TRANSPORTATION RESEARCH RECORD 1225

Headway and Speed Data Acquisition

Using Video
M. A. P. TayroR, W. YouNc, eNp R. G. THonlpsoN

Accurate knowledge of vehicle speeds and headways on trallÌc ment (such as a freeway) before this study, so there was an
networks is a fundamental part of transport systems modelling.
excellent opportunity to evaluate the system and suggest mod-
Video and recently developed automatic data-extraction tecñ-
niques have the potential to provide a cheap, quick, easy, and ifications to it. This equipment also made it feasible to inves-
accurate method of investigating traflic systems. This paper pre- tigate the relationship between vehicle speeds and location in
sents two studies that use video-based equipment to investigate the car parks.
character of vehicle speeds and headways. Investigation oÌ head-
rvays on freeway traffic allows the potential of this technology in
a high-speed environment to be determined. Its application to the
THE VIDEO SYSTEM
study ofspeeds in parking lots enabled its usefulneis in low-speed
environments to be studied. The data obtained from the video was
compared to traditional methods of collecting headway and speed Using film equipment to obtain a permanent record of vehicle
data. movements is not a new concept. However, considerable recent
developments have occurred in collecting data using video.
Digital image-processing applications offer the potential to In particular, ARRB has developed a trailer-mounted video
automate a large number of traffic surveys. It is, therefore, recording system (3). This relatively new equipment has until
not surprising that considerable interest has been directed at recently experienced only a limited range of applications. It
developing procedures for vehicle classification and number consists of a gas-operated extendible mast (with video camera
plate recognition (1 , 2). However, there appear to be several attached), a power supply, and other video accessories con-
obstacles preventing these systems from being used regularly cealed in an inconspicuous, covered trailer. A range of lenses
in traffic surveys. Some of these difficulties are the relatively is available that enable it to view various aspects and areas.
high cost of the systems, problems of accuracy, quantity of The extraction of the data from the video could be carried
data collected, and the time required to process the raw data. out by a human observer, but this is time consuming. A pref-
Fortunately, systems are being developed that can be used erable method is to extract the data automatically. A pro-
for specialized applications. This paper presents one such sys- cedure for carrying out this task has been developed by Trout-
tem. It is a video-based system that allows vehicle speed and beck and Dods (3). The position of the vehicle is recorded as
headway data to be extracted. it passes through a particular point on the screen. These points,
Video-based techniques offer the means to overcome many termed detection points, are identified in the data collection
of the difficulties of collecting information on speeds and stage. The movement of a vehicle through these points is
headways. The basic idea is to create a video recording of determined by a change in the intensity of the image and is
traffic, then extract data from it. The video approach has a recorded on computer tape for further analysis. The basics of
number of advantages including being unobtrusive and requir- VADAS are as follows. The image-processing system relies
ing a small labor component. Its major advantage is the pro- on the observation that most vehicles have a video illuminance
duction of a permanent, complete record of the traffic scene. level much greater than that of the road pavement on which
This may be reanalyzed at any stage, and provides an account they are traveling. The general luminance of the scene is
of each traffic event observed. monitored by locating a reference point just off the carriage-
The oft-quoted disadvantage of the video technique is the way. Thus, changes in ambient light levels due to cloud cover
large amount of time and effort needed for data extraction. The and the angle of inclination of the sun can be accommodated.
use of an automatic data-extraction (image processing) system Besides the reference point, VADAS allows the analyst to
can overcome this problem. The video vehicle detection system define up to L6 detection points within the video frame. The
(VADAS) (3) developed by the Australian Road Research illuminance level at each point is then monitored by VADAS,
Board (ARRB) offers a sophisticated, high-technology system and a change in luminance above a set threshold is taken to
useful in data extraction. VADAS was developed for use in mean the passage of a vehicle over that point. The basic steps
studies of the performance of unsignalized intersections, involved in data extraction using VADAS are the positioning
particularly roundabouts where complex traffic maneuvers of detection points on the screen and the adjustment of the
take place. It had not been applied to a high-speed environ- sensitivity threshold. Once the data have been logged, they
can be edited. The output from this process is a computer file
of times and detector numbers that can be used in subsequent
Department of
.Civil-Engineering, Monash University, Clayton,
data analysis. The speeds and headways can be calculated
Victoria, Australia 3168. from data on the times when vehicles were detected.
 131
Taylor et al.

