Front. Comput. Sci.

, 2014, 8(4): 619628

DOI 10.1007/s11704-014-3156-0

Road scene analysis for determination of road trac density

Omar AL-KADI
2

, Osama AL-KADI2 , Rizik AL-SAYYED1 , Jafar ALQATAWNA1

1 King Abdullah II School for Information Technology, University of Jordan, Amman 11942, Jordan
College of Engineering and Computer Science, Australian National University, Canberra, ACT 0200, Australia

c Higher Education Press and Springer-Verlag Berlin Heidelberg 2014



Abstract Road trac density has always been a concern

in large cities around the world, and many approaches were
developed to assist in solving congestions related to slow
trac flow. This work proposes a congestion rate estimation approach that relies on real-time video scenes of road
trac, and was implemented and evaluated on eight dierent hotspots covering 33 dierent urban roads. The approach
relies on road scene morphology for estimation of vehicles
average speed along with measuring the overall video scenes
randomness acting as a frame texture analysis indicator. Experimental results shows the feasibility of the proposed approach in reliably estimating trac density and in providing
an early warning to drivers on road conditions, thereby mitigating the negative eect of slow trac flow on their daily
lives.
Keywords road congestion, image texture, local binary
pattern, scene morphology

Introduction

Regular and lengthy delays in trac jams is not only a source

of nuisance and frustration by wasting motorists and passengers time, but can also be a major cause for economic losses
[13]. On the global level, the increase of air pollution and
carbon dioxide by wasted fuel contributes to weather global
warming. Down to the personal level, congestions may have
a negative impact on motorists health due to increased stress
and road rage, late arrival which may result in disciplinary
actions, passage obstruction of emergency vehicles, wear and
Received May 3, 2013; accepted February 20, 2014
E-mail: [email protected]

tear on vehicle due to accelerating and breaking requiring

more repairs, allocating more time for travelling which could
be otherwise saved for other productive activities. All of these
congestion collateral side-eects can be attributed to road capacity being below normal, free flow levels.
Trac congestions could be attributed to many possible
reasons. One of the most recurring circumstances is limited
road capacity, especially at peak hours. Other reasons could
be sudden increase in number of vehicles over a specific
length in a road, accidents causing of blockage to lanes, roadwork that narrow the flow which result in bottlenecks in parts
of the road, and bad weather conditions resulting in partial
or complete roadway obstruction. This gives strong motivation for developing techniques that could reliably measure the
healthiness state of real-time trac flow, in analogy to blood
pressure which gives indication of heart and circulatory system fitness.
Types of measurement technologies used for detecting
road trac density are twofold: intrusive and non-intrusive
detectors. The former depends on signals emitted or reflected
from passing vehicles, such as inductive loops, acoustic,
magnetic, radar detection, while an example of the latter is
video imaging detection which is categorised into tripwire
systems that provide count and speed information at a single
spot, and tracking systems relying on spatial information that
can measure true density instead of detectors occupancy [4
6]. The main focus of this work is to improve video imaging
tracking eciency, as it has the advantage of providing spatial trac information that are not related to a single point as
in intrusive detectors.
Examples of recent major prior research related to realtime trac analysis from video scenes for the purpose of congestion estimation include trac conflict evaluation system

620

Front. Comput. Sci., 2014, 8(4): 619628

that was developed by Oh et al. [7] which looks for three categories in video images acquired from a single camera, which
are trac speed, trajectory, and conflict. In another work [8],
Cheng and Hsu applied a time-varying adaptive system state
transition matrix in Kalman filter for vehicle tracking and
also utilised a regression analysis model for estimating traffic flow. Houben et al. extracted maximum phase congruency
and edges from stereo images and matched together with local matching algorithm, then processed by maximum spanning tree clustering algorithm to group the points into vehicle objects [9]. Lane trac characteristics, number of vehicles in each lane, and mandatory lane-changing fractions
in lanes with trac congestion were employed in this work.
Bishop and Casabona used GPS-enabled networked mobile
phones for collecting and analysing location based data set,
and followed by sample filtering and route machining for
trac congestion determination [10]. Others like Ozkurt and
Camci used a machine learning approach based on neural
networks for trac density estimation [11]. Based on maximum phase congruency and edges features matching and a
3D tracking, the system detects vehicles and determines their
dimensions. Jain et al. applied a simulation-based analysis on
a simple network topology where a local congestion protocol
that controls the flow of trac into near-congested regions for
preventing congestion collapse sustaining time variant trac
bursts [12]. Chen et al. developed a night-time vehicle detection and trac surveillance system, where vehicles head
and tail-lights are located and processed by a spatial clustering for analysing the spatial and temporal features of vehicle
light patterns [13]. While Yu and Chen used a simple consecutive temporal dierence approach for trac monitoring after pre-processing with a square neighbourhood algorithm for
compensation of camera disturbances [14]. Also, Marfia and
Roccetti proposed a short-term trac congestion forecasting
approach that requires no prior. knowledge on road condition,
by estimating the time interval of how long would a congestion last [15]. Another work implements a vehicle counting
method through blob analysis and tracking, and eventually
determining the speed of vehicles [16]. A review on computer vision techniques utilised in trac scene analysis can
be found in [17].
Since image texture has shown in previous work its usefulness in dierentiating between dierent image patterns
[18,19], it could as well assist in representing a better understanding of road trac scenes, if we take into account that the
vehicles distribution are the patterns of interest. Therefore,
and to the best of our knowledge, road trac density from
a texture perspective has not been classified before into ad-

