See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/262177463

Road traffic congestion in the developing world

Conference Paper · March 2012

DOI: 10.1145/2160601.2160616

CITATIONS READS
72 8,349

3 authors, including:

Ashlesh Sharma
New York University
13 PUBLICATIONS   144 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Ashlesh Sharma on 05 August 2014.

The user has requested enhancement of the downloaded file.

 Road Traffic Congestion in the Developing World

Vipin Jain Ashlesh Sharma Lakshminarayanan

Polytechnic Institute of NYU New York University Subramanian
[email protected] [email protected] New York University
[email protected]

ABSTRACT Transportation Institute’s 2011 Mobility report [1], congestion in

Road traffic jams continue to remain a major problem in most cities the US has increased substantially over the last 25 years with mas-
around the world, especially in developing regions resulting in mas- sive amounts of losses pertaining to time, fuel and money. Sã
sive delays, increased fuel wastage and monetary losses. Due to o Paulo, Brazil is known to experience the world’s worst traffic
the poorly planned road networks, a common outcome in many de- jams [32], where people are stuck for two to three hours everyday
veloping regions is the presence of small critical areas which are in traffic jams. The issue of traffic congestion has affected both the
common hot-spots for congestion; poor traffic management around developing and developed economies to different degrees irrespec-
these hotspots potentially results in elongated traffic jams. In this tive of the measures taken to curb the issue.
paper, we first present a simple automated image processing mech- A common feature across road networks in many urban regions
anism for detecting the congestion levels in road traffic by process- in the developing world is the presence of critical congestion ar-
ing CCTV camera image feeds. Our algorithm is specifically de- eas; we refer to a critical congestion area as one where a network
signed for noisy traffic feeds with poor image quality. Based on of roads converge and a large amount of traffic needs to traverse
live CCTV camera feeds from multiple traffic signals in Kenya the common congestion area. As per free-flow traffic theory [43], a
and Brazil, we show evidence of this congestion collapse behav- free flow traffic road segment can be associated with a traffic curve
ior lasting long time-periods across multiple locations. To partially where the traffic exit rate is a function of the traffic density in the
alleviate this problem, we present a local de-congestion protocol road segment. A free-flow road segment is known to exhibit a crit-
that coordinates traffic signal behavior within a small area and can ical density point where any traffic input that pushes the density
locally prevent congestion collapse sustaining time variant traffic beyond the critical value can trigger a “spiralling effect” that re-
bursts. Based on a simulation based analysis on simple network sults in the road segment operating at a low-capacity equilibrium
topologies, we show that our local de-congestion protocol can en- point. Worse still, small traffic bursts over short time periods can
hance road capacity and prevent congestion collapse in localized potentially trigger the spiralling effect resulting in a congestion col-
settings. lapse. Many critical congestion areas in developing regions have
poor traffic management systems that if any of these critical con-
gestion areas hits a congestion collapse, the road network can result
Categories and Subject Descriptors in a massive traffic jam for elongated time periods.
I.4.9 [Computing Methodologies]: Image Processing and Com- In this paper, our goal is to design mechanisms to detect the state
puter Vision—Applications; I.6.3 [Computing Methodologies]: of traffic congestion in and around critical congestion areas and
Simulation and Modeling—Applications also design simple preventive mechanisms to prevent critical con-
gestion areas from hitting congestion collapse. In this paper, we
General Terms describe a simple image processing algorithm that can be used to
analyze CCTV video feeds from traffic cameras to detect conges-
Algorithms, Measurement
tion levels in real time. Using this algorithm, we show evidence of
actual congestion collapse across multiple locations in São Paulo,
Keywords Brazil and Nairobi, Kenya. Specifically, we show congestion col-
traffic congestion, traffic detection, congestion collapse, simulation lapse scenarios that last for multiple hours at important junctions in
Nairobi and São Paulo. Our congestion detection image processing
1. INTRODUCTION algorithms have been specifically designed for highly noisy traffic
camera feeds and differ in spirit from conventional traffic image
Poor road traffic management is the primary reason for extended
processing techniques which typically rely on high quality traffic
periods of traffic congestion throughout the world. As per Texas
images [41, 44, 39].
To partially alleviate this problem, we propose a local de-congestion
protocol that coordinates traffic signal behavior within a possible
Permission to make digital or hard copies of all or part of this work for critical congestion area to prevent the critical tipping point behav-
personal or classroom use is granted without fee provided that copies are ior. The goal of the local traffic signal coordination is to maintain
not made or distributed for profit or commercial advantage and that copies the traffic density in the congestion area below the critical density
bear this notice and the full citation on the first page. To copy otherwise, to value. Our local de-congestion protocol coordinates the traffic sig-
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
nals that control the input flow within the congestion area and en-
DEV ’12, March 11-12, Atlanta, GA sures that the local traffic density does not cross the critical tipping
Copyright c 2012 ACM 978-1-4503-1262-2/12/03 ...$10.00.
point. Based on a simulation-based study across simple real-world
road network topologies, we show that our local de-congestion pro-
tocol can prevent congestion collapse in the face of bursty traffic
arrivals and can enhance the road network capacity during such
congestion. We believe that our local de-congestion protocol can
be deployed around potential congested areas resulting in prevent-
ing congestion collapses and thus, maintaining a healthy input and
output rate during congestion. Our local decongestion protocol is
primarily designed for small congestion areas and is orthogonal in
design to many of the conventional intelligent signaling solutions Figure 1: The first picture, taken at 5:15pm shows an empty
proposed in the transportation literature [17, 5, 9, 21, 14, 31, 3]. road in Rio de Janeiro, Brazil. Barely 5 minutes later, the road
While our solution is by no means optimal, an advantage of a lo- is completely jammed up.
calized approach is that the decongestion protocol is easily deploy-
able in critical congestion areas to enhance the operational capacity
during a congestion collapse.

