Headway Distribution Modeling with Regard to Traffic Status

Shengchao Yin, Zhiheng Li, Member, IEEE, Yi Zhang, Member, IEEE, Danya Yao, Member, IEEE,
Yuelong Su, Li Li, Member, IEEE

Abstract-Knowledge of vehicle headway distribution is very All these indicate that vehicle headway distribution is a
important for intelligent transportation and intelligent vehicle fundamental research for many traffic flow theory and
simulations. Various headway distribution models have been vehicle simulation issues.
proposed, but most of them only fit for a certain traffic situation. There have been several reports concerning development
To solve this problem, we study the dependence of headway
of headway models over the past decades. Generally, there
distributions on traffic status in this paper. Results show that
the log-normal distribution model is adequate in fitting are two types ofmodels: single distribution models and mixed
headway data when the traffic is in free flow status; while the models. The representative single distribution models include
log-logistic distribution model is more suitable in fitting Exponential distribution model [6], Gamma distribution
headway data when the traffic is in congestion status. This model, Weibull distribution model, Log-normal distribution
conclusion is useful in the traffic signal optimization algorithm, model [7] and etc. The representatives of mixed distribution
since it indicates that we should apply different delay estimation
models include Cowan M3 model [8], M4 model [9], the
models during different traffic status so as to design optimal
timing plan. Semi-Poisson model [10], Double Displaced Negative
Exponential Distribution (DDNED) [11] and etc. However,
the possible dependence of headway distributions on traffic
I. INTRODUCTION states needs further discussions.
To select an appropriate model for headway distributions
HEADWAY is usually defined as "the time, in seconds,
between two successive vehicles as they pass a point on in Beijing, China, we studied a large amount of headway data
the road way, measured from the same common feature of that are extracted from the stored traffic monitoring videos.
Particularly, the headway data are distinguished according to
both vehicles" e.g. [1]. In some literatures, it is defined as "the
elapsed time between the front ofthe leading vehicle passing different traffic status here.
a point on the roadway and the front ofthe following vehicle Results show that the log-normal distribution model is a
better choice when fitting headway data when the traffic is in
passing the same point", e.g. [2].
free flow status; while the log- logistic distribution model is
As a microscopic measure of traffic flow passing a point,
more suitable in fitting headway data when the traffic is in
accurate modeling vehicle headway distribution are useful in
congestion status. This conclusion is useful in intelligent
intelligent traffic and intelligent vehicle simulations [3-4].
vehicle design, since it indicates that we should apply
Additionally, an appropriate headway models based on
different car-following strategies during different traffic
different traffic pressures will help engineers to analyze the
status so as to avoid collision.
roadway capacity and evaluate the effectiveness of the
To give a more detailed description of this conclusion, the
applied traffic demand management strategy. Furthermore,
rest ofthis paper is arranged as follows: Section 2 presents the
headway models are essential for determining the dissipating
collection method for the data; Section 3 describes which
delay of mixed traffic flow at intersections and then the
model is more suitable for free and congested status. Finally,
adjusting timing strategies at the signalized intersections [5].
Section 4 concludes the whole paper.

Manuscript received December, 15th, 2008. This work was supported in

part by Hi-Tech Research and Development Program of China (863 Project)
II. DATA COLLECTION
2006AAIIZ208, 2006AAIIZ229, National Basic Research Program of
China (973 Project) 2006CB705506 and Natural Science Foundation of Two digital cameras were used to collect the traffic video
China 60774034,60721003,50708054,50708055.
data from the platform bridges crossing the busiest express
~. Yin is with Department of Automation, Tsinghua University, Beijing,
ChIna 100084 (e-mail: [email protected]). ways in Beijing. One of the camera was located on the
~. Li is with Department of Automation, Tsinghua University, Beijing, XueYuan Road around Northern 4th Circle Road, and the
ChIna 100084 (e-mail: [email protected]). other was located on the Middle ofNorthern 3rd Circle Road
D. Yao with School ofthe Information Science and Technology, Tsinghua
University, Beijing, China 100084 (e-mail: [email protected]). Haidian District; see Table I below. '
-y. Zhang with Department of Automation, Tsinghua University, Beijing, The well-known fundamental variables of traffic flow:
ChIna 100084 (e-mail: [email protected]). density, space-mean speed and flow rate are retrieved from
Y. Su is with Department of Automation, Tsinghua University, Beijing,
China 100084 (corresponding author, phone: +86 (10) 62795043; e-mail:
the video data and examined [12-14]. The aggregation time
[email protected]). window length for these three variables is 1 minute.
~. Li is with Department of Automation, Tsinghua University, Beijing,
ChIna 100084 (e-mail: [email protected]).

