Comparison of Computer Vision Approaches for Recognition of Scenarios Suspected of Being Mosquito Breeding Sites in Aerial Images Acquired by UAVs

Resumo

The use of unmanned aerial vehicles (UAVs) for acquisition of aerial images to support health
surveillance teams in activities of combatting the mosquito breeding sites has increased a lot in recent years.
However, it is still common the manual analysis of such images, requiring much time of the health workers. In this
work we investigate two state-of-the-art computer vision approaches which can be employed for recognition of
scenarios suspect of being potential mosquito breeding sites from aerial images acquired by UAVs. The first
approach, named as BoVW+SVM, is based on Bag of Visual Words (BoVW) technique combined with the
Support Vector Machine (SVM) classifier, while the second approach is based on a model of convolutional neural
network (CNN) known as YOLO (You Only Look Once). For conducting the experiments, in which the
approaches were compared in terms of the mAP-50 measure, we employed a dataset containing 230 images,
acquired in urban regions of the city of São Paulo, which contemplate real and simulated suspected scenarios
(gutters and roofs with accumulation of objects, open-air inorganic garbage containing old tires, old tires, pet
bottles, plastic and paper packaging and other open containers that can accumulate water). The results obtained
by YOLO were much superior to those obtained by BoVW+SVM, in terms of precision and processing speed,
demonstrating that this CNN model can be employed to compose a computer vision system for automatic
inspections in real time.