intro.tex

\chapter{Introduction}
\label{chap:intro}

The \emph{Aedes aegypti} is the main vector of several diseases caused by the arbovirus (acronym for arthropod-borne virus), such as dengue, zika, chikungunya and, more recently in Brazil, urban yellow fever~\cite{ruckert2017, web:MSyellowfever}.
Zika virus disease can be quite dangerous for pregnant women due to its correlation with the microcephaly, a congenital fetus brain malformation~\cite{petersen2016zika, web:who2016zika}.
Among these diseases, dengue is the one that causes the most deaths, with about 390 million people infected per year in the world~\cite{bhatt2013global}.
Yellow fever also has a high rate of mortality and chikungunya can incapacitate those infected for long periods.

By considering the high rates of lethality and eradication difficulty, arboviruses transmitted by \emph{Aedes aegypti} are one of the leading global health problems.
Wherefore, the World Health Organization (WHO) launched, in 2012, a comprehensive strategy for dengue control and prevention~\cite{world2012global}, whose one of the goals is to reduce disease cases by 25\% by 2020.
Unfortunately, combat tools are still limited: the dengue vaccine remains in the improvement phase, and the fumes against mosquitoes are ineffective~\cite{newton1992model}.
Thus, the current best form of combat is through the control and elimination of possible mosquito foci proliferation, which acts directly in the prevention of all these diseases.
Given that the \emph{Aedes aegypti} reproduces in clean and stagnant water, the main mosquito foci are open water bowls, gutters, tires, bottles, plant pots, and any container that can collect water.


As a result, monitoring and controlling the mosquito without proper technical support is expensive, time-consuming and therefore inefficient.
For that reason, allying the knowledge of an expert with a tool that accelerates the search for potential mosquito foci and towards a more precise work is extremely important in the current scenario.
Thus, using images and videos captured by an unmanned aerial vehicle (UAV), better known as a drone, with several sensors and camera may be a reasonable approach.
The objective is to identify objects with high potential of being a mosquito breeding site.
This technology has already been used by organizations to visually inspect difficult-to-reach sites in order to locate such breeding spots.
In this process, the acquired videos are examined by a specialist, which makes the procedure time-consuming and tiring, what may lead to failures.

In this sense, a possible solution to increase efficiency is to apply machine learning techniques to automate the analysis process, helping the specialist in the decision-making action~\cite{casfinal2018}.
After this fast analysis, potential breeding sites can be treated or removed by a team of agents, as usual.
It is known, through a local study~\cite{tun2009reducing}, that treating the most productive water container types in a region has a comparable result as treating them all in the same region.
This roughly cuts by half the number of containers to be treated while keeping similar effectiveness, drastically reducing the potential for epidemic development.
In the case of Nova Iguaçu, a city located in the state of Rio de Janeiro, the reservoirs listed with high potential were, according to~\cite{Lagrotta2006}: water tanks, glass and plastic bottles, buckets, tires, and external drains.
The initial goal then becomes to automatically recognize as many of these objects as possible in videos or images acquired from a UAV
to reduce the amount of images or videos the agents would need to visually evaluate.
In the future, we plan to expand this work and provide an intelligent decision support tool for agents, generating heat maps highlighting the places with more risk, thereby increasing the effective area of action.

\section{Main contributions of this work}
%
Throughout this work,
we describe the problem of automatic detection of potential mosquito breeding sites using aerial images acquired from a UAV
and propose a complete solution.
%
We construct a video dataset containing objects considered as potential mosquito breeding sites of the \Aedes in several scenarios.
%
We apply a methodology to calibrate the camera of our UAV and reduce the lens distortions that may be expanded for other drone models.
%
We train a state-of-the-art object detector using a small dataset in order to detect tires, considered big productive water containers.
%

\section{Dissertation organization}
%
In Chapter~\ref{chap:mosquito}, we do a review on the {\it Aedes aegypti}, including biologic aspects, transmitted diseases and sequelae, and preferred ground sites.
Also, we point out some statistics related to the theme and government plans to combat the transmitter.
In Chapter~\ref{chap:database}, we make a literature review of related themes, in order to see how several techniques of machine learning can be used to address the problem.
We also present a new dataset containing the main objects considered as potential foci of the \Aedes in several scenarios.
In Chapter~\ref{chap:system}, since we are interested in detecting specific objects, we describe the state-of-the-art algorithm employed in order to accomplish this task.
We discuss from classical object detections up to recent deep-learning-based models, particularly the one employed in this work.
The evaluation method, implementation details, and results for the method used are discussed in Chapter~\ref{chap:results}; and we finally conclude at Chapter~\ref{chap:conclusions}.