2. POOR TRAFFIC MANAGEMENT

Cities throughout the world have found themselves at the brink
of massive traffic explosion, hence curtailing their ability to manage
traffic. The situation has worsened for developing regions due to
the following reasons:
Unplanned cities: Roads tend to be narrow and poorly built.
As cities grow in an ad-hoc manner, no provision is made towards
scaling road capacities, eventually resulting into several bottleneck
roads, which remain congested for extended periods of time. Fur-
thermore, many developing countries have witnessed an explosive
growth in their vehicular population resulting in a failure of con-
ventional traffic management strategies.
Poor discipline: Drivers often are not trained sufficiently to fol-
low lane discipline. The impact of poor lane discipline, especially
at traffic junctions, deteriorates the already overcrowded junction
situation. Furthermore, drivers frequently jump red lights and block Figure 2: The traffic curve showing the relationship between
the intersection, causing further traffic congestion. These problems traffic density and operational exit rate.
are compounded by the fact that traffic law enforcement is poor,
thereby providing no incentive for drivers to follow the rules.
Alternate traffic means: Countries with fast growing economies
have witnessed a surge in the number of vehicles across major
3. CONGESTION COLLAPSE
cities. These cities seldom have efficient mass transit systems, forc- Fig.1 shows two pictures from a traffic feed from the Rio-Niteroi
ing people to operate private vehicles. This problem is compounded bridge in Brazil. The first picture taken at 5:15pm shows traffic
by the social stigma, where people view operating a private vehicle flowing smoothly, with absolutely no congestion. The second pic-
as a sign of prosperity, while public transport is viewed as being ture, taken barely 5 minutes later, shows that congestion has set in,
used by the lower echelons of society. leading to a traffic jam lasting for two hours. This illustrates the
Archaic management: Traffic junctions are often unmanned, basic fact that once congestion sets in, it takes a long time to be
thereby allowing drivers to drive in a chaotic manner. Even if a resolved.
junction is controlled by a cop or a traffic light, the traffic junc- To understand why traffic congestion sets in quickly, and also
tions are largely independent of any traffic management strategy, takes long to dissipate, we need to understand the concept of a traf-
only optimizing the respective junction traffic flow, in the direc- fic curve, as explained next.
tion of maximum traffic build up. Furthermore, these approaches
enhances traffic mismanagement in already congested roads, accel- 3.1 Traffic Curve
erating congestion collapse (see §3). In this discussion, the terms road and link are used interchange-
Tighter budgets: A significant amount of investment is required ably. There are several metrics that define traffic characteristics
to set up a traffic management infrastructure which can scale with such as speed, flow and density of a link. Consider any link with
the increasing traffic. Such an infrastructure not only involves mea- two points with traffic free-flowing across them. Each link is asso-
suring and analyzing real-time traffic data but also focuses towards ciated with a traffic density representing the number of vehicles in
enhancing congestion detection, solving real time congestion and the link. The operational free-flow exit rate of a link varies with the
forecasting congestion scenarios. In developing countries, ravaged traffic density. The traffic curve, illustrated in Fig. 2, captures the
by corruption and bureaucracy, there are multiple hurdles before variation between these two parameters. At high traffic densities
the money actually progresses towards such large initiatives. (signifying traffic jams), links have very low operational exit rates
In this paper, we specifically make the case that while increas- and at low densities, the exit rate varies linearly with the traffic
ing road capacity is useful, it is not the only way to mitigate traffic density. Each traffic link reaches an optimal capacity at a corre-
problems. By smart flow-control techniques in present infrastruc- sponding optimal operating density.
ture, it is possible to increase the operational capacity of the existent To formalize this notion, there are a few parameters that describe
road system. the state of a link l:
 • Capacity: This is the maximum number of vehicles that the
L
link can hold. It does not vary over time, and is represented
as M AXl .

• Buffer size: This is the number of vehicles on the link at Hbef Haft
any given point of time. However, traffic density is defined
as number of vehicles per unit of road. Thus, buffer size is
simply a multiple of traffic density(x link length). Hence, the
terms buffer size and traffic density are used interchangeably, Figure 3: A sample topology illustrating a 2:1 merge scenario.
represented as Bl .

