On-line Visual Tracking with Feature-based Adaptive Models

This thesis proposes a robust on-line tracking method by 1) enlarging the convergence range and 2) improving the observation memory of a phase-based appearanceadaptive “Wandering Stable Lost” (WSL) tracker, which was developed by Jepson, Fleet and El-Maraghi. The resultant tracker is demonstrated to be adaptive to temporary and permanent appearance changes in the object being tracked while at the same time it can deal with large scale and orientation changes of the object. Unlike the original phase-based WSL tracker, the new tracker can handle both partial and total occlusions when the object being tracked is temporarily unobservable. A set of “Scale Invariant Feature Transform” (SIFT) feature keypoints, which were developed by Lowe, are extracted from the object region to aid the original phase-based WSL tracker with longer past observations. The feature keypoints extracted from object region are matched against newly found keypoints in video sequences and then are passed into a Hough transform to filter out outlier mismatched pairs. The importance measures of those feature keypoints are learned using an on-line EM algorithm, which helps the tracker identify unreliable feature keypoints and concentrate computational resources on reliable ones. A deterministic gradient-based iterative tracking method is developed to use both the matched keypoints and the phase features to locate the object being tracked.