Semester

Fall

Date of Graduation

2008

Document Type

Dissertation

Degree Type

PhD

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Xin Li.

Abstract

Video modeling refers to the development of analytical representations for explaining the intensity distribution in video signals. Based on the analytical representation, we can develop algorithms for accomplishing particular video-related tasks. Therefore video modeling provides us a foundation to bridge video data and related-tasks. Although there are many video models proposed in the past decades, the rise of new applications calls for more efficient and accurate video modeling approaches.;Most existing video modeling approaches are based on explicit motion representations, where motion information is explicitly expressed by correspondence-based representations (i.e., motion velocity or displacement). Although it is conceptually simple, the limitations of those representations and the suboptimum of motion estimation techniques can degrade such video modeling approaches, especially for handling complex motion or non-ideal observation video data. In this thesis, we propose to investigate video modeling without explicit motion representation. Motion information is implicitly embedded into the spatio-temporal dependency among pixels or patches instead of being explicitly described by motion vectors.;Firstly, we propose a parametric model based on a spatio-temporal adaptive localized learning (STALL). We formulate video modeling as a linear regression problem, in which motion information is embedded within the regression coefficients. The coefficients are adaptively learned within a local space-time window based on LMMSE criterion. Incorporating a spatio-temporal resampling and a Bayesian fusion scheme, we can enhance the modeling capability of STALL on more general videos. Under the framework of STALL, we can develop video processing algorithms for a variety of applications by adjusting model parameters (i.e., the size and topology of model support and training window). We apply STALL on three video processing problems. The simulation results show that motion information can be efficiently exploited by our implicit motion representation and the resampling and fusion do help to enhance the modeling capability of STALL.;Secondly, we propose a nonparametric video modeling approach, which is not dependent on explicit motion estimation. Assuming the video sequence is composed of many overlapping space-time patches, we propose to embed motion-related information into the relationships among video patches and develop a generic sparsity-based prior for typical video sequences. First, we extend block matching to more general kNN-based patch clustering, which provides an implicit and distributed representation for motion information. We propose to enforce the sparsity constraint on a higher-dimensional data array signal, which is generated by packing the patches in the similar patch set. Then we solve the inference problem by updating the kNN array and the wanted signal iteratively. Finally, we present a Bayesian fusion approach to fuse multiple-hypothesis inferences. Simulation results in video error concealment, denoising, and deartifacting are reported to demonstrate its modeling capability.;Finally, we summarize the proposed two video modeling approaches. We also point out the perspectives of implicit motion representations in applications ranging from low to high level problems.

Download

Share

COinS