Tracking freeform objects with geometric primitives

Joshua J Zapf, Purdue University

Abstract

This thesis addresses the problem of tracking rigid freeform objects in 3D. The term freeform is used in this context to describe a non-polyhedral, free flowing shape, such as a car body. For the most part, the difficulty associated with quantifying the shape of a freeform object has precluded the use of local geometric primitives for tracking. Nonetheless, tracking with local geometric primitives offers several advantages over alternative approaches. In particular, this method is stable under varying lighting conditions, can elegantly handle occlusions, and will tolerate large changes in pose. To exploit these advantages, we propose a new method for extracting and matching local geometric primitives that can accommodate a large range of freeform objects. The key attributes that make this possible are a feature set based on line segments and elliptical arcs, a soft grouping mechanism for merging adjacent primitives, and a backtracking strategy for removing outliers. To verify the discriminative power of our method, we first show how our feature extraction and matching algorithms can be used in an object detection framework to locate instances of an object class in a set of digital images. We then show how our algorithms can be used in a recursive 3D tracking framework to estimate the six degrees of freedom pose of a general freeform object over time. Through several experiments, we demonstrate the robustness of our system under several challenging conditions, including fast motions, background clutter, and partial occlusion.

Degree

Ph.D.

Advisors

Park, Purdue University.

Subject Area

Electrical engineering

COinS