in these cases 30-50 pixels were set. The fitness function was at
the beginning only the correctly detected road pixels, but very
quickly it was recognized that the length/area weighting is
indispensable. The initial population size was 100 genes, the
number of generations was 100 to 200 thousand epochs. Within
an epoch just one gene was modified by the genetic rules
(mutation and crossover). Since the crossover seemed too
drastic modifier in the rectangle keypoints’ coordinates, it was
ignored in the later runs.
The initial randomly generated population has been laid on the
segmented image, as Fig. 5 demonstrates.
Initial population
Figure 5. Initial population on the segmented image
During the computation the fitness of all genes are calculated,
then a cumulated fitness are derived, which is excellent to
describe the population. The average running time for training
is between 1300 and 3600 s.
The best candidates of the final generation (Fig. 6) can be
linked by human operator in the later phase (so the unnecessary
oblique rectangles can be deleted).
The second test series was executed with the differential
evolution. This procedure implements an algorithm, where three
different genes must be selected, then the mutation modifies the
first gene. Because of the drastic effect of crossover, it has been
avoided also in this phase. The population size is roughly the
same as with GA, but the runtime is definitely lower: 40-50 s.
Even with 200 population size and 200 generations it requires
—50 s CPU time. The reason is that this philosophy modifies all
genes in a single run. The mutation factor was set for 1.0.
The training can be monitored by the cumulative fitness
function, which aggregates the scores for all genes within an
epoch. Fig. 7 shows a typical decrease during 150 generations.
Figure 6. Best 50 candidates of the final GA generation
Population scores
24: r
23%
2:
21H
20r
15- eg
50 100 A
Figure 7. Population cumulative fitness during the evolution
The differential evolution has a drawback, namely the
continuously mutated genes can be modified in a way that they
start resembling each other; in critical case the population
converges into one single position. To avoid that negative
feature, the training is monitored by the variances of the genes.
If there are enough variations within the population, the
variances of the coordinates are high. Resembling can be
detected by dramatical decrease in the variance diagram. A
healthy variance plot can be seen in Fig. 8.
Population variability
350r = - EE
300 -
250r
2007
150 i
1008
Bg 50 100 150
Figure 8. Variance plot to keep track on resembling
The differential evolution method is a global optimization
technique, so the uniformization can be “dangerous”. If one has
sti
Sk
re