vanced five stage conditions. It would be very advantageous

from a computer vision system perspective to have a manifold
view of road trac density that provides road trac control
authorities with a clearer view on the trac status quo, and
hence giving time to take appropriate measures to deal with
restricted flow conditions before it reaches to a complete stop.
Also the system needs to be low-cost due to the large number
of hotspots that would be monitored, and with minimal computational complexity to be able to work reliably in real-time.
In this work, a robust model that relies on both morphological analyses for estimation of average speed on roads
along with video scenes randomness acting as an image texture measure was developed for measuring actual trac flow.
First step in the developed model is to pre-process each acquired video frame for non-uniform image intensity adjustment. Secondly, the vehicles are delineated and tracked by
morphologically operating via motion segmentation, then estimating regional maxima, followed by connected component
extraction, and this phase terminates by image opening and
vehicle centroid determination. Thirdly, the average entropy
of the local binary pattern (LBP) is used as a supporting measure for global measurement of the degree of randomness
in the image texture; besides it is known for computational
simplicity and robustness to monotonic grey-scale changes
caused by illumination variation. Finally, the tracked vehicle speed is estimated and a simple multilevel thresholding
classification approach is used for determination of road flow
condition. The steps followed for the proposed model are illustrated in Fig. 1.

Fig. 1

Road trac density measurement model applied in this work

The paper is organised as follows. The cameras installation

and video acquisitions are explained in Section 2, followed
by the approach proposed for road congestion estimation in
Section 3. The experimental results and associated discussion
are presented in Sections 4 and 5, respectively. Conclusions
are given in Section 6.

Materials

In order to ensure consistency and replication of results, the

Omar AL-KADI et al.

Road scene analysis for determination of road trac density

camera type, specifications and installation procedure, and

how the trac videos were captured and made ready for subsequent processing are discussed next.
2.1 Camera specification and calibration parameters
The trac video scenes were captured using a JVC Everio
GZ-MG360 60 GB Hard Drive camcorder with 35x optical
zoom (800x digital zoom) having a high-performance Konica Minolta lens. The camera was mounted on a tripod for
stabilisation (i.e., reduction of camera shake) and appropriate elevation (i.e., better field of view) to achieve maximum
sharpness.
At a height of 1.5 meters from the surface of a 5 meters
high bridge, the camera tripod was fixed in the middle on one
of the bridge sides overlooking the monitored road beneath.
The camera was tilted by a 45 and focused down on the road
to monitor a distance of 25 meters which is 15 meters away
from the side of the bridge. Figure 2 illustrates the installation
protocol followed in all trac areas of study.

621

in the early diagnosis of trac flow conditions, and thus can

assist in studying, analysing and taking appropriate measures
to alleviate congestions. The time of video capture corresponding to each hotspot area (A) covering the roads of interests (R) locations which could have multiple branches (a,
b) are given in Table 1. This will facilitate results comparison
in terms of road condition, given that rush hours in Jordanian
cities range from 7:30 to 8:30 in the mornings and from 3:00
to 4:00 in the afternoons; o-peak otherwise.
Table 1
Road

