The International Journal of Multimedia & Its Applications (IJMA) Vol.10, No.

4/5, October 2018

A METHOD FOR TRACKING ROAD OBJECTS

Salma Kammoun Jarraya1,2
1
Computer Science Deapartment, Faculty of Computing and Information Systems
King Abdulaziz University, Jeddah, Saudi Arabia.
2
MIRACL-Sfax

ABSTRACT
In this paper, we present a new road traffic monitoring approach for a highway control and management
system called RoadGuard. This system copes with several challenges in this type of applications in order to
count and track road objects robustly. It adopts a novel approach for tracking road objects, first, to
determine continuously their positions on the road, and then it uses vehicle positions to estimate their
trajectories. The trajectory analysis provides vital information to control and manage highway traffic. In
this paper, our main contribution is a tracking method based on a coherent strategy where both region and
object information are used to establish objects correspondence over time. Our method operates in two
phases: a spatial analysis that uses a multilevel region descriptors matching in order to identify object
interactions and particular object states; and a continuous temporal analysis applied to cope with track
management issues. As demonstrated experimentally, the proposed method can detect, track and count
road objects accurately in highway videos that include several constraints. In addition, it produces
effective and stable road objects tracking.

KEYWORDS
Foreground segmentation, Tracking Target, SIFT, Control and management system.

1. INTRODUCTION
The alarmingly increasing numbers of car accidents stim- ulated several research efforts to find
counter measures. Among the explored solutions, computer vision software are being developed
for highway control and management. These software rely essentially on tracking road objects to
estimate the trajectory of moving objects over time. The information gathered by road object
tracking helps in identifying their behavior in the observed scene. In addition, it can be used to
collect statistical information about the traffic, which in turn can be used to control and manage
the traffic to prevent road congestion and accidents.

In this highway traffic control and management context, the work presented in this paper aims at
proposing a new method for tracking multiple rigid moving objects (i.e., road objects) with
different sizes and speeds in highway traffic videos. The videos are supposed to be acquired using
a stationary camera with a large field of view, thus independently of the camera position. Our
method relies on an automatic detection of moving objects. In addition, to handle the size and
speed differences of the moving objects, our method has the merit of automatically accounting for
possible state changes of the moving objects, interactions among them like occlusions,
appearances of new objects and/or disappearances of existing objects.

The remainder of this paper is divided into four sections. In Section 2, we describe a brief state of
the art in object tracking. Section 3 presents our proposed method. Section 4 highlights its
advantages through the results of a quantitative and a qualitative evaluation. Finally, Section 5
recapitulates the presented work and outlines its extensions.

DOI: 10.5121/ijma.2018.10501 1
 The International Journal of Multimedia & Its Applications (IJMA) Vol.10, No.4/5, October 2018

2. RELATED WORK
Tracking moving is one of the most challenging computer vision tasks. Several methods [24] [1]
[3] [23] were proposed to deal with object tracking. In our study of existing methods, we focused
on the modern methods called online trackers. More specifically, we did not consider pre-trained
and offline trackers for which pre-processing steps are required.
Online tracking is a hard problem where all the information in the sequence is needed, especially
in the initial frames [24]. The accuracy of these methods depends on both the constraints and
context of the final-end application. The constraints are pertinent to the sensors (single or
multiple, mobile or fixed), the observed scene (indoor and/or outdoor) and the tracked object(s)
(single or multiple, rigid or nonrigid). It is worth noting that non-rigid object tracking methods
allow tracking of both non-rigid and rigid objects (e.g. road objects) and can better deal with the
various challenges. These advantages come however with a high computational time and a lack of
genericity because these methods rely usually on silhouettes model that encodes the object nature
and/or shape; the high computational time is unacceptable in our real-time application.
Furthermore, existing rigid object tracking methods suffer from a low performance face to
different challenges. Nonetheless, the low computational time of rigid object tracking methods
motivated us to investigate this strategy for tracking road objects while improving their
performance.
In addition to the constraints stemming from the application context, the methods reported in the
literature differ in their object representation. According to the recent tracking survey papers [24]
[1], the proposed methods can be classified into two categories of approaches: Model based (cf.
[1] [3] [16] [17] and Features based (cf. [29] [6]). In both categories of approaches, the tracking
strategy relies on matching information provided by features/models over time.
Model-based methods can be successfully used as long as they have accurate models for the
different types of tracked vehicles. However, given their reduced timing complexity which is
needed in real-time applications, tracking road objects is usually achieved using the features
based approach. The method proposed in [26] requires no prior model, its main idea is to divide
occluded vehicles into many small fragments (or patches) that are then grouped according to the
clusters of motion vectors found by tracking feature points; as such, this method is applicable
only to a high point of view. The method proposed in [6] tracks vehicles in real-time using Salient
discriminative features (compactness, aspect ratio, and area ratio) and Euclidean distance to
measure the distance between the centroids of two objects. Because the evaluation of this method
[6] is only given for videos with low traffic, it is unclear how well it will perform for high traffic.
In [8], a graph-based vehicle tracking method is used for building the correspondence between
regions (compactness, aspect ratio). In [25], vehicle positions are predicted by Kalman Filter
using feature vector (velocity, centroid position); Kalman Filter is a fast tracker and can deal with
total occlusion. However, the performance of this method depends on a good quality of video
since it requires an initialization step to detect vanishing point detection. Besides the
aforementioned methods, several features-based methods [2][11][20][28] use descriptors points.
Despite the popularity of descriptors points in other applications [29], few works used them to
track road objects. Given their higher robustness compared to other features [29], in addition to
their speed, we revisit the original idea of local features based on descriptors points to apply it for
road object tracking.

