Classificação e detecção de pessoas em ambientes não controlados utilizando redes neurais convolucionais

Abstract

Detecting and counting people in complex environments is a challenge in Computer Vision. This problem becomes even more difficult when it comes to experiments in uncontrolled environments, presenting several obstacles that substantially influence the accuracy of the process. Total or partial occlusion, non-uniformity of illumination, noise, pose variation, scale variation are some of the most common problems to be addressed. One of the ways used to mitigate the influence of these factors is by training detection algorithms using as many labeled examples as possible in the various situations that can occur in everyday situations. Additionally, image preprocessing methods may be applied to mitigate some of the obstacles mentioned, notably the non-uniformity of illumination. The present work proposes to investigate solutions using Convolutional Neural Networks (CNN) for the detection and classification of faces in complex and uncontrolled environments. Variants of a classical CNN architecture called LeNet were evaluated, as well as the application of CLAHE preprocessing to the images of interest. In the empirical tests, the developed models obtained an average accuracy above 98%.