Classificação do número de pessoas nas motos em imagens de trânsito corretamente segmentadas

Abstract

A common traffic violation is the excess passengers on motorcycles. In order to detect this kind of violation, it is important to conduct a low error rate counting of passengers on motorcycles. One way to accomplish this is to count passengers using a classifier, having as input the characteristics of the image region containing the motorcycle and passengers (segmented region). In this work we verify if this is a good approach, given that the image is correctly segmented. We conclude that it is possible to obtain ROC area above 95% by employing a neural network and one of the models proposed in this work to represent the segmented region.