• Exit Rate: This is the rate at vehicles can exit a link, and merging never faces a congestion problem. If however, the sum of
is represented as Cl . As mentioned earlier, it depends on the input rates of L and Hbef is larger than Cl∗ (Haf t ), then the
M AXl , as well as Bl . buffer size of Haf t grows. If the buffer of Haf t grows beyond
Bl∗ (Haf t), then the exit rate of Haf t begins to drop thereby, trig-
The traffic curve behavior has been well documented by the In- gering the spiralling effect. Figure 3 shows how a small burst in
telligent Transportation Systems (ITS) community [43]. From the traffic beyond Cl∗ (Haf t ) is sufficient to drive the system to operate
curve, we see that for low buffer sizes, the exit rate increases as the at a low-capacity point. Once reached, the system will continue at
traffic density increases. Beyond a certain point, however, conges- this operational point and if the total input rate is greater than the
tion takes place, and the exit rate of the link reduces. The point operational exit rate, the congestion increases and spreads into the
at which this transition takes place has the highest exit rate, de- links Hbef and L.
noted by Cl∗ . The corresponding traffic density is denoted by Bl∗ .
Previous research has shown that Bl∗ is approximately equal to
M AXl
3
[43]. 4. DETECTING TRAFFIC CONGESTION
Consider the case when the traffic density is Bl∗ , and hence the The first step in mitigating traffic congestion is to estimate the
exit rate is optimal at Cl∗ . If a short burst of traffic enters the link amount of traffic on the link at any given point of time. A common
and temporarily pushes the traffic density Bl to be more than Bl∗ , method is to place sensors on the road and count the number of
the exit rate Cl will drop below Cl∗ ; this decreased exit rate will fur- times they are actuated by the passing wheels of a vehicle. This
ther increase Bl . This domino effect leads to the exit rate decaying approach suffers from four main problems: a) it is expensive to
rapidly, and we call it congestion collapse. deploy, as the sensors need to be partially embedded in the tarmac,
An important point to note is that even if the input rate is greater b) the sensors on the road are prone to theft, c) sensors need to be
than the maximum exit rate Cl∗ , congestion does not take place until placed at multiple entry and exit points on the road, to maintain
the traffic density becomes greater than Bl∗ . Consider a congested accurate counts, and d) even on a single stretch of road, the sensors
link whose Bl value is greater than Bl∗ . Even if the input rate need to be placed at regular intervals so as to estimate the density
is reduced to below Cl∗ , we are not guaranteed to get rid of the on different segments of the road.
congestion in a short period of time, because the exit rate Cl at a A number of highways constructed in the previous decade con-
traffic density of Bl > Bl∗ will be lesser than Cl∗ . This is the reason tain CCTV cameras to monitor the real-time traffic situation along
that once congestion collapse takes place, it stays that way for long the highway. Traditional methods of traffic estimation utilizes these
periods of time. CCTV camera images for vehicle counting [44] and base front [39]
Thus emphasis of the curve analysis lies in not letting the Bl estimation. Although, these techniques are highly problematic and
value of a link to get greater than Bl∗ − ǫ for a small constant ǫ. A erroneous in such images because of their highly noisy nature and
small burst of traffic ǫ, even for a short period of time, is sufficient decreased ability to isolate vehicle characteristics. Along with low
to push the link into congestion collapse, as illustrated in Fig.1. quality CCTV cameras, the process suffer from three major issues,
a) low camera resolution resulting in highly noisy images, b) traffic
3.2 A simple example camera’s limited field of view and c) light illumination from multi-
To better understand how congestion collapse happens, we con- ple reflecting sources distorting vehicle classification capabilities.
sider a simple example of 2:1 merge where two free-flow roads The detection mechanism is divided into two parts, a day time
merge; this is a common occurrence in road networks. A simple and a night time estimation methodology. Both mechanisms are
example is illustrated in Figure 3 where vehicles in L are merg- different due to the high environmental differences, which results
ing with the stream of vehicles on H. This simple example can into two different image processing techniques. Apart from the en-
be viewed at multiple granularity’s: two lanes in the same freeway vironmental differences, vehicle’s headlight and billboard illumina-
merging into a single lane or two separate free-flow roads merging. tion adds considerable noise to the image making vehicle counting
To visualize this problem from the perspective of traffic curves, difficult.
consider three links in the setup:(a)Hbef representing,a small seg-
ment of H (covering a short distance of up to 0.5 miles) before 4.1 Day Time Congestion Detection
the merge point; (b) a small segment L before the merge point; During the daytime, the underlying intuition is that when there
(c)Haf t , representing a small segment of H after the merge point. is no traffic on the road, it appears gray in color irrespective of the
Each of the links can be associated with their corresponding traffic natural day light. When the road is filled with traffic, the amount
curves. Since we are dealing with a discrete version approximation of visible gray(empty road) in the picture reduces because of the
using traffic curve, we should choose reasonable lengths to have majority of vehicles attributing a varied level of non gray color.
meaningful buffer values for the links. To perform traffic density estimation, we first use a simple poly-
We primarily concentrate on two specific parameters of Haf t : gon to manually mark the road segment area for the image anal-
Cl∗ (Haf t ) and Bl∗ (Haf t). If the sum total of the exit rates of Hbef ysis. For a given traffic camera feed on a road segment, this is
and L is always less than the optimal exit rate Cl∗ (Haf t), then the a one-time operation that explicitly specifies the region of interest
 No Congestion Heavy Congestion
80 80

70 70

60 60

50 50

40 40

30 30

20 20

10 10

0 50 100 150 200 250 0 50 100 150 200 250

Figure 5: Different Threshold levels (a) Original Image, (b)

Figure 4: Left: No congestion shot and histogram. Right:
Threshold Level(0.5), (c) Threshold Level (0.7), (d) Threshold
Heavy congestion shot and histogram.
Level (0.8)

for the analysis. Then, we convert the picture into an 8-bit gray- Similar to the day time estimation algorithm, the road segment
scale and analyze the pixels within the marked segment area. For is marked which in our scenario is typically along the middle of
each value (0-255), we plot a histogram for the number of pixels the image. The algorithm then calculates the amount of white light
that have each of the 256 different gray-scale values. We have veri- in the image, reading a pixel line and registering the pixel measure
fied that the gray of roads lies in the 135-165 range. Intuitively and towards the total white light in a weighted fashion. The weighted
backed by the analysis, if an histogram is constructed for the vary- factor is required as the observer goes towards the far point of view.
ing levels of gray in the picture, depending on the level of conges- Light from multiple vehicles becomes difficult to distinguish and
tion of the road, a histogram would be observed to have a smooth thus a same size segment contains a different number of vehicles
gray level area as compared to a high peak expected in an empty as compared to the near point of view. The traffic density in night
road. With increasing traffic, the peaks at 135-165 begins to re- time is given by:
duce and the drop-off at either side is more gradual. By examining
these histograms, we can easily estimate the traffic density on the
roads. Fig.5 shows snapshots from a traffic feed (along with the
X H
Density = count(i)
corresponding histogram) in Rio de Janeiro (Brazil) The image on i
H − (Pi /P∆ )h∆
the left shows no congestion, and we observe that the correspond- where i is the ith pixel line and count(i) represents the number of
ing histogram shows peaks in the road gray areas. Similarly, the white pixels in that pixel line. H denotes the actual height of the
figure on the right shows a congested road and the corresponding camera, Pi and P∆ represents the projection of the ith pixel line
histogram. Observe that the histogram is more evenly spread out, and the last pixel line on the camera approximately. h∆ represents
and does not peak in the gray areas as much as the case with low the observed height of complete road length in the image and is
congestion. given by
4.2 Night Time Congestion Detection
Night time congestion detection is a harder problem because of Xmax h∆
=
multiple extraneous factors. Absence of light eliminates typical Xmax + d H
vehicle feature estimation techniques. The next contender for vehi- where Xmax represents the farthest visible point on the road seg-
cle identification becomes headlight counting, which suffers from ment and d represents the nearest visible point of the road segment
light reflection/refraction and alternate light sources such as bill- measured relative to the position of the camera.
boards and traffic signal lamps. Here, we present a summary of our
night time vehicle identification technique explained in [38]. The 4.3 Density mapping function
density estimation algorithm estimates the white pixel distribution The final vehicle estimation is done by a simple density map-
in the image. We reduce the effects of reflected lightning from ping function which is used with the help of a semi supervised
our thresholding process which considerably reduces the amount machine learning algorithm. For each road segment and specific
of false light present in the image due to the different intensities of camera feed, we manually estimate the vehicular density for several
light source and the reflected/refracted light. Fig. 5 shows four im- day-time and night-time images. These density estimates provide
ages at different thresholding levels and the successive extraneous the known data for our learning algorithm. Hence, for every road
light reduction in a noisy image. segment, we have manually calculated vehicular density for a very
Since successive thresholding levels affects the image noisiness small number of images and the graded density measure analyzed
in a progressive manner, we thresholded multiple images to bal- from the image processing algorithm. The unknown data points
ance characteristics loss and identify a suitable thresholding level. are analyzed using the known data points by utilizing the Shepard’s
We found that a thresholding level of 0.8 is an ideal thresholding method of Inverse Distance Weighing [15]. Note that the mapping
level for our traffic image dataset.. The thresholding process is fol- function differs for night time and day time images. We have a
lowed by a gradient white light estimation across the link length for simple background filter function to distinguish between day time
calculating the traffic density. and night time images. Given a previously unseen image, we first
estimate the graded density measure from the image processing al-
gorithm and then use the manual data to analyze the graded traffic
density. We then compute the final vehicular density as a weighted
average of the vehicular densities with help from the manually clas-
sified images.

