Real-Time Long-Term Tracking with Adaptive Online Searching Model

Abstract

Kernelized correlation filter (KCF) based trackers have drawn great attention for their superiority in terms of accuracy and speed in visual tracking problem. However, these methods are not robust under scale changes, rotation and occlusion due to their having a fixed size filter. In this work, we take advantage of the KCF tracker to propose a novel algorithm for long-term visual object tracking which handle scale variation using Log-Polar Transformation and Phase Correlation. To detect exactly loss tracker moment when object partly for fully occlusion, this paper propose an effective technique combining PRS ratio and histogram distance. We also learn an online SVM classifier on consecutive and reliable samples to redetect objects in case of tracking failure due to heavy occlusion or out of view movement. Experimental results in several challenging tracking datasets from camera UAV show that our tracker achieves remarkable speed in real-time application at 40FPS while handling scale changes and occlusion better than many state-of-the art tracking algorithms.