Robust visual tracking in image sequences.

This thesis is concerned with the core computer vision challenge of obtaining efficient and robust visual tracking of objects over extended image sequences. Effective solutions to this problem are crucial for applications such as smart video surveillance, intelligent human machine interaction, and robotics. Most tracking algorithms can be classified into two major types, namely, probabilistic filtering algorithms and deterministic localisation algorithms. This thesis presents novel enhancements to both types of algorithm. The probabilistic filtering algorithms adopted in visual tracking are mainly based on Kalman filters and particle filters. Whereas Kalman filters are restricted to linear and Gaussian noise models, particle filters can propagate more general distributions, albeit only approximately. This is valuable in visual tracking, as simple models of noise do not suffice. Although particle filter trackers have been quite successful, they too have significant drawbacks. Several strategies are advanced in this thesis to overcome these limitations. Two alternative means are proposed for generating a proposal distribution, which is a key step in particle filtering. These increase the efficiency and robustness of the algorithm in the presence of sudden motion. The particle filter is also extended so as to accommodate multiple cues, such as colour and edge information, affording greater reliability. Additionally, an efficient kernel subspace method is introduced to capture a tracked object's appearance. Finally, a novel method is proposed for tracking the motion of an articulated structure. This significantly improves sampling efficiency, alleviating the curse of dimensionality in Monte Carlo sampling methods. The value of these enhancements is confirmed experimentally. The second part of this thesis concerns the mean shift algorithm, recently advanced as an alternative to stochastic trackers, that seeks the global mode of a suitable density function. A novel, multi-bandwidth mean shift procedure is presented along with a means of accelerating the algorithm. This improved tracker is applied to the problems of object localisation and visual tracking. We empirically show on various data sets that the proposed algorithm reliably finds the true object location when the initial position of the mean shift is far away from the global maximum, in contrast with the conventional mean shift algorithm which can often become trapped in a local maximum.Thesis (Ph.D.) -- University of Adelaide, School of Computer Science, 200