29-05-2014, 02:13 PM
Reliable Classification of Vehicle Types Based on Cascade Classifier Ensembles
Reliable Classification.pdf (Size: 744.45 KB / Downloads: 22)
Abstract
Vehicle-type recognition based on images is a chal-
lenging task. This paper comparatively studied two feature ex-
traction methods for image description, i.e., the Gabor wavelet
transform and the Pyramid Histogram of Oriented Gradients
(PHOG). The Gabor transform has been widely adopted to extract
image features for various vision tasks. PHOG has the superiority
in its description of more discriminating information. A highly
reliable classification scheme was proposed by cascade classifier
ensembles with reject option to accommodate the situations where
no decision should be made if there exists adequate ambiguity. The
first ensemble is heterogeneous, consisting of several classifiers,
including k-nearest neighbors (kNNs), multiple-layer perceptrons
(MLPs), support vector machines (SVMs), and random forest. The
classification reliability is further enhanced by a second classifier
ensemble, which is composed of a set of base MLPs coordinated by
an ensemble metalearning method called rotation forest (RF). For
both of the ensembles, rejection option is accomplished by relating
the consensus degree from majority voting to a confidence measure
and by abstaining to classify ambiguous samples if the consensus
degree is lower than a threshold. The final class label is assigned
by dual majority voting from the two ensembles. Experimental
results using more than 600 images from a variety of 21 makes
of cars and vans demonstrated the effectiveness of the proposed
approach. The cascade ensembles produce consistently reliable
results.
I NTRODUCTION
VISION-BASED vehicle make-and-model recognition
(MMR) is an important part of an intelligent traffic system
with a multitude of applications. For example, for vehicle
surveillance for high-security areas, conventional number-plate
recognition systems have to be augmented as a number plate
may be faked or tampered with. An MMR system would offer
valuable assistance to the police in identifying blacklisted vehi-
cles at toll stations or in their search for specific suspect vehicles
from the traffic surveillance image database [1], [2]. With the
increasing demand for security awareness and widespread use
of surveillance cameras, the need for vehicle identification and
classification technologies has become ever more relevant in the
recent years [3]–[5].
I MAGES C OLLECTION AND P REPROCESSING
The local police department of Dushu Lake Higher Educa-
tion Town in Suzhou provided a large collection of vehicle
images recorded with their traffic surveillance cameras over
a one-week period. The images were captured using charge-
coupled device cameras (SP-140N) installed at ten different
intersections, between 7:30 A . M . and 9:50 P. M ., encompassing
a wide range of weather and illumination conditions. From the
recorded images (> 50 000), 620 images of different vehicles
from 15 brands were selected, accounting for 21 classes, includ-
ing Audi, Buick (two classes), Changan, Chery (three classes),
Chevrolet (two classes), Citroen, Ford, Honda, Hyundai (two
classes), Mazda, Nissan, Peugeot, Volkswagen (two classes),
Toyota, and Wulin. Some classes describe different models
of the same brand of cars, e.g., the Hyundai Sonata and the
Hyundai Elantra, when the two models have quite different
appearances. If different models of a same brand shared similar
appearances, e.g., Audi A7/Audi A8, Buick Excelle/Buick Xt,
they would be included into the same class, although the effort
to include only one model of each vehicle brand has been made
in the data preparation process.
F EATURE D ESCRIPTIONS
Although many types of features could be defined for ve-
hicle images, geometrical features pertinent to edge are more
important. Examples of such geometrical features include the
square mapped gradients [9] and oriented-contour point [10].
Here, we introduce two well-developed descriptors appropriate
for expressing edge-rich information.
R ELIABLE C LASSIFICATION VIA C ASCADE
C LASSIFIER E NSEMBLES
Although plenty of supervised learning algorithms, such as
neural networks, kNNs, and SVMs [19]–[21], has been applied
to the various pattern recognition problems, few studies have
addressed the issue of classification reliability, which is about
the confidence in making a decision. Even if good accuracy
could be achieved from some of the existed methods, they may
still be insufficiently accurate to be used in practical situations
such as the recognition of a suspected vehicle during traffic
surveillance. High-confidence classifier is a top priority in such
scenarios. On the other hand, it is often not realistic to make
the assumption that all classes are known beforehand. For
example, the number of vehicle types on road is difficult to
estimate. Another simple case that highlights the importance
of reliability is the presence of missegmented samples from
the preprocessing stage (e.g., the segmentation of ROI) for
most of image recognition issues. A missegmented sample can
represent arbitrary fragments or combinations of proper class
patterns. Therefore, in practical classification systems, it is
desirable to have a reject option in addition to taking a hard
decision for class labels, i.e., an option to withhold a classifier
decision and transfer the ambiguity to human judgement.
CONCLUSION
Accurate and robust classification of vehicle images is still
a challenging task in intelligent transportation system and
surveillance. To find the best description for vehicle images,
many feature extraction methods have been attempted in pre-
vious studies, but all encountered with certain aspects of dif-
ficulties in dealing with various irregularities in the vehicle
images. First, we have strived to demonstrate that the classi-
fication can be improved by using highly discriminative image
features. This can be achieved by the utilization of two well-
known feature description methods, i.e., the PHOG and the
Gabor transform. Second, the theme of classification reliability
was emphasized, which is distinguishable from the previous
research on vehicle-type classification with accuracy as the
only subject. In this paper, a new scheme of cascade classifier
ensembles has been proposed with rejection strategies. Rather
than simply pursuing accuracy, the importance of reject option
was stressed to minimize the cost of misclassifications. The
cascade ensembles employ a serial approach where the second
ensemble is responsible for the patterns rejected by the first,
thus labeling some ambiguous samples as undecidable with
high confidence.