5. EVIDENCE OF CONGESTION COLLAPSE

Numerous sources [32, 2, 7, 46] have reported that traffic con-
gestion is a serious problem in many cities in the developing world,
leading to billions of dollars in losses annually. To understand the
prevalence of congestion collapse, we have analyzed several days
worth of round-the-clock traffic feeds from several locations.
The traffic feeds used in this evaluation primarily consisted of
CCTV camera feeds. Quite often, video camera feeds consists of
an multi-camera stream where feeds from multiple cameras are dis-
played cyclically with a known time period. Since we base our
work solely using camera images, we convert the video feeds to
images and aggregate them per camera in an automated manner.
The process is easy to perform and can be done in real-time, taking Figure 6: The famous Rio-Niteroi bridge in Brazil across the
an insignificant amount of time per image. The static images now Guanabra Bay.
obtained from the video feeds are fed to the evaluator which utilizes
the configured semantics and performs day/night time estimation to
analyze the vehicle density estimates.
The sources from which traffic data was collected from were:
Rio de Janeiro, Brazil: Rio-Niteroi bridge [22, 4], the junction of
Av pres Vargas and Av Rio Branco, Camerino, Rd Santana and Rd
Frei Caneca.
Mombasa, Kenya: Moi Ave (TSS Bldg), Kengeleni Junction,
and Sabasaba Traffic Lights [16].
Nairobi, Kenya: Pushorttam Place (Museum Hill), Barclays
Plaza (Uhuru Highway/Kenyatta Ave Rndbt), Ukulima House (Haile
Selassie/Upperhill Road Rndbt), National Bank Building (Haram-
bee Avenue), National Bank Building (Railways Roundabout), Na-
tion Centre (Kenyatta Ave/Kimathi Street), Heidelberg Building
(Mombasa Road), and a few others [16].
In total, we collected traffic image feeds from all three of the
above-mentioned places for a total of 80 hrs during the month of
July 2011. Different road segments were identified to exhibit con-
gestion on different timings. A bird’s eye view of the source image
set displays high congestion in Brazil region during evening which
continues to night primarily near 4:30PM to 8:30 PM. On the other
hand, traffic in Kenya was observed to be highly erratic and bursty
starting from 10AM to ending at 2PM. The evening time traffic was
relatively non congested for the Kenyan region.
These feeds show us that congestion collapse is a big problem
that occurs on a daily basis. Furthermore, congestion collapse lasts
for elongated periods of time (sometimes, spanning over several
hours) over long highway roads. In this section, we use camera
feeds from two different locations on the famous Rio-Niteroi bridge
(Brazil) to illustrate the problem of congestion collapse. This mas-
sive 8-lane bridge is over 13 kilometers long, and connects the two
cities across the Guanabara Bay. A picture of this impressive bridge
is shown in Fig. 6.
Fig. 7 shows pictures from location 1 on the bridge. The first pic-
ture in this set was taken at 5:15pm, and congestion sets in barely
5 minutes later (as illustrated in Fig. 1). The next five pictures Figure 7: Traffic feed from the Rio-Niteroi bridge in Brazil
are taken over a 2-hour period, and we observe that the congestion shows shots from a 2-hour long traffic jam (consider the on-
remains until 7:35pm. This illustrates the fact that once conges- coming lane only). Traffic congestion sets in barely 5 minutes
tion sets in, the operational capacity of the road reduces, thereby after the first picture was taken, as illustrated in Fig. 1.
leading to long periods of congested traffic. The erratic behavior is
observed on a daily basis throughout the analysis.
Fig. 8 shows pictures from a traffic feed from a different location
on the bridge. The first picture in this set was taken at 5:05pm, and
Figure 8: Traffic feed from a different location on the Rio- Figure 9: Tollboth congestion built up in a span of 30 mins.
Niteroi bridge shows shots from a traffic jam, that lasts for
about an hour.
Fig. 3, we can see that our goal is to decrease (or, stop) the incom-
ing traffic (through links L and Hbef ) when congestion takes place
shows an already congested road. The next five pictures are taken in link Haf t .
over a 45 minute period, and we observe that the congestion re- To accomplish this goal, we first need to monitor the traffic den-
mains until 6:00pm. This example also illustrates the fact that once sity on Haf t; we can use the techniques described in § 4 for this
congestion sets in, the operational capacity of the road reduces, purpose. We also place two emergency signals on L and Hbef a
thereby leading to long periods of congested traffic. Fig. 9 shows little distance before the actual merge point. When there is no con-
the congestion for a tollbooth where a smooth traffic input leads to gestion on Haf t, these signals are green, thereby allowing the free
a heavily congested output. Within a span of 30 minutes we see a flow and merging of traffic. As soon as the traffic density Bl (Haf t )
very high congestion build up. The example verifies that even with begins to approach Bl∗ (Haf t), we need to throttle down the incom-
a controlled flow input, congestion can build up in a relatively short ing traffic.
time. The signals at L and Hbef therefore change to red for certain du-
From our traffic feeds in Mombasa and Nairobi, we have ob- rations to allow the density Bl (Haf t) to reduce. The traffic density
served similar congestion problems that persist over long durations at an epoch t + 1 is related to the that at epoch of time t by the
of time. In summary, we conclude that unless the resources of in- following simple equation:
frastructure are properly managed, congestion will remain a long-
term problem. Bl (t + 1) = Bl (t) + input − output
. This equation is used to determine how much traffic can enter in
6. LOCAL DECONGESTION PROTOCOL an epoch from both L and Hbef ; which determines the amount of
In this section, we build on §3, and design a simple protocol time that the signals can be kept green. In order to calculate how to
that can be used to decongest a small network of links. While the divide the green time between the signals, a simple (yet, effective)
main goal is to demonstrate the fact that intelligent flow control of method is to divide the total permissible green time amongst them
traffic can mitigate congestion in localized settings, it can also be in a ratio that is proportional to their traffic densities (which in turn,
extended to a larger network of roads. We motivate and describe are monitored by sensors). This mechanism guarantees that the link
the protocol through two examples: a) a 2:1 merge, and b) a round- Haf t will never go into congestion collapse.
about which acts as a meeting point for four roads. This approach can be simplified as follows: Consider the com-
posite region of L and Hbef that lie after their emergency signals,
6.1 2:1 Merge combined with a small portion of Haf t that lies after the merge
We re-visit § 3.2 that introduced the 2:1 merge junction. Except point. This region behaves like a junction, and has its own traffic
in rare cases, merges between roads follow the free flow model, curve. By using just one camera, we can estimate the traffic den-
where there is no throttling of the input rate of traffic. However we sity in the entire junction, and use this composite traffic curve to
