Stereo Vision

1
Stereo Vision

• Static Stereo

2
Static Stereo Pipeline

• Image Acquisition
• Camera Modeling
• Feature Extraction
• Correspondence Analysis
• Intensity Based
• Feature Based
• Triangulation
• Interpolation (Approximation)

3
Standard Stereo Geometry
4
yL f YL
yL ZL
xL f XL
xL ZL
xR f XR
xR ZR
yR f YR
yR ZR
5
Constraints Assumptions

• Epipolar Constraint
• Uniqueness Assumption
• Compatibility Assumption
• Continuity of Disparities Assumption
• Compatibility of Features Assumption
• Disparity Disparity Gradient Limit
• Ordering Constraint

6
Intensity BasedCorrespondence Matching

• Block Matching
• Dynamic Programming

7
Block-Matching Method
8
Block-Matching Method
9
Block-Matching Method
10
Block-Matching Method
11
Dynamic Programming
(dis)similarity measure
Cumulative dissimilarity function
d (d1, d2, , dM)
12
Dynamic Programming
13
Dynamic Programming
dmin ? ?y(xL) ? dmax 1 ? xL ?y(xL) ?
M MIN(xL) ? ?y(xL) ? MAX(xL) MIN(xL) max
xL M, dmin MAX(xL) min xL,
dmax MINIMUM(xL, ?y(xL)) ? ?y(xL 1) ?
MAX(xL 1) MINIMUM(xL,d) max d 1, MIN(x 1)
14
Dynamic Programming
DSPy(x,d) DSPy(1,d) MSE(1,y,d) d ? ?
MIN(1) ? ? ? MAX(1) DSPy(x,d) DSP(x 1,
backTrace(x,d)) MSE(x,y,d) DSPy(x 1, dprev)
MINIMUM(x,d) ? dprev ? MAX(x 1)
15
Dynamic Programming
Gimelfarb algorithm
16
Dynamic Programming
Gimelfarb algorithm
17
Feature BasedCorrespondence Analysis

• FBCA by Zero-Crossing Vectors

18
Feature Based Correspondence Analysis
ADVANTAGES

• Ambiguities are reduced
• Less sensitive to photometric variations
• More accurate (sub-pixel accuracy)

19
FBCA by Zero-Crossing Vectors

• Edge detection using LoG operator
• If a point(i,j) is not zero-crossing
• e?(i,j) (0,0)
• If a point(i,j) is zero-crossing

20
FBCA by Zero-Crossing Vectors

• Find correspondence candidate pairs
• zero-crossing vector angles differ less than the
  threshold ?0.
• Binary assignment function
• If pixel pair is correspondence candidates
• M?L (i,j,?) 1 and M?R (i-?,j,?) 1
• Otherwise
• M?L (i,j,?) 0 and M?R (i-?,j,?) 0

21
FBCA by Zero-Crossing Vectors

• Global Disparity Histogram

22
FBCA by Zero-Crossing Vectors

• Disparity candidate multi-interval

0 lt a lt 1
23
FBCA by Zero-Crossing Vectors

• Local Disparity Histograms
• Placed image window

n? x n? image window n? ?2 ? ? ? is the
standard deviation in LoG operator
24
FBCA by Zero-Crossing Vectors

• Local Disparity Histograms

25
FBCA by Zero-Crossing Vectors

• Edge points with different ?

? 1.41
? 3.18
? 6.01
a 0.5
26
FBCA by Zero-Crossing Vectors

• The resolution where the the difference between
  the largest and second largest values in the
  local disparity histograms are larger is chosen.
• An assignment is chosen if and only if L(r,s) and
  R(r,s) exceed a given threshold and if
  and only differ slightly.
• The final scalar disparity

27
The End