Pedestrian identification through Deep Learning with classical and modern architecture of Convolutional Neural Networks

Abstract

This article, refers to the research carried out at the National University Micaela Bastidas (UNAMBA), whose specific objectives were: To determine in a first stage of learning the proportion of accuracy of a classical architecture of Convolutionary Neural Network (CNN) in the identification of UNAMBA peoples, to determine in a second stage the proportion of precision in a modern architecture of RNC and finally compare the first stage with the second, to find the highest proportion. The training was given with a quantity of 242 people. Therefore, 27,996 images had to be generated through the technique of Video Scraping and data augmentation, which were divided into 19,700 images for training and 8,296 for the validation. Regarding the results in the first stage, a modified model VGG16-UNAMBA is proposed, with which a ratio of 0.9721 accuracy was achieved; while in the second stage it is proposed to DenseNet121-UNAMBA, with which a proportion of 0.9943 accuracy was achieved. Coming to the conclusion that the use of deep learning allows UNAMBA staff to be identified in a high proportion of accuracy.