Performance
Analysis of Vehicle Detection
Techniques: A Concise Survey
Adnan Hanif
1
, Atif Bin Mansoor
2
, and Ali Shariq
Imran
3
1
Department of Avionics Engineering, Air University, Islamabad,
Pakistan
2
School of CSSE, The University of Western Australia, Perth,
Australia
3
Department of Computer Science, NTNU, Gjøvik,
Norway
ali.imran@ntnu.no
Abstract. Attention towards Intelligent Transportation System (ITS) has
increased manifold especially due to prevailing security situation in the past
decade. An integral part of ITS is video-based surveillance systems extracting
real-time traffic parameters such as vehicle counting, vehicle classification,
vehicle velocity etc. using stationary cameras installed on road sides. In all these
systems, robust and reliable detection of vehicles is significantly a critical
step. Since, several vehicle detection techniques exist, evaluating these tech-
niques with respect to different environment conditions and application sce-
narios will give a better choice for actual deployment. The paper presents a
concise survey of vehicle detection techniques used in diverse applications of
video-based surveillance systems. Moreover, three main detection algorithms;
Gaussian Mixture Model (GMM), Histogram of Gradients (HoG), and Adaptive
motion Histograms based vehicle detection are implemented and evaluated for
performance under varying illumination, traffic density and occlusion condi-
tions. The survey provides a ready-reference
fo
r preferred vehicle detection
technique under different applications.
Keywords: Vehicle detection Gaussian Mixture Model
Histogram of Gradients
Performance analysis
1
Introduction
With an ever-increasing vehicular traffic on urban cities roads, the significance of
Intelligent Transportation System (ITS) is inevitable. This system is to gather inputs in
real time from its traffic sensors which need to be reliable, robust and efficient. An
intelligent transportation system for vehicle detection may comprise of different types
of sensors including loop detectors, ultrasonic and supersonic sensors, or cameras. In
all these sensors vehicle detection using road-side surveillance cameras are most effi-
cient because of their wide area coverage and economical
installation
procedures [1–3].
Much research has been done during past decades by image processing and computer
vision community to assess different traffic parameters from stationary camera video in
a real-time. Today, ITS is benefitting most from video-based surveillance systems by
extracting and analyzing information useful for traffic
planning
and security with
diverse applications including vehicle counting, vehicle tracking, vehicle trajectory,
Post-print version of the paper by Hanif et al. in WorldCIST'18 2018: AISC, vol 746. pp 491-500
https://doi.org/10.1007/978-3-319-77712-2_46
vehicle classification, vehicle velocity, queue length, license plate recognition, traffic
density, traffic lane change etc. [4–7].
However, developing a reliable and efficient video-based vehicle detection system
is quite challenging and is a growing field of research. A promising video-based
detection system must handle environment dynamics efficiently. It must be adaptive to
changes in scene illumination and weather conditions. Jittering camera or noise con-
tamination due to wind are practical issues being faced in vehicle detection. Vehicle
shadows under sunlight are also quite challenging to address as long shadows cause
occlusion problems and thus incorrect classification in many cases. Similarly, at night
time headlights and low illumination poses accurate detection problems. Therefore,
detection of moving vehicles under such scenarios is an important yet demanding task.
During past decades many research projects have been done to detect vehicle and
extract different traffic parameters from stationary traffic surveillance camera video [8–
10]. Early research on
vehicle detection techniques from video-based data started in the
late 1970s. In 1984, University of Minnesota started a system, called Autoscope which
was a wide-area multi-spot video imaging detector [11]. Later on, computer vision,
wide-area detection systems developed for advanced vehicular traffic detection and
extraction of traffic flow parameters with reduced installation and maintenance cost.
A comparative study of vehicle detection between video cameras and loop detectors
were carried in 1990’s, funded by Minnesota Department of
Transportation.
The results
were favourable for vehicular detection through the wide-area stationary camera as the
video-based system was cost effective with several applications in traffic flow analysis
and management [12]. Today, almost all applications concerning traffic parameter
measurement require robust and reliable detection of a vehicle as a crucial step.
