Model-based pose estimation for rigid objects

Abstract

Determining the pose of objects appearing in images is a problem encountered often in several practical applications. The most effective strategy for dealing with this challenge is to proceed according to the model-based paradigm, which involves building 3D models of objects and then determining object poses by fitting their models to new images with the aid of detected features. This paper proposes a model-based approach for estimating the full pose of known objects from natural point features. The method employs a projective imaging model and incorporates reliable automatic mechanisms for pose initialization and convergence. Furthermore, it is extendable to multiple cameras without the need to perform multi-view matching and relies on sparse structure from motion techniques for the construction of object models offline. Experimental results demonstrate its accuracy and robustness. © 2013 Springer-Verlag.