Reconhecimento de indivíduos em vídeo utilizando estimativa de pose humana

• Autor(a) principal: Silva, José Carlos da
• Data de Publicação: 2020
• Tipo de documento: Trabalho de conclusão de curso
• Idioma: por
• Título da fonte: Repositório Digital da Universidade Federal Rural do Semi-Árido (RDU)
• Texto Completo: https://repositorio.ufersa.edu.br/handle/prefix/6647
• topic: OpenPose; Redes Neurais Artificiais; OpenCV; Reconhecimento de Indivíduos; Visão Computacional; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO

The need to identify people has always been present in society, being used in several areas, of which we can mention the security industry as a highlight. This task has become faster due to the use of computing, which has increased its use, with facial recognition being among the most accurate and used techniques. Although facial recognition is effective, it requires high resolution images with good lighting, in addition to requiring that, in most applications, the face be captured from the front. These requirements are not always met and, for many situations, we need to resort to other techniques for identifying individuals, using other parts of the body. In this work we discuss the possibility of recognizing individuals in images and videos, using some key points of their body, provided by an OpenPose library. The goal is to design an Artificial Neural Network capable of classifying the individuals present in an image using the body structure formed by these points. The methodology used consists in the definition of some proportion relations between the parts of the body, calculated from the distance between the key points, so that a consistent classification model can be created, with a set of weights capable of recognizing an individual . The tests performed showed good results for individuals who had their classifier trained, recognizing these with an accuracy of over 90 %. In tests with images of individuals who did not have a trained classifier on the network, the rate of classification errors was very high, where an unknown individual was classified as an individual with trained classifier.