next up previous contents
Next: Optimal Sampling for Metrological Up: Active Perception Previous: Enhanced 3D Representation using

Integrating Descriptions from Multiple Views

P. Tremblay, F.P. Ferrie The main goal of this research is to identify and possibly resolve ambiguities in the parts representation of data gathered through autonomous exploration. Conflicts arise during the model building due to limitations of the models used (e.g. superellipsoids) as well as physical constraints. In the former case, these conflicts are mostly due to the parts-segmentation process which is view-dependant while the parts-representation at the most abstract level is view-independant. Thus differences in interpretation often arise in practice. In the latter case, the absence of physical constraints in the modeling process allows the system to produce models which violate basic physical principles (for example, where the maximum-likelihood surface of one part cross and extend beyond measurements on the surface of another). By modeling these physical constraints, we wish to regularize the model parameters to make them consistent across different parts, in terms of space occupancy. To achieve this, the interactions between the different parts of an object are modeled in terms of each part's surface probability distribution function. Reasoning is then applied to obtain an explanation for conflicts in the parts segmentation and propose alternate models.

Thierry Baron
Mon Apr 7 12:54:24 EDT 1997