Integration of Multiple Range Maps through Consistency Processing

Abstract

This paper presents a method for modeling the surfaces of some 3D scene from a set of registered range maps. The integration of range maps into a unique accurate representation is made tricky mainly because of the presence of noise in the viewpoints positions and in the range estimates. In the present case, the scene is captured by a CCD camera system and the depth maps are estimated by a stereovision technique. This approach makes the problem of integration particularly thorny. In fact, the range maps are generally redundant but corrupted by noise and not always coherent with each other. The integration method presented in this paper is based on a fundamental principle: whatever the scene is, the range maps must be consistent with each other. This principle is used as a constraint to discard noise and increase the 3D data accuracy and to identify and remove the redundancies leading to a minimal accurate representation. This phase is realized through the detection of inconsistencies between the range maps of the different viewpoints, the identification and the removal of the most inconsistent points, and the fusion of the remaining redundant points. The process is repeated until the depth maps are coherent with each other. Finally, the facet model is built by incrementally integrating the coherent depth maps. This system is independent of the depth estimation part and can process any set of depth maps of any scene.