Performance Benchmarking of Visual Human Tracking Algorithms for UAVs

With the evolution of robotic systems, unmanned aerial vehicles (UAV) have become a target of interest for domains such as computer vision (CV) and artificial intelligence (AI), contributing to a variety of applications for surveillance, transportation and many more. A very hot topic that is the playground of the proposed benchmark is visual human tracking in images acquired by a camera mounted on a UAV. This target application troubles CV and deep learning (DL) research community in recent years and it has created serious demands for visual tracking algorithms. Some of the most important demands are high performance under hard visual tracking conditions and deployment in edge devices with limited computation resources. These two challenges are the main motivation of the presented paper, where 37 tracking algorithms have been benchmarked in visual object tracking (VOT) images. For each tracking algorithm two metric categories, relative to detection performance and hardware resources consumption, have been considered. The objective of the proposed paper is to highlight the most lightweight and high performance tracking algorithms for usage in UAV based applications.