have seen that this model easily leads to congestion collapse. perform traffic flow control.
To overcome congestion in a 2:1 merge, it is simple to under- This approach can easily be generalized to an n:1 junction in the
stand that we need to throttle the amount of traffic entering it. From same manner. We study the efficacy of this protocol in § 7, where
we show that congestion collapse can be easily prevented.
In this section, we provide an evaluation of our protocol, to demon-
6.2 Four-way Roundabout strate the efficacy of the decongestion protocol in alleviating con-
Similar in spirit to § 6.1, we explain the decongestion of a four- gestion collapse in road traffic networks. The specific sample topolo-
way roundabout. The roundabout has short segments on its in- gies we consider are: a) n:1 merge and b) four-way roundabout.
put/output links as a composite junction j, with its own traffic We evaluate our protocol across these topologies, and show how
curve. Each of the links that leads into the roundabout has an emer- our protocol avoids congestion collapse and provides considerable
gency signal that turns red as a reaction to congestion in the round- enhancements to the operational capacity.
about. When the traffic density of the junction Bl (j) is lower than
Bl∗ (j) − ǫ, all these signals are green. 7.1 Methodology
When the junction nears its congestion threshold density Bl∗ (j), The first step in performing the simulation is identifying the es-
the signals will have a reduced green time, determined by the actual sential parameters, a list of which is given below:
density Bl (j). This green time is divided amongst the input links
in a manner proportional to their traffic densities. As before, this 1. The number of links n in the network. Each link l has a
approach can be generalized to any junction with n input and output traffic curve that is defined by two values: Bl∗ and Cl∗ . The
links. minimum delay dmin a vehicle will face in the link is gov-
erned by the equation: Bl∗ = dmin × Cl∗ . Therefore by
6.3 Localized Congestion Areas specifying dmin and Cl∗ , we define all the characteristics of
The mechanism described above for the 2:1 merge and round- the traffic curve. The delay d that a vehicle will face in the
about cases can be generalized for local congestion areas. Consider link is calculated using dmin and the current traffic density
a localized congestion area which has a small network of roads and Bl . Furthermore, the rest of the traffic curve for the link can
a set of N input/output traffic signal points which dictate the input be approximated from these values.
and output flow out of this localized area. Based on free flow traffic
2. The number of junctions/merges in the network. Each junc-
theory [43], any free flow road segment can be associated with an
tion/merge is identified by which input links flow into it, and
approximate traffic curve as described in earlier examples. Given
which output links flow out of it. This determines the topol-
that the localized area is a high congestion area, we can assume that
ogy of the localized network.
most road segments are used at high capacity even in the presence
of traffic signals. Hence, there is constant flow across different 3. The number of sources (vehicle generation points) in the net-
directions within the localized region. In our local de-congestion work. Each source also defines which link it will send all its
protocol, we explicitly model a localized congestion area with an traffic into.
approximate traffic curve with a tipping point as specified by the
traffic curve. The exact parameters of the traffic curves and the 4. The number of sinks (end-points of vehicle journeys) in the
tipping point can be generated using simulation based analysis of network. Each sink also defines which link it receives all its
the road network topology for different arrival patterns. vehicles from.
Consider a local congestion area A with a traffic curve and a tip-
ping point ∆ and a set of N input /output signals T1 , . . . TN . Our 5. Vehicle routes are determined when a vehicle is generated by
local decongestion protocol works as follows. Consider two small a source. Each source is configured with a (possibly large)
safety thresholds ǫ1 > ǫ2 > 0 with a small separation between two set of pre-determined routes that a vehicle generated by it can
values. Let δ(A, t) represent the traffic density in the congestion take. The simulator verifies that these routes are legal (allow-
area A at time t. Our decongestion protocol uses CCTV camera able by the topology) before starting the simulation, and then
feeds at each of the input and output points to estimate the input randomly chooses one of these routes for every vehicle upon
flow Xi (t) and output flow Yi (t) at each signal i (both are mea- generation.
sured in terms of traffic density). Hence, the system measures the 6. Vehicle generation rates (burst rates) can also be configured
net
P flow into P the congestion area across all the signals at time t as for each of the sources. These rates are set over variable
i X i (t) − i Yi (t) and normalizes the net flow in terms of traf- length epochs of time.
fic density change within the area. Using this measure, the protocol
constantly measures δ(A, t) within the traffic area. If the local den- 7. The exit rate of free-flow links is determined by its traf-
sity is less than ∆ − ǫ1 , then the local decongestion protocol does fic curve and current density. Links feeding into a junc-
not alter traffic signal behavior. The local decongestion protocol tion/merge can emit vehicles at their maximum rate Cl∗ .
triggers when
Once the various parameters are identified, a parameter file for
δ(A, t) > ∆ − ǫ2 the topology is created. The simulator uses this parameter file and
When this event happens, the traffic density is close to the tip- simulates the functioning of the road network. We simulate each
ping point threshold. At this juncture, the local decongestion proto- topology for a period of 40 minutes. In the results presented here,
col provides feedback to the individual signals to increase the "red" each traffic source generates vehicles uniformly within epochs that
signal time and reduce the input flow. The traffic input rate at each last 10 minutes each; the burst rate can vary across epochs.
signal is kept as proportional to the estimated queue size based on The simulator records the time at which a vehicle was generated
the image feeds from the individual signals. The protocol continu- by the source, and the time it was absorbed by the sink. It can
ously provides feedback to increase the "red" signal time until the therefore calculate the travel time, along with the throughput (av-
density drops below ∆ − ǫ1 . The two safety thresholds are used erage number of vehicles that reach their destination) at any point
to emulate a simple hysteresis behavior. In practice, ǫ1 and ǫ2 are of time. In our graphs, we plot the throughput on the Y-axis, and
very small constants. time (in seconds) on the X-axis. The throughput at any second
is the total number of vehicles that have reached their respective
7. EVALUATION destinations in the 60 second period centered around that second.
 80
Queue-based Signal Control (Small burst)
No Coordination (Small burst)
70 Queue-based Signal Control (Large burst)
No Coordination (Large burst)

