Calibration and component placement in structured light systems for 3D reconstruction tasks.

Abstract

This thesis examines the amount of detail in 3D scene reconstruction that can be extracted using structured-light camera and projector based systems. Structured light systems are similar to multi-camera stereoscopic systems, except that a projector is use in place of at least one camera. This aids 3D scene reconstruction by greatly simplifying the correspondence problem, i.e., identifying the same world point in multiple images. The motivation for this work comes from problems involved with the helical tomotherapy device in use at the University of Minnesota. This device performs conformal radiation therapy, delivering high radiation dosage to certain patient body areas, but lower dosage elsewhere. The device currently has no feedback as to the patient's body positioning, and vision-based methods are promising. The tolerances for such tracking are very tight, requiring methods that maximize the quality of reconstruction through good element placement and calibration. Optimal placement of cameras and projectors for specific detection tasks is examined, and a mathematical basis for judging the quality of camera and projector placement is derived. Two competing interests are taken into account for these quality measures: the overall visibility for the volume of interest, i.e., how much of a target object is visible; and the scale of visibility for the volume of interest, i.e., how precisely points can be detected. Optimal calibration of camera and projector systems is examined as well. Calibration is important as poor calibration will ultimately lead to a poor quality reconstruction. This is a difficult problem because projected patterns do not conform to any set geometric constraints when projected onto general scenes. Such constraints are often necessary for calibration. However, it can be shown that an optimal image-based calibration can be found for camera and projector systems if there are at least two cameras whose views overlap that of the projector. The overall quality of scene reconstruction from structured light systems is a complex problem. The work in this thesis analyzes this problem from multiple directions and provides methods and solutions that can be applied to real-world systems.