Vehicle classification based on binary images of optical barriers using convolutional neural networks.

• Ano de defesa: 2022
• Autor(a) principal: Barreyro, Joaquin
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Axle count; Data augmentation; Intelligent transport systems; Praças de pedágio; Redes neurais; Sistemas inteligentes de transporte; Transfer learning; Vehicle classification; Veículos (Classificação)
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/3/3142/tde-27052022-081725/

Intelligent Transport Systems are tools that facilitate the acquisition, processing, integration, and availability of information to manage the road transport network. A part of these systems is dedicated to identifying and classifying vehicles. The regulatory and inspection agencies of transport in Brazil define a set of vehicle categories based on the number of axles and double wheels. There are different approaches to solving the problem of classifying vehicles. The best score among the solutions found varies between 95% and 99%. These strategies use discrete sensors installed on the pavement, activated by mechanical effort, in the vehicle wheel\'s passage. This approach has some drawbacks, such as the impossibility of counting suspended axles, time degradation, and reliability loss. Thus, it is necessary to continually develop and improve the tools and computational techniques that allow the automatic classification of vehicles traveling on highways. We propose developing methods and computational models to improve vehicle axle counting in traffic analysis systems - a classification method based on binary images of vehicle profiles extracted from a curtain composed of optical sensors. The proposed method uses a modified AlexNet convolutional neural network. The last layers of the network were adapted to the vehicle classification problem. To train the network, we use transfer learning and data augmentation techniques. The method was tested using 5329 images consisting of 11 vehicle categories. Results achieved 98.39% accuracy, indicating that replacing a set of sensors by processing data only from the optical barrier may be a feasible alternative.