Automatização do reconhecimento de buracos em rodovias usando inteligência computacional

• Ano de defesa: 2020
• Autor(a) principal: Schmidt, Leonardo de Abreu
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Santa Maria; Brasil; Ciência da Computação; UFSM; Programa de Pós-Graduação em Ciência da Computação; Centro de Tecnologia
• Palavras-chave em Português: Redes neurais artificiais; Máquinas de vetores de suporte; Detecçãode buracos; Pavimentação de rodovias; Artificial neural network; Support vector machine; VANETs; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: http://repositorio.ufsm.br/handle/1/22279

Potholes in asphalt pavements and highways are a well-known problem, which is aggravated by the growing development of autonomous vehicle technology. This technology, therefore, needs to incorporate efficient automatic pothole detection systems. In fact, there are different methods for automatic pothole detection, however, these methods still need to be further evaluated both in terms of their accuracy and speed of response in order to be used in real contexts. In order to contribute with a step in the direction of this evaluation, this work presents the proposal of a solution for the detection of potholes in images, being the detection of the existence of a pothole and the determination of the same in the image, the objective of this work. The work has two main parts, the first being an image classification architecture, which uses the Histogram of Oriented Gradients (HOG) technique as an extraction of image characteristics, and two classifiers; Artificial Neural Networks and Support Vector Machines. The objective is to optimize parameters and determine the best classifier for the problem. The second part presents a proposal for architecture to detect potholes in images, determining not only the existence but also the location of a pothole in the image. This task uses texture descriptors from Haralick in a system of grid applied to the images, added to the classification architecture of the first part. The processing times of these architectures were also evaluated. The results presented show that architecture using the Artificial Neural Network classifier is the best option for pothole detection, reaching 83% IoU (Intersection of Union) in addition to 73% accuracy and 93% precision. Processing time is also favorable to the Neural Network classifier reaching about 9 frames per second (FPS).