Robust Automatic Registration of Range Images with Reflectance

In proceedings of Central European Seminar on Computer Graphics for Students (CESCG 2007), Apr. 2007
 

Abstract

We tackle the problem of automatic matching, consistency checking and registration of multiple unknown and un- ordered range images. Utilizing robust visual features extracted from the supplied reflectance images, an efficient pairwise view matching scheme is used to build up a directed correspondence graph, nodes representing the input range images and edges labeled with relative pose estimates. Subsequently, a local and global consistency check eliminate false positive edges in the graph as these prevent the succeeding to a correct solution. Absolute poses are recovered by a breadth-first search (BFS), thereby, for each visited node, combining the weighted contributions of all encountered paths back to the root. Remarkably, the absolute alignments are accurately recovered from only the features. Thus, a subsequent fine registration step can be omitted. The framework is independent from object size and particular sensor model.

Stichwörter: pose consistency, surface matching

Bilder

Paper herunterladen

Paper herunterladen

Bibtex

@INPROCEEDINGS{koertgen-2007-registration,
     author = {K{\"o}rtgen, Marcel},
      title = {Robust Automatic Registration of Range Images with Reflectance},
  booktitle = {Central European Seminar on Computer Graphics for Students (CESCG 2007)},
       year = {2007},
      month = apr,
   keywords = {pose consistency, surface matching},
   abstract = {We tackle the problem of automatic matching, consistency checking and registration of multiple
               unknown and un-
               ordered range images. Utilizing robust visual features extracted from the supplied reflectance
               images, an efficient pairwise view matching scheme is used to build up a directed correspondence
               graph, nodes representing the input
               range images and edges labeled with relative pose estimates. Subsequently, a local and global
               consistency check eliminate false positive edges in the graph as these prevent the succeeding to a
               correct solution. Absolute poses are recovered by a breadth-first search (BFS), thereby, for each
               visited node, combining the weighted contributions of all encountered paths back to the root.
               Remarkably, the absolute alignments are accurately recovered from only the features. Thus, a
               subsequent fine registration step can be omitted. The framework is independent from object size and
               particular sensor model.}
}