Identificação de imperfeições em pavimentos flexíveis de rodovias por meio de imagens de vídeo e aprendizado de máquina profundo

• Ano de defesa: 2024
• Autor(a) principal: Vita, Diego Eduardo de
• Orientador(a): Melanda, Edson Augusto
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Engenharia Urbana - PPGEU
• Palavras-chave em Português: Análise de Imagens baseada em objetos; Aprendizado profundo; Detecção automatizada de objetos; Pavimentos flexíveis; Gerência de pavimentos
• Palavras-chave em Inglês: Object-based image analysis; Deep learning; Automated object detection; Flexible pavements; Pavement management
• Área do conhecimento CNPq: ENGENHARIAS::ENGENHARIA CIVIL::GEOTECNICA
• Link de acesso: https://hdl.handle.net/20.500.14289/21801

According to data from the National Transportation Confederation in 2023, almost 57% of Brazilian highways Brazilian highways have some kind of pavement-related problem, after analyzing the 43.843 km of state highways and 67.659 km of federal highways. By 2022, around 60% of all cargo transported in the country used the road system. and between 2019 and 2021 alone, there were 195,326 accidents. This shows the importance of the road system, but also, in a way, warns of the conservation conditions that roads must be in, and they also need to be maintained quickly and assertively. and assertiveness in maintenance, which is directly related to efficiency in assessment of sidewalk conditions. We can therefore understand the impact of the slowness in assessing the sidewalk of Brazilian roads. In this sense, great efforts have been efforts have been made to carry out an automated assessment, providing the speed that the the process needs. As a result, computational machine learning machine learning techniques have been used to train artificial intelligence algorithms capable of algorithms capable of identifying patterns through images and, therefore, diagnosing the situation the situation found on the roads using the images analyzed. This paper therefore presents structure, pathologies and sidewalk management. It also the main literature on deep neural networks, convolutional networks and object detection. networks and object detection, and then proposes automated inspection of road sidewalks using using deep learning algorithms, by means of video images. images. The results of this work show the identification of pathologies in 98% of the images selected for a real detection test. The method chosen for dataset enabled the model to process the training stage and then be applied to the test. applied to the test. In addition, the dataset was formulated after the initial detection section, which led to the best section, which led to the best labeling strategy for the case of detecting the pans and cracks. Finally, the average confidence result based on the detected images was promising when the size of the image bank used for training is analyzed. used for training.