Long-term reliable visual tracking with UAVs

Abstract

In the paper, we propose an effective long-term real-time tracking method to address the problem of robustness and tracking failure in visual tracking with UAVs. Most existing trackers only consider short-term tracking, therefore are unable to cope with partial and complete occlusion, which finally leads to object drifting or loss. Our method still follows the tracking-by-detection framework. However, after choosing kernelized correlation filter as the tracker baseline, we introduce the confidence of candidate patches to measure tracking reliability, and trigger redetection process with random forest and learned object model when needed. We further improve object update strategy to make the object model with memory more robust against object drift. Extensive experiment results on UAV videos show that our algorithm performs better than widely used TLD, KCF, and LCT methods.