Aerolevantamento com VANT no auxílio contra o abigeato

Abstract

Detecting people in digital images is a recurring task for the development of various applications. The present work has the objective of proposing a method for identifying people in aerial images with the use of UAVs to support the producer and the specialized agencies in the fight against rustling. The proposed solution executes processes of creation of georeferenced mosaics and classification of objects. The OpenDroneMap tool and the MapsMadeEasy Web utility were tested for mosaic generation. The GDAL open source library was also studied. The training of cascading descriptors was done using software that presents the parameters implemented in the OpenCV library in a graphical interface, thus facilitating their manipulation. The HOG descriptor used is available in the OpenCV library. For testing purposes, two Haar and LBP classifiers with 8, 15 and 20 stages were created, all using the same set of negative images. However, Haar and LBP classifiers were generated using a first set of positive images and later, other Haar and LBP classifiers were generated using a second set of positive images. Given the results it can be concluded that the aerial imaging tools generate satisfactory results. It has also been noted that cascades of classifiers effectively detect people, but also generate a high amount of false positives. With the tests performed, it may be noted in practice that effectively Haar classifiers require more time for training, but their hit rate is better. It was also observed that the number of stages of a cascade should be adequately sized, since classifiers with 20 stages obtained very poor results, generating only false positives.