Monitored roadway
Capture

Road

Capture

Road

Capture

location

time

location

time

location

time

A1-R1a

16:00

A2-R2a

13:00

A4-R3a

14:00

A1-R2a

16:00

A2-R2b

13:00

A4-R3b

14:00

A1-R2b

16:30

A3-R1a

15:15

A5-R1a

08:00

A1-R1b

16:30

A3-R1b

15:15

A5-R1b

08:00

A1-R3a

15:45

A3-R2a

15:00

A6-R1

12:30

A1-R3b

15:00

A3-R2b

15:00

A7-R1a

08:30

A1-R4a

15:00

A3-R3

14:00

A7-R1b

08:30

A1-R4b

15:45

A4-R1a

13:30

A7-R2a

08:30

A1-R5

15:45

A4-R1b

13:30

A7-R2b

08:30

A2-R1a

11:45

A4-R2a

14:00

A8-R1a

15:30

A2-R1b

11:45

A4-R2b

14:00

A8-R1b

15:30

2.2 Video acquisition and preparation

Fig. 2 Camera installation for trac video acquisition

Roads at designated study areas were composed of three

lanes in each direction. Each video was further trimmed
equally into two parts, one for incoming and another for the
outgoing vehicles, giving us the opportunity to investigate the
road condition for both directions from the same camera location without the need for a second camera.
The cameras were installed at carefully selected locations
deemed as busy intersections in the city of Amman, Jordan.
Monitoring the selected hotspots which are illustrated at the
marked locations on the area study map in Fig. 3, can assist

Fig. 3 Camera locations covering hotspots at major busy intersection in

Amman city, Jordan. Image courtesy of Google maps

In order to preserve quality, the acquired video files were initially processed at full size with a resolution of RGB 720480
pixels and 29 frames per second (fps), and the recording process was for duration of 3 minutes. This will result in a very
large file size that would be deemed computationally intensive. Thus a preprocessing stage would be required to reduce
the video size and frame rate while maintaining acceptable
quality, a process that would have minimal side eect on the
vehicle segmentation.
The captured videos from the fixed camcorder were in the
MOD file extension, a JVCs implementation of MPEG-2
transport stream, and for the successive video frames to be
appropriately read and processed, they were converted to AVI
format, a common multimedia container used by dierent
codecs and known for its good video quality and wide applicability. Xilisoft video converter http://www.xilisoft.com
was used to convert the recorded MOD video files to its corresponding AVI format after setting appropriate resolution,
frame rate and aspect ratio.
All acquired videos were converted to grey scale in order
to reduce processing time since we are interested in detecting
any car irrespective of its colour. Then each video frame rate
and size was adjusted from 29 fps, 720480 pixels having

622

Front. Comput. Sci., 2014, 8(4): 619628

a duration of 3 minutes to 15 fps, and 320240 pixels and

a 1 minute duration, respectively. The converted files had a
size reduction of 80% with unnoticed loss in video quality,
which resulted in a remarkable reduction in the video processing operation. Thereby, the videos are ready for the next
step for vehicle detection and road condition analysis. The
preprocessing and analysis program representing this work
was written in MATLAB Version 7.8 (R2009a) and tested
on a 3 GHz Intel CoreTM 2 Duo processor notebook with 2
GB RAM memory running Windows XP 32-bit operating
system.

Finally the state of the road denoted as congestion rate CR

can be estimated as:
CR = 1

VA
,
Vmax

(3)

where Vmax is the maximum road speed limit of 80 km/hr

when substituting the value of 800 in Eq. (2), which denotes
the number of detected vehicles, for a monitored distance of
25 meters for a duration of 1 minute.
3.2 Object recognition and scene analysis
3.2.1 Image intensity adjustment

Methodology

This section starts by explaining the applied mathematical

model for computing the average road velocity, then the video
image vehicle detection approach along with the global LBP
image average entropy is described, and eventually the road
condition is classified on a scale of 1 (stopped flow) to 5 (free
flow).
3.1 Representational model
Let  be an order set of acquired video frames vi , where i is a
certain point in the recorded time interval. We can define C to
be the set of univariate vectors c that represents the number
of detected vehicles in , where C = {c1 , c2 , c3 , . . . }, cm .
So the average road velocity VA could be estimated as:
VA =

D
,
FA

(1)

where D is a priori known representing a fixed monitored

distance that vehicles cross in meters, and F A as calculated in
Eq. (2), is the average object flow in seconds required for the
vehicles to cross D.
1  fr
,
n m=1 C(cm )
n

FA =

(2)

where fr is the frame rate and C corresponds to the total number n of m detected vehicles c per video frame.
For example in case A1-R1b in Table 1, suppose the monitored part of the lane D was 25 meters long and the video
displayed the frames at a rate of 15 per second for a duration
of 1 minute and the number of detected objects (i.e., vehicles)
was 740. The number of detected vehicles (F A ) in the video
can be calculated by substituting in Eq. (2), so the average
velocity (VA ) can be easily estimated by substituting in Eq.
(1) which yields 74 km/hr.

