Estudo da influência das etapas de segmentação, extração de características e classificação do alfabeto em linguagem brasileira de sinais a partir de sinais eletromiográficos de superfície

• Ano de defesa: 2020
• Autor(a) principal: Mendes Junior, Jose Jair Alves
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Curitiba; Brasil; Programa de Pós-Graduação em Engenharia Elétrica e Informática Industrial; UTFPR
• Palavras-chave em Português: Lingua de sinais - Brasil; Processamento de sinais; Sistemas de reconhecimento de padrões; Eletromiografia; Surdos - Meios de comunicação; Percepção de padrões; Sign language - Brasil; Signal processing; Pattern recognition systems; Electromyography; Deaf - Means of communication; Pattern perception; CNPQ::ENGENHARIAS::ENGENHARIA BIOMEDICA; Engenharia Elétrica
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/25098

Sign Language Recognition Systems are developed to ease the communication between deaf people and speakers, which identify and translate the signals to local language. Among the used techniques for this purpose, surface Electromyography (sEMG) signals have been highlighted.SEMG is the technique of acquisition of muscle signals and it has presented high accuracy and feasibility in applications for gesture recognition. This thesis presents the study of processing stages of sEMG to recognize the gesture for Brazilian Sign Language (Libras). Signs of 26 letters of alphabet in Libras were acquired for 12 volunteers using a commercial armband MyoTM. Posteriorly, it was evaluated the stages of segmentation, feature extraction, and classification in the process of pattern recognition. In segmentation, the window length and overlapping rate were parameters evaluated. In feature extraction stage, it was evaluated the contribution of each feature and for feature set (both presented in similar works and new proposed sets). It was evaluated different classifiers and their behavior for the data from all volunteers, for the contribution of each volunteer, and for the acquisition trials. Reduction dimensionality techniques were proposed to reduce the number of attributes in the classifier inputs. Accuracies above 90% were obtained, reaching to 99% with the following combinations: window length of 1.75 s, overlapping rate of 12.5% with the feature set of MFL, RMS, TTP, and MNP, classifier based in Random Forests and up to six dimensions obtained by t-SNE method and reduction of channels and features.