Login | Register

Feature binding of MPEG-7 Visual Descriptors Using Chaotic Series

Title:

Feature binding of MPEG-7 Visual Descriptors Using Chaotic Series

Azhar, Hanif (2010) Feature binding of MPEG-7 Visual Descriptors Using Chaotic Series. PhD thesis, Concordia University.

[img]
Preview
Text (application/pdf)
NR71130.pdf - Accepted Version
9MB

Abstract

Due to advanced segmentation and tracking algorithms, a video can be divided into numerous objects. Segmentation and tracking algorithms output different low-level object features, resulting in a high-dimensional feature vector per object. The challenge is to generate feature vector of objects which can be mapped to human understandable description, such as object labels, e.g., person, car. MPEG-7 provides visual descriptors to describe video contents. However, generally the MPEG-7 visual descriptors are highly redundant, and the feature coefficients in these descriptors need to be pre-processed for domain specific application. Ideal case would be if MPEG-7 visual descriptor based feature vector, can be processed similar to some functional simulations of human brain activity. There has been a established link between the analysis of temporal human brain oscillatory signals and chaotic dynamics from the electroencephalography (EEG) of the brain neurons. Neural signals in limited brain activities are found to be behaviorally relevant (previously appeared to be noise) and can be simulated using chaotic series. Chaotic series is referred to as either a finite-difference or an ordinary differential equation, which presents non-random, irregular fluctuations of parameter values over time in a dynamical system. The dynamics in a chaotic series can be high - or low -dimensional, and the dimensionality can be deduced from the topological dimension of the attractor of the chaotic series. An attractor is manifested by the tendency of a non-linear finite difference equation or an ordinary differential equation, under various but delimited conditions, to go to a reproducible active state, and stay there. We propose a feature binding method, using chaotic series, to generate a new feature vector, C-MP7 , to describe video objects. The proposed method considers MPEG-7 visual descriptor coefficients as dynamical systems. Dynamical systems are excited (similar to neuronal excitation) with either high- or low-dimensional chaotic series, and then histogram-based clustering is applied on the simulated chaotic series coefficients to generate C-MP7 . The proposed feature binding offers better feature vector with high-dimensional chaotic series simulation than with low-dimensional chaotic series, over MPEG-7 visual descriptor based feature vector. Diverse video objects are grouped in four generic classes (e.g., has [barbelow]person, has [barbelow]group [barbelow]of [barbelow]persons, has [barbelow]vehicle, and has [barbelow]unknown ) to observe how well C-MP7 describes different video objects compared to MPEG-7 feature vector. In C-MP7 , with high dimensional chaotic series simulation, 1). descriptor coefficients are reduced dynamically up to 37.05% compared to 10% in MPEG-7 , 2) higher variance is achieved than MPEG-7 , 3) multi-class discriminant analysis of C-MP7 with Fisher-criteria shows increased binary class separation for clustered video objects than that of MPEG-7 , and 4) C-MP7 , specifically provides good clustering of video objects for has [barbelow]vehicle class against other classes. To test C-MP7 in an application, we deploy a combination of multiple binary classifiers for video object classification. Related work on video object classification use non-MPEG-7 features. We specifically observe classification of challenging surveillance video objects, e.g., incomplete objects, partial occlusion, background over lapping, scale and resolution variant objects, indoor / outdoor lighting variations. C-MP7 is used to train different classes of video objects. Object classification accuracy is verified with both low-dimensional and high-dimensional chaotic series based feature binding for C-MP7 . Testing of diverse video objects with high-dimensional chaotic series simulation shows, 1) classification accuracy significantly improves on average, 83% compared to the 62% with MPEG-7 , 2) excellent clustering of vehicle objects leads to above 99% accuracy for only vehicles against all other objects, and 3) with diverse video objects, including objects from poor segmentation. C-MP7 is more robust as a feature vector in classification than MPEG-7 . Initial results on sub-group classification for male and female video objects in has [barbelow]person class are also presentated as subjective observations. Earlier, chaos series properties have been used in video processing applications for compression and digital watermarking. To our best knowledge, this work is the first to use chaotic series for video object description and apply it for object classification

Divisions:Concordia University > Gina Cody School of Engineering and Computer Science > Electrical and Computer Engineering
Item Type:Thesis (PhD)
Authors:Azhar, Hanif
Pagination:xxvi, 193 leaves : ill. (some col.) ; 29 cm.
Institution:Concordia University
Degree Name:Ph. D.
Program:Electrical and Computer Engineering
Date:2010
Thesis Supervisor(s):Amer, Aishy
ID Code:979417
Deposited By: Concordia University Library
Deposited On:09 Dec 2014 17:58
Last Modified:18 Jan 2018 17:49
Related URLs:
All items in Spectrum are protected by copyright, with all rights reserved. The use of items is governed by Spectrum's terms of access.

Repository Staff Only: item control page

Downloads per month over past year

Back to top Back to top