Saliency combined particle filtering for effective aircraft tracking

Abstract

Vision-based aircraft tracking has been considered for emerging real-world applications, such as collision avoidance, air traffic surveillance, and target tracking for military use. However, conventional tracking methods often fail in following aircraft due to 1) variations of object shape, 2) continuously varying background, and 3) unpredictable flight motion. In this thesis, we address the problems of vision-based aircraft tracking. To this ends, we propose a principled manner of improving color-based tracking algorithm by combining a biologically inspired saliency feature. More specifically, we exploit the integration of color distributions into particle filtering, which is a Monte Carlo method for general nonlinear filtering problems. To overcome the varying appearances which are usually from changing illumination and pose conditions, we update the target color model. Furthermore, we adopt a structure tensor based saliency algorithm to incorporate the saliency features into particle filter framework, which results in robustly assigning appropriate particle weights even in complex backgrounds. The rationale behind our approach is that color and saliency information are complementary, both mutually fulfilling and completing each other, especially when tracking aircraft in a harsh environment. Tests on real flight sequences reveal that the proposed system yields convincing tracking outcomes under both variations of background and sudden target motion changes.