In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Voi. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
(a) (b) (c) (d) (e)
Figure 4. Image processing stages: (a) Source image; (b) Retinex
transformation result; (c) result of retinex and contrast adjustment
stages; (d) overall image preprocessing result; (e) Texton map
After Retinex correction mean value of luminance becomes
equal to one. Before scaling L-component back to nonnal,
contrast adjustment is achieved by squaring it. Then, L-
channel is scaled back to nonnal (Figure 43(c)). This
operation helps to make detection of road defects easier even
in low-contrast images.
Colour correction uses conventional grey world algorithm. In
Lab colour space it consists of the shift of colour components
so as to make mean value of these components be equal to
zero, instead of scaling colour components in RGB space. In
the final stage bilateral filtration is used to smooth image
without loss of important details (Figure 4(d)).
Image segmentation
The hierarchical structures is a powerful tool to analyze data
in many applications. Several basic approaches to
construction of such multi-level image structure exist. The
first approach involves recursive segmentation. An image is
segmented in a large scale, and then segments are
independently split into pieces. Another approach involves
successive segmentation of an image at several scales. But in
this case large segments not necessarily represent
combinations of smaller ones; this fact limits the scope of
application of this method for segmentation.
In this work we used a method based on determination of
strength of the boundaries between segments by means of the
analysis of saddle points between density modes and merging
segments that are weakly separated. For segmentation of the
image in our work the hierarchical version of algorithm of
mean shift, proposed in (Paris, 2007) is used.
This algorithm provides fast hierarchical segmentation on the
basis of idea of the saddle point analysis. Results of this
hierarchical segmentation are shown in Figure 5, where
borders of segments at different levels of hierarchy are shown
in white.
Features calculation
A number of various features are used for classification of
segments. We use colour statistics, such as mean values of
CIE Lab components and mean values of RGB components,
colour variance, Lab components’ percentiles.
To account for shape information we calculate coordinate
statistics, such as mass centre, coordinate variance,
elongation, orientation, area of the segment. Usage of
information about neighbourhood of the segment is also very
informative for road defects detection. Accordingly distance
between mean values of colour components inside segment
(a) (b) (c) (d) (e)
Figure 5.Cascade classification stages, (a) - Input image, (b) -
ground truth image, (c) - 1 st cascade layer result, (d) - 2nd
cascade layer result, (e) - overall algorithm result.
and inside its neighbourhood are also included in the list of
features.
Texton histograms are also used in our system (Leung 1999).
These features are proven to be highly effective in
recognition task and are used nowadays in many detection
and recognition systems (Criminisi, 2006).
Previously created filter bank is applied to the image; filter
output vectors for each pixel are associated with the nearest
texton vectors from previously trained universal texton
dictionary. Then histogram of textons over the segment is
used as feature for classification task. Figure 4(e) illustrates a
resulting texton map, which is an image, where pixels are
labeled accordingly to corresponding textons.
4. ONLINE STAGE OF MAPPING PROCESS
At online stage automatic detection algorithm is applied to
parts of rectified road image. User examines results of
automatic detection on one image part and corrects detection
errors if needed. Then automatic detection is adapted to new
data. After that user goes on to the next part of the road and
again analyses and corrects results of automatic detection.
Accordingly automatic detector is continuously tuned in
order to capture specifics of particular road.
Our system provides various facilities for making process of
error correction easier for the user. The GUI contains a
control which lets user change segmentation level. Operator
is able to mark ground truth in a less detailed level and then
specify it in a more detailed one. It makes user work more
efficient.
Another facility allows controlling tradeoff between detection
rate and false positive rate individually for lane marking and
road defects. For example, user can increase detection rate of
road defects detection (thus increasing false positive rate) by
moving a slider. The change in detection rate is performed by
changing a threshold on classifier output for road defects on
the last cascade layer. This feature helps to significantly
reduce amount of manual work in the beginning of online
stage, when classifiers show instable performance.
5. LANE MARKING AND PAVEMENT DEFECTS
DETECTION ALGORITHM
Our approach is based on cascade classifiers. The idea of
cascades is derived from (Viola, 2002). General workflow of
cascades is the following. There is ordered set of classifiers,
where every subsequent classifier is more "complex" than the
preceding one ("complexity" of the classifiers is defined
depending on specifics of data or application). Input data