The video frames intensity values are adjusted by equalising

the contrast of each image frame prior to the vehicle detection process. This pre-processing process is deemed necessary when capturing video in outdoors environments where
varying illumination conditions are very common. The trac
videos should adapt to change in illumination during daylight
and low-light, e.g., early mornings and late afternoons, therefore this process could assist in reduction of possible glare
and shadows due to sunlight reflection and other changing
weather conditions and in dierent seasons, which could affect image con trast, and hence the accuracy of vehicle segmentation.
A homomorphic filtering technique was employed for correction of the non-uniform illumination in the acquired video
frames as defined in Eq. (4). It operates by assuming the scene
I(x, y) consists from an undesirable change in illumination L
due to varying lighting conditions at the time of video capture, and the scene reflectance R related to physical properties
of the objects, vehicles in our case. The illumination component L is considering as additive noise, and hence can be
filtered-out via a high-pass filter since it tends to have a gradual low-frequency change as compared to the abrupt highfrequency R component.
The filter can be applied by performing an inverse
Fourier transform to a high-pass filter H(u, v) as h(x, y) =
1 [H(u, v)] and then convolving the image scene I(x, y) in
the log domain we get the intensity corrected image:
f (x, y) = exp[h(x, y) ln I(x, y)].

(4)

3.2.2 Morphological processing

Spatial object motion segmentation
A simple, yet eective approach for detecting changes between successive video frames f (x, y, t1 ), f (x, y, t2 ), . . . ,
f (x, y, tn ) is applying the absolute dierence between each

Omar AL-KADI et al.

Road scene analysis for determination of road trac density

frame and a reference frame R(x, y), where R(x, y) =

f (x, y, tr ), and r = 1 denoting a zero state empty road trac
scene. Assuming that all video frames in the scene are appropriately registered and have the same size, spatial resolution
and illumination conditions, and stationary objects would be
eliminated and the moving image components would be preserved, described as:

1, if |R(x, y) f (x, y, tk )| > T

Di j (x, y) =
(5)

0, otherwise,
where k (k > 1) is the time of video frame capture, and T is
an appreciably dierent threshold set to the mid grey-level of
the each video frame according to the following function:
1
(max( f (x, y, tk )) + min( f (x, y, tk ))) .
(6)
2
Regional maxima
Regional maxima can be defined as a certain connected components of pixels with a certain height attribute, such that all
surrounding grey level pixels in the external boundary have
strictly lower values. This could be utilised in removal of
isolated low-valued points in the absolute dierence image
Di j (x, y). Isolated small structures are likely to be considered
as noise, thus regarded as an essential procedure for reduction of false positives in the segmentation process. The elimination of such case can be done by arranging a certain set
of n-connected regions (8 pixels was applied in this work)
in Di j (x, y) and ignoring detected structures that are smaller
in size. Nevertheless, small size and/or slow moving vehicles
might be erroneously eliminated, yet the appropriate selection of the threshold value T could contribute in mitigating
this drawback, taking into consideration the camera field of
view, depth of field, spatial resolution, focused distance, position and orientation.
Connected component extraction
Connectivity is an important concept that defines regions and
boundaries in images. To establish connectivity between two
pixels in an image, they have to be neighbours and their corresponding grey levels satisfy a criterion of similarity.
Let Y represent a connect component contained in a set A
and assume that a point of Y is known. Then,
T=

Xk = (Xk1 B) A,

k = 1, 2, . . . , n

(7)

yields all the elements of Y, where X0 = p (p is a specific

pixel in the image), B is a suitable structuring element, and
is the dilation (XOR) operation.
Image opening
Using a disk-shaped structure element c, which best resembles the shape of a car, with a radius R set empirically to

623

4 pixels, a morphological opening on the binary image was

performed for noise removal and other non-disk shape structures in the scene which are considered not related to vehicles
of interest. The opening of each frame fi , i = 1, 2, . . . , by a
structuring element s, denoted fos can be represented as [20]:
f os = ( f  s) s,

(8)

where  and denote erosion and dilation, respectively.

Image regions properties measurement
The area and centroid properties of each of the remaining objects in each frame are determined, and the object with the
largest area is the detected vehicle.
Assuming a square S QR , where Q = (q1 , q2 ) and R =
(r1 , r2 ) would enclose each of the detected vehicles or objects
W s , where s = 1, 2, . . . , k, such that qi  ri , where i = 1, 2,
and the area of the object A s could be simply computed as
(r1 q1 ) (r2 q2 ). The centroid CRW (cr1 , cr2 ) of A s can be
defined as:
n
1
cr1 =
W xi ,
n i=1
(9)
n
1
cr2 =
Wyi ,
n i=1
where n is the total number of pixels belonging to object W s
within area A s , W xi represents the x coordinate of the ith pixel
in A s , Wyi represents the y coordinate of the ith pixel in A s .
Finally the objects with the largest area would be considered the detected vehicles
Wdet = argmax(A seg ).

(10)

This process when applied to a single video frame is illustrated in Fig. 4.

