Date of this Version
Tracking-by-detection is a class of new tracking approaches that utilizes recent development of object detection algorithms. This type of approach performs object detection for each frame and uses data association algorithms to associate new observations to existing targets. Inspired by the core idea of the tracking-by-detection framework, we propose a new framework called detection-assisted tracking where object detection algorithm provides help to the tracking algorithm when such help is necessary; thus object detection, a very time consuming task, is performed only when needed. The proposed framework is also able to handle complicated scenarios where cameras are allowed to move, and occlusion or multiple similar objects exist.
We also port the core component of the proposed framework, the detector, onto embedded smart cameras. Contrary to traditional scenarios where the smart cameras are assumed to be static, we allow the smart cameras to move around in the scene. Our approach employs histogram of oriented gradients (HOG) object detector for foreground detection, to enable more robust detection on mobile platform. Traditional background subtraction methods are not suitable for mobile platforms where the background changes constantly.
Adviser: Senem Velipasalar and Mustafa Cenk Gursoy