This paper describes a method for determining an object's pose given
its 3D model and a 2D view. This 2D-3D registration problem arises
in a number of medical applications, e.g. image guided spine procedures.
Previous approaches often rely on a good initial estimate of the pose parameters
and an optimization procedure to refine this initial pose estimate, e.g.
the iterative closest point (ICP). However, such algorithms can identify
local minima as global minima, leading to registration errors, if the initial
pose is not carefully chosen. The specification of the appropriate initial
conditions, however requires user interaction and is time consuming.
We propose an approach where sample 2D views are generated from the 3D
model and matched against the given view (2D-3D registration). Additional
views are then generated in the vicinity of the best view and the procedure
is repeated until convergence. Results of estimating the coordinates
of a vertebrae spine bone from its 3D model, obtained from volumetric (CT
or MR) data, and a 2D view, as might be obtained from fluoroscopic data,
demonstrates that the pose can be reliably obtained without requiring extensive