Among the techniques used to compute descriptor points are Harris detector [5], KLT (Kanade-
Lucas-Tomasi) detector [14] [27] and SIFT descriptor (Scale Invariant Feature Transform) [12]
[13]. We performed a comparative study between these techniques according to a set of
invariance criteria (Translation, Scale Changes, Image Rotation, Illumination changes,Image
Locale Deformations, Affine Transformation), the results show that, unlike the other techniques,
descriptors from SIFT are invariant to different criteria. In addition, from a theoretical point of

2
 The International Journal of Multimedia & Its Applications (IJMA) Vol.10, No.4/5, October 2018

view, SIFT can produce a great number of descriptor points, give a local image measurement that
is robust to noises and to partial occlusions, and it can give distinctive points as well. Encouraged
by the advantages of descriptors points, we have decided to adopt SIFT descriptor to track road
object.

Within the application context of road traffic, the success of object tracking relies on the
management of frequently object state changes (the lifecycle) as well as object interactions. The
lifecycle of a road object starts by its appearance in the scene (state ’Entry’) and ends by its
disappearance (state ’Exit’). In addition, during its presence in a scene, a road object can be in a
normal state (’Normal’), a normal state with a high speed (’Normal HS’), stopped (’Stopped’),
restarting motion after stopping (’Re-moving’).

Furthermore, during a lifecycle, two types of interactions among road objects can occur. The first
interaction happens when two or many objects appear close to one another (’Merge’) causing
partial or total occlusion. The second interaction results from two or many objects fragmentation
(’Split’ ) after their merger.

In the literature, most of the proposed methods [28] [11] [2] track pre-selected (single and rarely
multiple) specific object(s), the method proposed by Rahman et al. [20] is the exception. This
latter proposes a multiple objects tracking detected automatically, their experiments show that the
proposed method is dedicated to track two object in simple indoor scene. In addition, states
changes of moving objects are not considered.

We have examined recent tracking road objects methods [6] [8] [25] and their challenges. Also,
we have considered the work of [19] which has a good established reputation demonstrated by the
number of times it has been cited. We summarize our observations in Table 1. In [8] and [6],
Entry and Exit states of objects and the counting step are managed through a very small region of
interest fixed manually: frame with five vehicles only two are tracked and counted. In [25], the
authors claim that the graph association and weight assignment can deal with the various object
states and interactions except Re-Moving and stopped objects. However, this paper gives neither a
quantitative evaluation nor an explanation on how this can be done.
Table [Link] road objects methods and their challenges

Challenge Paper [8] Paper [6] Paper [25] Paper [19] Our method
Cars and Cars and Cars Cars Road object
bus bikes

Appearance Yes{Entry} Yes{Entry} Yes{Entry} – Yes

[Entry, Re- {Entry/Re-Moving}
Moving]
Disappearance Yes{Exit} Yes{Exit} Yes{Exit} Yes{Exit} Yes
[Exit, Stop] {Exit/Stopped}
Merge Yes – – Yes Yes
Split Yes – – Yes Yes
Occlusion Yes{Partial} – Yes{Partial} Yes {Partial, Yes{Partial}
[Partial, total] total}
Counting Yes Yes – Yes Yes

As shown in Table 1, Kalman filter and vanishing point detection are used to deal with Entry, Exit
and occlusions in [25]. In [19], the proposed method is based on Kalman filter and vehicle
3
 The International Journal of Multimedia & Its Applications (IJMA) Vol.10, No.4/5, October 2018

contours to determine relation among objects. This kind of practice requires ”perfect” conditions
of video acquisition. We can see also that the works [8] and [19] deal with Merge objects using
spatial analysis which is not efficient for more than two merged objects. As recommended in [24],
combining local (e.g. SIFT) and global features (e.g. template matching) can resolve complex,
frequently merged objects.
From Table 1, we can conclude: (1) despite its popularity, SIFT-based tracking objects is not very
much applied in the context of road object, and (2) the challenges related to road object tracking
remain unsolved with a single method.

