ROAD DETECTION BY NEURAL AND GENETIC ALGORITHM IN URBAN
ENVIRONMENT
A. Barsi*
Dept. of Photogrammetry and Geoinformatics, Budapest University of Technology and Economics,
Budapest, Hungary — barsi@eik.bme.hu
Commission III, WG III/4
KEY WORDS: Segmentation, Genetic algorithm, Differential evolution, Road detection
ABSTRACT:
In the urban object detection challenge organized by the ISPRS WG III/4 high geometric and radiometric resolution aerial images
about Vaihingen/Stuttgart, Germany are distributed. The acquired data set contains optical false color, near infrared images and
airborne laserscanning data. The presented research focused exclusively on the optical image, so the elevation information was
ignored. The road detection procedure has been built up of two main phases: a segmentation done by neural networks and a
compilation made by genetic algorithms. The applied neural networks were support vector machines with radial basis kernel function
and self-organizing maps with hexagonal network topology and Euclidean distance function for neighborhood management. The
neural techniques have been compared by hyperbox classifier, known from the statistical image classification practice. The
compilation of the segmentation is realized by a novel application of the common genetic algorithm and by differential evolution
technique. The genes were implemented to detect the road elements by evaluating a special binary fitness function. The results have
proven that the evolutional technique can automatically find major road segments.
1. INTRODUCTION
The working group III/4 of ISPRS Commission III has
conducted a challenge of detecting man-made objects from
digital aerial imagery. These objects are urban objects (roads,
trees, cars etc.) and 3D buildings. This challenge has been
supported by adequate data sets: there was an aerial
photography campaign near Stuttgart, Germany, when color
images and LIDAR-point clouds were acquired. The raw data
were processed: the aerotriangulation was calculated, and
digital elevation model (DEM) was derived. All of these
products can be downloaded and used for developing object
detection methodologies.
The author is strongly interested in the use of current
developments of computer science in digital image
understanding. Among the modern methods neural network
based classification techniques and genetic algorithms can be
mentioned, which are in the focus of the current paper.
The general theory of the two artificial intelligence tools is
presented, followed by the details of the applied methodology in
image analysis.
2. NEURAL AND GENETIC ALGORITHMS
2.1 Neural networks as classifiers
The artificial neural networks have become widely used tool in
different data processing workflows, among them also in digital
image processing. There are several types of network; therefore
different categorization can be done, for example
* Corresponding author.
e by the architecture (layers, recurrent, delay, transfer
functions etc.),
e by the initialization, training, testing and validation
algorithms,
e by the data structure and configuration,
e by the used pre- and post-processing operations.
The most used neural networks in image processing have a
strong focus on classification, which is appropriate application
for back-propagation (BP) network, radial basis function (RBF)
networks, learning vector quantization (LVQ) networks and
support vector machines (SVM). Beside these supervised
classifiers, the unsupervised category is similarly important: e.g.
self-organizing maps (SOM), competitive learning networks.
[Beale et al, 2012]
The current paper presents applications of SVM and SOM
technologies. The used support vector machine classification is
based on the formula:
c- 2 au x) (1)
where s; is the support vector, x is the vector to be classified,
k() is the kernel function, o; is a weight, b is a bias, and c is
the result, which having a value more than zero means to be
classified in the category, otherwise it is rejected. The applied
kernel function was the radial basis function (RBF), which is a
nonlinear function, so it has the advantage to make right
decision even with complicated class borders. This supervised
classification method must be parameterized by data from
training areas [Beale et al, 2012].
CCT SCD X
TO mJ D ged