Personal and Ubiquitous Computing

https://doi.org/10.1007/s00779-020-01434-y

ORIGINAL ARTICLE

Examining queue-jumping phenomenon in heterogeneous traffic

stream at signalized intersection using UAV-based data
Afzal Ahmed 1 & Fatma Outay 2 & Syeda Ofaq Raza Zaidi 1 & Muhammad Adnan 3 & Dong Ngoduy 4

Received: 25 May 2020 / Accepted: 27 July 2020

# Springer-Verlag London Ltd., part of Springer Nature 2020

Abstract
This research presents an in-depth microscopic analysis of heterogeneous and undisciplined traffic at the signalized intersection.
Traffic data extracted from the video recorded using an unmanned aerial vehicle (UAV) at an approach of a signalized intersec-
tion is analyzed to study the within green time dynamics of traffic flow. Various parameters of Wiedemann 74, Wiedemann 99,
and lateral behavior models used in microscopic traffic simulation package, Vissim, are calibrated for the local heterogeneous
traffic. This research is aimed at exploring the queue-jumping phenomenon of motorbikes at signalized intersections and its
impact on the saturation flow rate, travel time, and delay. The study of within green time flow dynamics shows that the flow of
traffic within green time is not uniform. Surprisingly, the results indicate that the traffic flow for the first few seconds of the green
time is significantly higher than the remaining period of green time, which shows a contradiction to the fact that traffic flow for
the first few seconds is lower due to accelerating vehicles. Mode-wise traffic counted per second shows that this anomaly is
attributed to the presence of motorbikes in front of the queue. Consequently, the outputs of simulation results obtained from
calibrated Vissim show that the simulated travel time for motorbikes is significantly lower than the field-observed travel times
even though the average simulated traffic flow matches accurately with the field-observed traffic flow. The findings of this
research highlight the need to incorporate the queue-jumping behavior of motorbikes in the microsimulation packages to enhance
their capability to model heterogeneous and undisciplined traffic.

Keywords Vissim . Calibration . Microscopic simulation . Saturation flow rate . Karachi, Pakistan . Lane discipline

1 Introduction with weak or no lane discipline. Vehicles in such undisci-

plined traffic streams have different static and dynamic char-
Traffic flow modeling is widely used in the planning, design, acteristics resulting in higher heterogeneity in streams [4, 5].
and operations of transportation systems. Various traffic flow Modeling the traffic flow of heterogeneous and undisciplined
models and simulation packages have been developed so far. traffic streams is a complex procedure and still an unresolved
These mathematical models and simulation packages are de- problem.
signed to replicate the actual traffic flow phenomenon in a Numerous research studies have highlighted the nonstan-
controlled environment [1, 2]. Most of these models and sim- dard behavior of heterogonous traffic with poor lane disci-
ulation packages are developed for traffic streams with strict pline. Recently, Ahmed et al. [6] studied the characteristics
lane discipline and speed limitations [3]. However, most of the of heterogeneous and undisciplined traffic streams in Karachi,
South Asian and developing countries have traffic streams Pakistan. The study proposed a width-based cell transmission
model for heterogeneous traffic based on the observation that
the traffic flow of undisciplined streams is a function of the
* Muhammad Adnan width of the road rather than the number of lanes on the road.
[email protected]
Asaithambi and Basheer [7] studied the effect of weak lane
discipline in Chennai, India, and concluded that the lateral and
1
NED University of Engineering and Technology, Karachi, Pakistan longitudinal gaps between vehicles depend upon the class of
2
Zayed University, Dubai, United Arab Emirates vehicles. Nokandeh et al. [8] studied the effect of lane width
3
Hasselt University, Hasselt, Belgium on passenger car units (PCU) on an intercity highway in India.
4
Monash University, Melbourne, Australia
This study concluded that the lane widths highly affect the
 Pers Ubiquit Comput