In this paper, focusing on tracking road objects for a straightforward surveillance and security
application, our proposed method aims to track an unlimited number of rigid road objects
(multiple rigid moving objects) with different sizes and speeds. More specifically, our objective is
to propose an online tracking features based method that both is capable of overcoming various
challenges and has a low computational time. In addition, the proposed method must take into
consideration: (1) possible state changes and interactions of road objects, and (2) the appearance
of a new or old object and the disappearance of existing object. To handle all these challenges, we
propose a SIFT and Template Matching-based method to track road object at low and high point
of views, even in the presence of severe object interactions and significant object state changes.

3. PROPOSED METHOD
Our proposed method for tracking road objects is based on the two main steps : (1) Spatial
Analysis (AS) to manage objects’ states and interactions for each input frame; and (2) continuous
Temporal Analysis (CTA) to establish all objects tracked from the beginning up to an instant t
and to generate objects’ trajectories. Note that while AS produces object information based on
consecutive frames (at time t-1 and t), CTA produces object information that is ‘global’ in time
(from 0 to a frame t). Both processing rely on similarity measuring and matching.

[Link] Analysis (AS)

We adopt our fast and accurate moving object detection method described in [4] to obtain targets

∈ ∈
to be tracked. Let and denote respectively the segmented regions from frames

∈ ∈
and with cc {1,...,m} and c {1,...,n} , n and m, are respectively the number of region in
two successive [Link] STP, the Rtcc[cc {1,...,m}] and [c {1,...,n}] regions are used to

( { , }. ) and state ( { , } .
{ , } { , }
manage objects states and interactions for each input frame, thus produces region correspondence

attributes and multilevel region descriptors matching of Rtcc[cc {1,...,m}] and

). Spatial analysis takes into account both the spatial
∈ [c ∈
{1,...,n}]. Each region R is represented by a set of attributes (Z(R) = (β1...5(R),φ(R))). Where
β1...5(R) are 2D spatial attributes (cf. Figure 1) and φ128k (R) is a K-by-128 matrix, each row gives

normalized to unit length. Regions correspondences ( { , } .

{ , }
an invariant descriptor for one of the K key points. The descriptor is a vector of 128 values
∈ ∈
, [cc {1,...,m}, c {1,...,n}]
∈ ∈
) are initialized [Link] project βt1(Rtcc) [cc {1,...,m}] onto area fromβt2..5(Rt-1c) [c {1,...,n}],
thusprovides correspondence for regions in ’Normal’ states and/or in ’Split’ interactions. Region
in state ’Normal’ corresponds to the case where βt1(Rtcc) belongs to only one area. The Split
interaction corresponds to the case where βt1 of two or more Rtccbelong to one area. We
associate regions Rtccand according to equation 1.

ℎ !
. =# * (1)
$% &$ '! # ! ℎ !
, (,…
. =#
4
 The International Journal of Multimedia & Its Applications (IJMA) Vol.10, No.4/5, October 2018

Figure 1. 2D Spatial attributes (β1...5(R))

{ , }
A multilevel region descriptors matching is proposed for regions { , }

, }. == +1)[cc ∈{1,...,m}, c ∈{1,...,n}]. This step allows us

{ , }
withcorrespondences ( {

and ’Re-Moving’). We aim to select, for each region descriptors(. (1.../0 2 3), its match to
to cope with region interaction (’Merge’) and states (’Entry’, ’Exit’, ’Normal HS’, ’Stopped’

(./4(1 2 3) (equation 2).

