) the best
voi
1000 1500
road (red
1, but the
(sts a 9x9
Euclidean
chs, having
method, a
llelepiped)
ecognition.
rived from
x-classifier
ium values
>d a binary
[he binary
| the next
, Where the
the shorter
:d by these
. parameter
> computed
ectangle
During the
th random
arameter w
the binary
ned for this
oad pixels
e calculating ratio of the road and non-road pixels
divided by the rectangle's area, weighted by the
reciprocal logarithm of the road pixels.
The genetic algorithm technology needs a suitable fitness
definition, but all of the above mentioned have practical
considerations. The first fitness optimization is a maximum
search, but globally all road pixels can be categorized also by
small rectangles (short road segments) The second fitness
definition fixes this problem by a length/area based weighting,
but behaves too rough and converges drastically into a single
rectangle state. The last definition was found as the best (Eq. 5).
logl# mask pixels)
#roadpixels | 1 (5
#maskpixels )
fitness= | =
where the # operator means the number of pixels. The logarithm
can be also the natural logarithm function. This fitness function
supports the minimum search optimization.
Other, similarly suitable fitness functions can be defined and
their performance has to be evaluated.
The listed fitness functions can be implemented not only
considering genetic algorithms, but with differential evolution,
too.
The execution of the automatic road segment compilation must
be followed by a human evaluation and network creation.
4. RESULTS
4.1 Segmentation results
The support vector machine training results a support vector set
of 1701 elements. The training lasts 140 s, thereafter the
classification of the whole image requires only 56 s. The in-
sample confusion matrix is as follows:
SVM Classified
Road Non-road| 2 OE PA
$ [Read 1146 0|1146| 0,00%| 100,00%
F |Non-road 142 481| 623|22,79%| 77,21%
3 1288 481] 1769
CE 11,02% 0,00% 91,9796| 88,6096
UA 88,98%| 100,00% OA AA
Table 1. In-sample accuracy for SVM classifier
The overall accuracy is 91.9796, the average accuracy is
88.60%, the accuracy of road recognition (producer's accuracy)
is 100%. The output image is visually not satisfactory, because
a lot of roof pixels are also ordered to the road category.
The SOM solution has produced a label image containing 9
different label values. The most similar label images are the sh
and gh so their union is handled as road category. The
confusion matrix was derived in the same way (Table 2).
As one can see, the overall accuracy is less, than before:
89.77%, the average accuracy is similarly lower: 85.55%.
The hyperbox method was the last classification algorithm,
which was also evaluated. This technique is very fast, because
the relational operations are of very low level, so they can be
executed quickly. The error matrix is in Table 3.
The average accuracy is the highest at this method, it is 91.97%,
the overall accuracy is very close to this value with 91.63%.
The reason is the highest producer’s accuracy that means this
method could recognize the non-road pixels with the highest
rate.
SOM Classified i OE PA
Road Non-road OE PA
+ |Road 1144 2|1146| 0,17%| 99,83%
s Non-road 179 444| 623| 28,73% | 71,27%
2 1323 4461 1769
CE 13,53% 0,45% 89,77% | 85,55%
UA 86,47% 99,55% OA AA
Table 2. In-sample accuracy for SOM classifier
Box-m Classified I OE PA
Road Non-road OE PA
+ |Road 1041 105| 1146| 9,16%| 90,84%
Non-road 43 580| 623| 6,90%| 93,10%
2 1084 685] 1769
CE 3,97% 15,33% 91,63%| 91,97%
UA 96,03% 84,67% OA AA
Table 3. In-sample accuracy for hyperbox classifier
The best binary result (Fig. 4) of the hyperbox algorithm
therefore was propagated into the next phase.
Hyperbox classification
Figure 4. Result of the segmentation by the hyperbox technique
Since the segmentation results were not fully satisfactory, and
the genetic algorithm was intended to be tested, a set of
synthetic images were created. These synthetic images contain
only binary pixels labeling the road pixels.
4.0 Road detection results
In the detection phase two genetic technique were applied,
firstly a general genetic algorithm (GA), then the differential
evolution (DE).
The GA tests (11 in total) were run with rectangle width of 7
pixels. The mutation operation can be controlled by its buffer;