Detecção de embarcações por imagens nos rios da Amazônia

Abstract

Fluvial monitoring is an intensive and necessary task used as a support to avoid possible threats such as piracy and environmental aggression. Usually, this task is performed manually by a human operator, who analyzes video monitors for long duration, which leads the task to be long, tedious and prone to errors. In the Amazon rainforest, fluvial monitoring is even more challenging due to the large territory of the region and due to the fact that the largest river basin in the world is in this region. Remote sensing, usually based on aerial images obtained by satellite or unmanned/manned aerial vehicles is one possibility to deal with this problem. In this context, we propose in this work a computational method focused on detecting the movement of vessels in a typical scene of the Amazon rivers. This new method is necessary since current solutions are not suited to the deal with the Amazon scenario due to several reasons, for instance, the diversity of the background of the Amazon region. In addition, several methods available in the literature were designed in the context of open sea scenarios, where there are usually only vessels and water in the scenes. In this work, digital image processing techniques are employed to detect moving objects in the scene, while machine learning algorithms are used to indicate the probability about the class of the observed object, i.e. whether or not the detected object is a boat. This double detection process is intended to reduce false alerts provided by the proposed method. The experiments conducted in this work show that the proposed method achieved 79% of accuracy and 91% precision rate when all images are considered, as well as a 71% precision rate when only scenes with vessels are analyzed. These results indicate the effectiveness of the proposed method when compared to other strategies.