driving behaviors resulting in different PCU for different lane macroscopic simulation of Vissim for the homogeneous envi-
widths. Mallikarjuna et al. [9] studied the behavior of hetero- ronment. Matsuhashi et al. [25] simulated the traffic flow in
geneous traffic streams and proposed a method for extracting Ho Chi Minh City, Vietnam, using image processing and
lateral gaps in heterogeneous traffic streams. Ahmed et al. [10] Vissim. The study concluded that the presence of motorcycles
compared the travel time predicted through Google Live in the traffic stream interferes with other vehicles resulting in
Traffic in Karachi, Pakistan, with actual travel time collected the reduction of speed of the entire traffic stream. Zhang et al.
through probe car and probe motorbike. The study concluded [26] evaluated the feedback-based tolling system using
that there is a significant difference between travel time of Vissim. Park and Kwak [27] adopted a multi-criteria approach
motorbike and travel time estimated by Google Live Traffic, in TRANSIMS calibration. Traffic count and travel time
indicating that the behavior of motorbikes and cars is quite values obtained from the field were compared with the values
different in heterogeneous traffic streams. Kays et al. [11] from the output of microsimulation through TRANSIMS. The
studied the heterogeneous traffic of Dhaka and developed a results showed that the proposed approach accurately calibrat-
flow-density fundamental traffic flow diagram as input for ed the microsimulation package to represent actual field
Heterogeneous Cell Transmission Model (HCTM). Most of conditions.
the reported literature available on the modeling of heteroge- PTV Vissim is used for the microsimulation of selected
neous traffic is based on modeling of traffic flow at midblock, intersection in this research [28, 29]. Vissim is a user-
and very limited research is available for modeling of hetero- friendly microscopic behavior-based simulation package with
geneous traffic at signalized intersections. the ability to model vehicular as well as pedestrian move-
Signalized intersections are considered one of the most ments. Vissim is based on the psychophysical car-following
important traffic control measures in urban areas. The basic model proposed by Wiedemann. Lateral behavior parameters
parameters for the operation and design of signalized intersec- and lane changing parameters are also included in Vissim to
tions include saturation flow, headway, lost time, and traffic simulate the traffic stream with weak lane discipline [30–33].
composition [12–14]. Modeling of heterogeneous traffic flow, The existing literature contains numerous studies on the
especially at the signalized intersection, is more complex than microsimulation of heterogeneous traffic using Vissim.
the standard traffic streams. Microsimulation is considered an Various systematic methods have been proposed by re-
important and useful tool for modeling of heterogeneous traf- searchers for calibration of simulation parameters, which in-
fic. However, most of the simulation packages have been pro- clude trial and error methods, iterative estimation of parame-
duced for traffic in developed countries with proper lane dis- ters, and tuning of car-following and other driving behavior
cipline and lesser vehicle heterogeneity. Therefore, these sim- parameters. Also, various modifications like genetic algo-
ulation models need to be calibrated according to local condi- rithms, latent cubes, and image processing techniques were
tions to simulate heterogeneous traffic [15, 16]. The unsys- introduced for calibration to make the procedure of optimiza-
tematic calibration of simulation packages may result in large tion hassle-free and convenient. This research estimates the
errors [17]. Recently, several studies have proposed methods calibrated values of Vissim parameters using the analysis of
for systematic calibration of simulation packages for hetero- video recorded using an unmanned aerial vehicle (UAV).
geneous traffic [18]. Microscopic parameters such as lateral and longitudinal gaps
Microscopic traffic simulation packages have been widely at different speeds are essential inputs in modeling the behav-
used to model and simulate the flow of traffic on transporta- ior of drivers, which can be accurately estimated using aerial
tion networks. Hossain [19] calibrated well-known simulation visuals taken at the bird’s-eye view through a UAV.
package PTV Vissim for heterogeneous traffic of Dhaka, In contrast to other studies on Vissim calibration, which
Bangladesh, using field-collected data. Schultz and Rilett only compare average traffic flow or travel time during the
[20] proposed a method to calibrate motor vehicle distribution simulation period, this paper performs an in-depth analysis of
in Corridor Simulation (CORSIM). Milam and Choa [21] cal- the simulated results of the calibrated Vissim model. The pres-
ibrated the CORSIM package by proposing a genetic ence of motorbikes in the traffic stream, which can maneuver
algorithm-based parameter optimization technique. Menneni through spaces between cars, especially at signalized intersec-
et al. [22] developed a genetic algorithm for Vissim calibra- tions, makes the modeling phenomenon complex. Although
tion based on field-collected speed and flow data. Mathew and Vissim now includes many options for heterogeneous traffic
Radhakrishnan [23] proposed a general methodology to cali- with poor lane discipline, the queue-jumping phenomenon of
brate the microsimulation package using optimization tech- motorbikes is still unaddressed, which causes significant de-
nique by genetic algorithm for heterogeneous traffic in India. viation of the simulation results from the field-observed data.
A sensitivity analysis was conducted to select the parameters This research paper performs a detailed analysis of videos
for calibration. Both the car-following models Wiedemann 74 recorded using a UAV at a selected signalized intersection in
and Wiedemann 99 were calibrated at three intersections. Karachi. The following contributions are claimed in the
Fellendorf and Vortisch [24] validated the microscopic and existing literature through the present study:
Pers Ubiquit Comput

i. Within green time dynamics of traffic flow are studied. of selected intersection due to the amount of time required for
ii. The queue-jumping phenomenon of motorbikes is identi- manual traffic data extraction from the recorded video.
fied, and its effect on saturation flow rates and travel time is The video of the traffic stream at Johar Chowrangi was
studied. recorded from Tuesday 17 September 2019 to Thursday 19
iii. The behavior parameters of Vissim are calibrated deter- September 2019, which includes different peak and off-peak
ministically using actual field data obtained from a UAV. timings during the day. The continuous recording of the traffic
iv. The weakness of Vissim in modeling the queue-jumping stream was not possible due to the limitation of the flight time
phenomenon of motorbikes is highlighted, and its impact of the UAV. Several backup batteries were used during the
on simulation accuracy is quantified. video recording to ensure maximum recording during the field
survey. In total, traffic videos for about six hours were record-
ed using UAV, which contains 175 signal cycles. Figure 1
shows the selected approach for the data collection along with
2 Data collection and extraction a visual taken from the drone.

2.1 Video recording using UAV 2.2 Extraction of vehicular traffic flow

The accurate extraction of microscopic parameters such as There is a lack of accurate tools to automatically measure
longitudinal and lateral gaps requires georeferenced images traffic flow from the video due to heterogeneity and weak lane
at the bird’s-eye view, for which UAV is the best-suited discipline of traffic. Therefore, traffic flow and all other data
equipment. UAV has now been recognized as a useful tool were extracted manually from the recorded video. This re-
in real-time and offline traffic data collection [34–36]. DJI search evaluates within green time dynamics, for which the
PHANTOM 3 UAV was used in this research to record video time stamp at which a vehicle crosses the stop line was need-
of the traffic stream. PHANTOM 3 can record video at a ed. These time stamps can be aggregated to determine the
resolution of 4 k with a flight time of 15 min. Traffic data mode-wise flow per unit time (1 s, 2 s, 5 s, etc.). For this
was collected at a moderately busy signalized intersection purpose, a special C++ program was designed, which records
located in Karachi, known as Johar Chowrangi. This intersec- strokes of assigned keys on the keyboard along with the pre-
tion was selected after conducting a preliminary survey of cise time stamp. Six different keys were assigned to the vehi-
most of the signalized intersections in Karachi. The following cles classified in six different modes (car, bus, bike, HGVs,
criteria were established for the selection of the intersection: pickups, and rickshaw). Each mode was assigned to a differ-
ent research assistant, and the time stamps of all the vehicles
i. No bus stop within 200 ft. of the approach of the intersec- were recorded as the vehicles cross the stop line. The program
tion, as the busses stopping may have an impact on flow generates a log file and provides the count of the selected
rate and travel time of other vehicles. modes of vehicles for the unit of the time supplied as input
ii. No legal/ illegal parking at the upstream and downstream (e.g., 1 s, 2 s, or 5 s).
of the approach within 200 ft. of the intersection. The traffic stream at the selected intersection is representa-
iii. Well-channelized intersection with exclusive left and tive of the heterogeneous traffic observed in Karachi. In total,
right turning movements, physically separated from the 44,525 vehicles were counted of which 19,589 (44%) were
through movement. The proportion of left and right turn- motorbikes, 16,029 (36%) were cars, 447 (1%) were trucks,
ing vehicles is another factor that affects the flow and 1337 (3%) were loading vehicles, 6670 (15%) were rickshaws
speed of through vehicles. and chingchi, and 453 (1%) were busses. The traffic mix at
iv. Uniform width of the approach. The approach with var- this intersection is similar to the traffic mix reported in other
iable width could have an impact on flow rates and travel studies from Karachi and also confirms with the vehicle reg-
time. istration data of Karachi [6, 37, 38].

