| Resumo: | Recent technological progress made over the last decades in the field of Computer Vision has introduced new methods and algorithms with ever increasing performance results. Particularly, the emergence of machine learning algorithms enabled class based object detection on live video feeds. Alongside these advances, Unmanned Aerial Vehicles (more commonly known as drones), have also experienced advancements in both hardware miniaturization and software optimization. Thanks to these improvements, drones have emerged from their military usage based background and are now both used by the general public and the scientific community for applications as distinct as aerial photography and environmental monitoring. This dissertation aims to take advantage of these recent technological advancements and apply state of the art machine learning algorithms in order to create a Unmanned Aerial Vehicle (UAV) based network architecture capable of performing real time people tracking through image detection. To perform object detection, two distinct machine learning algorithms are presented. The first one uses an SVM based approach, while the second one uses an Convolutional Neural Network (CNN) based architecture. Both methods will be evaluated using an image dataset created for the purposes of this dissertation’s work. The evaluations performed regarding the object detectors performance showed that the method using a CNN based architecture was the best both in terms of processing time required and detection accuracy, and therefore, the most suitable method for our implementation. The developed network architecture was tested in a live scenario context, with the results showing that the system is capable of performing people tracking at average walking speeds.
|
|---|