3.3 LBP
LBP is a local image texture descriptor that is robust to illumination variations, where it forms labels to image pixels
by thresholding the neighbouring pixels with the centre value
and converting the result to a binary number. This computational simplicity gives another advantage in making it suitable for real-time analysis.
For a local neighbourhood N xy of P (where P > 1) number
of pixels g xy in a monochrome image I xy , a texture T (where
T N xy ) can be defined as a joint distribution of grey-level
values of g xy [21,22]:
T t(s(I x0 y0 I xc yc ), s(I x1y1 I xc yc ), . . . , s(I xP1 yP1 I xc yc )),
(11)

624

Front. Comput. Sci., 2014, 8(4): 619628

Fig. 4 Example of vehicle detection and tracking in a video frame of road A1R5. (a) Acquired frame, (b) intensity adjustment, (c) background
removal, (d) morphological processing and object detection, and (e) labeling detected vehicles on original frame for tracking subsequently speed
measurement

where I xc yc corresponds to the grey value of the centre pixel

of a local neighbourhood, and I x p y p (p = 0, 1, . . . , P 1) correspond to the grey values of P equally spaced pixels on a
circle of radius R (R > 0) that form a circularly symmetric
set of neighbours. s(x) is a threshold function that checks for
the sign of the dierence according as:

1, x  0,
s(x) =
(12)

0, otherwise.
If the coordinates of gc where (xc , yc ) the coordinates
of



2p
, yc
the neighbours g p are given by xc + R cos
P


2p
.
R sin
P
The spatial structure of the local image texture can be characterised by
LBPP,R =

P1


s(I x p y p I xc yc ) 2P .

In this work a P and R value of 8 and 1 was applied, respectively.

The average entropy measure eavg can be defined as:
L1


p(zi ) log2 p(zi ),

A scene in road A1R5 and its corresponding average entropy LBP

(13)

p=0

eavg =

Fig. 5
image

(14)

i=0

where z is a variable denoting LBP image grey levels and

p(zi ), i = 0, 1, . . . , L1 is the corresponding histogram, where
L is the number of distinct grey levels. An example of a generated LBP texture image for a road scene is given in Fig.
5. Then the average LBP entropy (eavg ) is computed for each
of the texture images and used for supporting the road flow
condition estimation.
3.4 Classification by multilevel thresholding
A simple approach was employed to determine the road flow
RF based upon dividing the congestion rate on a scale 5 level

scale, normalised from 0 to 1, by setting four dierent thresholds T s , (s = 0, 1, 2, 3) relevant to the road average velocity. The threshold describes how much the road capacity is
utilised, where the congestion rate is represented as 1 minus
the detected vehicles divided by the value of 800, which is
the number of detected vehicles equivalent to the maximum
road speed limit of 80 km/hr for a monitored distance of 25
meters for duration of 1 minute, which is indicated as:
Let T s = [0.1, 0.3, 0.6, 0.8];
if CR < T 3 then RF = 5;
else if CR  T 3 AND CR < T 2 then RF = 4;
else if CR  T 2 AND CR < T 1 then RF = 3;
else if CR  T 1 AND CR < T 0 then RF = 2;
else if CR  T 0 then RF = 1;
end if.
(15)
The feature vectors with the morphological operations are
used to determine the road congestion rate can be summarised
in Algorithm 1.

Omar AL-KADI et al.

Algorithm 1

Road scene analysis for determination of road trac density

Congestion rate estimation

for i = 1, 2, . . ., m do

% Number of video scenes

C = 0;

% Vehicle count initialisation

for j = 1, 2, . . ., n do

% Number of video frames

Hq = exp(hxy ln(Ixy ));

% Perform homomorphic filtering

D xy = |Hq Ir |;

% Absolute dierence image

if D xy > S then D xy = 1;
else D xy = 0;
end if
M = Dos

% Perform morphological opening

for k = 1, 2, . . ., l do

% Number of detected objects

arg max(M(k))

% Find region maxima

end
C = C + M;

% Compute number of detected vehicles

FA = f /C;

% Estimate average flow, where f is frame

rate

Table 2

Road condition estimation for 8 areas covering 33 dierent roads

Monitored Detected Detection

road

625

LBP Estimated Congestion

vehicles error /% entropy

speed

Road

rate /%

condition
Yellow

/kmhr1
A1-R1a

423

0.50

3.70

42.3

47

A1-R2a

626

0.32

3.98

62.6

22

Green

A1-R2b

710

0.42

3.49

71.0

11

Green

A1-R1b

740

1.10

3.93

74.0

07

Blue

A1-R3a

81

5.19

4.79

08.1

90

Red

A1-R3b

449

0.22

4.12

44.9

44

Yellow

A1-R4a

738

0.00

4.21

73.8

08

Blue

A1-R4b

687

2.23

4.18

68.7

14

Green

A1-R5

774

0.77

3.70

77.4

03

Blue

A2-R1a

214

3.38

4.48

21.4

73

Orange

A2-R1b

392

1.26

3.22

39.2

51

Yellow

A2-R2a

357

0.83

2.88

35.7

55

Yellow

A2-R2b

702

1.81

4.10

70.2

12

Green

A3-R1a

553

2.47

3.68

55.3

31

Yellow
Yellow

Va = D/FA ;

