Institute of Transportation Studies
UC Berkeley Traffic Safety Center
( University of California, Berkeley)
Year 2007 Paper UCB - TSC - TR - 2007 - 3
Pedestrian Counting Methods at
Intersections: a Comparative Study
Mara Chagas Diogenes  Ryan Greene- Roesel†
Lindsay S. Arnold‡ David R. Ragland  
 UC Berkeley Traffic Safety Center
† UC Berkeley Traffic Safety Center
‡ UC Berkeley Traffic Safety Center
  UC Berkeley Traffic Safety Center
This paper is posted at the eScholarship Repository, University of California.
http:// repositories. cdlib. org/ its/ tsc/ UCB- TSC- TR- 2007- 3
Copyright  c 2007 by the authors.
Pedestrian Counting Methods at
Intersections: a Comparative Study
Abstract
Resources for implementing countermeasures to reduce pedestrian collisions
in urban centers are usually allocated on the basis of need, which is determined
by risk studies. They commonly rely on pedestrian volumes at intersections.
The methods used to estimate pedestrian volumes include direct counts and
surveys, but few studies have addressed the accuracy of these methods. This
paper investigates the accuracy of three common counting methods: manual
counts using sheets, manual counts using clickers, and manual counts using
video cameras. The counts took place in San Francisco. For the analysis, the
video image counts, with recordings made at the same time as the clicker and
sheet counts, were assumed to represent actual pedestrian volume. The re-sults
indicate that manual counts with either sheets or clickers systematically
underestimated pedestrian volumes. The error rates range from 8- 25%. Addi-tionally,
the error rate was greater at the beginning and end of the observation
period, possibly resulting from the observer’s lack of familiarity with the tasks
or fatigue.
Diógenes, Greene- Roesel, Arnold, & Ragland 1
Pedestrian Counting Methods at Intersections: a Comparative Study
Submission Date: August 1, 2006
Word Count: 2976
Number of Figures and Tables: 6
Authors:
Mara Chagas Diogenes
Traffic Safety Center
University of California, Berkeley
140 Warren Hall # 7360
Berkeley, CA 94709
510- 643- 7625
maracd@ berkeley. edu
Ryan Greene- Roesel
Traffic Safety Center
University of California, Berkeley
140 Warren Hall # 7360
Berkeley, CA 94709
510- 643- 7625
ryangr@ berkeley. edu
Lindsay S. Arnold
Traffic Safety Center
School of Public Health
University of California, Berkeley
140 Warren Hall
Berkeley, CA 94720
510- 282- 5896
larnold@ berkeley. edu
David R. Ragland ( corresponding author)
Traffic Safety Center
University of California, Berkeley
140 Warren Hall # 7360
Berkeley, CA 94709
510- 642- 0655
510- 643- 9922 ( Fax)
davidr@ berkeley. edu
TRB 2007 Annual Meeting CD- ROM Paper revised from original submittal.
Diógenes, Greene- Roesel, Arnold, & Ragland 2
ABSTRACT
Resources for implementing countermeasures to reduce pedestrian collisions in urban centers are
usually allocated on the basis of need, which is determined by risk studies. They commonly rely
on pedestrian volumes at intersections. The methods used to estimate pedestrian volumes include
direct counts and surveys, but few studies have addressed the accuracy of these methods. This
paper investigates the accuracy of three common counting methods: manual counts using sheets,
manual counts using clickers, and manual counts using video cameras. The counts took place in
San Francisco. For the analysis, the video image counts, with recordings made at the same time
as the clicker and sheet counts, were assumed to represent actual pedestrian volume. The results
indicate that manual counts with either sheets or clickers systematically underestimated
pedestrian volumes. The error rates range from 8- 25%. Additionally, the error rate was greater at
the beginning and end of the observation period, possibly resulting from the observer’s lack of
familiarity with the tasks or fatigue.
TRB 2007 Annual Meeting CD- ROM Paper revised from original submittal.
Diógenes, Greene- Roesel, Arnold, & Ragland 3
INTRODUCTION
Road collisions are a major public health concern throughout the world. It is estimated that 1.2
million traffic fatalities occur each year worldwide. The problem is especially acute for
pedestrians, who face a significantly greater risk of death when involved in traffic collisions than
do vehicle occupants ( 1). Significant resources are focused on countermeasures that aim to
reduce the risk of pedestrian injury. Because resources are limited, risk analysis is necessary to
develop cost- effective countermeasures ( 2).
