DYNAMIC MOTION AND APPEARANCE MODELING FOR ROBUST VISUAL TRACKING
Open Access
- Author:
- Lim, Hwasup
- Graduate Program:
- Electrical Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- March 23, 2007
- Committee Members:
- Octavia I Camps, Committee Chair/Co-Chair
Constantino Manuel Lagoa, Committee Chair/Co-Chair
Mario Sznaier, Committee Member
William Evan Higgins, Committee Member
Sunil K Sinha, Committee Member
William Kenneth Jenkins, Committee Member - Keywords:
- visual tracking
caratheodory-fejer
appearance modeling
motion modeling
object tracking - Abstract:
- Visual tracking is one of the most active areas in computer vision and it has many promising applications such as human motion capture, human computer interface, and visual surveillance. The performance of visual tracking systems is often severely affected by appearance changes of the target, occlusion, scene clutter and sensor noise. Conventional approaches address these challenges by introducing simple dynamic models such as constant velocity of the target, and adaptive appearance models based on the assumption that the target features are continuous in the spatial and time domains. These approaches, however, have been shown in the literature to be fragile to occlusion and susceptible to incorrect observations. In addition, modeling dynamic target appearance is often faced with difficulties due to its high dimensionality and nonlinearity. This thesis presents efficient frameworks for modeling dynamic motion and appearance of the target object by employing tools from robust system identification theory and nonlinear dimensionality reduction techniques. The target motion and appearance are modeled as unknown operators that satisfy certain interpolation conditions. The unknown operators can then be identified by solving a convex optimization problem where high dimensionality and nonlinearity in appearance changes are addressed by efficiently projecting high dimensionality descriptors into low dimensional manifolds. The learned dynamic models are then used to accurately predict the location and appearance of the targets in future frames, thus preventing tracking failures due to model drifting, target occlusion, and scene clutter. The advantages of this approach are multiple: 1) It allows to treat the tracking problem from an input/output point of view, requiring very little a priori knowledge about the target. At the same time, it allows to naturally incorporate as much prior knowledge as it is available. 2) It decouples nonlinear appearance changes from linear dynamics facilitating the identification process. 3) It provides mechanisms to invalidate a priori assumptions and worst-case estimates of the identification error that can be used to determine how long the predictions are valid. 4) Finally, the use of reliable models allows for simpler computational complexity algorithms. The proposed approach was validated with several experiments conducted under various scenarios such as single and multiple target tracking, and static and moving camera environments, and finally, single and multiple camera environments. The advantages of the proposed approach are shown with illustrated examples in which robustness to severe occlusion and efficiency of appearance modeling are compared to conventional approaches.