Safety Science 91 (2017) 351360

Contents lists available at ScienceDirect

Safety Science
journal homepage: www.elsevier.com/locate/ssci

Studying critical pedestrian behavioral changes for the safety

assessment at signalized crosswalks
Wael K.M. Alhajyaseen Dr. Eng. a, Miho Iryo-Asano Dr. Eng. b,
a
Qatar Transportation and Traffic Safety Center, College of Engineering, Qatar University, P.O. Box 2713, Doha, Qatar
b
Institute of Industrial Science, The University of Tokyo, 4-6-1 Komaba, Meguro 153-8505, Tokyo, Japan

a r t i c l e i n f o a b s t r a c t

Article history: Pedestrian-vehicle conflicts are one of the most important safety concerns at signalized intersections
Received 12 December 2015 especially in urban areas. Conflicting vehicles and pedestrians select their maneuvers by predicting the
Received in revised form 3 August 2016 other user behavior. Sudden behavioral changes of pedestrians such as sudden speed change (accelera-
Accepted 5 September 2016
tion or deceleration) cannot be predicted by drivers, which can lead to safety hazards especially if these
Available online 14 September 2016
sudden behavioral changes occur near conflict areas. In this study, continuous pedestrian speed profiles
were analyzed at signalized crosswalks to investigate the existence of such behavioral changes. A
Keywords:
methodology is proposed to identify significant sudden changes of pedestrian speeds. The locations
Pedestrian speed
Pedestrian-vehicle conflicts
and timings of these sudden speed change events were analyzed, and the influencing factors such as
Pedestrian safety crosswalk geometry and signal timing were studies and evaluated. Five signalized crosswalks at three
Crosswalk geometry intersections in Nagoya City were videotaped for analysis. Individual pedestrian maneuvers were
Pedestrian traffic signal extracted with image processing software. Empirical analysis showed that sudden acceleration events
were observed at the entrance points to the pedestrian-vehicle conflict area, which highlights the signif-
icance of these events to pedestrian-vehicle conflict analysis. A multinomial logit model was developed to
estimate the probability of a pedestrian making a sudden speed change as a function of different influ-
encing factors. The results implied that the entering speed, necessary speed to finish crossing before
the onset of the pedestrian signal red phase, and crosswalk length have a significant impact on speed
change choices. This paper presents important suggestions to understanding pedestrian maneuvers in
detail from the viewpoint of safety.
2016 Elsevier Ltd. All rights reserved.

1. Introduction which leads to negligence of pedestrian provisions making pedes-

Vehicle and pedestrian flows in urban areas have increased
drastically over the years due to the increase in population density
and economic growth. This magnified the interactions between
pedestrians and vehicles to the degree that pedestrianvehicle
conflicts became the most severe traffic safety problem in urban
areas. Majority of pedestrian-vehicle crashes occurs at pedestrian
crossings where pedestrian maneuvers are significantly different
from those in ordinary walking spaces because of the impacts of
the crosswalk geometry, signal control, and presence of vehicles.
Policy makers around the world agree that pedestrian priority
and safety have always been their primary objectives. However
due to the increasing vehicle congestion in urban areas, the
emphasis is often on improving the efficiency of vehicle movement
Corresponding author.
E-mail addresses: wyaseen@qu.edu.qa, waelah1981@yahoo.com
(W.K.M. Alhajyaseen), m-iryo@iis.u-tokyo.ac.jp (M. Iryo-Asano).
trians vulnerable and prone to road crashes.
Most existing works have attempted to investigate variations in
the walking speed at other pedestrian facilities, such as walkways
and sidewalks. Few studies have addressed the issue of pedestrian
speed at signalized crosswalks, which have significantly different
operating and surrounding conditions compared to the others
(Alhajyaseen, 2014). Although pedestrians at signalized intersec-
tions are given priority during the green phase, turning vehicles
in conflict often do not give the right of way to pedestrians but
tend to compete for it. Furthermore, long crosswalks, a short avail-
able green time, and other reasons may cause pedestrians to rush
or suddenly change their speed without paying attention to the
surrounding conditions (Iryo-Asano et al., 2014). In addition to
other factors such as limited visibility and the intersection layout,
pedestrianvehicle accidents have become a major safety problem
in urban areas that has resulted in a high rate of fatalities (National
Police Agency in Japan, 2015). Many existing works on the mecha-
nism of such crashes have concentrated on the driver behavior,

