Loading...
| Friend's email | |
| Your name | |
| Your email | |
| enter code | |
This page was sent successfuly
- Type of Document: Ph.D. Dissertation
- Language: Farsi
- Document No: 51916 (19)
- University: Sharif University of Technology
- Department: Computer Engineering
- Advisor(s): Manzuri Shalmani, Mohammad Taghi
- Abstract:
- Video segmentation means grouping of pixels of the video sequences into spatio-temporal regions which exhibit coherence in both appearance and motion. Due to complexity and spatio-temporal variations, dynamic texture segmentation is a one of the most challenging task in video processing. The problem of dynamic texture segmentation has received considerable attention due to the explosive growth of its applications in video analysis and surveillance systems. In this thesis, two novel approaches have been proposed. The first proposed method is based on generative Dynamic texture models (DTMs) which represent videos as a linear dynamical system. Since DTMs cannot be used for complex videos which consist of co-occurring dynamic textures, dynamic texture mixture (DTM) and layered dynamic texture (LDT) were introduced. DTMs suppose that video sequences are composed of some dynamic textures from a finite set of dynamic textures. DTM is not a global generative model for video of co-occurring textures as is the case of the LDT. This model requires decomposing the video into a collection of small spatio-temporal patches, which are then clustered. The localized nature of this video representation is problematic for the segmentation of textures which are globally homogeneous but exhibit substantial variation between neighboring locations. LDTs suppose that the video sequences are superposition of finite linear dynamical systems (LDSs) outputs. A major limitation of these models concerns the automatic selection of proper number of their segments, i.e., the number of layers derived by video segmentation. Hence, the derived region segments are usually done subjectively using expert knowledge which is an important restriction for systematic database approaches. Dirichlet process mixture (DPM) models which has appeared recently as the cornerstone of non-parametric Bayesian statistics, is an optimistic candidate toward resolving this issue. Under this motivation, for solving the aforementioned drawback of DTMs and LDTs, in this report we proposed a novel non-parametric approach for DT model, formulated on the basis of a joint DPM and DT construction, named Infinite dynamic texture model (IDTM). This interaction causes the algorithm to overcome the problem. We derive the Variational Bayesian Expectation maximization (VBEM) inference for the proposed method. In order to compute the 1st order and 2nd order sufficient statistics of the hidden states, Forward-backward algorithm on Variational Bayesian LDS is used. Then experiments on different video sequences, taken from Dyntex and UCSD-Synthdb datasets, were performed. The second proposed approach, a novel per-pixel motion descriptor is proposed for motion detection in video sequences which outperforms the current methods in the literature particularly in severe scenarios due to the difficulties raised by illumination variations, occlusion, camouflage, sudden motions appearing in burst and environmental changes such as those on weather conditions, sunlight changes during a day, etc. The proposed descriptor is based on two complementary three-dimensional discrete wavelet transforms (3D-DWT) and a three-dimensional wavelet leader. In this approach, a feature vector is extracted for each pixel by applying a novel three-dimensional wavelet-based motion descriptor. Then, the extracted features are clustered by the well-known K-means algorithm. The experimental results demonstrate the effectiveness of the both proposed methods compared to state-of-the-art approaches in several public benchmark datasets. The application of the proposed method and additional experimental results for several challenging datasets are available online
- Keywords:
- Dynamic Texture ; Dynamic Texture Segmentation ; Nonlinear Dynamical System ; Bayesian Nonparametric Model ; Image Segmentation ; Variational Bayesian Inference ; Dirichelt Boundary Condition ; Three Dimentional Discrete Wavelet Transform (3D-DWT)
Digital Object List-
محتواي کتاب - view
Bookmark