HEADWAY STUDY Such correlations should be expected on arterial roads where

there are signalized intersections. There is no clear evidence
A convenient way to describe the inherent variability within one way or the other for freeway traffic flows.
a traffic stream is to consider it as a stochastic process and Another phenomenon of interest is the distribution of traffic
examine the headways between vehicles in a traffic stream. between lanes on a multilane road. The conceptual model is
Headways are defined as the time separation between the that traffic uses (indeed, under some highway codes, should
same point on two successive vehicles in a traffic stream. This use) the outside lanes of a carriageway unless overtaking,
definition holds for both single-lane and multilane traffic when vehicles will use the inside (median) lane(s). Under light
streams. Statistical distributions can be fitted to represent traffic volumes, this model works satisfactorily. But, as traffic
headways. The reciprocal of mean headway is equal to the volume increases, this flow regime breaks down. Vehicles in
mean flow rate. the outside lane that overhaul a slower vehicle in that lane
Headways are perhaps the most important of all traffic and wish to overtake find increasing difficulties in changing
variables since a Iarge amount of traffic research and simu- Ianes. For instance, Mahalel and Hakkert (8) examined traffic
lation exercises make use of headway models in the analysis flows on a twolane carriageway in Israel and found that the
of traffic problems. Buckley (4) stated that "headways are median lane carried more than 50 percent of the traffic flow
the building blocks with which the concepts of volume and for flow rates above 1,150 vph. Chen (9) found that the median
capacity, and indeed the entire traffic stream are con- lane attracted the largest volume of traffic on a. threelane
structed." The ability of traffic models to replicate traffic carriageway in Ohio, for total carriageway flows exceeding
systems realistically will be enhanced through more efficient 4,500 vph. This tendency for increased proportions of total
and accurate techniques of collecting headway data. flow to use the faster lanes as flow builds up leads to a greater
A number of theoretical models of headways exist. Schuhl tendency for bunching in that lane, and perhaps more insta-
(5) and Buckley (4) provided some of the earlier research in bility in the traffic flow.
this area. The most comprehensive theoretical study of head- Although the negative exponential distribution has not proven
way distributions was provided by Cowan (ó), who derived a to be a good descriptor of single-lane headways, there is a
family of four headway models. The first three members of strong possibility that it could represent multilane traffic flows
this family are the negative exponential distribution, the shifted since headways close to zero are feasible. Buckley (4) and
exponential distribution, and the mixed exponential distri- Gerlough and Barnes (10) both suggested that the negative
bution. These three models show increasing levels of sophis- exponential model could be used for multilane unidirectional
tication from the simple random traffic model (the negative traffic flow.
exponential distribution) through the shifted exponential dis- Given these observations, the data collection scheme was
tribution (which allows for the finite length of vehicles trav- designed to include the defined flow regime; traffic volumes
eling in a single lane by imposing a minimum headway) to in the identified flow regime were spread out over different
M3, which represents traffic as a mixture of free and following times of day and different days to account for the known
vehicles. Cowan fitted the models to data collected in a single variations in traffic flow levels over hours of the day and days
lane on an arterial road. He concluded that the M3 model of the week. Spreading the data-collection periods would pro-
(and the fourth member of the headway model family, M4) vide the means to avoid misinterpretations based on fluctua-
could provide a reasonable description of traffic flow. tions or deviations from the normal traffic pattern.
A number of problems emerge in the application of statis- Theories of traffic flow and models of traffic system per-
tical distributions to traffic headways. There are two theo- formance require comprehensive, accurate, and unbiased data
retical issues of importance. First, the models assume that for their appraisal. Much theoretical work has been directed
mean flow rate (q), and hence mean headway, are constant at the case of uninterrupted flow, in which the performance
over time (i.e., the traffic process is assumed to be stationary). and behavior of a traffic stream is dictated by the nature of
Although this may not be unreasonable for short time periods, the stream and the interactions between vehicles in the stream,
it can cause problems when collecting the large data sets needed free of outside influences. The traffic situation most closely
to estimate the parameters for the theoretical models and to approximating these conditions is probably freeway traffic.
fit the distributions, particularly when using conventional data- Data recording by conventional means, involving the place-
collection methods. Second, the headway distribution models ment of vehicle detectors on the carriageway, is difficult and
also assume that the headways are independent of each other. impractical for freeway operations. Further, their presence
Another difficulty is a practical one, of particular relevance on the roadway could modify driver behavior, producing biased
to studies of multilane traffic flows. This is the problem of results.
observing and assessing the fulI set of vehicle headways between
Ianes and within lanes. It is this difficulty that a video-based
data recording method can help to overcome. In the past, DATA COLLECTION
most observational studies of headways [e.g. , Cowan (ó)] have
concentrated on a single lane of traffic. The usual method of The primary objectives of the headway study were to examine
treating multilane traffic is to assume that vehicle arrivals are the use of VADAS for automatic image processing in a high-
independently distributed and that a superposition of single- speed traffic environment, as well as to fit theoretical headway
lane headways can be made. The validity of this approach models to freeway traffic data. In a series of experiments
could be questioned if vehicle arrivals are not independent. performed in 1986, data were collected using video cameras
Miller (7) and Mahalel and Hakkert (8) found some evidence located at particular vantage points, e.g., freeway overpasses
of correlations between vehicle arrivals in multilane traffic. and pedestrian bridges on the Mulgrave Freeway in the east-
r32 TRANSPORTATION RESEARCH RECORD 1225