978-1-4244-3504-3/09/$25.00 ©2009 IEEE 1057

 Locale Time

27th Jui. 2007

Friday
North 16:00-19:00
(No restraint)

Link 1 177
(Two Lanes) metre
17th Aug. 2007
Friday
South 16:00-19:00
(Restraint)

00
CD®0 13th Aug. 2007
Monday
East
6:30-9:30
(No restraint)

Link 2 193
(Three Lanes) metre

20th Aug. 2007

West Monday
6:30-9:30
(Restraint)

The headway data are retrieved and stored with a much

higher frequency 1/25 second. Table II. Fundamental statistical characteristics of the headway data
(Link 1)
Firstly, the headway data obtained in ordinary working
Strategy \Vithout OEDVP WithOEDVP
days are labeled with their sampling time; and those Peak Non-Peak Peak Non-Peak
Period
collected during peak hours are compared to those during Hours Hours Hours Hours
non-peak hours. Secondly, the headway data obtained in the Average Flow Rate (V/h.l) 569 669 617 567
Sample Size 867 916 877 834
days when a special Traffic Demand Management (TDM)
Mean of Headw'ays (s) 3.70 2.82 2.91 3.11
strategy had being applied are compared to that obtained in Std deviation of Headways (s) 4.21 1.61 1.76 2.10
the days without such TDM strategy. This special TDM Minimum Value (s) 0.62 0.64 0.35 0.65
strategy is called Odd-Even Day Vehicles Prohibition Maximum Value (s) 60.00 10.18 13.00 17.40

(OEDVP), which was tested from August 17th to August

20th, 2007 in Beijing. It is indeed a car-usage restriction Table III. Fundamental statistical characteristics of the headway data
(Link 2)
strategy, which restricts that: cars with odd license plate
Strategy \Vithout OEDVP WithOEDVP
number can be only used in odd days (Aug 17th and 19th) Peak Non-Peak Peak Non-Peak
Period
whereas cars with even license plate number can be only Hours Hours Hours Hours
used in even days (Aug 18th and 20th). This provides us a Average Flow' Rate (vih..l) 601 548 568 470
Sample Size 1423 1406 1459 1311
scarce chance to study the headways under different traffic
Mean of Headways (s) 2.34 2.65 2.67 2.96
loads. Std deviation of Headways (s) 1.14 1.81 1.79 2.30
Minimum Value (s) 0.80 0.30 0.30 0.40
III. PROPERTIES OF HEADWAY DATA Maximum Value (s) 18.20 17.00 13.04 28.20

A. Fundamental Statistics ofHeadways

The fundamental statistical properties of the headway B. Headway Distribution Modeling with Single Models
data under different traffic status are shown in Table II and We had tried all the proposed distribution models to fit
Table III, including average flow rate of every road, mean the empirical headway data and found the Log-normal and
of headways, minimum and maximum values of collected Log- logistic model yield best fitting results.
data. Generally, the Log-normal distribution model is written
as [15]

1058
 1
f(h) = .J2iah exp-
( [In h - Jl f
2a2
J (1) Fig.1 (on the 5th page) representatively compares the
performance of those distribution modelS. The curves in
Fig.1 show the estimated probability density function. The
where the variable h represents the possible value of
lognormal distribution yields least fitting error to the sample
headway data, Jl is location parameter and a is scale data collected in the period of non-peak hours and the
parameter. log-logistic distribution outperforms in the period of peak
The Log-logistic distribution model is presented as [15] hours.
expCn hO"- ,u) (2)
Secondly, the performances of these models are checked
for different headway data collected in Aug 17th and Aug
f(h) = 2

