Copyright © 2018 Authors. This is an open access article distributed under the Creative Commons Attribution License, which permits
unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
International Journal of Engineering & Technology, 7 (4.44) (2018) 177-180
International Journal of Engineering & Technology
Website: www.sciencepubco.com/index.php/IJET
Research paper
Rear-Approaching Vehicle Detection using Frame Similarity
base on Faster R-CNN
Yeunghak Lee
1
*, Israfil Ansari
2
, Jaechang Shim
3
1
Andong National University
2
Andong National University
3
Andong National University
*Corresponding author E-mail: yhsjh.yi@gmail.com
Abstract
In this paper, we propose a new algorithm to detect rear-approaching vehicle using frame structure similarity based on deep learning
algorithm for use in agricultural machinery systems. The commonly used deep learning models well detect various types of vehicles and
detect the shapes of vehicles from various camera angles. However, since the vehicle detection system for agricultural machinery needs to
detect only a vehicle approaching from the rear, when a general deep learning model is used, a false positive is generated by a vehicle
running on the opposite side (passing vehicle). In this paper, first, we use Faster R-CNN model that shows excellent accuracy rate in deep
learning for vehicle detection. Second, we proposed an algorithm that uses the structural similarity and the root mean square comparison
method for the region of interest(vehicles area) which is detected by Faster R-CNN between the coming vehicle and the passing vehicle.
Experimental results show that the proposed method has a detection rate of 98.2% and reduced the false positive values, which is superior
to general deep learning method.
Keywords: faster r-cnn; vehicle detection; structural similarity index; deep learning; agricultural machine.
1. Introduction
Many agricultural machinery systems are improving in perfor-
mance due to the development of the industry. And the dependence
of agricultural machinery on cultivating crops is increasing. The
characteristics of farm machinery currently in use are that they
move slower than cars. Since the use of agricultural machinery is
used not only in rural areas but also in urban suburbs, many crashes
and serious injuries occur every year. This paper proposes a new
rear approaching vehicle detection algorithm for agricultural ma-
chine systems to prevent accidents in agricultural machinery.
Vehicle detection is used in many places such as public safety, pub-
lic security, surveillance, intelligent traffic volume control, and au-
tonomous driving. There are two kinds of vehicle detection: front
view and rear view. The front view is the detection of a vehicle run-
ning in front of a traveling vehicle. The rear view is the detection
of a vehicle approaching from behind (or approaching with a fixed
camera). The rear approaching vehicle detection method is divided
into vision based and audio (or sound) based. In an audio-based
study, Chen [1] detected a rear-access car using the natural fre-
quency analysis of the car. Ananthanarayanan [2] extracted various
types of features using sound (conversation, music, wind, automo-
bile, etc.) as input data and developed a rear approaching vehicle
detection system by analyzing the characteristics. The advantage of
these systems is that they are low cost to manufacture and have low
computational complexity. However, due to the use of sound (fre-
quency), areas with noises can be highly susceptible to noise, and if
the intensity of the frequency is weak, it may be exposed to danger
because the rear approaching vehicle id detected very closely.
Vision based vehicle detection research has focused on front-run-
ning vehicle detection for use in autonomous vehicles. There is not
much research on vehicle detection (frontal) approaching a camera
installed behind a vehicle. In general, vehicle detection algorithms
to obtain object features has been used a scale invariant feature
transform (SIFT), histogram of oriented gradients (HOG), speed up
robust features (SURF), Haar cascade, and Bigausian edge detec-
tion (BED). Adaboost, Support Vector Machine (SVM), and K-
Nearest Neighbor (KNN) has been used to classify vehicles and
other objects. These features lack the generalization ability to detect
different objects. In addition, feature extraction is influenced by
complicated and various illumination changes, camera view angle,
and image complexity, and so on. These bad feature extractions can
reduce the object detection rate and be difficult to apply in real time
because of system degradation. Vision based rear approaching ve-
hicle detection research has been focused on vehicle detection in
blind spots of running vehicles. Chen [3] has used the one - dimen-
sional distance information and two intersection times to detect the
blind area approaching vehicle. The paper [4] detected vehicles in
the blind spot using cascade classification through Adaboost learn-
ing based on the modified census transformation feature vectors. In
the above studies, vehicles detected as traveling and approaching
vehicles with similar speed and at a short distance, so it is somewhat
distant from this research.
Deep learning technology is becoming more and more popular in
the field of artificial intelligence. Especially, CNN is widely used
in various fields such as image recognition, speech recognition, pe-
destrian detection, face recognition, etc. Unlike general feature ex-
traction systems, CNN uses raw images as inputs and extracts fea-
tures through a large amount of training with high flexibility and
generalization capabilities. It shows considerably higher object
classification accuracy rate than traditional feature extraction
method. CNN can be applied to object detection using region based
R-CNN [5] model, which was greatly improved object detection