ern suburbs of Melbourne. The freeway section was that DATA EXTRACTION
between the Springvale Road and Wellington Road inter-
changes (Figure 1). This section has two lanes for each direc- Two methods were used to extract numerical data from the
tion of traffic flow. It was selected because of the considerable video recordings. The first was manual extraction, and the
variations in the level and quality of traffic flow and the pres- second was automatic data extraction using VADAS. The
ence of the footbridge, which provided a convenient and manual method was included to provide a basis for evaluating
inconspicuous high point to record the flow. This freeway the automatic procedure.
section represented uninterrupted traffic, with no external
influences on the flow.
This survey involved the collection of nine sets of data Manual Analysis
spanning three separate time periods on each of three days.
The first six sets were about 15 min long, while the remaining A stopwatch timer was included on the video record to assist
three sets were about 30 min long. In all, a total of about 3 in manual data extraction. The recorded tape was played at
h of traffic flow was recorded, covering traffic demand from a quarter speed. An arbitrary mark was identified on the
8 a.m. to 5 p.m. Table 1 shows reference information for the screen (e.g., a corner cube reflector or ceramic lane deline-
data sets. Only the first six data sets were used in the analysis ator). Each time a vehicle passed the mark, the time was
described in this paper. noted. Different observers watched different lanes. The whole
The data covered a wide range of medium to high flows. process took up to 90 min for 15 min of video recording. This
Young et al. (11) proposed that different models should be process required extensive manipulation and was time con-
applied to traffic under different flow regimes and that these suming. The entire processing time (not including the time
regimes could be defined in terms of the flow-density diagram subsequently spent in data analysis) exceeded 3 hr (6 person-
for a given road. They proposed three separate regimes: low, hr) per data set.
medium, and high flows.

VADAS Analysis

Figure 2 shows the configuration of 11 detection points, in

two bands across the carriageway, used in the data extraction.
The detection points were a fixed distance apart, and the
distance corresponded to that between consecutive corner
cube reflectors in adjacent groups of the raised pavement
markers that form the lane delineators for the freeway.
Knowledge of this distance permits speed information to be
computed, once the validity ofthe techniques has been prop-
erly established. The present discussion deals solely with
headways.
Care was taken when positioning the detection points to
ensure that any change in ambient light conditions, e.g., cloud
cover or shadows from bridges or foliage, would affect all
(observol¡on s¡te) points equally.
Adjusting the sensitivity threshold for each detection point
was vital. The sensitivity was set sufficiently high to detect as
many vehicles as possible. This was found to be especially
important for those high-speed vehicles with similar lumi-
nance to that of the road surface. However, when the sen-
sitivity was too high, some vehicles were detected more than
o once. This phenomenon was especially pronounced for lighr
500
Scole rn metres

FIGURE I Freeway study area. A trl-gE rJ

Þ
õ
q

TABLE 1 DESCRIPTION OF MULGRAVE FREEWAY

DATA
Data
Set Date Time Direction
I 7/8t86 9'.74 a.m. to city
2 718/86 12:00 p.m from city E EEE E]
3 7/8t86 3:36 p.m. from city EI
4 20t8t86 L0:24 a.m. from city q
5 20t8t86 1,1'.46 a.m. to city
6 20t8t86 2:21 p.m. to city
FIGURE 2 Location of VADAS detector points.
 133
Taylor et al.

colored vehicles, which could trigger the detection point per- was generally attributable to the poor quality of the video
haps three times (hood-windscreen-roof). An optimum sen- camera used for the data collection. The accuracy of the
sitivity was thus chosen, on the basis that all vehicles would VADAS system has also been tested in intersection surveys
be detected. False points triggered were then detected in the with results of approximately 95 percent being obtained (3).
data-editing phase, using a series of logical tests of the time
differences between successive events at the one detection
point. The method is not perfect. Some false detections may RESULTS
still have been accepted while some genuine detections may
have been missed. However, the procedure should be at least The data extracted in this study were subjected to three
as good as manual observations based on viewing the video separate analyses: a general view of the traffic flow for each
record at normal speed. data set and each lane and an examination of lane usage
When the video signal at each detection point exceeds the characteristics.
upper and lower limits, an output is produced and read by
the VADAS microprocessor (3). A preset time delay is used
to prevent further detections at the detector. This was set at Traflic Flow Characteristics
70 ms for the freeway study to ensure that all vehicles could
be detected. This, however, led to some multiple detections, Comparisons were made of east- and westbound flows for all
as indicated above. Some problems were noted in recording nine data sets and the range of traffic flow conditions' The
motorcycles when these vehicles traveled between the detec- results of this comparison are shown in 'Iab\e 2. This tabie
tion points. shows a range of lane flows from 632 vph/lane to 1,476 vphl
The time taken to extract the data from all the video tapes lane and a range of total flows from I,420 vph to 2,676 vph-
was approximately 5 hr. This involved an initial period of An important result is the consistent observation of small
around 4 hr to set up and become proficient with the equip- minimum headways (approximately 0.3 to 0.5 sec) for single-
ment. The processing of the tapes was completed within L hr. lane flows. No distinguishable differences were found between
Obviously, if data was regularly processed using this tech- the characteristics of traffic in each flow direction. Figure 3
nology, the time required for familiarization could be sub- shows the coefficient of variation of headways plotted against
stantially reduced and the data extraction would become the traffic flow rate, with the observations scattered about the
quicker. Further time was required for editing the data. The horizontal line representing a unit coefficient of variation.
editing was carried out using a series of programs developed There is a possible suggestion of differences between lane
by Troutbeck and Dods (3). The editing for all the headway flows, with the outer (slower) lane showing coefficients typ-
data took approximately t hr. It was therefore apparent that ically less than unity, while the inner (faster) lane shows coef-
VADAS offers considerable time savings in headway data ficients generally exceeding unity. More data are needed to
collection once users become familiar with the operation of pursue this observation, but it is interesting to reflect that a
this equipment. unit coefficient of variation is one analytical property of the
The VADAS headway surveys described here were com- negative exponential distribution. A further comment on Table
pared with manual methods. The comparison indicated a 2 is the strong indication that trucks were largely confined to
VADAS accuracy level of approximately 90 percent, which the outer lane.