Risk is defined as the frequency of an undesired event or collision per unit of exposure.
Pedestrian volume is the exposure measure most frequently used in risk analysis. According to
Gårder ( 3) pedestrian risk is closely correlated with pedestrian volume, more so than vehicle
volumes. Although many state, regional, and local agencies have developed methodologies to
collect pedestrian volume data, there is no consensus on which method is best ( 4, 5). To improve
the risk monitoring process, it is necessary to define a systematic pedestrian counting method.
The two most frequent types of pedestrian counting methods are direct counts and
surveys. Direct counts involve direct observation of pedestrian activity at fixed locations, such as
crosswalks or intersections. Surveys indirectly capture pedestrian activity in a geographic area by
gathering travel data from a sample ( 6).
Pedestrian volumes at intersections are usually collected directly using either ( i) manual
counts, taken by collectors in the field, or ( ii) automated counts using specialized equipment.
Although motorized vehicles are commonly counted with automated devices, the technology for
counting non- motorized modes of transportation, especially pedestrians, is not very developed
( 7).
The accuracy of these counting methods directly affects the accuracy of the exposure
estimate and thus the value of the risk analysis at an intersection. However, few studies have
attempted to compare the accuracy of different counting methods. This paper aims to compare
the accuracy of three common pedestrian counting methods: ( i) manual counts using sheets; ( ii)
manual counts using clickers; and ( iii) manual counts using video cameras.
METHODS
The research was conducted at 10 different intersections in the city of San Francisco, California,
during the last two weeks of April and the first week of May, 2006. Field observers collected
pedestrian counts with either sheets or manual clickers. Counts were taken for four hours
between 1: 00 pm and 6: 00 pm, with a break of one hour. Video footage of the intersection was
recorded simultaneously with the field counts.
Two persons were contracted from a private consulting firm specializing in data
collection. One individual made the field observations, and the other operated the video recorder.
The contracted staff was the same for all data collection. Sheets were used at eight intersections
and clickers at two intersections. The selected intersections had different pedestrian flows, with
values varying between 12 and 262 pedestrian crossings per hour based on the video analyses, as
shown in Table 1. Figures 1 and 2 present the camera angles used at two of the study
intersections.
Before the start of data collection, the researchers supplied the field staff with the
following directions:
TRB 2007 Annual Meeting CD- ROM Paper revised from original submittal.
Diógenes, Greene- Roesel, Arnold, & Ragland 4
1. The data collection must be synchronized with the video. The person collecting the
data should begin to count the pedestrians when the video begins to run. During the
period that the tape is being changed, the observer should stop counting.
2. The field observer must note any problem or interruption in the data collection, such as
a break or lack of attention for any reason. These interruptions are important since the
main objective was to compare the accuracy of the methods.
3. The field observer must count only pedestrians who cross the street centerline ( e. g. the
middle of the crossing). He or she should not count bicyclists unless they are walking
their bicycle across the intersection.
4. The field observer must stand close to the crosswalk.
TABLE 1 Data Collection Schedule and Pedestrian Flow
Intersection Date Method Volume
( ped)
Period
( hours)
Flow
( ped/ hour)
France and Mission St. 04/ 17/ 2006 Manual with sheets 128 4 32
Admiral Ave. and Mission St. 04/ 18/ 2006 Manual with sheets 49 4 12
16th St. and Capp 04/ 19/ 2006 Manual with sheets 412 4 103
Geneva and Mission St. 04/ 20/ 2006 Manual with sheets 1046 4 262
Folson and 7th St. 04/ 21/ 2006 Manual with sheets 334 4 84
Harrison and 7th St. 04/ 24/ 2006 Manual with sheets 651 4 163
Market and Castro 04/ 25/ 2006 Manual with sheets 579 4 145
Market and Noe 04/ 26/ 2006 Manual with sheets 994 4 249
Harrison and 10th St. 05/ 03/ 2006 Manual with clickers 161 4 40
Santa Rosa and Mission St. 05/ 05/ 2006 Manual with clickers 338 4 85
FIGURE 1 Camera angle used at Admiral Ave. and Mission St.
