and the 3DC Armband employing LDA for classification and two cycles of training. Lighter is better.

Figure A3

Confusion Matrices for the Myo and the 3DC Armband employing LDA for classification and three cycles of training. Lighter is better.

Figure A4

Confusion Matrices for the Myo and the 3DC Armband employing the Raw ConvNet for classification and one cycle of training. Lighter is better.

Figure A5

Confusion Matrices for the Myo and the 3DC Armband employing the Raw ConvNet for classification and two cycles of training. Lighter is better.

Figure A6

Confusion Matrices for the Myo and the 3DC Armband employing the Raw ConvNet for classification and three cycles of training. Lighter is better.

Figure A7

Confusion Matrices for the Myo and the 3DC Armband employing the Spectrogram ConvNet for classification and one cycle of training. Lighter is better.

Figure A8

Confusion Matrices for the Myo and the 3DC Armband employing the Spectrogram ConvNet for classification and two cycles of training. Lighter is better.

Figure A9

Confusion Matrices for the Myo and the 3DC Armband employing the Spectrogram ConvNet for classification and three cycles of training. Lighter is better.

Figure 1

The proposed 3DC Armband. The system and the battery are held in the receptacles identified by 1 and 10 respectively. The label on each part of the armband corresponds to the channels’ order that are recorded for the dataset described in Section 4.

Figure 2

System-level concept of the multichannel wireless sEMG sensor: The sensor is built around a custom 0.13-$\mathsf{\mu }$m SoC that includes 10× sEMG channels, each of which encompasses a bioamplifier, a $\Delta \mathsf{\Sigma }$ analog-to-digital converter (ADC), and a 4th order decimation filter. The SoC, the nRF24L01+ low-power wireless transceiver, and the ICM-20948 9-axis IMU are interfaced with an MSP430F5328 low-power MCU.

Figure 3

(a) Two-sided view of the sEMG sensor with each part identified: The printed circuit board (PCB) has a flexible region to fold the two rigid parts on top of each other to save space. (b) The packaged SoC which is wirebonded directly on a PCB substrate. (c) The system folded in its final position beside a Canadian quarter coin (diameter of 23.88 mm).

Figure 4

(a) Analog bandwidth of the bioamplifier (in black), digital bandwidth of the decimation filter (in blue), Myo bandwidth comparison (in orange), and (b) noise spectrum of the bioamplifier. The input referred noise is of 2.5 $\mathsf{\mu }$V${}_{\mathrm{rms}}$ over a 500-Hz bandwidth.

Figure 5

(A) The system’s receptacle: The bottom of the unit is used to receive the main electrode, while the system is stored inside. A cover slides on to enclose the system. (B) The battery holder: This receptacle is used to house the power source of the armband and, as such, should be placed next to the system’s holder. Once the battery is placed, the cover can then slide on to protect the system. A standard electrode is placed on the bottom of this holder. (C) This holder houses a standard electrode. For the proposed 3DC Armband, eight such receptacles are required.

Figure 6

The two different armband configurations (left/right) employed in this work with the 3DC being either above or below the Myo armband with respect to the participant’s wrist. This figure also showcases the wide variety of armband positions recorded in the proposed dataset.

Figure 7

The eleven hand/wrist gestures employed in the proposed dataset.

Figure 8

Comparison of the signals recorded with the Myo Armband and the proposed 3DC Armband. The x-axis represents time in seconds, while the y-axis is the different channels of the armbands. The three gestures recorded in order are the chuck grip, Open Hand, and Pinch Grip. Note that these signals were not obtained using the Comparison Dataset recording protocol to show a wider array of gestures in a continuous way.

Figure 9

The raw ConvNet architecture employing 34,667 parameters. In this figure, Conv refers to Convolution and BN refers to Batch Normalization. While the input represented in this figure is that of the 3DC, the architecture remains the same for all considered systems.

Figure 10

The Spectrogram ConvNet architecture employing 95,627 parameters. In this figure, Conv refers to Convolution and BN refers to Batch Normalization. The input represented comes from the 3DC Armband with the channels on the x-axis and the frequency bins on the y-axis. Due to the Myo Armband associated input size, P4 and C5 were removed from the architecture when training on Myo’s data.

Figure 11

Comparison between the Myo and the 3DC Armband employing LDA for classification: The number of cycles corresponds to the amount of data employed for training (one cycle equals 5 s of data per gesture). The Wilcoxon Signed Rank test is applied between the Myo and the 3DC Armband. The null hypothesis is that the median difference between pairs of observations (i.e., accuracy from the same participant with the Myo or the 3DC Armband) is zero. The p-value is shown when the null hypothesis is rejected (significant level set at $p=0.05$). The black line represents the standard deviation calculated across all 22 participants.

Figure 12

Confusion Matrices for the Myo and the 3DC Armband employing linear discriminant analysis (LDA) for classification and four cycles of training. A lighter color is better.

Figure 13

Comparison between the Myo and the 3DC Armband employing Raw ConvNet for classification: The number of cycles corresponds to the amount of data employed for training (one cycle equals 5 s of data per gesture). The Wilcoxon Signed Rank test is applied between the Myo and the 3DC Armband. The null hypothesis is that the median difference between pairs of observations (i.e., accuracy from the same participant with the Myo or the 3DC Armband) is zero. The p-value is shown when the null hypothesis is rejected (significant level set at $p=0.05$). The black line represents the standard deviation calculated across all 22 participants.

Figure 14

Confusion Matrices for the Myo and the 3DC Armband employing the Raw ConvNet for classification and four cycles of training. A lighter color is better.

Figure 15

Comparison between the Myo and the 3DC Armband employing the Spectrogram ConvNet for classification: The number of cycles corresponds to the amount of data employed for training (one cycle equals 5 s of data per gesture). The Wilcoxon Signed Rank test is applied between the Myo and the 3DC Armband. The null hypothesis is that the median difference between pairs of observations (i.e., accuracy from the same participant with the Myo or the 3DC Armband) is zero. The p-value is shown when the null hypothesis is rejected (significant level set at $p=0.05$). The black line represents the standard deviation calculated across all 22 participants.

Figure 16

Confusion Matrices for the Myo and the 3DC Armband employing the Spectrogram ConvNet for classification and four cycles of training. A lighter color is better.