Johar Chowrangi was one of the few intersections meeting 2.3 Average travel time and speed
the above-established criteria, as most of the signalized inter-
sections do not have provision for exclusive left and right The drone was operated in such a way that the traffic stream
turning movements. Johar Chowrangi is a 4-leg signalized within 100 ft. of the stop line was visible in the recording. The
intersection, and only one of the approaches (from travel time of every 5th vehicle of the selected mode was
University Road to Johar Chowrangi) was selected to perform measured to travel the tarp length of 100 ft. The time stamp
detailed analysis. The selected approach is a 40 ft. wide arte- of the vehicle as it enters the segment was also recorded along
rial with exclusive well-channelized left turning and right with travel time. The recorded travel time was used to estimate
turning movements. The study is only limited to one approach the speeds of vehicles.
 Pers Ubiquit Comput

Fig. 1 Google image of the selected intersection and snapshot from recorded video

3 Calibration of Vissim driving behavior of commuters. Once these parameters are

provided as input, a simulation run needs to be carried out to
Calibration of a simulation model for local conditions needs analyze the output. For the simulation model of this research,
representation of heterogeneous traffic, identification of sen- the performance measures considered were traffic flow, travel
sitive parameters, and calibration of identified sensitive pa- time, and delays.
rameters [39, 40]. The accurate calibration of the simulation
model requires data for intersection geometry, signal phasing,
3.2.1 Identification of sensitive parameters
vehicle characteristics, traffic composition, and traffic de-
mand. The accuracy of a calibrated model is measured
For the calibration of Vissim parameters, it is necessary to
through the comparison of the values of various simulated
identify the parameters which have a significant effect on
variables with the field-measured values of those variables
the simulation of traffic flow [41–44]. The developed model
such as travel time, delay, speed, and average traffic flow.
of the intersection in Vissim was initially simulated with the
The framework of the calibration of Vissim is described in
default parameter values. As expected, the default parameters
this section.
resulted in the simulation result deviating significantly from
field-measured values. The simulated traffic flow and delay
3.1 Development of simulation model in Vissim using default parameter values are compared with the field-
measured values. The results of the comparison of simulated
For the development of the simulation model in Vissim, four values using default values of Wiedemann 74 and Wiedemann
main characteristics are considered, which include intersec- 99 are shown in Table 2.
tion geometry, vehicular characteristics, driving behavior, For sensitivity analysis, the default value of each parameter
and signal characteristics. For vehicle representation, Vissim was gradually changed up to ± 50%. Each parameter was an-
has a wide range of vehicles available in the database, includ- alyzed individually by keeping all other parameters constant at
ing car, bike, bus, tram, and trucks. However, heterogeneous the default value. The identified parameters with significant
traffic streams in Karachi also comprise rickshaw/chingchi effect include standstill distance, additive part of safety, and
and pickups. For the representation of rickshaws/chingchi multiplicative part of safety. For Wiedemann 99 model, the
and pickup, 3D models for rickshaw and pickup were sensitive parameters include CC1, CC2, and CC3. The sensi-
imported in Vissim, and its static and dynamic characteristics tive parameters for lateral behavior include minimum longitu-
were adjusted to represent actual field conditions, such as dinal speed, the lateral distance at 0 km/h, and 50 km/h.
desired speed, acceleration, and deceleration. Vehicle charac-
teristics modified in Vissim are listed in Table 1.
3.2.2 Average standstill distance or CC0

3.2 Calibration of Vissim parameters Average standstill distance is the distance that the following
vehicle desires to maintain from the leading vehicle when both
The process of calibration involves changing the characteris- the vehicles are in a stationary state. It is the distance between
tics, including speed distributions, safety distances, and the rear bumper of the leading vehicle and the front bumper of
Pers Ubiquit Comput

Table 1 Vehicle characteristics

and traffic composition Vehicle types Length (m) Breadth (m) Desired speed (km/h) Traffic composition (%)

Car 4.00 1.60 50 36

Bus 11.47 2.50 35 1
Motorbike 1.80 0.60 55 44
HGV 12.3 2.60 35 1
Pickup 4.70 1.69 40 3
Rickshaw/Chingchi 2.60 1.40 45 15

the following vehicle. CC0 in Wiedemann 99 also represents 0.74 s, which is 17.7% less than the default value of 0.9 s. The
the same behavior parameter of commuters. The default value default value of CC1 is based on the average behavior of
for the average standstill distance is 2 m in Vissim. The commuters, including midblock, while the values calculated
uniqueness of this study is that it uses the video recorded from in this research are measured during the green time at the
a UAV, which records the traffic stream at an angle perpen- signalized intersection.
dicular to the direction of traffic flow. The visuals taken at
bird’s-eye view yield more accurate values of gaps between
vehicles in comparison with video recorded at an angle. 3.2.4 CC2—following variation
Around 500 frames were extracted from the recorded vid-
eo, covering different traffic conditions. To measure standstill The following variation is the difference between maximum
distance between vehicles, the frames were extracted when the following distance and minimum following distance between
traffic signal was red. The extracted images from the videos two following vehicles. It is an additional distance other than
were georeferenced in ArcGIS. The standstill distance be- the safety distance that a driver wants to maintain from the
tween all the vehicles stopped at the signal was plotted in leading vehicle. Values for the following variations were cal-
ArcGIS. In total, 16,250 observations of standstill distance culated in different traffic conditions for 1000 vehicles, and
were taken. The average measured standstill distance was the difference between the maximum and minimum following
0.6 m, which is more than three times lesser than the standard variations was provided as an input for CC2. The average
values of this parameter. This also highlights the aggressive observed value of CC2 was 2.75 m, which is 31% smaller
behavior of drivers. than the default value of 4 m for the following variation.

