An iterative approach for fitting multiple connected ellipse structure to silhouette

Yi Da Xu, Michael Kemp

Research output: Contribution to journalArticlepeer-review

2 Citations (Scopus)
42 Downloads (Pure)


In many image processing applications, the structures conveyed in the image contour can often be described by a set of connected ellipses. Previous fitting methods to align the connected ellipse structure with a contour, in general, lack a continuous solution space. In addition, the solution obtained often satisfies only a partial number of ellipses, leaving others with poor fits. In this paper, we address these two problems by presenting an iterative framework for fitting a 2D silhouette contour to a pre-specified connected ellipses structure with a very coarse initial guess. Under the proposed framework, we first improve the initial guess by modeling the silhouette region as a set of disconnected ellipses using mixture of Gaussian densities or the heuristic approaches. Then, an iterative method is applied in a similar fashion to the Iterative Closest Point (ICP) ([Alshawa, 2007], [Li and Griffiths, 2000] and [Besl and Mckay, 1992]) algorithm. Each iteration contains two parts: first part is to assign all the contour points to the individual unconnected ellipses, which we refer to as the segmentation step and the second part is the non-linear least square approach that minimizes both the sum of square distances between the contour points and ellipse's edge as well as minimizing the ellipse's vertex pair (s) distances, which we refer to as the minimization step. We illustrate the effectiveness of our methods through experimental results on several images as well as applying the algorithm to a mini database of human upper-body images.
Original languageEnglish
Pages (from-to)1860-1867
Number of pages8
JournalPattern Recognition Letters
Issue number13
Publication statusPublished - 2010


Dive into the research topics of 'An iterative approach for fitting multiple connected ellipse structure to silhouette'. Together they form a unique fingerprint.

Cite this