More and more unmanned aerial vehicles (UAV), also known as drones, are present in the national airspace either with remote control or with fully autonomous on-board system. In the latter case, the sense (based on e.g. visual, radar or multi-modal input) and avoid capability is crucial in the integration of such devices. In my research I focus on the automatic identification of the UAVs with convolutional neural networks (CNN).
This topic is part of the project called Real Flight Demonstration of Monocular Image- Based Aircraft Sense and Avoid from the Computer and Automation Research Institute (SZTAKI) of the Hungarian Academy of Sciences (MTA). The main goal is to develop a monocular camera based on-board system, which processes image data in real-time and initiates collision avoidance if required.
The core of my solution for the task are deep neural networks (DNN). These artificial intelligence models require quite a lot of computational capacity. In contrast to this, there are significantly less resources available in embedded environments. The learning part could be done previously on strong graphical processing units (GPU), but the deployment phase of the net should also be fast for real-time processing. Further requirement for machine learning models is data. In this case the small size of the available dataset is another challenge to solve.
In my research I review the potential techniques for the problems recommended by the literature and I develop an adequate method for this specific system. Furthermore, I compare different learning techniques for convolutional neural architectures with respect to the given restrictions. Last but not least the evaluation of the solution is also part of my work. The results are set side by side for the different models.