Average exit rate (cars/minute)

60
80
50 Queue-based Signal Control (Small burst)
No Coordination (Small burst)
70 Queue-based Signal Control (Large burst)
40
No Coordination (Large burst)

Average exit rate (cars/minute)

60
30
50
20
40
10
30
0
0 500 1000 1500 2000 2500
20
Time (s)

10

Figure 10: 2 : 1 Merge with d = 300s. Small bursts can lead to 0

0 500 1000 1500 2000 2500
collapse, which can be prevented using intelligent signaling. Time (s)

Figure 11: 2 : 1 Merge with d = 450s. Small bursts lead to

Each graph plots two scenarios: a) when no signaling (or, local de- collapse, albeit slightly delayed (when compared to d = 300s)
congestion protocol) is used (free-flow model), and b) our protocol because of larger buffer sizes of the destination road. Signaling
(§ 6) is used. prevents congestion collapse.

7.2 n:1 Merge

The evaluation was conducted on this topology to examine the
effectiveness of the proposed signaling system in simple scenarios
in which merging of traffic across multiple links to a single point of
confluence. In this study, we repeat the experiments for n = 2, 4.
For the n:1 merge simulations, we set the the n input links to 80
have d = 160s and Cl∗ = 1.0 cars per second. The output link has Queue-based Signal Control (Small burst)
No Coordination (Small burst)
70 Queue-based Signal Control (Large burst)
Cl∗ = 1.0 cars per second. We repeat the simulation for the output No Coordination (Large burst)

Average exit rate (cars/minute)

60
link having d = 300s and d = 450s. There is a source at each of
50
the input links, and each source generates traffic at a constant rate
40
during the 40 minutes of the experiment. In addition, each source
inserts an additional burst between t = 10min and t = 20min. 30

We repeat the simulation for two different burst values as described. 20

10

7.2.1 2:1 Merge 0

0 500 1000 1500 2000 2500
Each source generates traffic at a constant rate of 30 cars per Time (s)

minute. The additional burst is either 10 cars per minute (large

burst) or 5 cars per minute (small burst). Fig. 10 plots the output Figure 12: 4 : 1 Merge with d = 300s. Small bursts can lead to
rate when the destination link has d = 300s. Congestion collapse collapse, which can be prevented using intelligent signaling.
occurs as soon as the output link density goes above its Bl∗ value,
which takes place about 10 minutes into the simulation. The output
rate continues to decay as the buffer of the output link continues
to operate in the congestion region. However if signaling is used,
then the flow into the output link is controlled by metering the cars
flowing in from the two input links. As a result, the output link
continues to operate at its maximum capacity of 60 cars per minute. 80
Similarly, Fig.11 plots the output rate when the destination link has Queue-based Signal Control (Small burst)
No Coordination (Small burst)
70 Queue-based Signal Control (Large burst)
d = 450s. The behavior is very similar to when d = 300s, with No Coordination (Large burst)
Average exit rate (cars/minute)

60
the difference being that when d = 450s, congestion sets in after a
50
larger delay, because of the larger buffer on the destination link.
40

7.2.2 4:1 Merge 30

Each source generates traffic at a constant rate of 15 cars per 20

minute. The additional burst is either 10 cars per minute (large 10

burst) or 3 cars per minute (small burst). Fig. 12 and 13 plot the 0
0 500 1000 1500 2000 2500
output rate with the destination link has d = 300s and d = 450s Time (s)

respectively. We observe that in both cases, congestion sets in as

soon as the additional burst is issued by the sources, and it lasts Figure 13: 4 : 1 Merge with d = 450s. Small bursts lead to
a long time. Smart signaling can keep the buffer sizes below the collapse, albeit slightly delayed (when compared to d = 300s)
breakdown region, thereby averting congestion. because of larger buffer sizes of the destination road. Signaling
prevents congestion collapse.
7.3 Four-way Roundabout
The roundabout topology can be simplified as follows: There are
8 links L1 · · · L8 , with four of them flowing in to the roundabout,
 80
Queue-based Signal Control (Small burst)
No Coordination (Small burst)
the road traffic networks. Numerous ways have been devised by
70 Queue-based Signal Control (Large burst)
No Coordination (Large burst)
various government authorities like promoting car-pooling systems