TABLE 2 SUMMARY DATA FOR THE HEADWAY DATA SETS FROM MULGRAVE FREEWAY
Data Set Number
Description
FuIl Carriageway
Sample size 550 356 668 364 381 375
Flow rate (vph) 2,200 1,424 2,672 1,456 r,524 1,500
Truck percentage 7.8 12.r 5.2 L1.2 9.4 8.5
Minimum headway (sec) 0.0 0.0 0.0 0.0 0.0 0.0
Maximum headway (sec) 11.5 17.4 10.i 17.9 22.O 17.8

Outer Lane
Sample size 245 198 300 204 206 209
Flow rate (vph) 984 796 1,200 820 828 840
Truck percentage 15.9 t9.l 9.7 18.5 13.0 t3.3
Minimum headway (sec) 0.3 0.2 0.2 0.5 U.J 0.6
Maximum headway (sec) 14.5 20.8 12.0 18.5 22.0 17.8

Median Lane
Sample size 305 158 368 160 175 166
Flow rate (vph) 1,220 632 1,472 640 700 664
Truck percentage 1.3 5-Z 1.6 1.9 5.1 2.4
Minimum headway (sec) 0.3 0.2 0.2 0.3 0.3 0.4
Maximum headway (sec) 18.2 39.5 19.9 41 .5 40.0 45.9
t34 TRANSPORTATION RESEARCH RECORD 1225

Distribution of Lane Flows

62ooo
c
Mahalel and Hakkert (8) reported Israeli data that showed
s
c
that a majority of vehicles used the outer lane at low volumes.
The reverse occurred at high volumes. This tendency has been
ì(¡,
rsoo

observed in other studies from various countries, e.g., Chen ;

(9),Lenz and Hotop (12), and Rorbech (73). A similar tend- € rooo
ency was noted in this study. One possible explanation is given .9
below, in terms of bunching tendency. Figure 4 shows the o
split of traffic between the two lanes of the carriageway. It -o soo
c
suggests that equal lane flows occurred at about 1,550 vph. o
J
This is consistent with the results of the other studies cited
above. Further data are needed to fully test this finding. o rooo 2000
Totol troffic ftow (q)(veh./h).

Bunching Tendency too

d)
c
-9
An examination of the coefficients of variation in headways E
(Figure 3) indicates some apparent differences between the 875
o,
patterns of traffic flow in adjacent lanes. In the outer lane, .ç
the coefficient was typically less than unity, while it exceeded .o 4-r.a------x-
unity for the median lane. This difference occurred at all levels 850 xF' --G---o
of traffic flow. Figure 3 shows these results. There was a o
(¡)
greater tendency for bunching in the median lane, with a (fr t4
c'--
bunching tendency exceeding that of Poisson flow. On the c
o
()
other hand, the vehicles in the outer lane were underdispersed
in relation to Poisson arrival patterns. Similar results were o
0-
reported by Mahalel and Hakkert (8). This difference is prob- o 1000 2000 3000
ably due to the presence of slower-moving vehicles (those Totot rroffic ftow (q)(veh./h).
with speeds well below the mean speed) in the outer lane.
Under medium-to-heavy flows, there are limited opportuni-
FIGURE 4 Division of traffic lanes.

ties for lane changing and overtaking. To maintain a reason-

ably constant speed and not be caught behind slow vehicles, This hypothesis is supported by the shift in the distribution
drivers prefer to change lanes at the earliest opportunity while of vehicles towards the median lane with increasing volume
still some distance behind the slow vehicle. The long sight (see Figure 4). At higher volumes, the level of service is
distances available on a freeway facilitate this maneuver. lowered to a point where the initiation of overtaking maneu-
vers becomes difficult. A large proportion of drivers then elect
to stay in the median lane to achieve a smoother flow.

Distributions of Headways
ø
o Researchers have given considerable attention to headways
ì
D and suggested many theoretical distributions. The problem
o
û, {,ô
t with many of these models is that they often require complex
o iterative procedures to fit them to obse¡ved data. Rational
c subdivision and nonlinear least squares techniques have to be
.9
used, which are lengthy and at best approximate. As a result,
.9
o only the common univariate models offered in the TRAN-
STAT distribution fitting package (14) were fitted ro the head-
b o5 way data. The effort to estimate the additional parameters of
c KeY
.9
'
o left lone mixed models is justified only if they substantially improve
.9 x right lone the fit to the data. Distribution fitting using TRANSTAT
q)
a overoll suggested that the exponential model best represented head-
()o ways for the combined lanes, while the log-normal model best
represented the individual lanes. Since a large number of the
o 1000 2000 3000 headway models fitted the data to a satisfactory degree, the
Troffic f tow (q) (veh./h) mixed models, other than M3, were not investigated.
These general conclusions \À/ere not valid for all data sets,
FIGURE 3 Relationship between coefÏicient of variation and indicating that further work is required to investigate the con-
trafïic flow. ditions under which the different models could apply.
Taylor et al. 135

