Realtime Visual Tracking of Aircrafts

Abstract

Aircraft tracking has applications in guided landing, automatic scoring in aerobatics and target tracking for military. Radar based tracking is expensive, gives away the position of the radar and is prone to jamming. However, vision sensors are low cost, passive and robust to jamming hence motivating their use for aircraft tracking. This paper proposes a modified KLT tracking algorithm and tests its performance by tracking aircrafts. Vision based aircraft tracking is challenging because of the higher speed, agility and distance of the aircraft from the camera. The proposed tracker uses a feature clustering criterion to track an object based on its multiple local features. The local features are continuously updated to make the tracker robust to changing appearance of the object. The proposed tracker is generic and can be used to track multiple objects of any kind by defining multiple feature clusters. Experiments were performed on real data collected during the Air Race 2006 which will be made publicly available. The algorithm achieved accurate tracking in the presence of clouds and background clutter as long as the object size was comparable to the feature window. However, tracking failed at very small scales of the aircraft and in the presence of background clutter while the aircraft was flying low, a scenario where radars also fail.