CityU Institutional Repository >
CityU Electronic Theses and Dissertations >
ETD - Dept. of Manufacturing Engineering and Engineering Management >
MEEM - Doctor of Philosophy >
Please use this identifier to cite or link to this item:
|Title: ||Towards optimal 3D visual tracking of rigid object motion based on depth information recovery|
|Other Titles: ||Ji yu hui fu jing shen de zui you hua san wei gang ti shi jue gen zong|
|Authors: ||Chen, Huiying (陳慧瑩)|
|Department: ||Department of Manufacturing Engineering and Engineering Management|
|Degree: ||Doctor of Philosophy|
|Issue Date: ||2008|
|Publisher: ||City University of Hong Kong|
|Subjects: ||Computer vision.|
Motion perception (Vision) -- Computer simulation.
|Notes: ||CityU Call Number: TA1634 .C435 2008|
vii, 115 leaves : ill. (some col.) 30 cm.
Thesis (Ph.D.)--City University of Hong Kong, 2008.
Includes bibliographical references (leaves -115)
|Abstract: ||This thesis investigates 3D visual tracking problems and evaluates possible ways to optimize tracking performance. In this study, some new methodologies are developed to fulfill the task of 3D visual tracking. The related optimization methods for enhancing the tracking performance are explored. Depending on whether the depth information (a distance from the target object to the vision system along the optical axis) is recovered or not, visual tracking can be divided into two main categories, 3D tracking and 2D tracking. The 3D tracking problem can be tackled with three different types of vision systems, i.e. passive vision, active vision and dynamic vision system. The 3D tracking methods for each of these are studied in this thesis. Three approaches to improve tracking performance are proposed. These approaches are rooted in ideas from deterministic frameworks to stochastic frameworks. Discussions are given for the realization of optimal 3D tracking in a broader sense as the complexity engaged in methodologies increases.
With passive vision, the depth information can be recovered with two or more simultaneous views. With a relatively simple configuration and almost predictable noises, 3D tracking problems in passive vision can be solved with the aid of parametric models. A tracking method is developed using the interaction matrix for passive vision. The interaction matrix describes a direct linkage between the 3D object motion field and the image motion field. Although it is easy to implement for tracking purpose, the interaction matrix suffers from singularity. This thesis develops and introduces nonsingular constraints to improve tracking performance. Based on the use of the interaction matrix, an error model is established, and tracking performance is optimized by minimizing the estimated tracking error. The optimal targeting area and the best focused location are defined with both the nonsingular constraints and the minimal error constraints. Those constraints actually provide suitable candidates for further research on the design of the dynamic vision.
With active vision, which consists of a camera and a projector in this study, the depth information can be recovered with a single view. In this study, a stochastic state estimation method, the particle filter, is employed to enhance the tracking performance. The importance density function, which is a crucial factor that directly influences the quality of particle filter, is studied. An ordinary choice of the importance density is to use the probability of the prior states. However, the importance density obtained from this method is independent of the latest measurement and it is sensitive to outliers. In this thesis, an enhanced importance density function is proposed by fusing the most current passive observation data into the prior information of the active system. This solution directly improves estimation accuracy and tracking performance in the 3D tracking.
With dynamic vision, the vision system configuration can be adjusted on line to improve tracking performance. The thesis explores optimal 3D tracking in dynamic vision via dynamic view planning. A view planning method with a newly developed criterion, the effective sample size, is presented to guide the reconfiguration control of the vision system for achieving optimal tracking. The proposed view planning method was based on the use of an improved particle filter, whose effective sample size is maximized. In the approach, the vision system has been designed and configured to achieve the largest number of effective particles, which actually minimizes the tracking error by revealing the system to a better swarm of importance samples and interpreting the posterior state in a better way. It has been proved that the new objective function, which maximizes the number of effective sampling, actually leads the optimization towards system configurations that minimize tracking errors.
This thesis also implements the aforementioned methodologies in three kinds of typical testbeds respectively. Implementation results have verified the effectiveness in improving tracking performance to achieve optimal tracking.|
|Online Catalog Link: ||http://lib.cityu.edu.hk/record=b2340709|
|Appears in Collections:||MEEM - Doctor of Philosophy |
Items in CityU IR are protected by copyright, with all rights reserved, unless otherwise indicated.