SPEED STT]DY In summary, considerable effort has been directed at the

determination ofvehicle speeds on urban arterial roads, rural
The primary objectives of the speed study were to determine roads (17), and residential streets (18). The impact of road
mean desired circulation speeds on parking circulators and design standards has also been investigated (19). All these
aisles, as well as comparing two methods of speed data col- studies used radar to collect the data.
lection: radar and video.
Initial understanding of the specific nature of the data to
be collected is an important component of any traffic survey SPEEDS IN PARKING NETWORKS
(15). This is crucial in speed studies since the term "vehicle
speed" is by no means a simple concept. It can refer to space As discussed, no speed studies have been directed at vehicle
mean speeds, design speeds, safe driving speeds, desired (free speeds in parking lots. Therefore, before considering the stud-
speeds), or spot speeds. In this study, several speed-related ies presented in this paper, it is necessary to introduce the
concepts are discussed. Space mean speed is the average speed components of a parking lot and the concept of desired speed
of vehicles taken over a length of road at a particular point as it relates to vehicles in parking systems. Parking road net-
in time, whereas time mean speed is the average speed of works can be likened to the urban road network where a
vehicles taken at a particular point in space over a period of hierarchy of roads are defined, each with different functions
time. (20). The lowest level of the hierarchy in a car park is the
The desired speed is the speed a driver will adopt if not aisle (A). It is the area of pavement used by cars to gain access
hindered by other vehicles. It is often referred to as the free to parking spaces. The upper end of the hierarchy is defined
speed and is measured by observing the speeds of isolated as the circulator (C). It is a roadway that provides access to
vehicles and vehicles at the head of a platoon of vehicles, parking aisles from ingress and egress points of the facility.
since vehicles in a platoon cannot adopt their desired speed. Parking should not be provided on circulator roads.
Methods to overcome bias in measuring free speeds are As mentioned previously, drivers in parking lots are often
described by Taylor and Young (15). The definition of desired constrained by other activities (e.g., looking for a parking
speed provides the impression that drivers are not hindered space or exit). These constraints may not allow drivers to
in their desire to travel at a particular speed. Drivers in park- reach a desired speed. Therefore, this paper introduces the
ing lots are constrained by the geometric characteristics of the concept of circulation and refers to the desired circulation
roads and their major purpose (looking for a parking space, speed in parking lots. In parking lots, many of the roads are
leaving the parking lot, etc.). Therefore, they may never reach short, and a driver may spend all the time accelerating along
their desired speed. Possibly a more appropriate term would the road rather than adopting a desired constant speed. The
be "desired circulation speed." The desired circulation speed determination of vehicle speeds in parking facilities is com-
is the speed a driver will adopt when moving through the pticated by the need to decelerate to park and carry out other
parking system. Vehiclés that are not hindered can adopt this maneuvers. The concept of a desired circulation speed may
speed. therefore not be appropriate since the vehicles may never
The spot speed is the speed of a vehicle at a particular point reach a steady speed. Thus, the first step in this study was to
along a road. This speed can be measured using a radar gun. study the character of speed profiles of vehicles traveling along
The interval speed is the speed over a small length (or interval) roads in parking facilities. The determination of the steadiness
of road and can be determined using video-based equipment. of the vehicle speeds was studied in both circulators and aisles.
It is calculated by dividing the time taken to traverse a small
length of road by the length of the road. The average spot
(interval) speed is the arithmetic mean of the individual vehi- DATA COLLECTION AND EXTRACTION
cle speed measurements. Since the speed measurements in
this study are taken at a spot (or interval), the average speed The site chosen to conduct the data collection was Vermont
is a measure of the space mean speed. However, since the South Community Shopping Centre, located on Burwood
characteristics of the roads considered limit overtaking, the Highway opposite the Australian Road Research Centre. This
average spot speed is also a reliable estimate of the time mean site was chosen because it provided separate circulators and
speed. parking aisles with unimpeded views of at least two aisles
from the video camera. The study was carried out on a Friday
morning, since this provided high volumes of traffic. In fact,
A REVIE\ry OF GENERAL SPEED STUDIES the car park was around 70 to 90 percent full when the readings
were taken. The survey was carried out on a day with perfect
Knowledge of the speed of vehicles in parking lots is important weather conditions: sunny with light winds. This was advan-
in ascertaining the safety of the system and the effects of tageous since poor driving conditions may have affected the
speed-control devices, and as an input into parking lot design results.
models (11). Little attention has been focused on this param- Data was collected in two parts of the car park (Figure 5):
eter. This is difficult to understand given the considerable a major circulator (Site 1) and a parking aisle (Site 2). The
conflict between pedestrians and vehicles in parking lots, which video recording system enabled changes in speeds along routes
is exacerbated by the relative speeds of vehicles and pedes- to be measured. By contrast, it is difficult to measure vehicle
trians. No reported studies of vehicle speeds in parking lots speeds at particular points along routes using a radar gun.
could be found. It is, however, informative to review studies Although radar can be used in conjunction with treadle sys-
of speed on other parts of the transport network (1ó). tems to obtain speed profiles Q1), the video allowed the
 TRANSPORTATION RESEARCH RECORD T225

PROJECT 4207
PATH D
Speeds in Cor Porks
1986
65m
VENTURE COLES
E
I
'1"
t-51qJÈÍ I

t-ffiil __l9;, l.-

_t
l-
Tn_ lrz älH l(,

lã3
--| IGÈ
+lt-lt I

FIGURE 5 Vermont South parking lot layout.