% Average speed, where D is monitored

distance

eavg = HLBP log(HLBP );

% Average LBP entropy, HLBP is histogram of LBP

A3-R1b

540

1.10

3.73

54.0

32

A3-R2a

53

1.42

4.30

05.3

93

Red

CR = 1 (VA /Vmax )eavg ;

% Calculate congestion rate using video

frame eavg

A3-R2b

118

1.72

4.07

11.8

85

Red

A3-R3

227

0.88

3.76

22.7

72

Orange

T s = [0.1, 0.3, 0.6, 0.8];

% Set threshold

A4-R1a

433

0.92

3.00

43.3

46

Yellow

if CR < T 3 then RF = 5;

% Free flow

A4-R1b

541

1.99

3.17

54.1

32

Yellow

else if CR  T 3 AND CR
< T 2 then RF = 4;

% Moderate flow

A4-R2a

279

0.36

3.74

27.9

65

Orange

else if CR  T 2 AND CR
< T 1 then RF = 3;

% Restricted flow

A4-R2b

306

2.00

3.82

30.6

62

Orange

else if CR  T 1 AND CR
< T 0 then RF = 2;

% Slow flow

else if CR  T 0 then RF = 1;

% Stopped flow

end if
end for
end for

Results

A4-R3a

192

2.67

3.80

19.2

76

Orange

A4-R3b

236

2.48

3.62

23.6

71

Orange

A5-R1a

690

0.00

3.40

69.0

14

Green

A5-R1b

357

3.48

4.03

35.7

55

Yellow

A6-R1

210

0.00

4.31

21.0

74

Orange

A7-R1a

186

1.59

3.29

18.6

77

Orange

A7-R1b

663

2.21

3.76

66.3

17

Green

A7-R2a

597

1.97

3.27

59.7

25

Green

A7-R2b

538

0.00

3.27

53.8

33

Yellow

A8-R1a

735

1.94

4.14

73.5

08

Blue

A8-R1b

360

4.00

3.33

36.0

55

Yellow

Running the system in eight dierent areas and for a total of

33 roads, Table 2 shows the classified road conditions after
estimating congestion rates as discussed in the methodology
section. Also the same table indicates the number of detect
vehicles per each road, the detection error which is the percentage of error as compared to ground-truth, average road
speed, and the LBP entropy which estimates the average randomness in the captured videos.

Table 3

In order to make the interpretation of Table 2 easier and

more visually appealing, the city roads trac conditions were
classified into five colours, as shown in Table 3, which determines the status of the roadway according to the proposed
congestion rate estimation algorithm. The colour code would
give the user a visual indication of the road status depending on the description of the road condition. The blue, green,

yellow, orange, and red colours refer to the free, moderate,

restricted, slow, and stopped flow conditions, while black
colour code refers to the no-data condition, namely, either
the road is clear where no vehicles are passing for a monitoring period of two minutes or a possible malfunction in the
camera, the lens is covered or camera not operating.
The general daytime road conditions in the monitored ar-

Road trac colour code interpretation

Road congestion rate /%

Colour code

Description

Blue

Free flow

Less than 10

Green

Moderate flow

Between 10 and 29

Yellow

Restricted flow

Between 30 and 59

Orange

Slow flow

Between 60 and 79

Red

Stopped flow

Greater than or equal to 80

Black

No data

Not available (0)

626

Front. Comput. Sci., 2014, 8(4): 619628

eas are illustrated in Fig. 6, and Table 4 indicates the road

conditions and corresponding peak status. Calculations related with daytime and rush hours, namely low number of
detected objects in rush hours could mean congestion while
in o-peak hours could mean empty or running road.

