Re-Identificação de pessoas em imagens digitais utilizando redes neurais siamesas e triplet baseadas em uma rede neural convolucional e um autoencoder

Abstract

In environments monitored by security cameras, the problem of identifying if a person being watched has ever been in the scene or not, independent of the system uses one or more cameras, is called person re-identification. This problem is considered challenging, since the images obtained by cameras are subject to many variations, such as lighting and perspective. In addition, people in pictures may undergo transformations and partial occlusions. This work aims to develop two approaches for person re-identification robust to these variations, through deep learning techniques. The first approach proposed uses a Siamese neural network architecture, composed of two identical subnets, this model receives two input images that may or may not be from the same person. The second approach consists of a triplet neural network, with three identical subnets, which receives a reference image from a certain person, a second image from the same person and another image from a different person. Both networks have identical subnets, formed by a convolutional neural network that will extract general characteristics from each image and an autoencoder network, responsible for dealing with the great variations that the input images may undergo. To analyze and compare the developed networks, three datasets were used, and the metrics chosen for analysis were accuracy and the CMC curve. Experiments carried out proved an improvement up to 71.05% in the results with the use of the autoencoder in the subnets. Also, the experiments showed a superiority of the triplet neural network developed in this work to the siamese neural network and other state-of-the-art methods.