Average exit rate (cars/minute)

60 [36], congestion pricing or time variable toll mechanisms [33, 35,
50 34] and car plate quotas [7]. Congestion control mechanisms in-
40 cludes traffic engineering and monitoring. The available real-time
30 information from deployed road sensors is leveraged and a predic-
20 tive analysis is applied to propose near future traffic patterns and
10 develop a situational solution to tackle the current traffic. Nowa-
0 days real-time traffic information is collected from various sources
0 500 1000 1500 2000 2500
Time (s)
like traffic cameras, road sensors and even cellphone providers [29,
18, 13]. Advance GPS systems mitigate this problem by find-
Figure 14: Roundabout with d = 60s: The small buffer size of ing less congested routes [28, 10]. Various mechanisms like vari-
the roundabout leads to congestion quickly taking place, which able speeding, ramp metering and lane specific signaling are ap-
can be prevented using signaling. plied on available real-time information to manage traffic conges-
tion. [41] proposes the usage of histogram based density discov-
80
ery scheme for road connectivity which is leveraged for LOURVE
Queue-based Signal Control (Small burst)
70
No Coordination (Small burst)
Queue-based Signal Control (Large burst)
[40]; LOURVE focuses on global route optimal scheme to calcu-
No Coordination (Large burst) late best paths on its defined overlay network. [42] develops a new
Average exit rate (cars/minute)

