Classificação de veículos baseada em Deep Learning para aplicação em semáforos inteligentes

Abstract

Currently, in the literature, many studies appear every moment in order to reduce human intervention and improve the quality of life, proposing new services, mechanisms through applications, technological innovations and automatic sensors. Aiming at urban mobility, traffic lights are services that are exploited. Algorithms of DL have been widely used for identification and classification of images for decision making in traffic, with the objective of detecting safety and public health vehicles. However, there is a lack of algorithms to classify images into intelligent traffic lights with high precision and quick response. In this research, an image detection system is proposed for different types of services such as security, health and public transport vehicles and common vehicles, integrated with an intelligent traffic light. In addition, a prioritization algorithm based on CTB is also proposed. The detection system is based on a DL algorithm, using an improved model from YOLOv3, which was called the Priority Vehicle Identification Network (PVInet). In addition, a design strategy for the PVInet model is proposed, which presents a high performance in terms of execution time. For the training of the PVInet model, a new BD was created that considers homogeneous images of Brazilian traffic vehicles, since the current BD available in the literature are heterogeneous images. Our solution proposal can help to reduce the waiting time for vehicles with priority on roads controlled by a traffic light, something that does not happen when compared to a current traffic light with fixed waiting times.