EMG Controlled Electric Wheelchair

In this work, a controller for an electric wheelchair using electromyography (EMG) was developed. The commercial Myo armband that contains 8 surface EMG sensors, and an Inertial Measurement Unit (IMU) was used to collect raw EMG signals from the forearm muscles of a user. A preprocessing stage to denoise and condition the raw EMG signals was implemented. Then, the following features were extracted from each channel: Mean Absolute Value, Slope Sign Change, Waveform Length, Root Mean Squared, and Hjorth Parameters. The extracted feature vectors were used as inputs to an ANN that classified the hand gestures. Five hand gesture were selected to control the motion of the electric wheel chair as follows: Spread fingers (go forward), wave right (turn right), wave left (turn left), fist (go backwards), and rest position (stop). In order to find the optimal set of hyperparameters for the ANN, the Keras Tuner was used. Once the network was trained the TensorFlow Lite (TFLite) file was transferred to the Raspberry Pi for real time use. The ANN performed with a 96% accuracy rate with a delay of only 400 ms. The Raspberry Pi is where all of the signal processing and hand gesture recognition was performed. The hand gesture detected was sent over a general-purpose input/output connection to the Arduino that then communicates with the DC motors through a Sabertooth motor controller to operate electric wheelchair.