Loading...
Publication
Deteção de incêndios florestais com recurso a deep learning e visão computacional
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| DM_João Alves | 10.53 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Nos países mediterrânicos do sul da Europa, os incêndios têm tendência a
agravar-se devido às alterações climatéricas, resultado do aumento global da
temperatura. Este trabalho apresenta um sistema de deteção automática de
incêndios que procurar minimizar este problema ao permitir a sua deteção
precoce.
O sistema processa imagens de ambiente florestal e ´e capaz de detetar a
presença de incendio em fase inicial, através do fumo ou das chamas. Permite
também estimar a área em ignição para que se possa avaliar a sua dimensão.
Foi desenvolvido com a finalidade de poder ser aplicado em camaras moveis,
como em drones, e estáticas, como em torres de vigia. Foi utilizado um modelo
de Deep Learning e um classificador de Machine Learning na tarefa de
classificação das imagens, e técnicas de Visão Computacional no processo de
deteção da área das chamas. Como modelo de Deep Learning, foi utilizada
a Deep Convolutional Neural Network Inception-V3 para extrair descritores
de uma imagem que, de seguida, são utilizados para treinar um classificador,
o Logistic Regression. Em relação à Visão Computacional, são aplicadas
técnicas de processamento da imagem ao nível da cor.
De forma a perceber quais os tipos de situações que podem influenciar na
tarefa de classificação, o dataset proposto é composto por imagens com metadados onde se encontram anotadas as características (e.g. chamas, fumo,
nevoeiro, nuvens, elementos humanos,...) presentes em cada imagem.
O sistema obteve uma precisão de classificação de 94.1% em cenários diurnos
e de 94.8% em cenários noturnos. Apresenta uma boa precisão na estimação
da área das chamas em comparação com outras abordagens na bibliografia,
reduzindo o número de falsos positivos.
In southern Mediterranean countries, fires are likely to worsen as a result of global warming. This work presents an automatic fire detection system that seeks to minimize this problem by allowing its early detection. The system processes images of the forest environment and is able to detect the presence of an early stage fire through smoke or flames. It also allows estimating the area under ignition so that its size can be evaluated. It was developed with the purpose of being able to be applied in mobile cameras, as in drones, and static, as in watchtowers. A Deep Learning model and a Machine Learning classifier were used in the image classification task, and Computer Vision techniques in the process of detecting the area of the flames. As a Deep Learning model, the Deep Convolutional Neural Network Inception-V3 was used to extract descriptors from an image that are then used to train a classifier, the Logistic Regression. Regarding Computer Vision, image processing techniques are applied at color level. In order to understand the types of situations that may influence the classification task, the proposed dataset is composed of images with metadata where are recorded the characteristics (eg flames, smoke, fog, clouds, human elements, etc.) present in each image. The system obtained a classification accuracy of 94.1% in daytime scenarios and 94.8 % in nighttime scenarios. It presents good accuracy in the estimation of the area of the flames compared to other approaches in the literature, reducing the number of false positives.
In southern Mediterranean countries, fires are likely to worsen as a result of global warming. This work presents an automatic fire detection system that seeks to minimize this problem by allowing its early detection. The system processes images of the forest environment and is able to detect the presence of an early stage fire through smoke or flames. It also allows estimating the area under ignition so that its size can be evaluated. It was developed with the purpose of being able to be applied in mobile cameras, as in drones, and static, as in watchtowers. A Deep Learning model and a Machine Learning classifier were used in the image classification task, and Computer Vision techniques in the process of detecting the area of the flames. As a Deep Learning model, the Deep Convolutional Neural Network Inception-V3 was used to extract descriptors from an image that are then used to train a classifier, the Logistic Regression. Regarding Computer Vision, image processing techniques are applied at color level. In order to understand the types of situations that may influence the classification task, the proposed dataset is composed of images with metadata where are recorded the characteristics (eg flames, smoke, fog, clouds, human elements, etc.) present in each image. The system obtained a classification accuracy of 94.1% in daytime scenarios and 94.8 % in nighttime scenarios. It presents good accuracy in the estimation of the area of the flames compared to other approaches in the literature, reducing the number of false positives.