Information Technology Reference
In-Depth Information
Fig. 1
Flowchart
of the maximum
likelihood
estimation
based
flat surface
detection
algorithm
y
]
C
(
x
,
y
)
,where
C
(
x
,
y
) represents the intensity structure of
x
c
(
x
,
y
)=[
x
,
y
2
be two eigenvalues of matrix
C
(
x
,
y
). A corner
the local neighborhood. Let
λ
1
and
λ
1
,
point can be detected if min(
2
) is larger than a pre-defined threshold.
Once holding the points of interest, we then apply the sum squared of differences
correlation method to match these corner features. Using the matched features, we
exploit the well-established epipolar constraints to further refine the correspondence
of features. The camera parameters are then used for recovering the scene geometry
[7]. As an example, Fig. 2(a) and (b) show the original images superimposed by the
extracted corner features using the Harris corner detector [19], (c) is the disparity
map and (d) refers to the estimated depth map according to the relationship:
D
=
fd
/
z
,where
D
is depth to be computed,
f
focal length,
d
introcular distance and
z
estimated disparity.
λ
λ
2.2
Iterative RANSAC Planar Surface Detection
RANSAC planar estimation is supposed to effectively work in the presence of data
outliers. This method starts from fitting a plane to a set of 3 points (considered as
inliers) randomly selected from the matched corner features. Other image points are
then evaluated using the Euclidean distances between these 3-D points and the fitted
plane. If the points fall in a pre-defined region, then they will be classified as inliers.
Otherwise, the points will be removed from the consideration of coplanarity. These
steps are repeated until a count limit is reached. In a classical RANSAC plane fitting
approach, the iteration is terminated by either a user-specified number or the number
of outliers falling below a pre-defined threshold. This heuristic trick cannot handle
general situations, where either under- or over-estimation usually appears.