In this review paper, recently published moving vehicle detection techniques from
video-based data captured through rectilinear stationary traffic surveillance camera is
presen
ted. Our paper primarily focuses on contemporary moving vehicle detection and
segmentation techniques, leaving out camera calibration approaches and vehicle
tracking methods. While many vehicle detection techniques are available, there has not
been any comprehensive survey focusing on their relative performance and detection
accuracy under practical scenarios. The paper provides a brief overview of vehicle
detection techniques and evaluates the performance of three major vehicle detection
techniques under varying illumination, traffic density, and occlusion conditions. It,
thus, provides a
ready-reference
for a preferred choice of vehicle detection technique in
actual deployment.
The rest of the paper is organised as follows: In Sect. 2, vehicle detection and
segmentation approaches are discussed. The comparative analysis of three major
vehicle detection techniques is presented in Sect. 3, and finally, the paper is concluded
in Sect. 4.
2 Vehicle Detection and Segmentation Techniques
With advancement in image processing and computer vision techniques, much of the
research
has
been
done in the field of
moving object’s regions
of
change detection
among
multiple captured image sequences [13–16]. We categorise the vehicle detection and
ð
-
Þ
ð
t
Þ
¼
X
segmentation techniques based
on
the approach used in each technique into three
methods
which are Background
Subtraction;
Feature
Extraction-based
and
Motion-based.
A sim-
ilar classification was done
by
[17, 18]
but
it lacks comprehensive comparative analysis
of
the methods presented here.
2.1 Background Subtraction Methods
One of the most widely used method for real-time moving vehicle detection and
tracking is the Background Subtraction (BS) method. In BS, moving objects are
extracted as
‘foreground’
from each frame by taking an absolute difference between the
current frame and the reference frame called ‘Background’ frame or model. This dif-
ference is then thresholded to filter out foreground objects. The essence of the method
lies in the accurate estimation of Background model for which both non-adaptive [19]
and adaptive [20–22] modelling techniques are available. Since non-adaptive methods
suffer from a change in illumination and climate conditions, adapti
ve modelling is
preferred [23].
Early adaptive methods developed by Wren et al. in [24] and Lo et al. in [25]
proposed to use moving average and temporal median of the last n frames as the
background model, respectively. For I be the intensity of pixel
ðx; yÞ
at time t and B is
the Background model estimated. Then the proposed foreground FG for each frame is
computed as: in moving average method as:
FG ¼
jI
ðx; y; tÞ
-
Bðx; y;
t
ÞÞ
j
[
Th
where, in the case of moving average method,
1
X
n
-
1
Bðx; y; tÞ
¼
n
I
x; y;
t i
i
¼
0
and, in the case of temporal median method,
Bðx; y; tÞ
¼
median
f
I
ðx; y;
t
-
i
Þ
g;
i
2
f
0;
. . .; n
-
1g
These methods however, require large memory buffers for its computation and the
threshold value Th is non-adaptive and is same for all pixels in frame.
To addre
ss this, Ridder et al. [26] used Kalman filter to model each pixel which
made their system less susceptible to lighting changes in the scene but poor to handle
bimodal backgrounds. A significant work in the field of adaptive background mod-
elling was done by Stauffer and Grimson in [27] by modelling each pixel value x at any
time t as a mixture of K Gaussian probability distributions,
where,
K
P
X
i
¼
1
x
i;t
:N
ð
X
t
j
l
i;t
;
r
i;t
Þ
i;t
1 1 1
T
1
N
X
t
j
l
i;t
;
r
i;t
¼
ð
2
p
Þ
D
=
2
r
1=2
exp
-
2
ðX
t
-
l
i;t
Þ
r
-
X
t
- l
i;t
i;t
Each pixel in the image scene is classified either as part of the foreground (moving
vehicle) or background based on the knowledge of the Gaussian distributions of its
pixel model. For
l
i
and
r
i
be the mean and standard deviation of the K
th
Gaussian pixel
model, then pixel x
t
can be classified so as whether,
x
t
-
l
i;t
=
r
i;t
[
2:5
Later theoretical framework of this approach along with useful corrections is pre-
sented in [28].
As an extension to mixture of Gaussian models, authors in [29] presented a
combination of background as well as foreground model of each pixel, with back-
ground based on Gaussian Mixture Model (GMM) and foreground based on object
size, position, velocity, and
colour
distribution
models. In this method, each pixel of the
scene can be treated as part of the background, foreground or noise. Yet, velocity
model for each foreground objects are to initialise by providing an a priori estimate of
object velocity through a learned model of typical traffic direction and speed.
