Fast Moving Object Detection & Tracking.pdf

Objects that move over a distance exceeding their size within the exposure time, often rotating at high speeds. Common in sports videos.
- Appearance: In a single frame, FMOs appear as semi-transparent streaks larger than their actual size.
- Proposed Method: A three-algorithm pipeline for FMO detection and tracking, capable of recovering the object's appearance and axis of rotation despite its blurred appearance.
- Evaluation: Existing trackers are inadequate for FMO localization. The proposed method outperforms them on a new annotated dataset.
- Applications: Temporal super-resolution (synthesizing realistic videos with higher frame rates) and highlighting (extracting FMO properties for sports analytics).
- Related Work: Previous methods struggle with FMO blur and rotation. Most deblurring techniques assume local linear motion, lack consideration for rotation, or require large neighborhoods unsuitable for small objects.
- Problem Definition: FMOs are objects moving large distances relative to their size during a single frame's exposure, possibly rotating along an arbitrary axis with unknown angular speed. The video frame is formed by the object's blurred appearance and the partially visible background.