ì!ï l:r
_l^* L__l
Ë e¡onn
I

vrD€o¡ vtoEo 4 vtDEoã vtDEo6

t t kt f r
t t t
f
*s:ã
tì:--1ì--.---
t-.'-...,#
F l.-{
l.ì
h
N l-+-__

ñ
L
Ei*t--.
r
LJ
lrt;¡\.+
M
J e
t::" RìÙ-.'
hl ÉT LI H
t E
I
E

t t L
t þ
t I H
L
t Ë
t t
b
t F f t
t tr t
t t
t t

FIGURE é Speed profïles for vehicles traveling along circulator (path A).

flexibility to adjust the location of detectors and to relate speed, it was assumed that any headway between two cars of
vehicle speeds to trip characteristics (e.g., whether or not the less than 5 sec represented car following. In such cases, the
vehicle was parking). The trailer housing the video camera second car was rejected.
was positioned and stabilized perpendicular to the vehicles'
movement path. The camera was then focused and raised to
the top of the mast (10 m). After the camera had begun RESULTS
recording, the aisles and circulatory roads were marked by
witches' hats at 5-m intervals so they could be located when The vehicles moving along the circulators were divided into
the video tape was processed. four groups (Figure 5): those vehicles that did not enter an
The extraction of the data from the video could be carried aisle (Path A), those that enrered Aisle 3 (path B), and those
out by a human observer, but this is time consuming. A pref- that exited Aisles 1 (Path C) and 2 (path D).
erable method is to extract the data automatically. VADAS The speed profiles of the vehicles that did not enter an aisle
was used to extract the vehicle headways. In this system, the (Path A) are shown in Figure 6. The average (space) speeds
speeds are calculated by measuring the times the recording of these vehicles fluctuated between Z1,.2kmlhr and25.2km/
points are activated and dividing the difference into the dis- hr. The general evenness of the speed profiles indicates that
tance between detection points. To determine the desired the drivers do in fact have a desired circulation or cruising
Taylor et al, 137

speed, which fluctuates at about 23.2kmlhr. The vehicles in For the aisles, these speeds were around 14 to 15 km/hr; for
this lane are moving toward another part of the parking lot the circulator, tliey were around 2l to 25 km/hr.
or are using the parking lot as an alternate route to bypass A difference means test was used to test the difference in
the adjacent arterial road system. The speed profiles for those speed of vehicles turning from a circulator into an aisle. These
vehicles exiting an aisle (Paths C and D) showed that there two means showed no significant difference, even though they
is a pattern of steady acceleration toward the desired speed were not on the same road. The mean desired circulation
similar to those of vehicles moving through the system (1ó). speed of a car turning the corner from a circulator into an
Vehicle speeds in aisles varied considerably. There are two aisle was 13 km/hr.
types of vehicles that use aisles: those that find a parking
space and those that are looking for a space. Vehicles of the
first type slow down as they approach the parking space (Fig-
ure 7). Vehicles of the second type travel along the aisle COMPARISON OF RADAR AND VIDEO DATA
looking for a parking space or leave the parking lot (Figure
8). A comparison ofthe speed profiles for both parking aisles Video has a considerable advantage over radar in that it is
showed that there is a slight difference in both the shape of
located perpendicular to the vehicle's motion. Thus, it can
the speed distribution and the mean velocity for each aisle. record speed readings along a length of road easily, and var-
This could be due to a number of aisle characteristics, includ- iations in the speed of the vehicle along the road can be
ing differences in the grades between these aisles and differ- measureci. The raciar unit is iocateci aheaci of the car, which
ences in the proportions of the stalls that were used in each
creates problems in locating the position of the vehicle. Fur-
aisle.
ther, radar and video measure different speeds. Video mea-
An important point to establish in the interpretation of sures a speed over a given length of road (space), whereas
speed profiles is whether or not the driver reaches the desired
radar measures it at a particular point (spot). Comparison of
circulation speed. By looking at the evenness of the profiles, the two methods therefore involves the assumption that the
it is reasonable to conclude that desired speeds are reached. speed measured by video is constant over the length chosen.
Comparison between the two approaches was required, there-
fore, to determine if they produced similar results.

Data Recording Technique

Vehicle speeds were also measured with radar using a Kurston

H.R. 4 handgun, which had been adjusted to measure low
speeds accurately. The radar unit was set up in direct line
with the vehicle's path. Two observers were required to collect
the speeds of the cars using the radar. One paid particular
attention to when the car passed the data collection point and
noted the speed at that instance; the other observer recorded
the speed of the cars on paper as well as noting the time of
occurrence and occasionally noting an identifying character-
o 5 lo 15 20 25 þposition
(metres)
istic of the car. This enabled the radar measurements to be
matched to the same vehicle's speed recorded using the
FIGURE 7 Speed profiles for vehicles parking in aisles. camera.

MEAN VELOCITY PROFILE: AISLE 2

FIGURE I Speed profiles for vehicles in aisles.