0"[1+ expCn:- ,u)]

20th with Odd-Even Day Vehicles Prohibition (OEDVP).
The analyzing process is same as applied to headway data
without OEDVP: Table VI shows the analyzing results of
In this paper, the Kolmogorov-Smimov (K-S) test is
K-S test and the corresponding estimated parameters are
chosen to measure goodness-of-fit of the selected headway
shown in Table VII. The K-S test also supports our
models to the observed headway data. For each examined
conclusion drawn from visual comparisons of the curves.
sample value x , the K-S test compares the proportion of
Fig.2 (on the 5th page) compares the performance of
values less than x with the expected number predicted by
different distribution models, The curves indicate that the
the hypothesized distribution, and use the maximum lognormal distribution provided better fit to the sample data
difference over all x values is its test statistic. It can be collected in the period of non-peak hours and peak hours
written as follow mathematical Equation (3). totally. The K-S test also supports our conclusion drawn
T(x) = max(IF(x) - G(x)l) (3) from visual comparisons of the curves.
Where F (x) is the proportion of X values less than or
Table VI Comparisons of goodness-of-fit of distribution models.
equal to x and G (x) is the standard hypothesized Average Log-Logistic Gamma Lognormal
cumulative distribution function evaluated at x. So the Strategy Period
Flow Distribution Distribution Distribution
Rate
smaller value T (x) is, the better performance the (v/h..'1)
K-Stest K-Stest K-Stest

Peak
examining distribution model is. 617 0.0291 0.0433 0.0200
Hours
Firstly, headway data collected in Jun 27th and Aug 13th 8-17
Non-Peak
567 0.0293 Not Pass 0.0186
(without OEDVP) are investigated, including peak hours Without Hours
OEDVP Peak
and non-peak hours. Table IV tabulates the analyzing Hours
568 0.0153 Not Pass 0.0153
8-20
results of K-S test for log-logistic distribution, Gamma Non-Peak
470 0.0214 Not Pass 0.0166
distribution and lognormal distribution models; the Hours

corresponding estimated parameters of passing K-S test

models are shown in Table V. Table VII Estimated parameters for passing K-S test distribution models.
Log-Logistic Distribution Lognormal Distribution
Strategy Period
Table IV Comparisons of goodness-of-fit of distribution models. mu sigma mu sigma
Average Log-Logistic Gamma Lognormal Peak
0.885 0.315 0.882 0.550
Flow Distribution Distribution Distribution Hours
Strategy Period 8-17
Rate Non-Peak
K-Stest K-Stest K-Stest 0.925 0.326 0.923 0.569
(v/h.:l) Without Hours
Peak OEDVP Peak
569 0.0199 Not Pass 0.0421 0.819 0.303 0.821 0.533
Hours Hours
7-27 8-20
Non-Peak Non-Peak
669 0.0279 Not Pass 0.0239 0.860 0.345 0.862 0.618
Without Hours
Hours
OEDVP Peak
601 0.0225 Not Pass 0.0277
Hours
8-13
Non-Peak
548 0.0198 Not Pass 0.0158
Table VIII Statistical results of suitable models under different pressures
Hours Data Iun 27th Aug 13th Aug 17th Aug 20th
Non- Non- Non- Non-
Period Peak Peak Peak Peak
Peak Peak Peak Peak
Table V Estimated parameters for passing K-S test distribution models.
Average
Log-Logistic Distribution Lognormal Distribution Speed 21.9 46.3 23.5 48.4 53.1 55.7 35.3 51.9
Strategy Period
mu sigma mu sigma (Km.o'h)
Peak Suitable 1 2 1 2 2 2 lor 2 2
1.096 0.314 1.112 0.578 Model l-Log-Logistic Model; 2-Lognormal Model
Hours
7-27
Non-Peak
0.681 0.354 0.685 0.618
Without Hours
OEDVP Peak
As shown in Fig.3 (on the final page), an useful and
0.885 0.315 0.882 0.550
8-13
Hours interesting conclusion can be drawn in term of the above
Non-Peak
0.925 0.326 0.923 0.569 comparison results: whether to implement OEDVP or not,
Hours
the lognormal distribution model fit for headways on
general purpose lanes under most circumstances, except for
headway data of peak hours in Jun 27th and Aug 13th

