MedianStruck for long-term tracking applications

In this paper, we propose a mutual framework that combines two state-of-the-art visual object tracking algorithms. Both trackers benefit from each other's advantage leading to an efficient visual tracking approach. Many state-of-the-art trackers have poor performance due to rain, fog or occlusion in real-world scenarios. Often, after several frames, objects are getting lost, only leading to a short-term tracking capability. In this paper, we focus on long-term tracking, preserving real-time capability and very accurate positioning of tracked objects. The proposed framework is capable to track arbitrary objects, leading to decreased labeling efforts and an improved positioning of bounding boxes. This is especially interesting for applications such as semi-automatic labeling. The benefit of our proposed framework is demonstrated by comparing it with the related algorithms using own sequences as well as a well-known and publicly available dataset.