TRB 2007 Annual Meeting CD- ROM Paper revised from original submittal.
Diógenes, Greene- Roesel, Arnold, & Ragland 5
FIGURE 2 Camera angle used at Market and Castro ( still from video tape)
Field data were entered into a Microsoft Access 2000 database. For quality control, all
database tables were compared with the original field data sheets.
Manual with sheets
The field observer received a sheet with three fields: ( i) direction of travel; ( ii) pedestrian gender;
and ( iii) age. The observer was instructed to use his best judgment to assign the pedestrian to one
of seven age categories.
At the top of the sheet, the observer was instructed to write the following information: ( i)
name of the intersection; ( ii) his/ her name; ( iii) date of the data collection; and ( iv) period of the
data collection ( check box) – divided in periods of 30 minutes. The field observer was told to
concentrate on accurately counting the number of pedestrians, even if it meant leaving gender
and age fields blank in crowded intersections.
To improve the analysis, after the fourth day ( April 20), the field observer was asked,
when possible, to take note of any distinguishing characteristics that would allow an individual
to be identified in the video, i. e., clothing color, hair color, parcels or suitcases, exact time, and
so on. This information made it possible to determine when the field observer missed or over-counted
pedestrians, and to determine whether the manual data collection was properly
synchronized with the video.
Manual with clicker
On May 3 and May 5, the field staff collected pedestrian counts using a manual clicker. The
observer clicked once for every pedestrian crossing the intersection, regardless of direction. At
the end of every 10- minute period, the observer noted the count on the clicker on the data sheet
provided.
TRB 2007 Annual Meeting CD- ROM Paper revised from original submittal.
Diógenes, Greene- Roesel, Arnold, & Ragland 6
Manual with Video
The intersections were videotaped using a camera set up on a flatbed truck parked opposite the
crosswalk being studied. The camera recorded an image of the crosswalk at an angle that allowed
both directions of pedestrian travel to be captured. Video tapes were replaced after each hour.
Researchers involved in the study carefully analyzed the video tapes in order to obtain the
most reliable results possible. The researchers tried to identify each pedestrian counted by the
field observer. This task was only possible for the days that the field observer noted individual
pedestrian characteristics.
The tapes were viewed in variable time, and sometimes viewed more than once if the
results were in doubt. On average, one hour of video tape required three hours of video analysis.
During the analysis, the researchers paid attention to whether the field counts were synchronized
with the videotape and looked for any discrepancies between the field observations and the video
images.
DATA ANALYSIS
The purpose of the data analysis was to compare the accuracy of the methods. Because it was not
possible to know the exact number of pedestrians on the roadway at any given time, inter-reliability
between the methods was used as a proxy for accuracy. The counts derived from the
video tapes were assumed to be closest to the actual pedestrian volume.
The comparison used the relative difference between the counts taken through each
method to calculate the error:
NPv
NPi NPv
Error

= ( 1)
where NPi is the number of pedestrians counted in the field and NPv is the number of pedestrians
counted using the video images. The error was calculated for each interval of data collection ( 30
minutes for the sheets and 10 minutes for the clickers), as well as for the total number of
pedestrians counted at each intersection.
Synchronization of the field counts and video taping was a major issue identified during
the video analysis, despite the fact that field staff were directed to synchronize the counting
methods. Sometimes the field observer began counting slightly before or after the video camera
began recording. When this occurred, it was difficult to compare the counts obtained through
each method. To improve the results of the comparison study, counts taken in periods when the
field observer was not synchronized with the video were not included in the calculation of the
intersection error.
Comparisons of the accuracy of pedestrian gender and age identification were also made,
but not included in this paper. The researchers concluded that it was not possible to precisely
identify the gender or age of the pedestrians from the video images because of low image
resolution.
RESULTS
In the first week of data collection, the field observer did not follow all of the instructions he was
given and did not consistently collect data for four- hour periods. For example, he sometimes
TRB 2007 Annual Meeting CD- ROM Paper revised from original submittal.
Diógenes, Greene- Roesel, Arnold, & Ragland 7
started counting late; failed to take note of his breaks; and counted bicycles as pedestrians.