3.2.3 CC1—headway time

3.2.5 Additive and multiplicative part of the safety distance
Headway time (CC1) is the distance that a driver desires to
maintain at a certain speed from the leading vehicle measured The additive and multiplicative parts of safety distance are
as time gap in seconds. For calculation of headway time, a nondeterministic values and cannot be calculated from field
datum point was considered on the selected approach, and the data. Therefore, after calibrating all other parameters, the
time difference between two vehicles while passing through model was simulated using the calibrated values of other pa-
the datum point was taken as headway time. One thousand rameters. The values of additive and multiplicative parts were
vehicles were observed for the extraction of CC1 in different calibrated iteratively until the minimum difference between
traffic conditions. The average observed headway time was observed and simulated flows was achieved.

Table 2 Comparison of simulated results using default parameters values of Wiedemann 74 and Wiedemann 99

Time Actual Simulated outflow Error Simulated outflow Error Actual Simulated travel Error Simulated travel Error
intervals outflow using default (%) using default (%) travel time using default (%) time using default (%)
volume Wiedemann 74 Wiedemann 99 time Wiedemann 74 Wiedemann 99
(veh/h) (veh/h) (veh/h) (sec) (sec) (sec)

0–900 1100 835 24 806 26.7 59 96 62.7 126 113.5

900–1800 1030 750 27 715 30.5
1800–2700 1144 902 21 846 21
2700–3600 1032 755 26.8 694 32.7
 Pers Ubiquit Comput

3.2.6 Lateral distance at 0 km/h and 50 km/h Radhakrishnan [23], Siddharth and Ramadurai [46], and
Arkatkar et al. [47] for different cities of India, this research
The calibration of lateral distance is significant in representing estimates most of the sensitive parameters. Furthermore, the
the weak or no lane discipline, as observed in Karachi and calibrated values are different from each other in different
many other similar traffic streams. By default, the values of cities. The difference in calibrated values could be attributed
lateral distances at the speed of 0 km/h and 50 km/h are the to either iterative estimation of the parameters or the type of
same (1 m), which shows that vehicles keep the same lateral data used in these studies. The iterative estimation of behavior
distance when the lane is followed. Whereas when the lane parameters may result in significantly different values than the
discipline is weak, and the lanes are not followed, then the actual values, as the objective function tries to find values of
lateral distance may change with the change in speed. The use behavior parameters that minimize the error of simulated flow
of images taken from a drone increases the accuracy of the or travel time.
lateral distance estimation.
The lateral distances between adjacent vehicles at 0 km/h
were calculated when vehicles were in a stationary condition 4 Results and discussion
during the red time at the signalized intersection. Five hundred
frames were analyzed for extraction of lateral distance at One of the main contributions of this research is to study the
0 km/h. For lateral distance at 50 km/h, 500 frames were within green time dynamics of traffic flow. The results of the
extracted when the speed of vehicles was 50 km/h. The aver- calibrated model and analysis of flow dynamics within green
age lateral gap at 0 km/h was observed as 0.5 m, which is 50% time are discussed in this section.
lesser than the default value of 1 m. This value also shows that
vehicles ignore lane and stand close to each other, which is a 4.1 Traffic flow dynamics within green time
commonly observed behavior in such nonstandard traffic
streams. Similarly, the lateral distance at 50 km/h was estimat- The higher heterogeneity of traffic streams in many South
ed to be 0.8 m, which is also 20% lower than the standard Asian countries is a result of the higher number of motorbikes,
value of 1 m. which behave differently in the traffic stream. This research
focuses on highlighting the different behavior of the hetero-
3.2.7 Minimum longitudinal speed geneous and undisciplined traffic stream so that the modeling
and simulation of such traffic streams can be improved.
The minimum longitudinal speed is one of the important pa- To understand the within green time dynamics, the video
rameters which represent the lateral movement of recordings for the Johar Chowrangi intersection were ana-
vehicles when the vehicles come to a halt. For calculation of lyzed for traffic flow dynamics during the green time at the
minimum longitudinal speed in heterogeneous traffic, the highest possible resolution. The time stamp of each vehicle in
speed of every 5th vehicle in traffic stream was calculated, sub-second precision was recorded as it crosses the stop line
and the minimum value was taken as the calibrated value. In by six observers, counting each mode simultaneously. A spe-
total, 500 observations of longitudinal speed were taken out of cial computer program was developed, which can determine
which minimum speed of 2.5 km/h was taken as an input for the cumulative flows in the desired time intervals (e.g., 1 s, 2 s,
Vissim. 5 s) using the recorded time stamps of passing vehicles.
Figures 2, 3, 4, 5, 6, 7, 8, 9, and 10 show within green time
3.3 Comparative analysis of calibrated parameters dynamics by plotting measured traffic flows against the green
time (30 s) of the selected approach for selected cycles (30 cy-
Table 3 compares the values of calibrated parameters in this cles). Traffic flows cumulated per second during the green
study with the default values of Vissim parameters. The cali- time for the selected cycles are shown in Fig. 2. The selected
brated values show a significant difference with the standard approach of the signalized intersection is a four-lane
values, indicating that the driving behavior of drivers in local approach with a total width of 40 ft. Figure 2 shows a similar
traffic should be properly represented to improve the accuracy trend of traffic flow within each cycle. The highest flow is
of the simulation model. observed within the first few seconds of each green period,
The comparison of the calibrated parameters of Vissim for which then gradually slows down. The highest per second
local heterogeneous traffic with other studies on Vissim cali- flow is equivalent to 25,200 veh/h or 6300 veh/h/ln. A higher
bration is shown in Table 4. The values of calibrated param- fluctuation is observed when the traffic is counted per second
eters by Buck et al. [45] for traffic in Germany are significant- during the green time.
ly different from the values estimated in this study, and other To reduce the higher fluctuation, as observed in Fig. 2, the
studies reported from India for heterogeneous traffic. In com- flows were cumulated every 2 sec during the green time in
parison with the calibration studies conducted by Mathew and Fig. 3. Figure 3 shows that the fluctuation of traffic
Pers Ubiquit Comput

