| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 6.45 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
A reparação de defeitos presentes no asfalto de estradas constitui uma importante tarefa para aumentar o conforto e a segurança dos utilizadores das rodovias. Além disso, permite evitar níveis intensos de desgaste, no asfalto, que acarretam também gastos de reparação mais elevados. Em muitos casos, os responsáveis pela manutenção das rodovias pavimentadas percebem ou observam os defeitos nas rodovias somente quando ocorre um grave acidente. E isto poderia ser evitado se houve uma maior preocupação com deteção e uma reparação rápida dessas irregularidades nas rodovias. Um modo mais económico e rentável, é impedir que essas falhas nas rodovias não cheguem a serem um obstáculo. Para isso, é necessária a identificação dos defeitos antes mesmo de acontecer um acidente para custear o preço das manutenções e dos problemas de insegurança nas rodovias. Automatizar esse processo com técnicas de reconhecimento de objetos pode trazer uma maior segurança no trânsito, em razão de uma melhor organização das rodovias que precisam de manutenções. O objetivo deste trabalho é construir um algoritmo de reconhecimento de defeitos em imagens e vídeos digitais de asfaltos por meio de Deep learning. Utilizaremos o Yolo como nosso detetor de objetos, pois consegue prever objetos em poucos segundos. O nosso conjunto de dados é composto por 613 imagens, distribuídas por 317 imagens de rachaduras e 296 imagens de buracos. Apesar de o reconhecimento de defeito ser considerada uma tarefa difícil, obtivemos uma acurácia média dde 96% em nosso melhor modelo.
Repairing defects present in road asphalt is an important task to increase the comfort and safety of road users. In addition, it allows to avoid intense levels of wear on the asphalt, which also lead to higher repair costs. In many cases, those responsible for paved road maintenance notice or observe road defects only when a serious accident occurs. And this could be avoided if there was greater concern with detecting and quickly repairing these irregularities on the highways. A more economical and profitable way is to prevent these road failures from becoming an obstacle. For this, it is necessary to identify the defects even before an accident occurs to defray the cost of maintenance and insecurity problems on the highways. Automating this process with object recognition techniques can bring greater safety in traffic, due to a better organization of roads that need maintenance. The objective of this work is to build a defect recognition algorithm in digital images and videos of asphalts through Deep learning. We will use Yolo as our object detector, as it can predict objects in a few seconds. Our dataset consists of 613 images, spread over 317 crack images and 296 hole images. Although defect recognition is considered a difficult task, we obtained an average accuracy of 96% in our best model.
Repairing defects present in road asphalt is an important task to increase the comfort and safety of road users. In addition, it allows to avoid intense levels of wear on the asphalt, which also lead to higher repair costs. In many cases, those responsible for paved road maintenance notice or observe road defects only when a serious accident occurs. And this could be avoided if there was greater concern with detecting and quickly repairing these irregularities on the highways. A more economical and profitable way is to prevent these road failures from becoming an obstacle. For this, it is necessary to identify the defects even before an accident occurs to defray the cost of maintenance and insecurity problems on the highways. Automating this process with object recognition techniques can bring greater safety in traffic, due to a better organization of roads that need maintenance. The objective of this work is to build a defect recognition algorithm in digital images and videos of asphalts through Deep learning. We will use Yolo as our object detector, as it can predict objects in a few seconds. Our dataset consists of 613 images, spread over 317 crack images and 296 hole images. Although defect recognition is considered a difficult task, we obtained an average accuracy of 96% in our best model.
Description
Mestrado de dupla diplomação com a Associação Educativa Evangélica - UniEvangélica
Keywords
Yolo Deteção de objetos Deep learning