Prior Works
Discussion
Myo Armband
In order to gather sEMG signals and create an interactive-safe environment, we decided to utilize the Myo Armband for purposes of data collection and interface. What the Myo Armband shows for in mobility, it lacks in accuracy. For comparison, traditional methods of sEMG signal collection samples between 500 - 1000 Hz while the Myo Armband samples at 200 Hz (while the Myo Armband is advertised to run at 200 Hz, experiments demonstrate a slightly lower frequency around 188 Hz). Therefore, while convenient for our design, the Myo Armband can be considered an area of improvement. Recently, a research paper released a comparison of their 3D printed sEMG armband with the Myo Armband and was able to demonstrate significant improvement, especially in terms of frequency, sampling at 1000 samples per second.[16] While the said paper’s sEMG armband is not commercially available, it demonstrates the capabilities of improving mobile sEMG collection methods.
Neural Network Architecture
More recently, research on the Myo Armband has revolved around determining the best-fit neural network to recognize gestures at a real-time rate from sEMG signal analysis. The many different models and algorithms include: gated recurrent unit (GRU), stacked sparse auto encoders (SSAE), latent Dirichlet allocation (LDA), and convolutional and artificial neural networks (CNN and ANN). [2] [3] [4] [5] [6] [7] These papers manipulated many factors to determine the performances of each model given unique parameters; more specifically, they focused on making changes to attributes within the training set. The performances between unsupervised vs supervised sets and different time vs frequency features were explored. Given the results of these projects, models such as LDA and SSAE were not considered within our own neural network design. The networks that demonstrated the strongest classifications across these papers were CNNs and ANNs. Common characteristics of these models included batch normalization layers, a rectified linear unit activation function (ReLu), a sigmoid activation function, and a softmax activation function. [12] Each of these models were able to perform at >95% accuracy when classifying dissimilar large sEMG signal gestures. While CNNs demonstrated strong performances with unsupervised datasets, for the purposes of our project, we decided upon an ANN with 6 time-domain features.
Gesture Set
All previous designs consistently looked to differentiate from, essentially, the same gesture set. The main gestures include iterations of wrist flexion (waves), multiple finger taps, closed hand, and open hand.

