21-07-2014, 01:43 PM
Eigen-based traffic sign recognition
Eigen-based traffic sign.pdf (Size: 1.04 MB / Downloads: 55)
Abstract
This study’s purpose is to introduce eigen-based traffic sign recognition. This technique is based on invoking the
principal component analysis (PCA) algorithm to choose the most effective components of traffic sign images to classify an
unknown traffic sign. A set of weights are computed from the most effective eigen vectors of the traffic sign. By using the
Euclidean distance, unknown traffic sign images are then classified. The approach was tested on two different databases of
traffic sign’s borders and speed limit pictograms that were extracted automatically from real-world images. A classification
rate of 96.8 and 97.9% was achieved for these two databases. To check the robustness of this approach, non-traffic sign
objects and occluded signs were invoked. A performance of 71% was achieved when occluded signs are used. When signs
were rotated 10 degrees around their centre, the performance became 89% when traffic signs’ outer shapes were used and for
rotated speed limit pictograms the result was 80%
Introduction
Currently, the driver support systems (DSS) are under
development by most vehicle manufacturing companies. The
reason why DSS are critical in intelligent vehicles is because
in the event of the driver being intoxicated or feeling drowsy
or lethargic, sometimes a slight miss-concentration from
him may cause deadly accidents. Therefore DSS play an
important role to prevent road accidents.
Traffic sign recognition is challenging as visibility is affected
by weather conditions such as fog, rain, clouds and snow [1]. The
colour information is very sensitive to the variations of light
conditions such as shadows, clouds and the sun [1–3]. It can
be affected by the illuminant colour (daylight), illumination
geometry and viewing geometry [4]. The presence of objects,
such as buildings or vehicles, similar in colour and shape
to the road signs in the scene under consideration can also
affect recognition efficiency. Signs may be found disoriented,
damaged or occulted. If the image is acquired from a moving
car, then it often suffers from motion blur and car vibration.
Fig. 1 depicts these potential traffic sign problems.
The process of traffic sign recognition is divided into two
steps; first, a traffic sign is detected in the image, later this
traffic sign is recognised using a shape recognition
algorithm. Further the algorithm is extended to classify
traffic signs based on information given in the pictogram.
Principal component analysis (PCA), which is also called
the Karhunen –Loeve transform, was invented by Karl
Pearson in 1901 [5]. The method, which aims to reduce the
volume of information to be handled, transforms a number
of possibly correlated variables (redundant information) into
a smaller number of uncorrelated variables (principal
components). It is used in different fields including
Relevant work
In recent years, research in traffic sign recognition has grown
rapidly because of the real need for such systems in future
vehicles. Performance indexes headed for by these systems
include high recognition rates, real-time implementation,
PCA for images
PCA, also called orthogonal linear transformations, is a
technique that transforms the data into a new coordinate
system based on its variance [11]. Eigen-based systems,
which are based on PCA, are used to solve computer vision
problems. Eigen faces developed by Sirovich and Kirby
[12] are used by Turk and Pentland [13] for face
recognition. The training of the eigen-based system for
images can be summarised as follows:
1. Convert the training images into a set of column vectors,
calculate the mean image of these column vectors and
Colour segmentation
Colour segmentation is carried out by a shadow and
highlight invariant algorithm [15] in which RGB images are
converted into HSV colour space. The HSV colour
space is chosen because the Hue feature is invariant to
shadows and highlights. The values of H, S and V are
normalised into [0,255]. Although normalised H is used as a
priori knowledge to the algorithm, normalised S and V are
invoked to specify and avoid the achromatic subspaces in
HSV colour space [16]. When the H value of colour of the
pixel in the image under consideration is within the specified
colour of the traffic sign, and its S value indicates that this
colour is not in the achromatic area, then the corresponding
pixel in the output image is set to white, otherwise it is set to
black. Fig. 2 depicts results of colour segmentation.
Normalisation
Candidate object’s centre of gravity is first computed to
specify its location in the image and to extract it from test
image. This centre of gravity is then used to represent the
centre of the normalised image.
Experiments and results
To evaluate the performance of the eigen-based traffic sign
recognition, the database presented in the previous section
was used for training and testing of the PCA. A number of
experiments were conducted to evaluate the performance of
the proposed approach. Each of the experiments described
later was implemented for the traffic sign borders and for
the speed limit pictograms. Non-traffic signs images are
Conclusions and future work
In this paper, the eigen-based traffic sign recognition was
presented. The eigen-based system is well adapted to the
traffic sign recognition system. It is invoked to solve theproblem of classifying traffic signs according to their shapes
and interiors. The system was tested on two different
databases representing the outer shapes of traffic signs and
pictograms of speed limit signs. Images of traffic signs in
these databases were extracted from the original images,
normalised and centred in order to fit the requirements of
the PCA algorithm. The testing detailed in this paper
produced a high classification rate of traffic signs. The
performance was about 96.8% for traffic sign shapes and
97.9% for speed limit pictograms. The approach showed
high robustness to two of the most important cases in traffic
sign recognition. In the case of occluded signs, the
performance of classification achieved by this approach is
71%. In the case of rotated signs, this approach achieved
correct classification of 89% when rotated signs outer
shapes were used. For rotated speed limit pictograms, the
classification rate was 80%