Performance-oriented view planning for automated object reconstruction

Abstract

There is a growing demand for 3D virtual models of complex physical objects in a wide range of applications. Thus far, the task of reconstructing high quality object models with range cameras (a shape measurement sensor) has been a manual, time consuming process. There is a growing demand to automate the model building process. The key open problem is to find an accurate, robust and efficient view planning algorithm—that is, the process of determining a suitable set of sensor viewpoints and imaging parameters for a specified reconstruction task with a given imaging environment. This thesis presents and analyzes the performance of a model-based approach to view planning for automated object reconstruction. The view planning approach is “performance-oriented” in the sense that it is based on a set of explicit quality requirements expressed in a model specification and incorporates performance models for the range camera and positioning system. A new theoretical framework for the view planning problem is presented as an instance of the set covering problem with a registration constraint and is formulated as an integer programming problem. The theoretical framework facilitates a more intuitive understanding of the nature of view planning. It also provides an explicit mathematical formulation amenable to a well understood set of exact and approximate solution techniques. The measurability impact of pose error is studied in detail and counter-measures are presented to mitigate its effects. Methods for sparse sampling of object surface space and viewpoint space are presented and their performance is characterized.