Title

Eigenvector-based Dimensionality Reduction for Human Activity Recognition and Data Classification

Date of Award

2013

Degree Type

Dissertation

Degree Name

Ph.D.

Department

Electrical and Computer Engineering

First Advisor

Boubakeur Boufama

Second Advisor

Rachid Benlamri

Keywords

Mathematics, Computer Engineering, Computer science

Rights

CC BY-NC-ND 4.0

Abstract

In the context of appearance-based human motion compression, representation, and recognition, we have proposed a robust framework based on the eigenspace technique. First, the new appearance-based template matching approach which we named "Motion Intensity Image" for compressing a human motion video into a simple and concise, yet very expressive representation. Second, a learning strategy based on the eigenspace technique is employed for dimensionality reduction using each of PCA and FDA, while providing maximum data variance and maximum class separability, respectively. Third, a new compound eigenspace is introduced for multiple directed motion recognition that takes care also of the possible changes in scale. This method extracts two more features that are used to control the recognition process. A similarity measure, based on Euclidean distance, has been employed for matching dimensionally-reduced testing templates against a projected set of known motions templates. In the stream of nonlinear classification, we have introduced a new eigenvector-based recognition model, built upon the idea of the kernel technique. A practical study on the use of the kernel technique with 18 different functions has been carried out. We have shown in this study how crucial choosing the right kernel function is, for the success of the subsequent linear discrimination in the feature space for a particular problem. Second, building upon the theory of reproducing kernels, we have proposed a new robust nonparametric discriminant analysis approach with kernels. Our proposed technique can efficiently find a nonparametric kernel representation where linear discriminants can perform better. Data classification is achieved by integrating the linear version of the NDA with the kernel mapping. Based on the kernel trick, we have provided a new formulation for Fisher's criterion, defined in terms of the Gram matrix only.