PhD thesis [pdf], 13/04/2012
Visual object tracking aims at following objects in image sequences. This is a fundamental problem in computer vision and a pre-requisite for numerous applications. This dissertation categorizes the vast number of related tracking approaches into four tracking paradigms and presents four novel methods to effectively constrain object tracking.
Object tracking can be formulated in numerous ways. One line of research applies a pre-learned object model robustly to an image sequence. However, the sheer number of possible appearance changes make objects hard to render in advance. Adaptive methods are proposed in order to learn the object’s appearance on-the-fly. Alternatively, objects can be treated as outliers to a scene model which might be easier to learn, especially with a fixed camera. Furthermore, objects can be discovered using segmentation techniques. In this thesis, these paradigms are combined for improved tracking results.
The first part of this thesis examines adaptive model-free object tracking. Different constraints about the nature of objects are incorporated in order to alleviate the drifting problem. In a first approach we refine an initially weak object model during tracking with visual constraints. Objects are larger entities that often move independently from their surroundings. This motivates a second approach that includes robust motion segmentation to gather training data more robustly.
The second part of this thesis examines object tracking in a specific scene with a pre-learned person model. The goal is to constrain the problem spatially and to adapt the person model to each location in the specific scene. In the first method, simple local detectors are learned by means of a person model, adaptive tracking and multiple cameras. In a second approach, the local statistics of a person model are robustly adapted using scene assumptions about the size of the object, predominant background and smoothness of trajectories.