R_Match =.>(1 2 3, ./4(1 2 3? = 1 *(2)
5
!@ 2A %BC D E 03

(Des Match >0). Decision to select matched descriptors from ./4(1 2 3 is given by equation
There is matching (RMatch = 1) between two regions in case of at least one descriptor match

3.
Des_Match2i3 = 1 * (3)
G
A % + L #ℎ2 3 = 1 2AM 213 N 0.6 ∗ AM 223 ∈ {1, . . . , R1}
∈
In our work, SIFT descriptors matching is based on dot products (DPi[ i {1,...,k1}]) between
unit vectors of descriptors (equation 4). Generic rules of the multilevel region descriptors
matching is presented by Algorithm 1.
.../0 2 3)*(./4(1 2 3))
(1
DPi= sort(arcosine((. (4)

Three level matching levels are proposed: the first one is between ./(1 2 3 and
./4(1 2 3S## ∈ 1, . . . , , # ∈ 1, . . . , ! T to identify regions with state ’NormalHS’ in case
(1 2 3 matches to only one ./4(1 2 3 or prevent merging interaction (’Merge’) in case
of./
(1 2 3 matches to two or more ./4 U , (,…, V W (lines 1 to 4, Algorithm 1). The second
of ./ (1

one is between ./(1 2 3S## ∈ 1, . . . , , # ∈ 1, . . . , ! Tand Stopped(h).φ.

∈
Structure Stopped(h).φcorrespond to region of stopped objects in previous frames, thus, if they

to 8, Algorithm 1). The third matching is between ./4(1 2 3 and ./X(1 2 3 to identify
match, Rtcc[ cc {1,...,m}] are in state ’Re-Moving’, otherwise they are in state ’Entry’ (lines 5

2′Z[ 3S# ∈ 1, . . . , ! ] (lines 9 to 13,

Algorithm 1). ./X(1 2 3] corresponds to SIFT descriptors ofβt2..5(
stopped objects, otherwise means disappearance of
) projection onto current
frame.

More precisely, attributes of objects in states ’Entry’, ’Split’ and ’NormalHS’ are updated

objects in state ’Exit’ U\ #R !]^_` # a^ . b d 2 3e "" . W^ 2 . ""∗3 are

c
according to equation 5 . Objects in state ’Stopped’ are controlled by Stopped(Oj=1...h).φ and

killed.

5
 The International Journal of Multimedia & Its Applications (IJMA) Vol.10, No.4/5, October 2018

' U\ #R !]^_` # a^ . b d 2 3e "" . W^ 2 . ""∗3\ℎ !

c

h
g \ #R !]^_` # ^ . b d 2
c
3 " ## *
g
g \ #R !]^_` # ^ . b d 2i
c ...j
3 "i ...j
2 3
(5)

f \ #R !]^_` # ^ . b 2.3 " .2 3

cd

Algorithm 1 Multilevel region descriptors matching

./(1 2 3: Regions descriptors at t-1 [cc ∈{1,...,m}]

Input:

./((1 2 3): Regions descriptors at t [c ∈{1,...,n}]

-

Stopped(Oj).φ: structure of objects in state ’Stopped’ [j ∈{1,...,h}]

-
-

.
Output:

.
- : Regions states structure at t
- :Regions matching structure at t

IfR_MatchU.>(1 2 3, ./4(1 2 3k WThen

- Stopped(Oj).φ:structure of objects in state ’Stopped’

If R_MatchU.>(1 2 3, ./4(1 2 3k W Then

. =#
. = $l ′
1.

, (,.., V k
2.
Else if R_Match =.>(1 2 3, ./4(1 U W ?
1. . = #1, #2, … , #m
2. . = L ] ′
End if
Else
If R_MatchU.>(1 2 3, && nU^ W. oWS## ∈ 1, . . . , T Then
1. . = && n2^ 3.
2. . = − L p !]′
Else
1. . =∗
2. . = Z! @′
End If
If R_Match =.>((1 2 3, .>X(1 2 3? S# ∈ 1, . . . , ! T Then
1. && n2^ q 3. =#
2. && n2^ q 3. = && n′
Else
1. . =∗
2. . = Z[ ′
End If
End

3.2 Continuous Temporal Analysis (CTA)

CTA establishes all objects tracks (TrackingObject{Oi} [ i {1,...,ObjectCount}]) from the ∈
beginning of the video stream until an instantt. It relies on , , . and , , . S## ∈
1, . . . , , # ∈ 1, . . . , ! T to generate objects trajectories.
6
 The International Journal of Multimedia & Its Applications (IJMA) Vol.10, No.4/5, October 2018

∈
For each frame, LTP rule feedbacks objects (TrackingObject{Oi} [ i 1,...,ObjectCount]) and
their corresponding regions to update tracked object attributes. Spatiotemporal attributes and
c
descriptors of tracked object 2b d = 2i 2..5 2^ 3, , o2^ 33 are updated according to
region/object association. The association between objects and their corresponding regions is
based essentially on . .

