Jiang_Junfeng.pdf (14.46 MB)
Download file

Video content analysis using the video time density function and statistical models

Download (14.46 MB)
thesis
posted on 22.05.2021, 14:37 by Junfeng Jiang
As an interesting, meaningful, and challenging topic, video content analysis is to find meaningful structure and patterns from visual data for the purpose of efficient indexing and mining of videos. In this thesis, a new theoretical framework on video content analysis using the video time density function (VTDF) and statistical models is proposed. The proposed framework tries to tackle the problems in video content analysis based on its semantic information from three perspectives: video summarization, video similarity measure, and video event detection. In particular, the main research problems are formulated mathematically first. Two kinds of video data modeling tools are then presented to explore the spatiotemporal characteristics of video data, the independent component analysis (ICA)-based feature extraction and the VTDF. Video summarization is categorized into two types: static and dynamic. Two new methods are proposed to generate the static video summary. One is hierarchical key frame tree to summarize video content hierarchically. Another is vector quantization-based method using Gaussian mixture (GM) and ICA mixture (ICAM) to explore the characteristics of video data in the spatial domain to generate a compact video summary. The VTDF is then applied to develop several approaches for content-based video analysis. In particular, VTDF-based temporal quantization and statistical models are proposed to summarize video content dynamically. VTDF-based video similarity measure model is to measure the similarity between two video sequences. VTDF-based video event detection method is to classify a video into pre-defined events. Video players with content-based fast-forward playback support are designed, developed, and implemented to demonstrate the feasibility of the proposed methods. Given the richness of literature in effective and efficient information coding and representation using probability density function (PDF), the VTDF is expected to serve as a foundation of video content representation and more video content analysis methods will be developed based on the VTDF framework.

History

Language

eng

Degree

Doctor of Philosophy

Program

Electrical and Computer Engineering

Granting Institution

Ryerson University

LAC Thesis Type

Dissertation

Thesis Advisor

Xiao-Ping Zhang