1 INTRODUCTION
On-road vehicle detection based on computer vision
is becoming more and more important in vehicle
safety, especially in the vehicle assistant driving and
automatic navigation system [1] [2]. However, there
are many problems in vehicle detection algorithm
based on computer vision, such as incapability of
real-time detecting vehicle, low rate of detection,
high rate of false alarm, especially with the condition
of the light change. The vehicle detection rate will
be greatly influenced by vehicle obscured conditions
[2] [3].
Classifier performance is evaluated from three
aspects, that are detection rate as high as possible,
false alarm rate as low as possible and detection
speed as fast as possible. These three aspects restrict
each other, so we need to find a balanced point to
meet the requirements of practical application. The
feature extraction and classification algorithm de-
termines the quality of the classifier. Many features
have been used in current research, including the
Haar-like feature [4] [5], HOG feature [6] SIFT fea-
ture, SURF feature [7], BRIEF feature [7] [8] [9],
edge histogram [10], etc. For the selected these fea-
tures, we need to use the appropriate classification
algorithms to build robust classifiers, such as neural
networks [11], support vector machine (SVM) [6],
Adaboost classification algorithms [4] [5] [12] and
so on.
In order to solve the problem of high false alarm
rate and low detection rate of single classifier, Viola
and Jones put forward to a testing framework [13]
which can handle images very quickly. In this
framework, the input image background region is
quickly discarded by the front layers. More compu-
ting resource is used in the potential area of vehicles.
The application of the simple image processing me-
thod can extract candidate regions rapidly to speed
up the testing process.
However, the cascade structure demonstrated by
Viola in the vehicle detection has obvious flaws
[13]: On one hand, the cascade classifier which is
trained by discrete Adaboost using Haar-like features
can refuse simple sub-window quickly, behind a few
layers, it needs more linear combination of weak
classifiers to decrease the false alarm rate in com-
plex scenarios, which easily leads to excessive learn-
ing and time-consuming. On the other hand, as rec-
tangle feature limitations and the weakness of
relatively weak classifier, the performance of each
layer classifier is not greatly improved while using
more weak classifiers after some layer.
In [14], the author combines Haar-like feature
with HOG feature. However, introduction of HOG
A method of on-road vehicle detection based on comprehensive feature
cascade of classifier
Xiaole Li
College of Information Science and Engineering, Hunan University
Degui Xiao
College of Information Science and Engineering, Hunan University
Chen Xin
College of Information Science and Engineering, Hunan University
Huan Zhu
College of Information Science and Engineering, Hunan University
ABSTRACT: To detecting on-road vehicles rapidly and efficiently, we propose a robust method to improve
the accuracy of on-road vehicle detection rate and reduce false alarm rate. Firstly, we have enhanced the fea-
ture expressive force by combining Haar-like feature with BRIEF feature. Some improvements have been
done for achieving the robustness under the lighting and road conditions changes. Secondly, we have im-
proved the performance of weak classier based on Gentle Adaboost algorithm. Experimental results show that
the detection rate increased by 2.6% compared with the traditional cascade structure of classifier, and the false
alarm rate reduced in some degrees.
KEYWORD: Vehicle Detection; BRIEF; Haar-like; Adaboost;
资源评论