Despite this, the video tapes were analyzed for the entire counting period ( four hours) in order to
determine the average hourly pedestrian volume ( Table 1).
The results of the comparison reveal that the field observer systematically counted fewer
pedestrians than were observed on the video recordings. The average error calculated for the
manual counting using sheets was 15%, varying from 9% to 25%, as shown in Tables 2. For the
manual counting with clickers, the average error was 11%, varying from 8% to 15% ( Table 3).
Given the variation in the results, it is not possible to determine which method, with sheets or
clickers, is the most accurate.
TABLE 2 Comparison of Counting Methods ( Video vs. Sheets)
Date
4/ 17/ 2006 4/ 18/ 2006 4/ 19/ 2006 4/ 20/ 2006 4/ 21/ 2006 Period 4/ 24/ 2006 4/ 25/ 2006
Error Error Error Error Error Error Error
1: 00 to 1: 30 Not Counted Not Counted - 27% - 28% - 16% - 7% - 22%
1: 30 to 2: 00 150%* Not Counted - 18% - 6% 0% - 2% - 17%
2: 00 to 2: 30 - 13% 0% 3% - 23% - 17% - 29%
2: 30 to 3: 00 - 14% 0% - 28% - 2% - 12%
- 16%**
- 26%
4: 00 to 4: 30 - 13% - 22% - 42% - 14% - 8% - 8% - 27%
4: 30 to 5: 00 - 21% 86%* - 67% - 15% - 10% - 11% - 17%
5: 00 to 5: 30 Not Counted Not Counted - 25% - 16% - 5% - 3% - 25%
5: 30 to 6: 00 Not Counted Not Counted - 49% 3% - 8% - 10% - 31%
Error
( Total) - 15% - 11% - 21% - 12% - 10% - 9% - 25%
* Not included in the total, because it was not synchronized with the video
** In this period, the field observer failed to record the counts in half hour periods
TABLE 3 Comparison of Counting Methods ( Video vs. Clickers)
5/ 3/ 2006 5/ 5/ 2006
1: 00 to 2: 00pm 1: 00 to 2: 00pm
Error ( 10 min) - 11% - 43% - 13% 0% 0% 0% 0% 0% - 19% 17% - 8% 100%
Error ( hour) - 11% 2%
2: 00 to 3: 00pm 2: 00 to 3: 00pm
Error ( 10 min) - 25% - 67% 0% 100% - 50% 0% 0% - 14% 25% - 31% - 8% 9%
Error ( hour) - 23% - 5%
4: 00 to 5: 00pm 4: 00 to 5: 00pm
Error ( 10 min) 0% 17% 33% - 25% - 11% 0% 50% - 25% - 41% - 33% - 40% - 88%
Error ( hour) 0% - 32%
5: 00 to 6: 00pm 5: 00 to 6: 00pm
Error ( 10 min) - 20% 0% 38% - 33% 0% 20% - 30% 6% - 64% - 15% - 8% - 88%
Error ( hour) 0% - 21%
Error ( 4 hours) - 8% - 15%
An in- depth analysis of the data revealed that error was often greater at the beginning and
end of the data collection period. Possible explanations for this finding include: ( i) the observer’s
TRB 2007 Annual Meeting CD- ROM Paper revised from original submittal.
Diógenes, Greene- Roesel, Arnold, & Ragland 8
lack of familiarity with the intersection and the counting method at the beginning of the data
collection; ( ii) the long counting periods, which may have caused the observer to become
fatigued and lose attention; and ( iii) lack of synchronization with the video that was not possible
to identify.
It was assumed that the observer would have more difficulty counting at intersections
with high volumes of pedestrians, increasing the error value. However the results revealed that
pedestrian flow did not influence the error, since the correlation ( R ² = 0.1) between them was
weak. Figure 3 presents a graph with the relationship between the error and the pedestrian flow.
0%
5%
10%
15%
20%
25%
30%
0 50 100 150 200 250 300
Pedestrian Flow
( Ped/ h)
Error
FIGURE 3 Relationship between the error and the pedestrian flow
DISCUSSION
The most significant results of this study were that pedestrian counts taken in the field were
systematically lower than counts taken by observing video recordings, and that the accuracy of
field counts did not seem to be strongly related to pedestrian flow. These results stem from the
fact that the collection of field counts using either sheets or clickers is very difficult to control,
and requires planning and organization during the counting day ( 5).
