An Exploration into Model-Free Online Visual Object Tracking

Abstract

This thesis presents a thorough investigation of model-free visual object tracking, a fundamental computer vision task that is essential for practical video analytics applications. Given the states of the object in the rst frame, e.g., the position and size of the target, the computational methods developed and advanced in this thesis aim at determining target states in consecutive video frames automatically. In contrast to the tracking schemes that depend strictly on specic object detectors, model-free tracking provides conveniently flexible and competently general solutions where object representations are initiated in the first frame and adapted in an online manner at each frame.

We first articulate our motivations and intuitions in Chapter 1, formulate model-free online visual tracking, illustrate outcomes on two representative object tracking applications; drone control and sports video broadcasting analysis, and elaborate other relevant problems.

In Chapter 2, we review various tracking methodologies employed by state-ofthe-art trackers and further review related background knowledge, including several important dataset benchmarks and workshop challenges, which are widely used for evaluating the performance of trackers, as well as commonly applied evaluation protocols in this chapter.

In Chapter 3 through Chapter 6, we then explore the model-free online visual tracking problem in four different dimensions: 1) learning a more discriminative classier with a two-layer classication hierarchy and background contextual clusters; 2) overcoming the limit of conventionally used local-search scheme with a global object tracking framework based on instance-specic object proposals; 3) tracking object affine motion with a Structured Support Vector Machine (SSVM) framework incorporated with motion manifold structure; 4) an efficient multiple object model-free online tracking approach based on a shared pool of object proposals.

Lastly, as a conclusion and future work outlook, we highlight and summarize the contribution of this thesis and discuss several promising research directions in Chapter 7, based on latest work and their drawbacks of current state-of-the-art trackers.