Table 3 Calibrated values of

Vissim parameters Parameter Default value Calibrated value

Average standstill distance 2.00 m 0.6 m

Additive part of the safety distance 2.00 0.10
Multiplicative part of the safety distance 3.00 0.35
Minimum longitudinal speed 3.60 km/h 2.5 km/h
Lateral distance at 0 km/h 1.00 m 0.50
Lateral distance at 50 km /h 1.00 m 0.80
CC0 1.50 m 0.6 m
CC1 0.90 s 0.74 s
CC2 4.00 m 2.75 m

significantly reduces when traffic is counted per 2 sec instead 3600 veh/h/ln. Figure 7 shows the motorbike flows cumulated
of 1 sec. The highest flow for a 2-sec interval is equivalent to for 5 s. The highest flow for 5-s motorbike count is
19,800 veh/h/ln. Figure 4 shows the vehicular flow cumulated 12,240 veh/h or 3060 veh/h/ln.
at 5 s to reduce fluctuation and to get a better understanding of The variation in the flow of cars during green time is shown
green time flow dynamics. The highest flow for 5 s is ob- in Figs. 8, 9, and 10. It can be observed from Figs. 8, 9, and 10
served as 16,560 veh/h or 4140 veh/h/ln. that flows for cars are lesser during the initial 5 s and the
Mode-wise traffic flow of two major transportation modes maximum number of cars for approximately all studied cycles
(cars and motorbikes) was explored in detail to investigate the flow after the initial 5 s. Figure 8 shows the flow of cars during
reason for higher traffic flows shown in Figs. 2, 3, and 4. each second of green time, and it reveals that the highest flow
Figure 5 shows that the highest flow for motorbikes is ob- of cars for almost all the studied cycles is observed after the
served within the first 5 s of the green time. The comparison first 5 s of green time. The 5-s lag in attaining maximum flow
of Fig. 5 and Fig. 2 indicates that the peak flow within the is due to the slow acceleration capabilities of cars in compar-
green time observed in Fig. 2 is attributed to the flow of mo- ison with motorbikes. The highest flow for cumulative per
torbikes. However, it is observed that flow for motorbikes for second flow is observed as 14,400 veh/h or 3600 veh/h/ln.
most of the observed cycles rapidly decreases after the initial This is still a significantly higher number than the standard
5 s of green time. This indicates that motorbikes do not follow saturation flow rate, which is taken around 1800–2000 veh/h/
FIFO (first-in-first-out) rule, and motorbike riders tend to ln. The significantly higher flow of cars is due to the lane
jump the queue using the available spaces between other big- behavior of drivers, as 6–7 lanes of cars can be observed in
ger vehicles. Motorbikes are small in size and can easily ma- reality on a four-lane road segment. Similarly, Fig. 9 and
neuver through the gaps to jump the queue of vehicles. This is Fig. 10 show the cumulative car flow for 2 s and 5 s, respec-
a common phenomenon observed at traffic signals. The mo- tively. The highest flow of cars aggregated at 5 s is observed
torbike riders wait at or beyond the stop line to move quickly as 10,080 veh/h or 2520 veh/h/ln.
once the red signal changes to green. Figure 6 shows the Figures 2, 3, 4, 5, 6, 7, 8, 9, and 10 show within green time
motorbike flows aggregated for 2 s. The highest flow for a dynamics of traffic flow aggregated for different time intervals
2-sec motorbike count is equivalent to 14,400 veh/h or for cars and motorbikes for selected signal cycles. In total,

Table 4 Comparison of calibrated parameters with other studies

Location Calibrated parameters

CC0 CC1 CC2 Standstill Additive part of Multiplicative part of LC at 0 km/h LC at 50 km/h Minimum longitudinal
distance safety distance safety distance speed km/h

Berlin and Karlsruhe, 2.5 1.3 – 2.1 – – – – –

Germany [45]
Jaipur, India [23] 0.8 0.8 0.32 1.3 0.28 0.16 0.30 0.42 –
Chennai, India [46] – – – 1.0 0.20 0.78 0.62 – –
Delhi, India [47] – – – 1.35 0.25 0.35 – – –
Karachi, Pakistan 0.6 0.74 2.75 0.6 0.10 0.35 0.50 0.80 0.80
(This research)
 Pers Ubiquit Comput

Fig. 2 Flow variation per second during the green time

traffic videos for about six hours were recorded using UAV, of green time with an average flow rate of 8000 veh/h, which
which contains 175 signal cycles. The mode-wise traffic flow further increases until the fifth second and then start decreas-
averaged for each second of all the cycles is shown in Fig. 11, ing. On the contrary, the flow of cars is almost negligible
which clearly shows the flow dynamics within green time. during the first 4 sec of green time, and it starts increasing
The flow of motorbikes can be observed from the first second from the fifth second after the signal turns green. The

Fig. 3 Flow variation during green time cumulated for 2 s

Pers Ubiquit Comput

Fig. 4 Flow variation per 5 sec during the green time

maximum flow of cars is observed between 7 and 10 s at aggregated at 5 s show a smooth variation in comparison with
around 10,000 veh/h after which the flow starts decreasing. the average per second flows shown in Fig. 11. Figure 12
Other modes of transport, such as rickshaws, pickups, trucks, shows that the average flow of cars during the first 5 sec
and busses, have a significantly lower flow during the first few was 984 veh/h, whereas the average flow of motorbikes was
seconds of the green time. The highest flow of rickshaws is 10,464 veh/h during the same period. The highest flow of cars
observed at the 17th second. Other modes show a random was observed during the second interval (5–10 s) of green
flow throughout the green time. The average flow during time at 8208 veh/h. The exceptionally higher values of flow
green time aggregated at 5 s is shown in Fig. 12. The flows during the first 10 s are recorded as the vehicles stop close to

Fig. 5 Motorbike flow variation per second during the green time
 Pers Ubiquit Comput

