The AdaBoost algorithm for vehicle detection based on
CNN features
Xiaona Song
1,2
Ting Rui
1
Zhengjun Zha
3
Xinqing Wang
1
Husheng Fang
1
1
College of Field Engineering, PLA University of Science and Technology, Nanjing, China
2
College of Mechanical Engineering, North China University of Water Resources and Electric Power, Zhengzhou, China
3
Hefei Institute of Intelligent Machine, Chinese Academy of Sciences, Hefei, China
songxiaona1@126.com, rtinguu@sohu.com, zhazj@iim.ac.cn, ruwaye@126.com
ABSTRACT
Vehicle detection is the key technology of Intelligent
Transportation System(ITS), Self-Driving Cars and Active
Driving Assist. Existing methods have some problems with the
robustness and real-time performance. This paper proposes the
AdaBoost algorithm for vehicle detection based on CNN features.
We set up a shallower CNN than traditional one to extract features
to reduce the complexity of the network. On this basis, AdaBoost
classifier with the ability for rapid classification is used to
complete the vehicle detection. The combination of shallow CNN
and AdaBoost algorithm improves the robustness and real-time
performance of vehicle detection. The result of the experiments
proves that the algorithm not only improves the real-time
performance but also the recognition rate.
Categories and Subject Descriptors
I.4 [Image Processing abd Computing Vision]
:
Scene
Analysis –Object recognition
General Terms
Algorithms
Keywords
vehicle detection ,CNN, AdaBoost, shallow CNN
feature-extractor
1. INTRODUCTION
Nowadays, the Intelligent Vehicle and Active Driving Assist have
become the research hotspots. The application of these technology
are all based on vehicle detection. The technology of vehicle
detection has achieved great progress and the detection precision
has been improved through several decades of development.
Generally, there are two kinds of vehicle detection: non-vision
sensor inspection and vision inspection. The vehicle detection
based on computer vision gains popular and is taken more and
more attention by researchers because of its low price and
convenient installation. But because of the influence of complex
environment, imaging angle ,illumination change and the variety
of vehicle, the vehicle detection is still a challenging problem.
As one method of deep learning[1,2] , Convolutional Neural
Network (CNN)has been the research hotspot of the domain of
image recognition and it is very excellent in the automatic feature
extraction. In the ImageNet Large-Scale Visual Recognition
Challenge 2014 (ILSVRC14)
[3]
, the GoogLeNet based on CNN
got the best result, and the network is 22 layers deep when
counting only layers with parameters (or 27 layers if also count
pooling). Even the binary classification of vehicle detection, in
order to have a high detection precision , the network may need at
least 3 or 4 layers. The vehicle detection demanding for real
time is rigorous, complex network is harmful to the improvement
of detection speed. In 2001, Viola et. proposed a real time face
detection based on AdaBoost learning algorithm
[4]
, and this
approach is approximately 15 times faster than any previous
approach while achieving high detection accuracy.
To meet the requirements of real time and robustness, an
AdaBoost algorithm
[5]
based on CNN features are proposed in this
paper. Firstly, the features of vehicle are extracted from CNN. The
CNN is only used for extracting feature other than recognition, so
we can setup a shallower network than traditional using. By this
way, the complexity of network is reduced evidently. Then, the
AdaBoost algorithm is used to complete the final classification in
the feature space extracted by CNN. This method makes full use
of advantages of CNN and AdaBoost, and makes the algorithm
have good performance on robustness and rapidity.
2. RELATED WORK
Vehicle detection based on vision includes the methods based on
geometrical feature, template, and machine learning . Soo Siang
Teoh et al.
[6]
use the edges and symmetrical characteristics of
vehicles to detect the vehicles. Minkyu Cheon et al.
[7]
use
shadow regions that appear under vehicles to detect the vehicles.
Tan F et al.
[8]
use shape template to detect side-view car. Zehang
Sun et al.
[9]
use the evolutionary gabor filter to extract features and
use SVM to train the features to get a classifier for vehicle. Han
Feng et al.
[10]
firstly determine the area which is possible the
vehicle and then extract the HOG feature, at last use SVM to
verify the result.
The model of these methods are shallow, and these shallow model
can’t represent the complicated problems for the situation of
limited samples, and the generalization ability is constrained in
complex classification. What’s more, all these methods are based
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee p
rovided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. Copyrights for
components of this work owned by others than ACM must be honored.
Abstracting with c
redit is permitted. To copy otherwise, or republish, to
post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from Permissions@acm.org.
ICIMCS '15, August 19-21, 2015, Zhangjiajie, Hunan, China
© 2015 ACM. ISBN 978-1-4503-3528-7/15/08
$15.00
DOI: http://dx.doi.org/10.1145/2808492.2808497
资源评论
