Interface cérebro-computador para detectar sonolência usando a abordagem de aprendizagem baseada em instâncias

• Ano de defesa: 2016
• Autor(a) principal: Jalilifard, Amir
• Orientador(a): Pizzolato, Ednaldo Brigante
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação - PPGCC
• Palavras-chave em Inglês: Brain computer interface; EEG classification; Drowsiness detection; K-Nearest neighbors; EEG signal processing; Kd-trees; Random forest; Feature selection
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/10703

Brain-Computer Interface (BCI) is a way to establish a communication between brain and computers. It allows the users to control a computer system and even an environment without moving a muscle or it allows the computer to record and analyze the user’s neuropsychological brain activities. Clearly, the range of BCI applications has increased in the past decade due to the use of modern machine learning and signal processing methods. Among various applications of BCI, lately, the use of EEG records for driver safety has been considered by some researchers. Drowsy driving is a major cause of many traffic accidents. The aim of this work is to develop an automatic drowsiness detection system using an efficient k-nearest neighbors (K-NN) algorithm. First, the distribution of power in time-frequency space was obtained using short-time Fourier transform (STFT) and then, the mean value of power during time-segments of 0.5 second was calculated for each EEG sub-band. In addition, standard deviation (SD) and Shanon entropy related to each time-segment were computed from time-domain. Finally, 52 features were extracted. Random forest algorithm was applied over the extracted data, aiming to choose the most informative subset of features. A total of 11 features were selected in order to classify drowsiness and alertness. The Kd-tree algorithm was used as the nearest neighbors search algorithm so as to have a fast classifier. Our experimental results show that drowsiness can be classified efficiently with 91% accuracy using the methods and materials proposed in this paper. We also compared the classification results obtained by K-NN (as an instance-based learning algorithm) with four well-known classifiers including decision tree, support vector machine, logistic regression and Naive Bayes.