Fig. 6 Motorbike flow variation per 2 s during the green time

each other. This behavior is also reflected through the small 4.3 Actual and simulated traffic flow comparison
values of lateral and longitudinal gaps presented earlier. After
the queue dissipation within the first half of green time, the The simulation results showed that the difference between the
flow of cars and motorbikes reduces, whereas the flow of simulation volume and actual volume for 15 min is within the
trucks and busses is observed at a similar rate throughout the acceptable limits. However, the comparison of actual and sim-
green time. ulated per second flows shows a significant difference.
Figure 13 shows the comparison of per second flows obtained
4.2 Validation of the calibrated simulation model from Vissim and actual field measurements. The mean square
error between the Vissim and actual field flow is 4.64. This
The model was validated using the calibrated parameters for difference in per second flow is due to the rapid discharge of
the traffic data from a different time at the same intersection. motorbike in the initial few seconds of green time of the sig-
The results for validation at the selected intersection for both nal, while the queue discharge rate for motorbikes and cars are
Wiedemann 74 and Wiedemann 99 model are shown in similar in the simulated results obtained from Vissim.
Table 5. When the flow is aggregated for a cycle of 15-min interval
The results of the simulation with calibrated Wiedemann (as shown in Table 5), the overall simulated traffic flows show
74 parameters show that there is a very slight difference be- a good correlation with the field-measured flows. However,
tween average field-measured flows and the simulated flow. this difference in simulated flows at the micro level may sig-
Results for simulated flows by Wiedemann 99 are also accept- nificantly affect the travel times, which is also reflected from
able. Most of the studies on calibration only compare the the results presented in Table 5.
simulated and actual traffic flows.
Results in Table 5 show that despite reasonably reproduc- 4.4 Actual and simulated travel time comparison
ing traffic flows, the simulated travel times are still different
than the field-observed travel time. To investigate the reason The observation from recorded videos showed that motor-
of this difference in travel time, the simulation results were bikes in the traffic stream tend to acquire the position near to
analyzed in detail and compared with the field-observed the stop line irrespective of their arrival time in the queue. The
values in the following subsections. motorbike riders jump through the queue, despite arriving late
Pers Ubiquit Comput

Fig. 7 Motorbike flow variation per 5 s during the green time

Fig. 8 The flow of cars per second during the green time
 Pers Ubiquit Comput

Fig. 9 The flow of cars aggregated for 2 s during the green time

in the queue to wait near the stop line.This behavior of mo- traffic stream. However, it was found that Vissim does not
torbike riders is exclusive for heterogeneous traffic and is also simulate this behavior of motorbike riders and the queue po-
observed in other cities and countries with a similar type of sition of motorbikes in Vissim is dependent upon their arrival

Fig. 10 The flow of cars aggregated for 5 s during the green time
Pers Ubiquit Comput

Fig. 11 Average of mode wise per second traffic flows

time. For testing this hypothesis, classified travel times of manually with a stopwatch of 0.01 s accuracy. Travel times
motorbikes and cars were extracted from field video recording for 768 motorbikes and 580 cars were extracted. Travel time
and recording of the Vissim simulation. for cars and motorbikes were averaged for each signal cycle.
The recorded videos were analyzed, and travel time for Similarly, average travel time for cars and motorbikes were
every 5th motorbike and every 5th car were extracted extracted for each cycle from Vissim. Figure 14 shows a

Fig. 12 Average of mode-wise traffic flow aggregated for 5 s

Table 5 Validation of calibrated Vissim model

Time Actual Simulated outflow Error Simulated outflow Error Actual Simulated travel Error Simulated travel Error
intervals outflow using calibrated (%) using calibrated (%) travel time using (%) time using (%)
(s) volume Wiedemann 74 Wiedemann 99 time calibrated calibrated
(veh/h) (veh/h) (veh/h) (sec) Wiedemann 74 Wiedemann 99
(Sec) (Sec)

0–900 1036 1015 2.04 990 4.5 59 67.85 15 70.34 19.2

900–1800 1111 1056 5.07 1018 8.7
1800–2700 1141 1099 3.75 1063 6.98
2700–3600 1085 1060 2.33 1066 1.76
 Pers Ubiquit Comput

30000
actual flow
25000
Traffic flow (veh/hr) VISSIM flow

20000

15000

10000

5000

0
0 50 100 150 200 250 300
Green time for the selected approach during different cycles (sec)

Fig. 13 Vissim and actual per second traffic flow comparison

comparison of classified travel time extracted from recorded motorbike riders at the signalized intersection and simulates
videos and Vissim simulation. The figure shows that the travel traffic based on standard FIFO principle. Whereas in reality,
times extracted from Vissim simulation videos for both motorbikes skip the queue to reduce their travel time.
motobikesbikes and cars are approximately the same. Travel
time of cars measured from the field recorded videos and
Vissim is approximately the same. However, the actual travel 5 Conclusions and recommendations
time of motorbikes is significantly lower than the travel time
of motorbikes simulated for each cycle through Vissim. This research studied the microscopic behavior of heteroge-
The lower travel times of motorbikes indicate that the neous and undisciplined traffic at the signalized intersection.
Vissim does not consider the queue-jumping behavior of Within green time traffic flow dynamics were studied in detail

100
90
80
Travel time (seconds)

70
60
50
40
30
20
10
0
0 5 10 15 20 25 30 35 40 45 50
Signal Cycles
Actual motorbike travel time Simulated motorbike travel time-VISSIM
Actual car travel time Simulated car travel time-VISSIM

