Unmanned Aerial Vehicles (UAVs), although hardly a new technology, have recently gained a prominent role in many industries being widely used not only among enthusiastic consumers, but also in high demanding professional situations, and will have a massive societal impact over the coming years. However, the operation of UAVs is fraught with serious safety risks, such as collisions with dynamic obstacles (birds, other UAVs, or randomly thrown objects). These collision scenarios are complex to analyze in real-time, sometimes being computationally impossible to solve with existing State of the Art (SoA) algorithms, making the use of UAVs an operational hazard and therefore significantly reducing their commercial applicability in urban environments. In this work, a conceptual framework for both stand-alone and swarm (networked) UAVs is introduced, with a focus on the architectural requirements of the collision avoidance subsystem to achieve acceptable levels of safety and reliability. The SoA principles for collision avoidance against stationary objects are reviewed and a novel approach is described, using deep learning techniques to solve the computational intensive problem of real-time collision avoidance with dynamic objects. The proposed framework includes a web-interface allowing the full control of UAVs as remote clients with a supervisor cloud-based platform. The feasibility of the proposed approach was demonstrated through experimental tests using a UAV, developed from scratch using the proposed framework. Test flight results are presented for an autonomous UAV monitored from multiple countries across the world.

Artificial Intelligence is evolving at an accelerating pace alongside the increasing number of large datasets due to vast number of image data on the Internet. Unnamed Aircraft Vehicles (UAVs) are also a new trend that will have a huge impact over the next years. The use of UAVs arises some safety issues, such as collisions with dynamic obstacles like birds, other planes, or random thrown objects. Those are complex and sometimes impossible to avoid with stateof-the-art algorithms, representing a threat to the applications. In this article, a new video dataset of collisions, entitled ColANet, aims to provide a base for training new Machine Learning algorithms for handling the problem of avoiding collisions with high efficiency and robustness. It is also shown that using this dataset is easy to build new neural network models and test them.

No abstract