Diseño e implementación de una interfaz para la traducción de caracteres del alfabeto dactilológico a palabras utilizando sensores musculares y reconocimiento de patrones.

Abstract

This research developed a system aimed to people with hearing disabilities who use the fingerprint alphabet for their communication. This device allows the translation of the conventional alphabet symbols represented in signs to letters using supervised machine Learning models. The implementation, processing and classification is executed on an embedded FPGA platform PYNQ-Z1, which is developed under the Linux open source system compatible with the Python program and its different bookstores. The process is initially developed with the adquisition of data through the myo armband muscle sensor. The electromiographic signals (EMG) are then processed through the Multiresolution Analysis of the Discrete Wavelet transformed with its mother wavelet Daubechies 4 and the level of decomposition 2. Once the signals are processed, useful information is extracted through the extraction of characteristics Mean Absolute Value (MVA), Root Mean Squared (RMS) and Waveform Length (WL) of each of the data subsets. Subsequently, with the characteristics obtained, the data set is prepared for the Learning models, supervised and selected: Multi Layer Perceptron neural network, Decision tree and Random Forest. The percentage of well-classified data for the three models is 91%, 82%, and 94% respectively. The design of the translation interface is user friendly as it consists of 4 buttons clearly identified with the function they fulfill.