138 TRANSPORTATION RESEARCH RECORD 1225

Data Analysis time to collect and code speed data. The VADAS system
would be considerably faster once the users were familiar with
Data analysis was carried out by comparing the speeds of the the equipment.
vehicles provided by the radar and the video at the same point.
These methods ideally should produce similar results, since
they are both measuring the same vehicle speeds. Two tests AN APPRAISAL OF THE DATA COLLECTION
were conducted to compare the two data sets. One test com- AND ANALYSIS TECHNIQUE
pared the overall means of each sample, while the other com-
pared the individual data between the two samples. A number of points should be considered when collecting
The comparison of the means is shown in Table 3. The data using the video:
significance of the difference of the means is shown using the
probability value. It is the significance level associated with o The position of the sun and its intensity play a significant
a significant ditference in the means. The first probability role in the quality of the recorded image. Glare tends to
value indicates the difference in the location of both speed confuse distant images, and the contrast between the image
distributions by comparing the sample means. It can be seen being observed and the background must be great enough to
that there is little difference at the L percent level. The second register a difference when the data is analyzed.
probability value refers to the paired means test. It indicates . It may be difficult to determine movements associated
whether the difference in the mean differences of individual with vehicles obscured behind larger vehicles. The positioning
speeds is not equal to zero. At the 1 percent level this is in of the camera should consider this possibility.
fact the case for all but Path B. There is therefore little evi- o The area the video can cover is restricted by the lens used
dence to support the hypothesis that the individual speeds in the camera; however, distortion error may result from the
measured by the two methods are the same. However, the curvature of the camera lenses. Consideration should be given
overall mean from the two data sets shows no difference. This to the need to have a site that will allow the video to cover
is consistent with arterial road speed studies (3). the full study area.
Another comparison of individual speeds was made using o Parallax errors will also result from the use of the camera
regression. If, as assumed, the video and the radar give similar since the camera image is two-dimensional while the actual
results, the linear regression should give a 45'line with an situation is three-dimensional. This two-dimensionality could
intercept at the origin. The results of four comparisons are cause errors in the distance traveled between the two markers
given in Table 4. on the screen. Corrections must be made.
The correlation between the two data sets is high, indicating o Although the extendible mast allows the camera to look
a strong relationship between the radar and the video read- down on the traffic flow for near sites, vehicles at distant sites
ings. However, as with the pairs Êtest, the individual readings must be viewed somewhere between a top view and a profile.
are not directly comparable. The video tended to underestimate This may cause a problem since the profile of a vehicle is not
the low radar speeds and overestimate the high radar speeds. square and, since vehicles do not always follow the same path,
The lack of agreement between the individual readings dis- the detector points may be activated by the hood or bonnet,
cussed earlier was therefore confirmed. or missed altogether. In such a case, some of the detector
Comparison of the radar and the video speed readings indi- points may have to be adjusted to detect the same point on
cated that the overall mean speed did not vary but the indi- the vehicles consistently.
vidual speed readings were not consistent. However, a sig- o Dark cars and motorbikes were very difficult for the data
nificant difference in statistical testing may have little practical logger to detect because a change in brightness is required to
impact. The magnitude of the differences in the mean speed activate the detector. Another effect of dark vehicles is that
using the two measuring techniques is 1.17 km/hr with a stan- the blimp may not be activated when the first part of the car
dard deviation of 3.30. The accuracy of both the radar and passes the blimp.
the video is approximately 1 km/hr. In the light of the mea- o Quite often, a car passed through a detector and caused
surement error in the two techniques, a L . 17 km/hr difference the detector points to flash until the car was totally past. This
is, for practical purposes, small. Due to the small sample sizes caused a succession of times to be recorded. In this case, the
of these surveys, further studies would be needed to confirm first time would be considered the correct time.
these initial conclusions. o The sensitivity of the data logger was continually adjusted
A comparison of the time taken to collect and extract speeds by the brightness of the sun. This was a problem in cloudy
using video and radar was also undertaken. The radar survey conditions, for the brightness changed every time the clouds
required two people for the duration of the survey (4 hr) to cleared. For the purposes of data logging, it would thus be
collect the data. The video, once set up, can be left unmanned, preferable to collect data on a bright, sunny day. This woutd
and therefore requires only one person for an initial period also cause more of the darker cars to be detectable due to
(approximately t hr) to ensure that the trailer and unit are the sun's reflection off their body panels.
positioned and recording correctly. The data-extraction times
for VADAS were similar to that of the headway data (around
6 hr) but would be considerably lower if the equipment was CONCLUSIONS
used regularly. The radar data was quickly transcribed into
the computer (around t hr) due to the small amount of data. Video-based data collection systems allow data to be collected
It therefore appears that both methods require about the same that have previously been unavailable using traditional tech-
Taylor et al. 139

TABLE 3 COMPARISON OF VIDEO AND RADAR DATA

Video Radar Prob. Value

Standard Standard
Path Type Mean Deviation Mean Deviation Mean Paired Means

A 25.9 7.5 24.7 6.7 0.33 0.0032

B 1,4.6 4.4 14.5 4.6 0.49 0.4700
c 13.4 3.1
D 22.9 5.5 20.3 5.3 0.18 0.0001
Aisles 14.9 4.3 t4.4 4.3 0.27 0.00só

