Sky Region Obstacle Detection and Tracking for Vision-Based UAS Sense and Avoid

Abstract

A customized detection and tracking algorithm for vision-based non cooperative UAS sense and avoid aimed at obstacles approaching from above the horizon is presented in this paper. The proposed approach comprises two main steps. Specifically, the first processing step is relevant to obstacle detection and tentative tracking for track confirmation and is based on top-hat and bottom-hat morphological filtering, local image analysis for a limited set of regions of interest, and multi-frame processing in stabilized coordinates. Once firm tracking is achieved, template matching and state estimation based on Kalman filtering are used to track the intruder aircraft and estimate its angular position and velocity. An extensive experimental analysis is presented which is based on a large set of flight data gathered in realistic near collision scenarios, in different operating conditions in terms of weather and illumination, and adopting different navigation units onboard the ownship. In particular, the focus is set on flight segments at a range between 3 km and 1.3 km, since the major interest is in understanding algorithm potential for relatively large time to collision. System performance is evaluated in terms of declaration range, probability of correct declaration, average number of false positives, tracking accuracy (angles and angular rates in a stabilized North-East-Down reference frame) and robustness with respect to track loss phenomena. Promising results are achieved regarding the trade-off between declaration range and false alarm probability, while the onboard navigation unit is found to heavily impact tracking accuracy.