Next: Recognizing Volumetric Models Up: Active Perception Previous: Building Volumetric Models

Multiple View Integration

Authors: [tex2html_wrap4168]G. Soucy, F.P. Ferrie

Investigator username: ferrie

Category: perception

Subcategory: active perception

Problems in robot vision such as object manipulation and trajectory planning require that multiple views of data (e.g. time series, multiple sensor, etc.) be integrated into a consistent interpretation of the scene. Our approach to this problem is based on a surface reconstruction algorithm that combines information from different viewpoints. It does so without explicitly computing correspondence and without invoking a global rigidity assumption. Motion parameters (rotations and translations) are recovered locally under the assumption that the curvature structure at a point on a surface varies slowly under transformation. The recovery problem can thus be posed as finding the set of motion parameters that preserves curvature across adjacent views. This might be viewed as a temporal form of the curvature consistency constraint used in image and surface reconstruction.

To reconstruct a 3-D surface from a sequence of overlapping range images, one can attempt to apply local motion estimates to successive pairs of images in pointwise fashion. However this approach does not work well in practice because estimates computed locally are subject to the effects of noise and quantization error. This problem is addressed by invoking a second constraint that concerns the properties of physical surfaces. The motions of adjacent points are coupled through the surface, where the degree of coupling is proportional to surface rigidity. We interpret this constraint to mean that motion varies smoothly from point to point and attempt to regularize local estimates by enforcing smooth variation of the motion parameters. This is accomplished by applying a regularizing filter to the local estimates as a second stage of processing.

By combining curvature and motion consistency constraints we have been able to derive a multiple view reconstruction algorithm that can accurately reconstruct three-dimensional surfaces from a sequence of range images.

Next: Recognizing Volumetric Models Up: Active Perception Previous: Building Volumetric Models