Fig. 8 Time series LBP entropy for (from top to bottom) A1-R3a (stopped),
A1-R1a (restricted), A1-R2a (moderate), A1-R2b (moderate), A1-R3b (restricted), and A1-R5 (free), respectively

Fig. 6 Road flow conditions for 33 roads in eight dierent busy areas in
Amman city at dierent times recorded at daytime
Table 4

Peak conditions for monitored roads in Table 2

Colour code

Peak cases /%

Free

50.00

Moderate

42.86

Restricted

54.54

Slow

85.71

Stopped

66.67

As a second measure that can assist in improving road condition estimation, the LBP entropy can measure the degree
of randomness in the monitored roads without counting the
number of vehicles in the assigned two minutes period. For
instance, Figure 7 shows how the LBP average entropy can
be used to dierentiate between restricted and moderate flow
road conditions for road A1-R1. Also when applied to six different roads in area A1 as illustrated in Fig. 8, it could assist
in distinguishing between stopped, restricted, moderate and
free flow road conditions.

two other statistical-based texture methods, namely greylevel co-occurrence matrix (GLCM) and run-length matrix
(RLM), and with non-textural processed video frames for investigating robustness. The errors for all cases are shown in
the boxplot of Fig. 9.

Fig. 9 Boxplot of vehicle error detection for the proposed LBP method
as compared to two other statistical-based texture method and non-textural
processed video frames

The proposed method was also benched marked against

Fig. 7 Congested LBP entropy A1-R1a (restricted flow ) vs normal LBP

entropy A1-R2a (moderate flow)

Discussion

Road trac usually exhibit stochastic behaviour due to conditions beyond our control, such as accidents, bad weather,
and urgent road works that could aect the general trac
flow. Thus this work tends to provide a system that can automatically indicate the severity of road congestion, and hence
improving the road flow by giving drivers an option to use
an alternative way in case if the trac situation is congested.
As humans tend to be more attentive to colours, and are also
mainly used in trac lights to indicate road status and control
trac flow, this work utilises colour codes as well to further
discriminate between dierent levels of congestion and to fa-

Omar AL-KADI et al.

Road scene analysis for determination of road trac density

cilitate road conditions interpretation compared to numerical

percentages.
It is expected that road conditions towards city centre in
rush hours in early mornings to be moderate to slow flow,
while the roads head out of the city centre to be free flow, and
vice versa for the afternoon rush hours. Examining the bellshaped diagram in Fig. 6, most of the road conditions were
restricted having a percentage of 33.33% from the 33 monitored roads. While the slow and average flow conditions had
a nearly similar percentage in the twenties, and 9.00% and
12.00% for the stopped and free flow, respectively. This indicates that most of the roads at the rush hours have a Gaussian
distribution with restricted flow being the average condition,
which is not unusual for that time of the day. Thus this system can assist drivers in avoiding stopped flow roads to an
alternative road with a better condition, and hence mitigating
the eect of road trac congestion in Amman city.
There were a number of odd cases, such as A7-R2a and
A7-R2b were classified as free flow although in peak time,
this is because these roads are heading outside of the city
centre. While road A2-R1a as slow flow although o-peak,
which could be possibly due to road works. Table 4 shows
also that the stopped, slow, and restricted flow conditions occurred mainly in the peak (rush hour) times that match when
visually checking the captured road videos.
According to the proposed algorithm described in the
methodology section, the classified road condition as stopped
had always a high LBP value above 4, indicating a high randomness in the captured video, i.e., the presence of many vehicles in the road scene. However there were other roads that
exhibited entropy values above 4, such as the cases A1-R4a,
A1-R4b, A2-R1a, A2-R2b, A5-R1b, A6-R1, and A8-R1a
when the road congestion condition was not stopped. This indicates that video scene randomness has to be integrated with
the morphological road scene analysis for ecient road flow
estimation. LBP average entropy is considered as a second
measure for checking road conditions that relies on general
image structure, and the more vehicles in the scene the more
chaotic the image becomes and hence would give higher entropy, and vice versa. This can be applied also to distinguish
stopped flow from empty roads, situations as in ocial holidays, late nights or very early mornings when usually roads
are empty for long periods without cars passing. In both conditions the algorithm running on videos fed by monitoring
camera assess no movements giving a stopped road condition,
whereas the entropy can measure the degree of randomness in
the video scene represented by the number of stationed vehicles superimposed on the empty road scene background. The

627

edges of the vehicle structures would increase the randomness of the road background and hence give an indication of
road occupancy. In other words, a higher entropy value would
be noticed as compared when the road is empty. This is apparent from the comparison results in Fig. 9 where the LBP average entropy had an improved performance as compared to
other statistical textural methods and to the non-textural processed video frames. This implies that the decrease in the vehicles detection error in respect to the other compared methods had a positive eect on the estimated average road speed,
and hence the accuracy of the estimated congestion rate.
As there is no system free from disadvantages, the system currently does not indicate explicitly the reason for the
increase in congestion (i.e., is it due to an accident, road
work, double parking, etc.), and other trac conditions and
their possible eect on the approach proposed, such as bad
weather, would be investigated in future work. Knowing the
reasons behind congestion would give road trac control authorities more information on the problem in order to take
fast and suitable measures to deal with this situation. Also the
level of road noise detected by the surveillance cameras microphones could be employed for improving congestion detection.

