VASC Seminar Announcement
            =========================

Date: Monday, 8/6/01
Time: 3:30-4:40
Place: NSH 3005

Speaker: Karl Kluge
Title: Monocular Vision-Based 2-D Map Building in Hallway Environment

Monocular vision is the only viable sensing modality for mapping in
certain robot applications which impose constraints on robot size,
weight, power consumption, and/or detectability. The problem addressed
in this talk is restricted to mapping the 2-D locations of significant
vertical edges in indoor corridor environments. The research described
covers three key aspects of monocular vision for mapping: feature
detection, feature matching, and solving for structure and motion.
The restriction to hallway environments is exploited in the matching
process through the use of the cross-ratio projective invariant for
four parallel coplanar lines. Use of the cross-ratio invariant allows
exhaustive search of the space of possible feature matches in
low-order polynomial time. The 2-D structure from motion problem is
formulated for a calibrated camera in Cartesian coordinates. Once the
camera rotations are specified, the 2-D problem is linear in the case
of perspective projection (each observation generates a constraint
that the feature lies along the line of sight from the camera focal
point through the image column the feature projected to). Optimization
of the camera rotations is done using standard iterative
techniques. Experimental results using images of a typical hallway are
presented to evaluate the reliability of the matching technique and
the quality of the structure reconstruction.