Fig. 14 Vissim travel time and actual travel time comparison

Pers Ubiquit Comput

with a specific focus on the behavior of motorbikes, which has Vissim simulation showed that Vissim simulation follows
the highest proportion in the traffic mix. The analysis of within the standard FIFO principle, which in reality is violated by
green time flow dynamics shows that the flow of traffic is not motorbikes riders. The Vissim-simulated travel times for all
uniform within the green time. The flow of motorbikes is the the modes were equal. However, motorbikes jump the queue
highest during the first few (1–5) seconds of the green time, through the available spaces, which significantly reduces the
and it is significantly lower after the initial few seconds of travel time of motorbikes in comparison with other modes.
green time. On the contrary, the flow of cars and other modes This phenomenon of queue jumping by motorbike riders is a
is lower during the initial few (1–5) seconds and gets its common practice in the similar heterogeneous traffic streams,
highest value after the first few seconds. The highest overall observed in many South and East Asian countries. The im-
traffic flow is observed at the beginning of green time when pacts of this behavior on flow rate and travel time have been
vehicles are queued close to each other. During the red time, quantified in this research. By incorporating the queue-
vehicles tend to move close to each other to occupy the avail- jumping behavior of motorbikes at signalized intersections
able gaps, to proceed as soon as the signal turns green. This and congested midblock segments, the accuracy of the simu-
results in a significantly higher number of lanes on the ap- lation packages and traffic flow models can be improved.
proach of a signalized intersection, which results in higher The existing traffic flow models and simulation packages
flow when the queue discharges at the beginning of the green lack in modeling the queue-jumping phenomenon of motor-
time. In contrast, the vehicles approaching during green time bikes at signalized intersections in heterogeneous and undis-
move at a distance, resulting in a comparatively lower number ciplined traffic streams. Future studies could be focused on
of lanes. Therefore, the flow rates are found to be decreasing developing mathematical models to represent this phenome-
with green time. However, if the queues are significantly lon- non. A new parameter could be introduced in the simulation
ger and are not entirely discharged during the green time, then packages to represent this phenomenon based on the mathe-
a consistent higher flow could be observed. matical model of queue-jumping behavior. Furthermore, a
The dynamics of traffic flow within green time explored in plug-in or add-on can be developed to add this capability to
this research shows that, contrary to the established concept, the existing simulation packages.
the traffic does not flow at a constant rate during the green
time, especially in heterogeneous traffic with weak lane disci- Acknowledgments Authors are thankful to research assistants Amna
Anwar, Anoosha Shoukat, Ain ul Huda Khan, Kanwar Kumar,
pline. This revelation may be useful in the design and opera-
Farhana, and Waqas Ahmed for their help in data collection and data
tion of signalized intersections. In the absence of information extraction.
about the actual flow rates, engineers and practitioners use the
standard value of saturation flow rates for the design and op- Funding information This collaborative research in UAE was supported
eration of traffic signals. The existing performance and the by the Zayed University Research Cluster grant # R17075. In Pakistan,
the research was carried out at Exascale Open Data Analytics Lab,
impact of any proposed improvement are also modeled using
National Center for Big Data & Cloud Computing (Funded by the
the standard values of saturation flow rates. The design, oper- Higher Education Commission of Pakistan).
ation, and performance of pre-timed and adaptive traffic sig-
nals can also be improved if the system is provided with ac-
curate information about the capacity flow dynamics within References
green time.
This research also calibrated the parameters of Vissim for 1. Das AK, Biswal MK, Chattaraj U (2020) Development of a model
simulating the heterogeneous traffic with weak lane disci- for heterogeneous traffic simulation. In: Transportation research.
pline. The UAV-recorded video provides a better option for Springer, pp 699–706
2. Vayalamkuzhi P, Amirthalingam V (2016) Influence of geometric
estimating the values of behavior parameters. The calibrated design characteristics on safety under heterogeneous traffic flow. J
parameter values are significantly lower than the default pa- Traffic Transp Eng (Eng Ed) 3(6):559–570
rameters values and values reported in the literature for stan- 3. Qian ZS et al (2017) Modeling heterogeneous traffic flow: a prag-
dard traffic. Like most of the existing studies, the calibrated matic approach. Transp Res B Methodol 99:183–204
Vissim accurately reproduced the average flow rates for the 4. Arasan VT, Arkatkar SS (2010) Microsimulation study of effect of
volume and road width on PCU of vehicles under heterogeneous
given simulation horizon. However, the detailed comparative traffic. J Transp Eng 136(12):1110–1119
analysis of simulated and actual traffic conditions revealed 5. Mallikarjuna C, Rao KR (2009) Cellular automata model for het-
that the simulated flow from Vissim could not reproduce the erogeneous traffic. J Adv Transp 43(3):321–345
significantly higher traffic flows at the beginning of green 6. Ahmed A et al (2019) Width-based cell transmission model for
time. Vissim models the queue discharge at a constant rate. heterogeneous and undisciplined traffic streams. Transp Res Rec:
0361198119838841
The comparison of simulated and field-measured travel 7. Asaithambi G, Basheer S (2017) Analysis and modeling of vehicle
times shows that Vissim accurately simulated travel time for following behavior in mixed traffic conditions. Transp Res
cars. The manual extraction of mode-wise travel time from Procedia 25:5094–5103
 Pers Ubiquit Comput

