Tracking, Learning and Detection over a Large Range of Speeds

In this thesis we propose an algorithm which allows detection and tracking of objects that appear in videos as fast moving which can possibly slow down. Object is fast moving (with respect to a camera) if its projected trajectory is larger than its size in one frame. In a single frame, such objects are often barely visible and appear as semi-transparent streaks. The detection part of the algorithm can discover previously unseen fast moving objects. The long-term tracking part is able to continuously track objects even when they are no longer fast moving. For the method evaluation we introduce FMOv2 dataset. The results show that the proposed method outperforms existing trackers when objects are fast moving. We demonstrate several applications of fast moving object detection and long-term tracking, such as temporal superresolution, highlighting, speed estimation and other.