Experimental assessment of vision-based sensing for small UAS sense and avoid

This paper presents first results of an experimental flight-test campaign aimed to gather data for performance assessment of non-cooperative Sense and Avoid architectures for small Unmanned Aircraft Systems (UAS). The attention is here focused on vision-based approaches. An innovative sensing technique is proposed which exploits a Deep Learning (DL) network as the main processing block of the detector algorithm, and a multi-temporal strategy for track generation and confirmation. Both the detection and tracking phases foresee ad-hoc solutions to deal with the presence of intruders either above or below the horizon. Two customized small quadcopters, equipped with high-resolution color cameras, are used to reproduce in flight low-altitude, near-collision scenarios characterized by different speed and height above ground, thus being able to act simultaneously as ownship and intruder. Results demonstrate the capability of the DL-based detector to provide maximum declaration range around 300 m and 100 m, above and below the horizon, respectively. The tracker can robustly produce firm track of the intruder while rejecting many false positives, particularly occurring in below-the-horizon scenarios.