Iterative Image Registration Technique with Application to Stereo Vision

1
Iterative Image Registration Technique with
Application to Stereo Vision

• Based on Bruce D. Lucas
• Takeo Kanade

2
Image Registration Problem

• Align a template image F(x) to an input image
  G(x)
• Goal Find the warping w that minimizes the
  difference between F(w(x)) and G(x)

3
Existing Techniques

• Differs in
• Measure of difference L1norm, L2 norm,
  Normalized Correlation
• Search strategy for w exhaustive search,
  hill-climbing search, coarse-fine search

4
Comparison
G(x) is NxN Region of possible w is of size MxM
5
Lucas-Kanade Method1-Dimensional case
(translation only)

• w(x) xh
• Find horizontal disparity h between 2 curves F(x)
  and

6

• Newton-Raphson Iteration
• Alternative form

7
Generalization

• To Multiple Dimensions

• To Arbitrary Linear Transformations
• w(x) xAh ? G(x) F(xAh)

8
Convergence

• h lt 0.5wavelength
• Improve range of convergence by smoothing the
  image
• Coarse-fine (multiresolution) processing
• w(x) speeds up the convergence

9
Application to Stereo Vision

• Depth extraction from stereo vision system
• Finding objects
• Matching objects in the 2 views
• Determining camera parameters
• Calculating depth

10
Applying registration technique

• Relative positions and orientations of cameras
  are known (depths are unknown)

• Depths of objects are known (camera parameters
  are unknown)