OBJECT DETECTION AND TRACKING ALGORITHMS BASED ON KLT FEATURE TRACKER FOR A SENSE AND AVOID SYSTEM

This paper addresses strategies of image processing algorithms under evaluation in the context of TERSA project, aimed at the development of a Sense and Avoid system for a tactical MINIUAV (MTOW<30 kg) in a joint collaboration between UNIPI and Sky Eye Systems. Firstly, the general sequence of image processing operations carried out by the sense subsystem is explained. Afterwards, particular focus is posed on the notorious Kanade-Lucas-Tomasi (KLT) feature tracker, whose improved implementation is thoroughly described. The algorithm finds correspondences between frames of corner points extracted from the object of interest, using Harris corner detector. A random sample consensus (RANSAC) approach is subsequently employed in order to infer the motion of the object centroid from the set of corresponding feature points pairs. The algorithm was tested on aeronautical footage taken from the web, portraying aircraft recorded by onboard cameras mounted on other aircraft, so as to be representative for the purposes of a SAA operation. Preliminary results show that the implementation runs in real time, is robust to mild lightning variation and successfully tracks non cooperative intruders in complex scenarios (scattered clouds, mountains, and urban ground pattern), with moving background. Finally, it is shown how the algorithm can be applied in order to augment the robustness of background-subtraction based object detectors, by providing motion compensation via the computation of an homography warping matrix using pairs of corresponding features points.