Tracking articulated motion in videos

The main goal of this research is to provide an insight of human or pedestrian tracking based on feature and develop a framework using statistical modelling. Understanding and tracking human motion is very important and key to explain human activity. Most important video content based analysis depends on it. Exemplar based human motion tracking techniques have been very successful for human body motion analysis. However, their accuracy strongly depends on the similarity of both camera viewing angle and scene properties between training and testing images captured by a static camera with a similar viewing angle observing only one individual. In this thesis three major aspects of pedestrian tracking are discussed. Firstly, the contour feature based tracking has been employed. The silhouettes of the person in question are extracted and fed to a Bayesian filter. Secondly, an online trained model has been proposed for tracking framework. Some major features like, colour, HOG and foreground extraction etc. have been used and exploited to propose an online tracker. Finally, a novel body pose based human tracking model is proposed for pedestrian tracking. Specifically, this method attempts to exploit the curvature information of different body poses in tracking framework to overcome general tracking problems. Results show that poselet-based features are more suitable for tracking than just detecting the person over the frames. Performance has been evaluated in a rich evaluation fiamework. Three public datasets like, HumanEva, PETS and Muhavi are chosen for their unique characteristic. Each method has been implemented carefully and they are tested on these datasets. Finally, they are evaluated based on a standard metric.