1059
(without OEDVP); the log-logistic distribution model is REFERENCES
better for fitting headway data of peak hours without [1] The Transportation Research Boards, Highway Capacity Manual,
OEDVP. More details can be found in Table VIII. National Research Council, Washington, D. C. 2000.
[2] P. G. Michael, F. C. Leeming, and W. O. Dwyer, "Headway on urban
Based on what we have discussed above, different traffic
street: Observational data and an intervention to decrease tailgating,"
loads yield different vehicle space mean speeds, which also Transportation Research, Part F, vol. 3, no. 2, pp. 54-64, 2000.
affect the headway data statistics. The variation trend of [3] M. Krbalek, "Headway in traffic flow: Remarks from a physical
mean headway values in term of mean speed formulates a perspective," Physical Review E, vol. 64, no. 661, pp. 1-7,2001.
[4] Dr. V. Thamizh Arasan and R. Z. Koshy, "Headway distribution of
"U"-type curve; see Fig.4 (on the final page). According to heterogeneous traffic on urban arterials," Journal ofthe Institution of
flow-density, speed-density and headway-speed statistical Engineers, Indian Civil Engineering Division, vol. 84, pp. 210-214,
results, it is interesting that traffic pattern changes from 2003.
[5] Y. Su, Z. Wei, S. Cheng, D. Yao, Y. Zhang, L. Li, Z. Zhang, and Z. Li,
unstable phase to stable phase around 35Km1h, which is the "Departure headways ofmixed traffic flow at signalized intersections:
nadir point of the "U"-type curve, see Fig.5 (on the final distributions, simulations and validations," Transportation Research
page). BoardAnnual Meeting, Washington DC, USA, CD: 08-0479,2008.
[6] R. 1. Cowan, "Useful headway models," Transportation Research,
vol. 9, no. 6, pp. 371-375, 1975.
[7] I. Greenberg, "The log-normal distribution of headways," Australian
IV. CONCLUSION Road Research, vol. 2, no. 7, pp. 14-18, 1966.
[8] O. Hagring, "The use of the Cowan M3 headway distribution for
Headway distribution model is fundamental to traffic modeling roundabout flow," Traffic Engineering and Control, vol. 37,
flow theory and simulation research. Based on the vehicle no.5,pp. 328-332,1996.
headway data collected at urban roadways, the adaptability [9] S. Hoogendoorn and P. Bovy, "A new estimation technique for
vehicle-type specific headway distributions," Transportation
of several typical headway distribution models was Research Record, vol. 1646, pp. 18-28, 1998.
examined. Our research indicated that headway patterns [10] P. Wasielewski, "Car following headways on freeways interpreted by
between lower traffic pressure and higher traffic pressure the semi-Poisson headway distribution model," Transportation
are significant different, each model is endowed with some Science, vol. 13, no. 1, pp. 36-55, 1979.
[11] 1. D. Griffiths and J. G. Hunt, "Vehicle headways in urban areas,"
strength and weakness under certain conditions. In our
Traffic Engineering Control, vol. 32, no. 10, pp. 458-462, 1991.
examinations, the lognormal distribution model fits for [12] G. Zhang, Y. Wang, H. Wei, and Y. Chen, "Examining headway
headways on general purpose lanes under lower traffic distribution models with urban freeway loop event data," Traffic
pressure circumstances, no matter we implement OEDVP Flow Theory, Washington, D. C., pp. 141-149,2007.
[13] R. Mahnke, 1. Kaupuzs and I. Lubashevsky, "Probabilistic
or not; while the log-logistic distribution model is better for
description of traffic flow," Physics Reports, vol. 408, no. 1-2, pp.
fitting headway data of peak hours without OEDVP, there 1-13,2007.
being higher traffic pressure. Helpful information in the [14] H. Cho, and Y. Wu, "Modeling and simulation of multiphase traffic
selection of appropriate headway distribution models under flow," Proceedings of International Conference on Computational
different traffic pressures is provided. Methods in Science and Engineering, pp. 1036-1039,2007.
[15] M. Evans, N. Hastings, B. Peacock, Statistical Distributions, 3rd
edition, John Wiley & Sons, New York, NY, 2000.

