SFDrones
SFDrones: Self-flying Drones with Visual Navigation, Interaction and Self-supervised Multi-layered Scene Understanding

We are a specialized team within the SpaceTime Vision and Robotics Laboratory focused on developing efficient vision and machine learning solutions for relatively small and low-cost fully UAVs, capable to perform autonomously, without any GPS access and using mainly monocular vision cameras, high level tasks such as: navigation and exploration of unknown environments, full scene understanding at different levels of interpretation (detecting objects and activities, semantically segmenting different regions, building a 3D semantic scene model), autonomously finding and following different targets, as well as returning home safely. All our solutions are focused on self-supervised methods in which small size student networks learn from different types of neural or analytical teacher procedures. Moreover, our solutions are developed and tested both in virtual, digital twin environments as well as the real world, for obtaining the best possible efficiency, robustness and accuracy.

Our mission is to create artificial intelligence for self-flying drones that is efficient, trustworthy, highly accurate and self-supervised.

SFDrones should be able to perform the following completely autonomously:

• • Fly safely by themselves in challenging unknown environments using as few sensors as possible (e.g. monocular vision only)

  • Have a full understanding of the scene in space and time, at the level of geometry and semantics: recognize 3D structures, understand objects movements, activities and interactions

  • Discover unknown objects, detect anomalies and abnormal events

  • Find, track, follow precisely a given target

  • Understand commands given by specific users from a distance using cameras only (e.g. visual gestures, different visual, thermal or hyperspectral patterns)

  • Return home using only cameras (mainly RGB vision), with no GPS or other communication from the base.