Note that attributes of objects in a merging region cannot easily be obtained since several objects
share the same region. To deal with this problem, we use template matching based sum of squared
difference to find 2D spatial attributes of each object, then, we compute their SIFT descriptors.
Sum of squared difference is implemented using FFT (Fast Fourier Transform ) based correlation.

4. EXPERIMENTAL RESULTS
In order to validate our contributions, we experimentally evaluated the proposed method to track
road objects. We carried out two series of experiments whose results are presented in the second
sub-section. We first clarify the experimental conditions, the used data set and validation
conditions and techniques.

We used two road traffic sequences1recorded in typical conditions (HighwayIIand HighwayIII).

‘HighwayII’ shows a dense traffic of road objects that share some characteristics (color, size, ...).
The distance between road objects is often very small, which produces partial occlusions and,
hence, frequent interactions (‘Merge’ and ’Split’). ‘HighwayIII’ shows a dense traffic of road
objects of different speeds often very fast, and with different types and sizes.

In the first experiment, we evaluated quantitatively and qualitatively the accuracy of our method
for road objects tracking. Quantitative evaluations need Ground-Truths (G-T) which can be
viewed as the correct answer for what exactly the algorithm is expected to produce; they are used
to evaluate the obtained results quantitatively. However, since Ground-Truth tracks are not
available for these sequences despite their celebrity, we had to develop a semi-automatic software
to produce, for each sequence, several road object tracks from typical sequence parts. Four Parts
from HighwayIIand five parts from HighwayIIIcover several challenges (frequent ’Merge’ and
’Split’ between road objects, dense traffic includes road objects in different size and speed,
several states changes of road objects at the same time, high road objects speeds).

The evaluation is made through the calculation of the rates of Centroid Error [22] [18] [15] with
regard to Ground-Truth (GT) of parts in both sequences (4 parts for ’HighwayII’ and 5 parts for
HighwayIII). The Centroid Error rates are computed by the Euclidean distance (between two
centroids) according to a twopass matching scheme: the first pass matches the system track to GT
(distanceSy) to find false positive tracks , and the second pass matches the GT to system track
(distanceTrack) to find false negative tracks. In typical results, the Centroid Error rates from the
two passes are the [Link] addition to the above quantitative metric, we also considered in our
evaluation a second metric ’Two-pass many-to-many system to ground truth track matching’ [10]
to measure how the system can deal with ‘Merge’ and ‘Split’ interactions. A GT/system track is
matched to the system/GT track if there is both temporal overlap and spatial overlap. Temporal
overlap is with respect to the duration of the system track. Spatial overlap is based on the
centroid of the system lying inside the bounding box of the ground truth track. If multiple GT-
matches, then this system track has ‘Merge Error’ equal to matched GT-tracks. If multiple
system-matches, then this GT track has ’Split Error’ equal to matched system- tracks.

1 [Link]
7
 The International Journal of Multimedia & Its Applications (IJMA) Vol.10, No.4/5, October 2018

In the second experiment, we evaluate our results by comparing our systemperformance with the
following similar works: a Kalman filter based method [9], a well-known and referenced paper
[19] and a recent proposed method [25]. In addition, the effectiveness and accuracy of the
proposed method is demonstrated, in the experiment 3, through a Highway Control and
Management System, called RoadGuard[7]. Semantic results of RoadGuardare based on
counting and tracking moving vehicles starting by detecting the real moving objects. We next
present and discuss results of the (1) Experiment 1: Quantitative and Qualitative evaluations, (2)
Experiment 2: Comparison with related works, and (3) Experiment 3: RoadGuard

[Link] 1: Quantitative and Qualitative evaluations

As we can see in Figure 2, the four HighwayIIparts echoed a very low average
distanceSys/distanceTrackrate per frame. Table 2 summarizes the average
distanceSys/distanceTrackand FPT /FNT rates for each sequence part: the distanceSys/ -
distanceTrackechoed respectively between 0 and 6.163 pixels while FPT and FNT are between 0
and 6.19 percent.

Figure 2. Average distanceSys/distanceTrack curves of the four parts from HighwayII

We have performed an experimental study to know how our system can deal with ’Merge’ and
’Split’ interactions. ’Merge Error’ and ’Split Error’ are computed for 11 tracks from HighwayII.
Temporal overlap and Spatial overlap curves for the 11tracks are depicted in figure 3. For each
track, both measures are computed firstly (A) from GT-Track-Matching and secondly (B) from
System-Track-Matching.

