ABSTRACT
This article presents a system for the automated tracking of non-rigid anatomic structures in two-dimensional image sequences, which was primarily applied to X-ray image sequences of the vocal tract. In this particular application articulatory organs have to be measured to investigate the complex dynamic characteristics of human speech production. Of particular interest is a robust boundary detection of non-rigid organs such as lips and tongue. To solve this ill-posed detection problem under the presence of transparently superimposing structures, varying textural appearances of organs and noise, a two-level system is proposed. At the lower level, several edge-, region-, and motion-based image operators are combined to exploit their respective benefits and concomitantly compensate for their deficiencies. For the sake of precision, the result of these operators are not represented as larger tokens, such as line segments, but remain pixel-related cues or image evidences. At the higher level, an active contour-based component allows for the introduction of a priori knowledge about the object to be detected.
