Complex geometric primitive extraction on graphics processing unit

Abstract

Extracting complex geometric primitives from 2-D imagery is a long-standing problem that researchers have had to deal with. Various approaches were tried from Hough transform based methods to stochastic algorithms. However, serial implementations lack sufficient scalability on high resolution imagery. As sequential computing power cannot pace up with the increase in size of datasets, researchers are compelled to exploit parallel computational resources and algorithms. In this study, we have merged parallelization capability of GPUs with inherent parallelism on genetic algorithms to cope with the problem of detecting complex geometric primitives on high resolution imagery. We have implemented ellipse detection on commodity graphics processing unit and showed that our GPU implementation achieve high speed-up relative to state of the art CPU by experimental results.