7B-2-4
Table 1. Performance of road segmentation
Test
No image pairs
Successfully segmented road
1.
24
100%
2.
64
93%
3.
24
96%
Results presented in Table 1 shows that the segmentation
algorithm presented (Section 2.2) works well. In most of the
cases the road and the points on the road boundary are extracted
correctly. Unfortunately, the measurement of road widths is
inaccurate. The main reason of low accuracy is the presence of
the errors in the parallax measurement. The errors in
determination of the parallax are caused by:
• Presence of multiple road boundary paintings e.g.,
highways.
• Presence of cars, especially long vehicles
• Bad visibility of the road width caused by cars parked
on the roadsides.
• Low respectively high saturation of the images.
• Very strong shadows.
• Perspective effect i.e. foreshortening, (Nalwa, 1993)
3 DISCUSSION
The practical test of automatic measurement was compared to the
model based on classical errors propagation and to semi
automatic measurements. The classical errors propagation model
was based on simulation studies of MMS (Gajdamowicz, 1994).
It was assumed that the positions of the camera was establish
with double differential GPS method with position errorr Xo =
oy 0 = crZo = 0.02 m. For the camera system the terrestrial
photogrammetry model (normal case) was assumed. The cameras
were situated 1.6 m from each other and 2 m above the road
surface. The camera system was considered to point little
downward with the pitch angle a> = -22^ (k, q> - 0^). The
accuracy of angular measurements was considered to be
a („= <J V - a K = 0®r. The road was assumed to be a planar
surface (XY) with the Y-axis parallel to the direction of
photography, the X-axis perpendicular to Y, and the Z-axis
pointing up, and corresponding to heightAssuming no errors in
the system the 12 m road width (in distance of 15 m from the
cameras), can be measured with a precision oicrdT = 0.11 m.
The precision of the road width measurement is proportional to
the distance from the cameras and to the distance of the road
boundaries from the road centreline (Fig 6).
Figure 6. Theoretical precision of the stereo measurements (from
Gajdamowicz, 1994)
The road width was measured on more than 300 image pairs.
From those images 112 were selected as representative ones
(Table 1). The discrepancy values (e) were calculate (Eq. 3)
e = measured value (Automatic) ~ 8 iven value (Semi Automatic)
(3)
Then, the accuracy of the method was determined by calculating
the root mean square error (discrepancy).? (Hallert et.al, 1967).
The s values in Test 1, Test 2 and Test 3 were 50.50 m, 64.02
and 55.98 m respectively. Such huge errors are mainly due to
gross errors i.e. high discrepancy values caused by incorrect
parallax determination (Section 2.3). To estimate the accuracy of
the algorithm for automatic road measurement some data pre
processing is required, namely, rejection of outliers. To
overcome the problem of outliers a simple rejection rule based on
robust median and median deviation (rather than mean and
standard deviation) was used. Such method is called “Hubber
skipped means” and applies the following rule: Reject
everything, which is more than 5.2 median deviations away from
the median, and take the mean from the reminder (Hampeit al.,
1986). The median value was calculated from absolute values of
discrepancy (e).
Let us closely analyse the test 2 in Table 1, (Figure 7). In a set of
64 measurements one can easily notice extreme outliers that
reach from 20 m up to 380 m. Because of such errors the
discrepancy respectively root square errors are always very large.
The results of measurement in 55 stereo pairs were analysed
(Figure 7). The semi-automatic method was considered as a
reference. The total amount of outliers in the set of 55
measurements was 32%. After rejection of outliers the average
road width measured automatically was 8.18 ra&dA = 0.83 m
where semi-automatic method resulted in an average road width
of 8.36 m, (JdSA ~ 0-29 m.
In comparison with the theoretical valuecr £ /7' = 0.11 m, the
precision of the automatic and semiautomatic measurement is
much lower. In case of the automatic method the main reason for
low precision is because of errors in the matching procedure.
Even if least square matching was used, computation of