TABLE 4 REGRESSION RELATIONSHIPS BETWEEN RADAR AND VIDEO DATA

lntercept Slope

Sample Standard Standard

Path Type Size Estimate Error Sign Estimate Error Sign Correlation

A 62 3.92 1.34 0.005 0.80 0.0s 0.00 0.90

B l4 0.01 1.30 0.99 1.00 0.09 0.99 0.96
D r3 -u.16 i.55 û.63 û.95 U.U / v,¿J
Aisles 10 0.01 t.20 0.98 0.95 0.0ó 0.25 0.95
Total 99 1.58 0.5ó 0.006 0.88 0.03 0.00 0.9s

niques (e.g., headways on multilane freeways and desired 7. A. J. Miller. An Empirical Model for Multilane Road Traffic
Flow. Transportation Science,Yol. 4,1970, pp. 164-186.
circulating speeds in car parks). This technology provides the
8. D. Mahalel and A. S. Hakkert. Traffic Arrival Patterns on a
traffic researcher with an unobtrusive method of observing Cross Section of a Multilane Highway. Transportation Research,
traffic movements. Video tapes of traffic events contain much Vol. A 174, No. 4, 1983, pp.263-270.
more information than manual recorders can collect' The 9. C-C. Chen. Lane by Lane Speed-Volume Studies on a Six-Lane
additional information becomes invaluable when modeling Freeway. Proc., 13th ARRB Conf., Vol. 13, No. 7, 1986, pp.
1-13.
complex traffic movements. The flexibility of being able to 10. D. L. Gerlough and F. C. Barnes. Poisson and Other Dßtribu-
position the on-screen detector points allows a wide range of tions in Trffic. Eno Foundation for Transportation, Sâugatuck,
traffic parameters to be collected. Conn.,1971.
The VADAS system offers a reasonably fast method of 11. W. Young, M. A. P. Taylor, and P. Gipps. Mícrocomputers ín
Traffic Engineerirzg. Research Studies Press, Letchworth, 1988.
collecting and extracting traffic survey data. Editing programs
72. K. H. Lenz and R. Hotop. Some Results Derived From Auto-
are necessary to ensure that the data extracted are accurate. matic Traffic Monitoring. Strassenv erkehrs technik I, 797 0.
Considerable time savings occur when operators become 13. J. Rorbech. The Multilane Traffic FIow Process. Ph.D. disser-
experienced with this technology. The technology also offers tation. Ministry of Public Works, Copenhagen,7974.
reasonable accuracy compared with traditional techniques, 14. R. G. Thompson, M. A. P. Taylor, and W. Young. TRANSTAT:
provided weather is appropriate and an adequate vantage A Microcomputer Program for Fitting Probability Distributions
to Traffic Data. Proc., 14th ARRB Conf . , 1988 (in press).
point can be found. 15. M. A. P. Taylor and W. Yotng. Traffic Analysß: New Tech-
nology and New Solutions. Hargreen Publishing Co., Melbourne,
1988.
REFERENCES 16. W. Young. An Exploratory Study of Speeds in Parking Lots.
Report AIR 466-1. Australian Road Research Board, 1987.
17. M. A. P. Taylor and W. Young. The Distribution of Speeds and
1. M. R. Wigan. Some Investigations of Image Processing Tech- Gaps in Traffic Studies.In Traffic Flow Theory (P. G. Gipps,
niques Applied to Road Problems. Proc., lst International ed.), Department of Civil Engineering, Monash University, 1984.
Cõnference on the Applications of Advanced Technology in 18. M. Armour. Vehicle Speeds in Residential Streets. Proc., llth
Transportation Engineering, ASCE, San Diego, Calif., i989, pp. ARRB Conf., Vol. 11, No.5, 1982, pp.90-205.
424-431. 19. J. R. Mclean. Two-Lane Road Traffic FIow and Capacit¡,. R¿psrt
2. R. Troutbeck and M. R. Wigan. Workshop on Visual Data Acquï AIR 359-6. Australian Road Resea¡ch Board, 1982, chap. 10.
sition and Applicanoz. Report ARR 1.51. Australian Road Research 20. Car Parking Facility: Off-Street Parking. Standards Association
Board, Vermont South, Victoria, 1988. of Australia, AS 2890.1, 1986.
3. R. Troutbeck and J. Dods. Video-Based Traffic Data Collec- 21,. M. A. P. Taylor and L. M. Rutherford. Speed Profiles at Slow
tion. In Trffic Analysis: New Technology and New Solution Points on Residential Streets. Proc., 13th ARRB Conf.,Yol. 13,
(M. A. P. Taylor and W. Young, eds.), Hargreen Publishing No. 9, 1986, pp.65-77.
Co., Melbourne, 1988.
4. D. J. Buckley. Road Traffic Headway Distributions. Proc., lst
ARRB Conf.,Yol. l,1962, pp. 153-183.
5. A. Schuhl. Poisson and Traffic. Eno Foundation for Transpor-
tation, Saugatuck, Conn., 1955, pp. 59-74.
6. R. J. Cowan. Useful Headway Models. Transportation Research, Publication of this paper sponsored by Committee on Applications of
Emerging Technology.
Vol. 9, No. 6,1975, pp.371.-375.