4. MODEL DRIVEN FEATURE EXTRACTION AND
MATCHING ON SEQUENTIAL IMAGES
This is a critical part of acquiring the reliable external
constraints from image sequences. The approximate positions
of road centerlines on the image can be predicted by the
projection of the model. The processing window can be
constrained in a local area.
4.1 Model Driven Feature Extraction
Firstly, A new oriented edge detector 1s introduced to detect
edge candidates and compute the edge magnitude and
orientation (Tao et al., 1996). Secondly, a feature filtering
algorithm is proposed to filter the undesired feature points
based on the following three constraint components
represented by a group of rules:
eSingle edge constraint: Rule(1) The orientation of the
desired edge should be perpendicular to the predicted line with
an acceptable tolerance of +20°. Rule(2) The magnitude of the
edge should surpass a certain threshold.
eDual-edge constraint: road centerlines are painted by
white or yellow markings with a certain width. Based on this
knowledge, a dual-edge constraint is designed. A dual-edge is
composed of a pair of edges locating on the same row of the
image. Rule(3) The orientations of these two edges must be
opposite to each other. Rule(4) The distance between these two
edges should be within a few pixels (2-5 pixels). Rule(5) The
average gray value of the zone between these two edges must
surpass a threshold and must be brighter than that of the
surrounding areas.
eShape constraint an extracted edge should be located
along a smoothed line and may not have big difference from
the predicted line. Rule(6) If the distance between a dual-edge
with its nearest one is beyond a range, this dual-edge will be
removed as a blunder (continuity constraint). Rule(7) The
position of the center point of the dual-edge should locate
along a smoothed line (smoothness constraint ).
Rule (7) has been implemented by a modified Hough transform
(Tao et al., 1996). The distance between the center point of
dual-edge and the predicted line 1s projected onto the Hough
space. After the accumulation of the number of the distances,
the majority of the distance range is determined and then the
blunders are recognized. After feature filtering, the output
feature point is the center point of the dual-edge.
4.2 Constrained Matching Range
Because the orientation parameters and the height parameter of
each camera station are known, the corresponding point on the
image B of a given point on the image A can be estimated.
Along with the eppipolar line constraint, the searching range
for matching on the image B can be located (Tao et al., 1996).
In figure 3, the dotted straight line means the eppipolar line.
Three eppipolar lines corresponding to three image pairs pass
through the point Po . The box shows the constrained matching
range. This method provides an effective way to address the
enduring problem of matching images with large geometric
discrepancies.
860
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
4.3 Spatio-temporal Consistency Matching
In order to make use of sequential image information, a scheme
of spatio-temporal consistency matching is proposed. The
scheme is designed as follows (in figure 3, assuming the
current left image is the master image):
1. For each extracted point on the master image, perform
matching on the other three images within the constrained
window. Kanade's SSD correlation criteria (Kanade et al.,
1992) are used in our matching.
2. For the matched point on the next left image and the current
right image, locate their best matches on the next right image
in their constrained matching ranges, respectively (the arrow
indicates its own constrained matching range in figure 3).
3. If two of three matching points on the next right image are
closer than 1.5 pixels, the matching point pair is determined. In
figure 3 case, Po and P; are considered as a matching point
pair because the distance between the point “A“ and the point
“e“ is closer than 1.5 pixels.
next left image next right image
To est I « PEN eppipolar
P; à | line
| A i | EN !
LE A EA I
| efx 1 pero | constrained
\ ° Ï p ;
Lon Hie: Lec I "| matching
B ITS des ro Mk Tange
b Po :N | 4 7
Ó Dt "
bu b tn pod ce apod
current left image current right image
Figure 3. Spatio-temporal matching
Each image will be taken as the master image, and the above
process will be executed in a loop. Finally, using
photogrammetric intersection, the corresponding 3D point of a
matching point pair can be computed and will act as an
external energy source (as shown in figure 2b).
S. EVALUATION OF THE APPROACH
5.1 Computational Aspects
Several parameters must be set: a desirable trade-off can be
achieved based on the trajectory accuracy outputting from the
Kalman filtering algorithm. If the accuracy of trajectory is low,
then e; —0.3 (e2—1-e;) is chosen, otherwise e;—0.5. To keep a
smooth centerline shape, a=0.7 and B=0.5 are used. w;=0.5,
@>=1.0 and @w3=0.8 are selected for representing historical
effects of the external energy.
Two main parts of the approach involving time-consuming
aspects are: (a) The computation of the inversion in equation
(7) by LU decomposition needs O(m) time. In our
implementation, control vertices are seeded every 2.0~2.5
meters along the model. The dimension (m) of the inversion is
(n
ca
co
ral
pr
Ux
en
Se
CO
ro
fa
ro
ga
bl
ap
ro
di
te
m