Another technique uses shadows underneath vehicles as the information to detect
vehicles [30, 31]. Traffic video normally captured through a camera set up on a low
place such as the roadside, sidewalk, etc. is used to determine the size of each vehicle
based on the distance between both ends of the front and rear tires. The shadows are
segmented for vehicle detection using
statistical parameters
which
automatically
update
both background subtraction image and binarization threshold.
In [32], frames are subtracted from an adaptive background model which is based
on Kalman filtering after dividing the frame into small non-overlapped blocks.
A change in gray levels in each block is used for detection of any candid
ate vehicle
part. Then, Principal Component Analysis (PCA) is applied to two histograms of each
candidate to produce the low-dimensional feature. Later, a classifier based on support
vector machine classifies each block either as part of vehicle or not. Finally, a paral-
lelogram shape represents the vehicle by combining all classifier results.
2.2 Vehicular Feature Based Method
s
This method segment moving objects from background image by detecting vehicle’s
inherent visual features like its colour, edges, contour, texture or body part such as head
lights [33–35]. Since the method does not require a vehicle in motion, it can detect
stationary vehicle as well. These feature-based methods are less prone to occlusion and
perform better even for overlapping vehicles, however, for detection, prior information
is required for modelling and therefore, differing feature-based methods result in dif-
ferent computational complexity.
A trainable system for certain class of vehicle detection w
ithout using motion,
tracking to handcrafted models in unconstrained, cluttered scenes [36] was a break-
through. The system using a training data of positive and negative example images as
input, first transform the images to Haar wavelet
representation
and then uses a support
vector machine classifier to detect in-class and out-of-class static patterns. For vehicle
detection in video sequences, the system is augmented with Kalman filtering and
detected feature density is modelled and then propagated through time for accurate
detection. The system produces appreciable results when applied even to face and
people detection.
However, producing a variety of trainable images or models is a mammoth task. An
approach in [37] uses computer graphics (CG) model to generate different target
vehicles instead of real images for vehicle detection and its classification. The method
uses eigenspace technique to obtain local-feature used for subsequent detection and
classification. The technique performs well even if parts of the vehicle are occluded, or
vehicle translates due to veering out of the lanes. Moreover, it does not require seg-
mentation of vehicle areas from input images.
Another vehicle recognition sys
tem proposed by [38] uses image’s curvelet
transform and standard deviation of curvelet coefficient matrix in different scales for
feature extraction. Curvelets having
time-frequency
localization properties show a high
degree of directionality and anisotropy. The approach uses k-nearest neighbour clas-
sifier along with different scale information as a feature vector. Recently, [39] used an
image descriptor generated from the statistical parameter of the curvelet-transformed
sub-bands, for vehicle verification with the hypothesis (candidate) during its detection.
A
statistical
approach to detection problem proposed by [40] proves robust not only
towards partial occlusions but also reduces the computational overhead. For automatic
detection, local-features within three significant subregions of image individually
generate PCA weight vector and an
Independent
Component Analysis (ICA) coefficient
vector which are used to model the low-frequency components of eigenspace and
high-frequency components of the residual space. This improves detection
tolerance
towards variations in the illumination and vehicle pose.
Another approach [41] for vehicle detection in wide area motion imagery (WAMI)
uses Histogram of Gradient (HoG) and Haar descriptors to construct an optimal kernel
for the purpose of classification. Here, a cascade of boosting classifier is used to select
Haar features from a huge feature set which combined with HoG descriptors, train the
final classifier. Results show better classification with fusion of HoG+Haar with
Generalized Multiple Kernel Learning (GMKL).
2.3 Motion Based Methods
Optical flow, a computer vision tool, can also be used for detection of objects in motion
[42]. The vehicular motion observed from a static camera seems as pixels in the image
to be moving. In optical flow, movement of each pixel is calculated by measuring
temporal changes of the pixel, and their correlation in an image sequence [43] and the
vector field of this motion is referred as Optical flow. Motion based vehicle detection
methods trace these flow vectors in 2-
D which are produced due to vehicle motion
velocity vectors in an image sequence. This approach can even detect independently
moving vehicles from the camera. However, optical flow is
computationally
expensive
due to its
iterative
algorithm and is very susceptible to noise which makes this approach
less suitable for real-time video processing without specialised hardware.