8
 The International Journal of Multimedia & Its Applications (IJMA) Vol.10, No.4/5, October 2018

Table [Link] AVGdistanceSys (AVGdS) /AVGdistanceTrack(AVGdT) and FPT / FNT rates

for HighwayII(HII)

HII part AVGdS AVGdT FPT FNT

HII Part 1 4.403 5.140 6.19 0

HII Part 2 6.618 6.163 0.83 0

HII Part 3 2.700 4.070 3.51 0

HII Part 4 0 3.283 0 0

There is a ’Merge Error’/’Split Error’ in the case of multiple GT-matches/systemmatches. More

explicitly, if a curve from GT-Track-Matching/ System-Track Matching show more than peak
with temporal overlap greater than 0.5. Our system achieves an average ’Merge Error rate’ of
9.09 percent per 11 tracks and an average ’Split Error rate’ of 0 percent.

Figure [Link] overlap and Spatial overlap of 11 tracks from HighwayII

The five HighwayIIIparts (figure 4) showed a low average distanceSys /distanceTrackrate per
part, respectively between 1.928 and 9,682 pixels. FPT and FNT are between 0 and 20,44 percent.
The average distanceSys/-distanceTrackrates per frame are given in table 3.

9
 The International Journal of Multimedia & Its Applications (IJMA) Vol.10, No.4/5, October 2018

Table [Link] AVGdistanceSys (AVGdS) /AVGdistanceTrack(AVGdT) and FPT / FNT rates for
HighwayIII(HIII)

HII part AVGdS AVGdT FPT FNT

HIII Part 1 1.928 7.547 5.77 0

HIII Part 2 8.795 0.685 3.85 0

HIII Part 3 3.181 9.682 10.08 9.52

HIII Part 4 7.361 5.247 0.60 20.44

HIII Part 5 1.505 5.245 1.67 0

Figure [Link] distanceSys/distanceTrackcurves of each frames part 5 from HighwayIII

Figure5 present qualitative results on frames from HighwayI for tracking road objects (car,
bike and person). Overall, our method produces good results in presence of merged objects,
appearance of new object and disappearance of objects. In addition, Figure 5 presents the
results of counting road objects (first column) in presence of Merged/Split objects. As we can
see in the detection results (second column), merged objects appear as one object and a

10
 The International Journal of Multimedia & Its Applications (IJMA) Vol.10, No.4/5, October 2018

stopped object does not appear in the detection. Nevertheless, our tracking method counts
correctly the existing road objects.

Figure [Link] of counting road objects in presence of Merged/Split objects

4.2 Experiment 2: Comparison with related works

A complementary quantitative evaluation was performed by comparing our results (Average
distanceSys/distanceTrack) with the results of Xin [9] on frames from HighwayII. Method of
[9] is based on Kalman Filter, which is a probabilistic prediction rule to visual tracking. This
method establish Kalman filter motion model with the features centroid and area of moving
objects. Note that, the frame level tests are chosen from those containing objects in merge and

∈
split [Link] illustrated in Figure 6, Kalman filter presents similar average distanceSys/

∈
distanceTrack rates ( [1..4]) for the first frames part. However, they give high average
distanceSys/ distanceTrackrates per frame ( [16..20]) in the second part which includes
objects in merge and split states. In fact, Kalman filter fails to provide useable results in the

∈
presence of objects in merge and split states. On the same frames, our tracking method shows
low average distanceSys/distanceTrackrates per frame ( [3..4]).

For our application domain, the lack of open access to the codes, datasets and detailed
descriptions of algorithms hinders the elaboration of a fair comparisonwith several methods
and on large datasets.

Figure [Link] results with Kalman filter

Nevertheless to put our method in its context, a reasonable comparison or the performance can be
deduced knowing the hardware, number and size of frames, and number of videos used in the
evaluations. The performance can be analyzed in terms of the rates of Merge error, Split error,
and Mean processing speed (MPS).
Table [Link] with related works
11
 The International Journal of Multimedia & Its Applications (IJMA) Vol.10, No.4/5, October 2018

Paper Merge Split error MPS Accuracy

error
[25] - - 10 f/s 91.5%
[19] 2.5% 1.6% 10.99 f/s 96%
Our 9.09% 0% 11.02 f/s 97.92%
According to this evaluation (see Table 4), our method records the best accuracy with 97.92% for
large number of frames (56 higher than [19]) because it is the only method that deals with stopped
and Re-moving objects. In addition, our method is faster than the method of [19] and [25] which
are considered in the literature as real time methods. The method of [25] achieves the best rate of
Merge error 6.59% fewer than our rate for frames with high size and quality; but our method
records a smaller rate for Split error.