http://dx.doi.org/10.1016/j.ssci.2016.09.002
0925-7535/ 2016 Elsevier Ltd. All rights reserved.
352 W.K.M. Alhajyaseen, M. Iryo-Asano / Safety Science 91 (2017) 351360
assuming that it is the most critical factor. In reality, road users
anticipate other users behavior in order to avoid collisions. Thus,
widely varying pedestrian behaviors and trajectories may result
in misunderstanding their decisions, which can lead to safety prob-
lems. Sudden changes in pedestrian behavior while they cross,
such as abrupt speed changes, cannot be predicted by the driver,
which can cause severe conflicts. There is no reference in the liter-
ature for studying the microscopic behavioral changes of pedestri-
ans at crosswalks from real world observations. Detection of the
sudden changes in pedestrian behavior is also demanding for
autonomous vehicles (Hashimoto et al., 2016). They need to ensure
safety by accurate estimation of potential conflicts with pedestri-
ans (Hashimoto et al., 2016). The existing methodologies to detect
pedestrians simply consider the existence of surrounding pedes-
trians (Dollar et al., 2012), while they do not consider the pedestri-
ans sudden behavioral changes that may cause hazardous conflict.
Therefore, understanding the characteristics of such behavior is
important also for development of safer avoidance algorithms for
autonomous vehicles. It is important to mention that assuming a
conservative reaction of autonomous vehicle whenever there is a
pedestrian approaching a crosswalk, will significantly deteriorate
the mobility levels and reduce capacity. What we are looking for
in this study, is to develop a methodology that can predict realistic
pedestrian behaviors and as a result accurate arrivals at conflict
areas which can be used to take informed and appropriate deci-
sions by drivers or autonomous vehicles. This will improve mobil-
ity levels will insuring safety.
In previous studies, authors analyzed pedestrian speeds at sig-
nalized crosswalks and found that the walking speed increases
with the elapsed time since the onset of the pedestrian green
(PG) phase (Zhang et al., 2013). Furthermore, empirical analysis
showed that pedestrians hurry when entering crosswalks as their
green light flashes and then tend to significantly decrease their
speed while crossing (Iryo-Asano et al., 2014). A methodology
was proposed to estimate pedestrian travel speeds in the first
and second halves of the crosswalk as a function of the crosswalk
geometry and signal timings to estimate the needed crossing time
when setting the pedestrian signal times (Iryo-Asano et al., 2014).
Previous works on pedestrian speed profiles divided them into
consecutive constant travel speeds (first and second halves of the
crosswalk), which is sufficient for analyzing the pedestrian signal
time settings but not for assessing the safety of pedestrianvehicle
conflicts. Studying the instantaneous pedestrian speeds and identi-
fying possible sudden changes in pedestrian behavior while cross-
ing is crucial for pedestrian safety analysis at crosswalks. The
existence of such changes in the pedestrian speeds should be ver-
ified so that surrogate safety measures such as post-encroachment
time can be quantitatively estimated to provide a reliable assess-
ment of pedestrianvehicle conflicts and their severity. This paper
is a continuation of previous efforts (Iryo-Asano et al., 2014; Iryo-
Asano and Alhajyaseen, 2014) aiming to investigate pedestrian
behavioral changes at signalized crosswalks in terms of speed,
and to identify influencing factors.
2. Literature review
In order to analyze the interactions between pedestrians and
vehicles at intersections, it is important to gain better insights into
the behavior of both pedestrians and vehicles. Most existing stud-
ies have concentrated on vehicle maneuvers, including speed,
speed change, and gap acceptance behaviors; they assumed that
vehicles are the main contributing element in pedestrianvehicle
crashes. Meanwhile, severe conflicts occur when road users fail
to predict and properly react to other users decisions. The widely
varying behaviors and maneuvers of vehicles and pedestrians as
well may lead to misunderstanding their decisions, which can
result in severe conflicts.
Alhajyaseen et al. (2013a, 2013b, 2012a, 2012b) intensively
analyzed turning vehicle maneuvers at intersections in Japan,
including the paths, speed profiles, and gap acceptance. They found
significant variations in vehicle paths and speeds at conflict points
with pedestrians to be functions of the vehicle approaching speed,
crosswalk location, and intersection geometry. These variations
