A model-based tracking framework for textureless 3D rigid curved objects

Abstract

This paper addresses the problem of tracking textureless
rigid curved objects. A common approach uses polygonal
meshes to represent curved objects inside an edge-based tracking
system. However, in order to accurately recover their shape,
high quality meshes are required, creating a trade-off between
computational efficiency and tracking accuracy. To solve this
issue, we suggest the use of quadrics calculated for each patch
in the mesh to give local approximations of the object contour.
This representation reduces considerably the level of detail of the
polygonal mesh while maintaining tracking accuracy. The novelty
of our research lies in using curves to represent the quadrics’
projection in the current viewpoint for distance evaluation instead
of comparing directly the edges from the mesh and detected
edges in the video image. In our tracking framework, we also
include a method to calculate the measurable Degrees of Freedom
(DoF) of the target object. This is used to recover the pose
parameters when the object has less than 6DoF. Experimental
results compare our approach to the traditional method of
using sparse and dense meshes. Finally, we present a potential
Augmented Reality application of the proposed method.