4.3 Experiment 3: RoadGuard

RoadGuard is a highway control and management system we have implemented on a standard PC
hardware. The control phase of RoadGuardis based on tracking vehicles in a defined Region Of
Interest (ROI) and the emergency area. The management phase of the RoadGuard is based on
counting vehicles in highways in order to obtain statistical information like the date and time of
overloaded highways. The counting process is done in the ROI. The counter is incremented for
each road object enters the ROI and decremented after its disappearance. In order to confirm the
important effect of tracking road objects for highway control and management system, we
integrate our proposed method in the RoadGuardprocess. ROI is obtained automatically by
method of [21] (cf. figure 7). RoadGuarduses ROI to detect vehicles stopped in road.

Figure [Link] of three highway sequences

RoadGuardachieves suitable rates for counting vehicles in ‘HighwayII’ (4 parts) and

‘HighwayIII’ (5 parts) (figure 8 and figure 9)when compared with the GT count (success rates:
90% for ‘HighwayII’ and 82% for ‘HighwayIII’) .

Figure [Link] results for four parts from HighwayII

12
 The International Journal of Multimedia & Its Applications (IJMA) Vol.10, No.4/5, October 2018

Figure 9. Counting results for five parts from HighwayIII

5. CONCLUSIONS
In this paper, we proposed a novel points-based method using the SIFT technique to track road
objects in highway videos. Our method operates in two phases: Firstly, a a spatial analysis uses
a multilevel region descriptors matching in order to identify object interactions and particular
object states. Secondly, a continuous temporal analysis is applied to cope with track
management issues. Our preliminary experimental evaluation on real highway sequences
shows that our method can track multiple rigid moving objects (i.e., road objects) with
different sizes and speeds in traffic videos of highways. Moreover, the proposed method has
the merit of automatically accounting for possible state changes of the moving objects,
interactions among them like occlusions, appearance of new objects and disappearance of
existing objects. Finally, this experimental study has demonstrated the practical usefulness of
our contributions through a highways control and management system, called RoadGuard,
which gives a good quality of results. Future works will focus on further interpretation of road
objects’ trajectories to identify suspect events.

REFERENCES
[1] Z. Al-Ameen, G. Sulong, A. Rehman, M. Al-Rodhaan, T. Saba, & A. Al-Dhelaan,
(2017)“Phasepreserving approach in denoising computed tomography medical images”. CMBBE:
Imaging & Visualization, 5(1):16–26.
[2] Y. Cheng-bo, Z. Jing, L. Yu-xuan, & Y. Ting, (2011)“Object tracking in the complex
environmentbased on sift”. 3rd International Conference on Communication Software and Networks,
pages 150–153.
[3] X. Gao, Y. Su, X. Li, & D. Tao, (2010)“A review of active appearance models”. IEEE Transactions
on Systems, Man, and Cybernetics, Part C (Applications and Reviews), 40(2):145–158, March.
[4] M. Hammami, S. K. Jarraya, & H. Ben-Abdallah, (2013)“On line background modeling for
movingobject segmentation in dynamic scenes”. Multimedia Tools and Applications, 63(3):899–926.

