| Abstract: |
The idea of a domain surface is presented. With this idea, an object scanning algorithm from an RGB-D
camera can get a simplified mesh. Object scanning usually consists of surface reconstruction and fusion in
two steps. Range data is native to surface construction. However, the constructed surface always requires
massive data. It is not convenient to directly apply it to another application. To fuse two surfaces correctly,
it is critical to have a precise registration of two views. A domain surface can solve the two main problems
simultaneously. For the scanned scene, the system will find some domain surface to approximate the described
surface. Thus, the object model is being simplified naturally. Traditionally, a registration problem is always
solved as a six degree of freedom transformation. Resolving a robust solution from two dependent factors
of the rotation and translation by a non-linear form is not straightforward. Usually, an iterative closest point
(ICP) algorithm is adopted to find an optimized solution. However, the solution is based on the initial guess,
and it is often trapped into a local minimum. From the normal of the mapped pair domain surface, it can
estimate the rotation matrix by a linear SVD method. After the rotation is known, the shift of the feature
points can more easily recover the translation. The idea of a domain surface is robust and straightforward for
surface reconstruction and registration. With the help of this idea, a simplified mesh constructed from range
data becomes easier. |