Reconhecimento de gestos estáticos utilizando redes neurais convolucionais

Abstract

Human gestures can complement the way we communicate and express ideas, they can be used as specific actions in a certain type of human-machine interaction, such as when manipulating a computer in front of a camera. For this process to work correctly, it is necessary to follow some steps, such as capturing images, separating and defining the objects of interest, extracting relevant information and classifying data. This dissertation proposes a method of gesture recognition using convolutional neural networks. The method uses image processing techniques, such as color segmentation, application of morphological filters and contour extractors and, for the extraction of characteristics and classification, the use of convolutional neural networks is adopted. The images used are obtained from two main image datasets with 24 ASL gestures. An own base built during the development of this work and another base already available in the literature. The results were obtained and compared with some evaluation metrics and proved to be promising for the classification of images of static gestures. These results are compared with other CNN architectures and another methods known from the literature.