The level of observer attention is one aspect of field data collection that is difficult to
control. In this study, the observer may have become distracted at intersections with little
pedestrian activity, but may have been more focused in areas with high activity that demanded
his attention. It is also possible that the error was related to the observer’s unique characteristics
and motivation. Future studies should use multiple field observers to determine how the
characteristics of the observers, such as their experience and background, affect the quality of the
pedestrian counts. However, given the budgetary constraints of most transportation agencies, it
may be difficult to ensure that field observers have high- level training and experience.
It was expected that manual counts taken with clickers would have very low error
because this method allows the observer to keep his attention on the intersection and does not
demand that he identify and record pedestrian characteristics. No significant difference was
TRB 2007 Annual Meeting CD- ROM Paper revised from original submittal.
Diógenes, Greene- Roesel, Arnold, & Ragland 9
found in the relative accuracy of manual counts using clickers and manual counts using sheets;
however, more research is needed to compare the methods.
Although this study suggests that field counts may be less accurate than counts taken with
video images, it is often necessary to use field observers to record detailed pedestrian
characteristics and behaviors. It is difficult to identify these characteristics on video recordings
without adequate image resolution and a well- selected camera angle.
This study suggests that video recordings should be used in situations where the accuracy
of the count is of primary importance. However, users of this method should be aware that
obtaining an accurate count from video can be very time consuming and requires meticulous
attention to the video analysis. Overall, the choice of pedestrian counting method depends on the
data collection needs and available resources.
TRB 2007 Annual Meeting CD- ROM Paper revised from original submittal.
Diógenes, Greene- Roesel, Arnold, & Ragland
10
ACKNOWLEDGEMENTS
The authors would like to thank Phyllis Orrick, Jill Cooper, Joseph Zheng, and others at the UC
Berkeley Traffic Safety Center for their ongoing support and contributions. Their comments and
suggestions helped make this paper possible. This work was partially funded through fellowship
support from the Brazilian Foundation for the Coordination of Higher Education and Graduate
Training ( CAPES) to Mara Diogenes and from the Eisenhower Transportation Fellowship
Program to Ryan Greene- Roesel.
TRB 2007 Annual Meeting CD- ROM Paper revised from original submittal.
Diógenes, Greene- Roesel, Arnold, & Ragland
11
REFERENCES
1. Peden, M., R. Scurfield, D. Sleet, D. Mohan, A. A. Hyder, E. Jarawan and C. Mathers.
World Report on Road Traffic Injury Prevention. World Health Organization, Geneva, 2004.
2. Høj, N. P. and W. Kröger. Risk Analyses of Transportation on Road and Railway from a
European Perspective. Safety Science, Vol. 40, No. 1- 4, 2002, pp. 337- 357.
3. Gårder, P. E. The impact of speed and other variables on pedestrian safety in Maine.
Accident Analysis & Prevention, Vol. 36, No. 4, 2004, pp. 533- 542.
4. Schneider, R., R. Patton, J. Toole and C. Raborn. Pedestrian and Bicycle Data Collection in
United States Communities: Quantifying Use, Surveying Users, and Documenting Facility
Extent. FHWA, Office of Natural and Human Environment, U. S. Department of
Transportation, Washington, D. C., 2005.
5. Schweizer, T. Methods for counting pedestrians. In The 6th International Conference on
Walking in the 21st Century. Walk21- VI " Everyday Walking Culture". Zurich, Switzerland,
2005.
6. Schwartz, W. and C. Porter. Bicycle and Pedestrian Data: Sources, Needs, and Gaps.
Publication BTS00- 02. Bureau of Transportation Statistics. U. S. Department of
Transportation, Washington, D. C., 2000.
7. Dharmaraju, R., D. A. Noyce, and J. D. Lehman. An Evaluation of Technologies for
Automated Detection and Classification of Pedestrians and Bicycles. In The 71st ITE Annual
Meeting. Compendium of Technical Papers. Institute of Transportation Engineering,
Chicago, IL, 2001.
TRB 2007 Annual Meeting CD- ROM Paper revised from original submittal.