Year

2010

Degree Name

Master of Computer Science

Department

School of Computer Science and Software Engineering - Faculty of Informatics

Abstract

A Posture detection system aims to identify and localize any specific types of postures in images and video sequences. Unlike human or pedestrian detection where only one class of objects is required to be detected, posture detection is designed to detect multiple classes of postures. It remains a challenging problem because human bodies are complex and articulated with very diversified appearances. Posture detection often relies on a good generalization of the variations from large quantity of training examples that cover different situations. In this thesis, we devise a new posture detection framework that combines the histogram of gradient (HOG)-based feature with a novel manifold-based open-set classifier designed to achieve a better generalization. In this framework, each posture class is represented by a complex manifold that lies in the high-dimensional visual input space. The manifold is learned using Kernel PCA. Classification of a new observation is achieved by comparing it to each trained posture manifold. In addition, a new greedy Kernel PCA approximation algorithm is proposed to speed up the learning of the posture manifolds. The approximation algorithm seeks to remove the redundant training samples in the kernel space while best retaining the accuracy of kernel mapping, resulting in a new kernel PCA model that provides almost identical learning and classification ability to the original kernel PCA with significantly lower computational cost. Both the detection framework and approximation algorithm were tested on 2D and 3D artificial datasets and real human and posture datasets. The results have shown that the approximation algorithm is effective and the proposed framework can provide accurate and efficient detection of different postures with a relatively small training set.

02Whole.pdf (3394 kB)

Share

COinS