Design and prototype development of real time AI enabled audio controlled assistive wheelchair mechanism

Abstract

Most assistive technologies use audio as the input for controlling electronic devices. Speech is one of the easiest forms of data to control for both disabled and non-disabled people to access systems. Audio data consists of noise, which can be challenging to classify or extract information from. In this paper, an assistive system is designed for audio classification and implemented on a Field-Programmable Gate Array (FPGA) board. Electromagnetic attenuators are designed to filter out noise signals from the useful information or signal. The Electromagnetic Interference (EMI) filtered signals are further processed to classify commands, assisting devices in calculating the performance of all multiclass SVM classifiers.  The classifiers are designed and optimized in MATLAB R2021a, and the work is extended to implement on Vivado 2020.1. The Hardware Description Language (HDL) code, generated from MATLAB, is used for model development on the Nexys 4 Artix 7 board. The built-in microphone on the mentioned FPGA board is used to collect audio signals as input data for the learning modules. The accuracy of the machine learning module depends on the training dataset. The larger training data set will ensure the accuracy of the proposed module. The accuracy of the proposed audio signal-based assistive device is 87% for eight different commands to control the wheelchair.