SIZE ESTIMATION IN NON-RIGID OBJECT TRACKING.PDF

Abstract

Tracking of moving, non-rigid objects in video sequences is an important and challenging task in the field of computer vision and artificial intelligence. It is one of the strategies used for extraction of information from objects of interest. Being an active research area, efforts are being made to bring innovations and make it more efficient. Many computer vision applications such as surveillance, augmented reality, perceptual user interfaces, smart rooms, driver assistance, object-based video compression, gesture recognition and man machine interface etc are related to realtime object tracking. The primary task in implementation of a tracker is the target representation and localization. The target representation and localization deals with representing the target in some color space and estimating its size and location in the subsequent frames. The next step is the filtering and data association process, which deals with the dynamics of the tracked object. A wide variety of tracking algorithms have been introduced in literature and implemented. But most of those tracking approaches focus only on the dynamic behavior of the target. Tracking of the spatial characteristics like size has not been given due attention. xvi A basic requirement for a practical tracking system is to adjust the tracking model in real time when appearance of the tracked objects change. Since the scales of the targets vary irregularly, systems with fixed size tracking window often fail. More so, the available techniques incorporating scale update mechanism also end up giving erroneous estimates. As a matter of fact, the problem of simultaneous size estimation and localization are inter-dependant. A compromise on one of the factors could degrade the performance of the other component. In this work a novel approach for size estimation of non-rigid objects has been proposed which also enhances the positioning accuracy. The tracker is primarily based on Mean Shift tracking principles. For better size estimation, a size cost function has been incorporated and augmented with a penalty factor scheme. To further improve on the size and position, background modeling has been embedded in the algorithm via the double bounding box scheme. This modeling helps in delineating and segregating target from the background and results in improved performance.