On data augmentation techniques for the automatic detection of mosquito breeding grounds using videos

Abstract

This work discusses data augmentation techniques for detecting mosquito breed- ing grounds using videos recorded by a drone. Firstly, a study regarding mosquito- related diseases is presented in order to propose a computer vision system capable of automatically detecting disease-related objects, such as water tanks, tires, and bottles. A database composed of six aerial videos containing breeding-related ob- jects is devised, including its planning and execution (recording and annotation) stages. However, due to the difficulty of obtaining extensive records of real sce- narios, artificial data augmentation techniques are presented. This work addresses three methods of inserting images of the objects into videos in order to increase the number of objects in the training set. Finally, a convolutional neural network detec- tor is used to evaluate these techniques, indicating that artificial data augmentation reduces overfitting, improving the overall detection performance by the proposed network.