Conclusions

A new video scene analysis method was introduced for road

trac congestion estimation. The method combines both
morphological operations for vehicle detection and speed
estimation, and LBP entropy for video scene randomness
measurement. The road trac congestion is classified onto a
5-level scale that represents the various possible conditions
of road flows. Results show the usefulness of this method
when applied to dierent road flows at peak and o-peak
times.
Acknowledgements The authors would like to thank Ruba Almumani,
Basma Daaja, Rawan Abuosbaa, and Amani Madi for their assistance in
acquiring the road trac videos used in this work.

References
1.

Villagra J, Milanes V, Perez J, Godoy J. Smooth path and speed planning for an automated public transport vehicle. Robotics and Autonomous Systems, 2012, 60: 25226

2.

Towards a new culture for urban mobility. Green paper of European

Commission, 2007, 16

3.

Hartgen D, Fields M G, Moore A. Gridlock and growth: the eect of

trac congestion on regional economic performance. Reason Foundation, 2009, 18

628

Front. Comput. Sci., 2014, 8(4): 619628

4. Schrank D, Lomax T, Eisele B. 2011 Urban Mobility Report and Appendices. Texas Transportation Institute, 2011

nighttime vehicle detection and trac surveillance. IEEE Transactions

on Industrial Electronics, 2011, 58: 20302044

5. Lomax T, Turner S, Shunk G, Levinson H S, Pratt R H, Bay P N,

Douglas G B. Quantifying congestion. Transportation Research Board,
1997

14.

Yu Z, Chen Y P. A real-time motion detection algorithm for trac monitoring systems based on consecutive temporal dierence. In: Proceedings of the 7th Asian Control Conference. 2009, 15941599

6. Pattara-Atikom M, Pongpaibool P, Thajchayapong S. Estimating road

trac congestion using vehicle velocity. In: Proceedings of the 6th
International Conference on ITS Telecommunications Proceedings.
2006, 10011004

15.

Marfia G, Roccetti M. Vehicular congestion detection and short-term

forecasting: a new model with results. IEEE Transactions on Vehicular
Technology, 2011, 60: 29362948

16.

Brahme Y B, Kulkarni P S. An implementation of moving object detection, tracking and counting objects for trac surveillance system. In:
Proceedings of the 2011 International Conference on Computational
Intelligence and Communication Networks. 2011, 143148

17.

Buch N, Velastin S A, Orwell J. A review of computer vision techniques for the analysis of urban trac. IEEE Transactions on Intelligent Transportation Systems, 2011, 12: 920939

18.

Al-Kadi O S. Combined statistical and model based texture features for

improved image classification. In: Proceedings of the 4th International
Conference on Advances in Medical, Signal and Information Processing. 2008, 14

19.

Al-Kadi O S. Supervised texture segmentation: a comparative study.

In: Proceedings of the 2011 IEEE Jordan Conference on Applied Electrical Engineering and Computing Technologies. 2012, 15

20.

Gonzales R C, Woods R E. Digital Image Processing. Prentice Hall,

2002, 299300

21.

Ojala T, Pietikainen M, Maenpaa T. Multiresolution gray-scale and rotation invariant texture classification with local binary patterns. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 2002, 24:
971987

22.

Guo Y, Zhao G, Pietik M. Discriminative features for texture description. Pattern Recognition, 2012, 45: 38343843

7. Oh J, Min J, Kim M, Cho H. Development of an automatic trac conflict detection system based an image tracking technology. Transportation Research Record, 2009, 4554
8. Cheng H Y, Hsu S H. Intelligent highway trac surveillance with
self-diagnosis abilities. IEEE Transactions on Intelligent Transportation Systems, 2011, 12: 14621472
9. Houben Q, Diaz J C T, Warzee N, Debeir O, Czyz J, Insticc. Multifeature stereo vision system for road trac analysis. In: Proceedings
of the 4th International Conference on Computer Vision Theory and
Applications. 2009, 554559
10.

Bishop B, Casabona J. Automatic congestion detection and visualization using networked GPS unit data. In: Proceedings of the 47th Annual Southeast Regional Conference. 2009, 79

11.

Ozkurt C, Camci F. Automatic trac density estimation and vehicle

classification for trac surveillance systems using neural networks.
Mathematical&Computational Applications, 2009, 14: 187196

12.

Jain V, Sharma A, Subramanian L. Road trac congestion in the developing world. In: Proceedings of the 2nd ACM Symposium on Computing for Development. 2012, 11

13.

Chen Y L, Wu B F, Huang H Y, Fan C J. A real-time vision system for