Visual tracking of non-rigid objects

Luka Čehovin (2010) Visual tracking of non-rigid objects. MSc thesis.

PDF Download (4Mb)

Abstract

In this thesis we study the field of visual tracking of non-rigid, articulated objects. For this task a typical visual model, used mostly for the description of rigid objects has to be extended in a way that it can adapt to the deformations of such objects. In our work we present an extension that is based on a hierarchical approach towards visual model construction. It is based on a hierarchical combination of local and global visual information. The resulting visual model extends the existing visual models that use a set of local features connected with geometrical constraints. This set represents the bottom layer of the presented visual model. Using local features the visual model builds a multi-modal representation of the object’s appearance that represents the top layer of the model. Based on this information an area of the object in a frame is determined, and based on this area, the local feature set is updated with new features. In the thesis our work is first placed into a research context by describing recent published work on visual tracking of non-rigid objects. Next, the proposed visual model is described in detail together with its integration in a simple tracker. The performance of the tracker is assessed in various experiments using nine different video sequences. Advantages and disadvantages of the tracker are shown in comparison of the tracker with three different state-of-the-art visual trackers. The thesis is concluded with a discussion in which some theoretical and practical limitations of the presented visual model are laid out as well as some ideas for further development.

Actions (login required)