wish
s not
hs in
ation
, also
ment
sis of
out a
| that
point
ment
PACE
ns of
ions.
tours
mera
| and
1 this
| EPI
shirt
The third step detects edge-like features in the EPI. Currently these are positive and negative
gero-crossings [Marr 1980] of the difference of Gaussians in the plane of the EPI. The zero-
crossings indicate places in the EPI where there is a sharp change in image intensity, typically
at surface boundaries or surface markings. Figure 15 shows the features detected in the EPI.
These are drawn here in cartesian image coordinates, but recall that the analysis maps these
(knowing the camera principal point and attitude) to lines of sight. Figure 16 shows the .
representation of these features in the space of lines of sight.
/, fEE-
/ À
^L: >
wud ^" Lad ^ am] | awd | Ged | Geel | asd | tu
U-T Ledgals for slice Y=30 U-T Ledgois for slioe Y «90 :
Fig. 15. Edge features in EPI Fig. 16. Lines of sight representation
The fourth step fits linear segments to the lines of sight (actually, to their Euclidean projections).
It does this in two passes. The first pass partitions the connected lines of sight at sharp corners
by analyzing curvature estimates along the path. The second pass applies Ramer's algorithm
[Ramer 1972] to recursively partition the remaining segments into linear pieces. Figure 10 shows
the line segments derived from the edges in Figure 15.
The fifth step builds a description of the line segments that links together those that are
collinear. The intent is to identify sets of lines that belong to the same feature in the world.
By bridging gaps caused by occlusion, the program can improve its estimates of the features’
locations as well as extract clues about the nature of the surfaces in the scene. The dashed
lines in Figure 10 show those linear features that are linked together. Line intersections indicate
temporal occlusions. For each intersection, the feature with the smaller slope is the one that
occludes the other.
The sixth step computes the x-y-z locations of the world features corresponding to the EPI
features. The world coordinates are uniquely determined by the location of the epipolar plane
associated with the EPI and the equation of the feature line in that EPI. To display these
three-dimensional locations, the program plots the two-dimensional coordinates of the features
in that particular epipolar plane. Figure 17 shows the epipolar plane coordinates for the features
shown in Figure 10. The shape and size of each ellipse depicts the error associated with the
feature’s location (these are determined by the eigenvalues and eigenvectors of the covariance
matrix of the linear fit).
*^ i
1 N ‘4 I~, i d ",f un
3
val
3 ent
eu 7
4 0 ' 4
1 . "n 9.5]
a po vot. ]
3 NW :
À LAM
ze] NM ; 7
3 led features have 243
ae accuracy (in inches): 1
1: (0.06, 0.009) 3
2.40, rm
1 4— cames puis: = (0.86, 0.03) 1
1 3: (0.25, 0.04) 1
Sw i Vt 2 TTTTUUTrrUprrPTDUCTUYYYY!Y rrr SN”
A TT Aw y p Y Ai Tr de TT A Tr Al TY du TT EZ , HV Yt EN oO T a EA AA AR ARP aan le ne
X-Z Points for slice Y=30 X-Z Points for slice Y=30
Fig. 17. (x, z) locations, 99% confidence ellipses Fig. 18. Free space for 2 features
125