13
 The International Journal of Multimedia & Its Applications (IJMA) Vol.10, No.4/5, October 2018

[5] C. Harris & M. Stephens. A combined corner and edge detector (1988)“Alvey vision
conference”.Volume: 15, Issue: Manchester, pages 147–151.
[6] D.-Y. Huang, C.-H. Chen, W.-C. Hu, S.-C. Yi, & Y.-F. Lin (2012),“Feature-based vehicle
flowanalysis and measurement for a real-time traffic surveillance system”. Journal of Information
Hiding and Multimedia Signal Processing, 3(3):282–296.
[7] S. K. Jarraya, A. Ghorbel, A. Chaouachi, & M. Hammami (2011)“Roadguard - highway controland
management system”. In L. Mestetskiy and J. Braz, editors, VISAPP, pages 632–637. SciTePress.
[8] J. C. Lai, S. S. Huang, & C. C. Tseng (2010)“Image-based vehicle tracking and classification onthe
highway”. In The 2010 International Conference on Green Circuits and Systems, pages 666–670.
[9] X. Li, K. Wang, W. Wang, & Y. Li, (2010)“A multiple object tracking method using kalman filter”.
InThe IEEE International Conference on Information and Automation, pages 1862–1866.
[10] M. B. Lisa, W. S. Andrew, Y. li Tian, J. Connell, & A. Hampapur, (2005). “Performance evaluation
of surveillance systems under varying conditions”. IEEE Int. Workshop on Performance Evaluation of
Tracking and Surveillance.
[11] Y. Liu, X. Wang, J. Yang, & L. Yao, (2011)“Multi-objects tracking and online identification basedon
sift”. International Conference on Multimedia Technology (ICMT),, pages 429–432.
[12] D. Lowe, (1999)“Object recognition from local scale-invariant features”. The Proceedings of the
Seventh IEEE International Conference on Computer Vision, pages 1150–1157.
[13] D. G. Lowe, (2004)“Distinctive image features from scale-invariant keypoints”. International Journal
of Computer Vision. Volume 60 Issue 2, pages 91–110.
[14] B. Lucas & T. Kanade, (1981)“An iterative image registration technique with an application tostereo
vision”. In Proceedings of the 7th international joint conference on Artificial intelligence, pages 674–
679.
[15] T. Nawaz, A. Ellis, & J. Ferryman, (2017)“A method for performance diagnosis and evaluation
ofvideo trackers”. Signal, Image and Video Processing, 11(7):1287–1295.
[16] Y. Ouyang, (2017)“Structural sparse coding seeds–active appearance model for object tracking”.
Signal, Image and Video Processing, 11(6):1097–1104.
[17] G. Phadke& R. Velmurugan, (2017)“Mean lbp and modified fuzzy c-means weighted hybrid
featurefor illumination invariant mean-shift tracking”. Signal, Image and Video Processing,
11(4):665– 672, 11.
[18] G. Pingali& J. Segen, (1996)“Performance evaluation of people tracking systems”. IEEE Workshop
on Applications of Computer Vision, pages 33–38.
[19] R. Rad & M. Jamzad, (2005)“Real time classification and tracking of multiple vehicles in
highways”.Pattern Recognition Letters, 26(10):1597 – 1607.
[20] M. S. Rahman, A. Saha, & S. Khanum, (2009)“Multi-object tracking in video sequences based
onbackground subtraction and sift feature matching”. In Fourth International Conference on
Computer Sciences and Convergence Information Technology, pages 457–462.
[21] N. B. Romdhane, M. Hammami, & H. Ben-Abdallah, (2011)“A comparative study of vision-
basedlane detection methods”. Advances Concepts for Intelligent Vision Systems, Volume 6915,
pages 46–57.
[22] A. Senior, A. Hampapur, Y.-L. Tian, L. Brown, S. Pankanti, & R. Bolle, (2001)“Appearance models
for occlusion handling”. IEEE Int. Workshop on Performance Evaluation of Tracking and
Surveillance.
[23] K. Shanmugapriya& R. S. M. Malar, (2017)“A multi-balanced hybrid optimization technique totrack
objects using rough set theory”. Signal, Image and Video Processing, 11(3):415–421.
[24] A. W. M. Smeulders, D. M. Chu, R. Cucchiara, S. Calderara, A. Dehghan, & M. Shah,
(2014)“Visualtracking: An experimental survey”. IEEE Trans. Pattern Anal. Mach. Intell.,
36(7):1442–1468.
14
 The International Journal of Multimedia & Its Applications (IJMA) Vol.10, No.4/5, October 2018

[25] J. Sochor& A. Herout, (2014)“Fully automated real-time vehicles detection and tracking with
lanesanalysis”. In The 18th Central European Seminar on Computer Graphics, pages 666–670.
[26] B. Tamersoy& J. K. Aggarwal, (2010)“Counting vehicles in highway surveillance videos”. In 20th
International Conference on Pattern Recognition, pages 3631–3635.
[27] C. Tomasi& T. Kanade, (1991)“Detection and tracking of point features technical report cmu-cs91-
132. Technical report”, 1-22.
[28] Y. Yan, J. Wang, & C. Li, (2011)“Object tracking using sift features in a particle filter”. IEEE 3rd
International Conference on Communication Software and Networks (ICCSN), pages 384– 388.
[29] Y. Yang & Q. Cao, (2013), A fast feature points-based object tracking method for robot
[Link] Journal of Advanced Robotic Systems, 10(3):170.

AUTHOR
Salma Kammoun Jarraya received a Ph.D in Computer Science from Sfax University, Tunisia. She is a
researcher in the MIRACL laboratory (Multimedia, InfoRmation systems and Advanced Computing
Laboratory). Currently, she is Assistant Professor in computer science, CS Department, Faculty of
Computing and Information Technology, King Abdulaziz University, Jeddah, KSA. Her research interests
include computer vision, video and image processing. She has served on technical conference committees
and as reviewer in many international conferences and journal

15