Abstract:
This thesis develops single-object tracking algorithms that are accurate, robust, and real-time-capable. We present a method that can track homogeneous regions that undergo arbitrary deformations efficiently. Additionally, we present an edge-based tracker that can cope with occlusions, is very accurate, and virtually drift-free. The tracker is evaluated on a challenging new dataset and performs on par with the current state of the art in deep-learning but is at least 40 times faster.