Reconhecimento das configurações de mão da língua brasileira de sinais - LIBRAS em imagens de profundidade através da análise de componentes principais e do classificador k-vizinhos mais próximos

Abstract

According to Brazilian Institute of Geography and Statistic IBGE – Instituto Brasileiro de Geografia e Estatística (Censo 2010), Brazil has 9.7 million Brazilians with some degree of hearing impairment or deafness, more than five per cent of the population. For the majority of those persons the main natural communication method is the Brazilian Sign Language (LIBRAS – Língua Brasileira de Sinais), instead of the spoken Portuguese. The computer-aided recognition of signs aims to expand the social and digital inclusion of the deaf community by means of those signs translation into audio or text format. This work presents one of the global LIBRAS parameters recognition, the hand gestures, using k-nearest neighbor’s classifier and 2D²PCA dimensionality reduction technique. A robust and representative data set of daily conditions, with 12.200 depth images from 61 hand gesture, captured by Kinect® sensor, was built. Initially the images were segmented in a preprocessing step, which isolated the region of the right hand from the rest of the body. Seeking to eliminate the data set redundancy, the images were submitted to a dimensionality reduction by (2D) dimensional 2PCA technique determining the most representative forms of data from original pixels linear combination. The classifier k-nearest neighbors was the technique in the final stage in the hand gesture automatic recognition. This classifier could correctly categorize 96.31% of test samples (k = 1 and 10x10 feature matrix). Six hand gesture sets were correctly classified obtaining 100% successful rates