8. Nokandeh MM, Ghosh I, Chandra S (2016) Determination of 28. Stevanovic J, Stevanovic A, Martin PT, Bauer T (2008) Stochastic
passenger-car units on two-lane intercity highways under heteroge- optimization of traffic control and transit priority settings in
neous traffic conditions. J Transp Eng 142(2):04015040 VISSIM. Transp Res C Emerg Technol 16(3):332–349
9. Mallikarjuna C, Tharun B, Pal D (2013) Analysis of the lateral gap 29. Lownes NE, Machemehl RB (2006) VISSIM: a multi-parameter
maintaining behavior of vehicles in heterogeneous traffic stream. sensitivity analysis. In: Proceedings of the 2006 Winter
Procedia Soc Behav Sci 104:370–379 Simulation Conference. IEEE
10. Ahmed A, Mehdi MR, Ngoduy D, Abbas M (2019) Evaluation of 30. Tang T et al (2009) A new dynamic model for heterogeneous traffic
accuracy of advanced traveler information and commuter behavior flow. Phys Lett A 373(29):2461–2466
in a developing country. Travel Behav Soc 15:63–73 31. Fellendorf M (1994) VISSIM: a microscopic simulation tool to
11. Kays HI et al (2017) H-CTM for simulating non-lane-based hetero- evaluate actuated signal control including bus priority. In: 64th
geneous traffic. Transp Lett:1–9 Institute of Transportation Engineers Annual Meeting. Springer
12. Hauer E, Ng JC, Lovell J (1988) Estimation of safety at signalized 32. Fellendorf M, Vortisch P (2001) Validation of the microscopic
intersections. Transp Res Rec 1185:48–61 traffic flow model VISSIM in different real-world situations. In:
13. Shao C-Q, Liu X-M (2012) Estimation of saturation flow rates at 80th Annual Meeting of the Transportation Research Board,
signalized intersections. Discret Dyn Nat Soc 2012 Washington, DC
14. Liu HX, Wu X, Ma W, Hu H (2009) Real-time queue length esti- 33. Bloomberg L, Dale J (2000) Comparison of VISSIM and CORSIM
mation for congested signalized intersections. Transp Res C Emerg traffic simulation models on a congested network. Transp Res Rec
Technol 17(4):412–427 1727(1):52–60
34. Ke R et al (2018) Real-time traffic flow parameter estimation from
15. Chu L, et al (2005) Evaluation of traffic delay reduction from au-
UAV video based on ensemble classifier and optical flow. IEEE
tomatic workzone information systems using micro-simulation. in
Trans Intell Transp Syst 20(1):54–64
Transportation Research Board 84th Annual Meeting, Washington
35. Coifman B et al (2006) Traffic flow data extracted from imagery
DC
collected using a micro unmanned aerial vehicle. Appl Adv
16. Ikeanyi OU et al (2014) Pattern of injuries from motorcycle acci-
Technol Transp:298–303
dents in Abia state, Nigeria: Thet influence of government regula-
36. Ke R et al (2016) Real-time bidirectional traffic flow parameter
tion. J Med Investig Pract 9(1):24
estimation from aerial videos. IEEE Trans Intell Transp Syst
17. Park B, Schneeberger J (2003) Microscopic simulation model cal- 18(4):890–901
ibration and validation: case study of VISSIM simulation model for 37. Taslim U, Mubashshir S, Arshad, SAlam J, Rehman I (2017)
a coordinated actuated signal system. Transp Res Rec 1856(1):185– Karachi traffic chaos: Jamming the wheels of economy. Karachi,
192 Pakistan. Available from: http://www.kcci.com.pk/research/wp-
18. Esfahani, H.N. and Z. Song, A new method for microsimulation content/uploads/2017/09/Karachi-Traffic-Chaos-Jamming-the-
model calibration: a case study of I-710. 2019 Wheels-of-Economy.pdf
19. Hossain M (2004) Calibration of the microscopic traffic flow sim- 38. Ahmed A, Mehdi MR, Rizvi SMA, Fatima T (2019) Evaluation of
ulation model VISSIM for urban conditions in Dhaka city. existing performance and potential for optimization of traffic sig-
University of Karlsruhe Germany nals in Karachi. Arab J Sci Eng 44(10):8747–8759
20. Schultz GG, Rilett LR (2005) Calibration of distributions of com- 39. Arasan VT, Koshy RZ (2005) Methodology for modeling highly
mercial motor vehicles in CORSIM. Transp Res Rec 1934(1):246– heterogeneous traffic flow. J Transp Eng 131(7):544–551
255 40. Maini P, Khan S (2000) Discharge characteristics of heterogeneous
21. Milam RT, Choa F (2002) Recommended guidelines for the cali- traffic at signalized intersections. in Transportation Research
bration and validation of traffic simulation models. In: Eighth TRB Circular E-C018: 4th international symposium on highway capacity
Conference on the Application of Transportation Planning 41. Manjunatha P, Vortisch P, Mathew TV (2013) Methodology for the
MethodsTransportation Research Board. Texas Department of calibration of VISSIM in mixed traffic. in Transportation research
Transportation; Corpus Christi Metropolitan Planning board 92nd annual meeting.Transportation Research Board
Organization; Federal Highway Administration; and Federal Washington, DC, United States
Transit Administration 42. Ge Q, Menendez M (2012) Sensitivity analysis for calibrating
22. Menneni S, Sun C, Vortisch P (2008) Microsimulation calibration VISSIM in modeling the Zurich network. in 12th Swiss Transport
using speed-flow relationships. Transp Res Rec 2088(1):1–9 Research Conference
23. Mathew TV, Radhakrishnan P (2010) Calibration of 43. Perrone L, et al VISSIM: A MULTI-PARAMETER SENSITIVITY
microsimulation models for nonlane-based heterogeneous traffic ANALYSIS
at signalized intersections. J Urban Plann Dev 136(1):59–66 44. Rrecaj AA, MBombol K (2015) Calibration and validation of the
24. Fellendorf M, Vortisch P (2010) Microscopic traffic flow simulator VISSIM parameters-state of the art. TEM J 4(3):255
VISSIM. In: Fundamentals of traffic simulation. Springer, pp 63– 45. Buck HS, Mallig N, Vortisch P (2017) Calibrating Vissim to ana-
93 lyze delay at signalized intersections. Transp Res Rec 2615(1):73–
25. Matsuhashi N, Hyodo T, Takahashi Y (2005) Image processing 81
analysis of motorcycle oriented mixed traffic flow in Vietnam. In: 46. Siddharth, S.P. and G. Ramadurai, Calibration of VISSIM for
Proceedings of the Eastern Asia Society for Transportation Studies. Indian heterogeneous traffic conditions. Procedia Soc Behav Sci,
Citeseer 2013. 104(0): p. 380–389
26. Zhang G, Wang Y, Wei H, Yi P (2008) A feedback-based dynamic 47. Arkatkar S et al (2016) Methodology for simulating heterogeneous
tolling algorithm for high-occupancy toll lane operations. Transp traffic on expressways in developing countries: a case study in
Res Rec 2065(1):54–63 India. Transp Lett 8(2):61–76
27. Park BB, Kwak J (2011) Calibration and validation of TRANSIMS
microsimulator for an urban arterial network. KSCE J Civ Eng Publisher’s note Springer Nature remains neutral with regard to jurisdic-
15(6):1091–1100 tional claims in published maps and institutional affiliations.