60
algorithm FlowScan, to cluster road segments instead of clustering
50
the moving traffic to identify congested traffic routes.
40
Texas Transportation Institute (TTI) [26] has developed several
30
applications that are widely used in local, state and federal levels
20
for controlling the signal timing information [20]. McTrans [17]
10
project at University of Florida provides a host of applications re-
0
0 500 1000 1500 2000 2500 lated to road traffic management. HCS+ [11] and Traffic Network
Time (s)
Study Tool [30] provides signal timing optimization based on a va-
riety of objective functions. Contram [8] is another software which
Figure 15: Roundabout with d = 20s: The extremely small provides real time traffic monitoring and optimized real time traffic
buffer size of the roundabout leads to congestion immediately routes. It is currently deployed in places like Stockholm, Kent and
taking place, which can be prevented using signaling. other large cities. The Sensor Project [24] enables efficient data
collection so that traffic management can be done in a better way,
by the use of sensors to collect the data. There are also applications
and the other four flowing out of it. These links have identical which monitor existing real time traffic and predict traffic mod-
traffic curves, with d = 160s and Cl∗ set to 1.0 cars per second. els which can prevent traffic congestion. Applications like IBM
The roundabout is associated with a traffic curve, with parameters Traffic prediction tool [12] and the combination of Caliper Prod-
Cl∗ = 1.0 cars per second, and the experiment is repeated for d = ucts [6](Transmodeller and TransCAD) are used to simulate and
60s and d = 20s. The four incoming links are each served by a predict traffic models.
source, while the four outgoing links each flow into a sink. Each Use of third party mediums to disseminate traffic information is
source generates traffic at the rate of 15 cars per minute during the being actively pursued. Danilo [45] explores the idea of cellular
entire experiment. Each source generates an additional burst from networks to convey information to users of information. Akinori
t = 10min to t = 20min of 10 cars per minute (large burst) or 3 et al., [37] proposes creation of well formed maps which could be
cars per minute (small burst). used by mobile traffic information services. However, in our case,
From Fig. 14 and 15 we observe that if signaling is not used, the we intend to create monitoring and signaling stations which can
roundabout gets congested very quickly. This congestion collapse take decisions on their own without user intervention and hence,
lasts till the end of the simulation. However if we use smart signal- regulate traffic.
ing to control the flow of traffic from the four input links into the
roundabout, we can keep the roundabout operating at its optimal ca- 9. CONCLUSION AND FUTURE WORK
pacity, thereby avoiding congestion collapse. We verified that this
behavior is repeated with different traffic burst rates, and different Road traffic congestion is a central problem in most developing
traffic curves for the links. In particular, we observe that when the regions. Most urban areas have poorly managed traffic networks
roundabout has a smaller buffer capacity (d = 20s, as illustrated in with several traffic hot-spots or potential congestion areas. In this
Fig.15), even a very small burst over a very small period of time is paper, we study the problem of road traffic congestion in high con-
enough to congest the roundabout. gestion hot-spots in developing regions. We first present a simple
image processing algorithm to estimate traffic density at a hot-spot
using CCTV camera feeds. Based on analysis of traffic images
8. RELATED WORK from live traffic feeds, we show evidence of congestion collapse
The problem of traffic congestion has been prevalent in both de- which last for elongated time periods. Based on the free-flow traf-
veloping and developed countries. Variety of solutions have been fic curve behavior of links, critical road segments when exposed to
developed in the previous decade [23, 19, 25, 27], but as mentioned short bursts in traffic can result in the specific segments operating
in § 2, developing countries suffer from an additional set of con- at low-capacity levels for long time periods. To partially alleviate
straints hampering these predefined solutions making the problem this problem for small congestion areas such as traffic hot-spots, we
tougher. Congestion prevention and congestion control are the two develop a local de-congestion protocol that controls the flow of traf-
approaches to solve the traffic congestion problem. Congestion pre- fic into near-congested regions, thereby preventing collapse caused
vention focuses on reducing the number of incoming vehicles in by short bursts of traffic. Our hope is that localized de-congestion
 mechanisms are potentially easier to deploy in real-world settings [35] Using pricing to reduce traffic congestion. http://www.cbo.gov/
and can enhance the traffic flow at critical hot-spots in road traf- ftpdocs/97xx/doc9750/03-11-CongestionPricing.pdf, 2009.
[36] Dynamic ridesharing:carpooling meets the information age.
fic networks. We believe that this represents only a first step in http://ridesharechoices.scripts.mit.edu/home/
the development of low-cost, deployable strategies for alleviating wp-content/papers/APA\_TPD\_Webinar\_Aug2010.pdf, 2010.
congestion in developing regions. [37] A. Asahara, S. Shimada, and K. Maruyama. Macroscopic structural
The future work lies towards deploying a real time proof of con- summarization of road networks for mobile traffic information services. 7th
International Conference on Mobile Data Management, 2006.
cept to analyze instantaneous traffic density. The paper discussed [38] V. Jain, A. Sharma, A. Dhananjay, and L. Subramanian. Traffic density
a means to detect and curb congestion in a localized setting. Al- estimation for noisy camera sources. In TRB 91st Annual Meeting, Washington
though, the solution is feasible to affect local congestion, it is still D.C., January 2012.
not able to curb the congestion extending for miles due to the local- [39] N. K. Kanhere, S. T. Birchfield, W. A. Sarasua, and T. C. Whitney. Real-time
detection and tracking of vehicle base fronts for measuring traffic counts and
ized focus of the approach. The analysis can thus be improved with speeds on highways. In Transportation Research Record: Journal of the
multiple sequential cameras along a highway which in addition to Transportation Research Board, volume 1993, pages 155–164.
localized congestion control analyzes the congestion buildup from [40] K. Lee, M. Le, J. Harri, and M. Gerla. Louvre: Landmark overlays for urban
the starting point to the ending point. With the aggregate image vehicular routing environments. In Vehicular Technology Conference, 2008.
VTC 2008-Fall. IEEE 68th, pages 1 –5, sept. 2008.
data, the congestion control strategy can make global decisions and [41] K. Lee, J. Zhu, J.-C. Fan, and M. Gerla. Histogram-based density discovery in
affect congestion control on a reasonably sized scale. establishing road connectivity. In Vehicular Networking Conference (VNC),
2009 IEEE, pages 1 –7, oct. 2009.
[42] X. Li, J. Han, J.-G. Lee, and H. Gonzalez. Traffic density-based discovery of
10. REFERENCES hot routes in road networks. In Proceedings of the 10th international conference
[1] 2011 Urban Mobility Report by Texas Transportation Institute. on Advances in spatial and temporal databases, SSTD’07, pages 441–459,
http://mobility.tamu.edu/ums/. Berlin, Heidelberg, 2007. Springer-Verlag.
[2] Beijing’s stalled traffic. http://www.bjreview.com.cn/forum/txt/ [43] H. Lieu. Traffic flow theory. US Department of Transportation, 62, 1999.
2009-04/18/content_191683.htm. [44] C. Ozkurt and F. Camci. Automatic traffic density estimation and vehicle
[3] The Berkeley Highway Laboratory. classification for traffic surveillance systems using neural networks. In
http://www.its.berkeley.edu/bhl/. Mathematical and Computational Applications, volume 14, pages 187–196.
[4] Brazil rio traffic feeds. [45] D. Valerio, A. D. Alconzo, F. Ricciato, and W. Wiedermann. Exploiting cellular
http://www.rio.rj.gov.br/web/riotransito. networks for road traffic estimation: a survey and a research roadmap.
[5] California Partners for Advanced Transit and Highway. [46] H. Yang and S. Yagar. Traffic assignment and signal control in saturated road
http://www.path.berkeley.edu/. networks. Transpn. Res.-A, 29A:2:125–139, 1995.
[6] Caliper. http://www.caliper.com/.
[7] China’s car plate quota. http://chinaautoweb.com/2010/12/
to-tackle-traffic-jam-beijing-sets-new-car-plate-quota-limits-out-of-towners/.
[8] Contram. http://www.contram.com/.
[9] Dynamit. http://mit.edu/its/dynamit.html.
[10] Garmin. http://www.garmin.com/.
[11] Highway Capacity SoftwareTM .
http://mctrans.ce.ufl.edu/hcs/downloads/index.htm.
[12] Ibm traffic prediction tool.
http://domino.research.ibm.com/comm/research.nsf/
pages/r.statistics.innovation.traffic.html/.
[13] Inrix. http://www.inrix.com/.
[14] Intelligent Transportation Systems Decision Support System. http:
//www.path.berkeley.edu/PATH/Research/projects.html.
[15] Inverse distance weighted interpolation.
http://www.ems-i.com/smshelp/Data_Module/
Interpolation/Inverse_Distance_Weighted.htm.
[16] Kenya traffic feeds. http://home.co.ke.
[17] McTrans: Moving Technology. http://mctrans.ce.ufl.edu/.
[18] Navteq. http://www.navteq.com/.
[19] Octo telematics. http://www.octotelematics.com/solutions/
traffic-management.
[20] PASSERTM : Sigmal Timing Optimization Software.
http://ttisoftware.tamu.edu/.
[21] Quadstone Paramics. http://www.paramics-online.com/.
[22] Rio-niteroi bridge feeds. mms://cameras.ponte.com.br/pontelive.
[23] Road traffic solutions. http://www.roadtrafficsolutions.com/.
[24] Sensor project. http://www.transport-research.info/web/
projects/project_details.cfm?id=4172&page=outline/.
[25] Sitraffic concert. http://www.itssiemens.com/.
[26] Texas Transportation Institute. http://www.tti.org/.
[27] Thales traffic management solution. http://www.thalesgroup.com/
Portfolio/Security/Road_traffic_management/?pid=1568.
[28] Tomtom. http://www.tomtom.com/livetraffic/.
[29] Traffic land. http://www.trafficland.com/.
[30] TRANSYT-7F.
http://mctrans.ce.ufl.edu/featured/TRANSYT-7F/.
[31] Vii california. http://viicalifornia.org/.
[32] The world’s worst traffic jams. http://www.time.com/time/world/
article/0,8599,1733872,00.html.
[33] Variable tolls in france. www.trb.org/Conferences/RoadPricing/
Presentations/Lecoffre.ppt, 2003.
[34] Driving change in stockholm. http:
//www.ibm.com/podcasts/howitworks/040207/index.shtml,
2008.

View publication stats