Aplicação de algoritmos computacionais para estudo de tráfego utilizando técnicas de visão computacional e deep learning

Abstract

The combination of image processing and data mining has proven to be increasingly effective in solving real-world problems, with these systems being used on a large scale in several fields of study. Given this scenario, one of the main factors that influence the dimensioning of pavements is the traffic it requests during its service life. Thus, the detection and the classification are parameters of great relevance for traffic monitoring and road sizing. Given this perspective, the present work aims to develop a prediction model using neural networks, capable of controlling the detection and identification of vehicles defined by the DNIT project. In this work, a model wastrained with five classes and had the following evaluation metrics as the final result: “Precision = 86%”, “Recall = 84.37%”, “mAP = 51%” and “F1-Score = 75 .37%”. The computational algorithms were based on the You Only Look Once (YOLO) method, and the experimental results were generated from images of urban traffic in Maceió-AL provided by the Municipal Superintendence of Transport and Traffic. Although the final model is not yet ready to beapplied in traffic studies, the present work shows the potential that Computer Vision and Deep Learning techniques have for the transport area.