1060
 0.4 - - Histogram of Colleoted Data - - Histogram of Colleoted Data
----- Normal Distribution ----- Normal Distribution
0.35
- - Log-Logistio Distribution -e- Log-Logistio Distribution
.......... Gamma Distribution .......... Gamma Distribution
-e- Lognormal Distribution - - Lognormal Distribution
0.3

0.25
:c-
.~

~ 0.2

0.15

5 10 3J 40 6J
Data Data

0.4 - - Histogram of Colleoted Data - - Histogram of Colleoted Data

----- Normal Distribution ----- Normal Distribution
0.35 - - Log-Logistio Distribution -e- Log-Logistio Distribution
.......... Gamma Distribution .......... Gamma Distribution

0.3 -e- Lognormal Distribution - - Lognormal Distribution

0.25
:c-
'in
ffi 0.2
o
0.15

8 10 12 14 16 10 12 14 16 18
Data Data

Fig.I. Nonnal, Log-Logistic, Gamma and Lognonnal distribution fitted to headway data (without OEDVP)

0.35 r----.-----.-----.---;::::::==::::!::::====::::!::::====:!:::==:;l
- - Histogram of Colleoted Data - - Histogram of Colleoted Data
0.35
----- Normal Distribution ----- Normal Distribution
0.3 - - Log-Logistio Distribution - - Log-Logistio Distribution
.......... Gamma Distribution 0.3 .......... Gamma Distribution
-e- Lognormal Distribution -e- Lognormal Distribution
0.25
0.25

:c-
.~
0.2
i' 0.2

o 0.15 ~
0.15

0.1
0.1

10 12 6 10 12
Data Data

0.35
- - Histogram of Colleoted Data --Histogram of Colleoted data
----- Normal Distribution ----- Normal Distribution
0.35
0.3 - - Log-Logistio Distribution -e- Log-Logistio Distribution
.......... Gamma Distribution .......... Gamma Distribution
-e- Lognormal Distribution 0.3
-e- Lognormal Distribution
0.25

0.25

:c-
.~
0.2
:c-
.~ 0.2
~ I
~
0.15
,I
0.15

0.1
0.1

0.05 :

10 15 20 25 6 10 12
Data Data

Fig.2. Nonnal, Log-Logistic, Gamma and Lognonnal distribution fitted to headway data (with OEDVP)

1061
 0.045,...---,....---,....---....----r-----r----...,.---...,.------.
High Traffic ~ K-8 Test Value
0.04 Pressure (Log-logistic Distribution)
~ K-8 Test Value
(Lognormal Distribution)
0.035

CD 0.03
:J
~ 0.025
-;
~ 0.02
U)

~ 0.015

0.01 Low Traffic

Pressure
0.005

OL.-_ _........._ _.....L..._ _.....a.._ _

----IL...-_ _..L-_ _........._ _........_ _---"

20 25 3J 35 40 45 &J 55 6J
Average of Space Mean Speeds (Km/h)
Fig.3. Comparison of headway distribution models with regard to different traffic status.

4.00

.33

3.00 2.94

Speed<=10 10<Speed 20<Speed 30<Speed 40<Speed 50<Speed 60<Speed

<=20 <=30 <=40 <=50 <=60

Speed Classification (Km/h)

Fig.4. Mean headway values under different speed classifications.

60.0

~
550.0
C
"g4O.0
~
Q"
rl:J30.0

40 60 SpeedClassl catlon(Km/h)
Dens.i!y_(v/km)

o
o
CD

Stable ~ Unstable
40
~ ~ $ ~ ~ m $
Averelle of Spece M.n Speeds (KmIh)
Density (v/km)
Fig.5. Mean headway values and headway distribution models under different traffic stratus.

1062