significantly affect the conflict area with pedestrians and the sever-
ities of the conflicts. As a complement, they analyzed pedestrian
behavior microscopically (concentrating on crossing decision and
speed) considering the impacts of the crosswalk geometry and sig-
nal timing parameters (Iryo-Asano et al., 2014; Iryo-Asano and
Alhajyaseen, 2014). They concluded that pedestrian maneuvers
widely vary depending on many factors, such as the timing of
the crossing, signal indication, pedestrian speed, origin and desti-
nation, probability of encountering a turning vehicle, crosswalk
geometry (length and width), presence of other pedestrians on
the crosswalk, and cultural background. Such differing behaviors
make it difficult for drivers to correctly predict pedestrian deci-
sions, which increases the probability of improper maneuvers that
put pedestrian safety at risk. Koh and Wong (2014) reached similar
results; in their empirical analysis, they found that pedestrian
crossing speeds significantly differ during the pedestrian green
(PG) phase compared to the pedestrian flashing green (PFG) phase.
Schmitz (2011) confirmed that the type of pedestrian signal signif-
icantly affects the pedestrian behavior and crossing speed. He
found that pedestrian countdown timers increase the pedestrian
crossing speed. Other studies (Supernak et al., 2013; Yang and
Sun, 2013; Xu et al., 2013) have confirmed that aspects of the
crosswalk layout such as the length, width, and presence of refuge
islands have a significant effect on pedestrian compliance to
signals.
Wang et al. (2011) studied the pedestrian compliance to signals
and found that the probability of pedestrians violating a signal is
influenced by the waiting time or delay, personal characteristics
(e.g., age and gender), trip purpose, and traffic conditions (e.g.,
pedestrian flow rate and vehicular traffic volume). In a recent study
by Hashimoto et al. (2016), a model based on Dynamic Bayesian
Network is proposed to estimate pedestrian stop-go decision when
approaching crosswalks. They claimed that the model is able to
recognize the pedestrian crossing decision in a few seconds from
the traffic signal and pedestrian position information. However
pedestrian behavior while crossing including potential velocity
changes is not analyzed. They highlighted the importance of such
models for the active safety systems which is highly needed in traf-
fic condition with growing demands in autonomous driving and
Connected Vehicles.
Several other studies have analyzed the pedestrian average
walking speed as an important factor for the operation and design
of pedestrian facilities. Navin and Wheeler (1969) and Fruin (1971)
were the first to perform macroscopic pedestrian speed analysis
and have been followed by many researchers. The effects of age,
gender, and other attributes have been addressed in several studies
(e.g., Avineri et al., 2012) on providing slow walkers such as elderly
people with the necessary clearance time (LaPlante and Kaeser,
2007). However, the pedestrian speed at crosswalks is important
to not only determining the clearance time but also analyzing
the probability of conflict with turning vehicles and the severity
of these conflicts.
Several studies have analyzed pedestrian speeds at various
facilities (e.g., sidewalks, walkways, and crosswalks) by consider-
ing the seasonality, gender, and age (Montufar et al., 2007;
Tarawneh, 2001; Lam and Cheung, 2000). They all yielded the same
conclusion that pedestrian speeds are significantly higher at cross-
walks than at sidewalks and walkways. They suggested that the
 W.K.M. Alhajyaseen, M. Iryo-Asano / Safety Science 91 (2017) 351360 353

surrounding conditions of signalized crosswalks, such as the pres-

ence of turning vehicles and signal control, significantly affect the
pedestrian behavior and crossing speed.
Lee and Lam (2008) developed a simulation model to evaluate
the performance of a crosswalk. Their model includes the stop/
cross probability during the pedestrian flashing green phase and
the interaction among pedestrians to represent the reduction in
speed of crowds. However, their developed model does not con-
sider the effects of the crosswalk length, conflict with turning traf-
fic, and pedestrian signal timing.
All of the previous findings highlighted the significance of
understanding pedestrian decisions and identifying influencing
factors. However, previous studies analyzed the speed profiles of
pedestrians at crosswalks by dividing them spatially into several
constant crossing speeds and then investigating the influencing
factors. This procedure hides important information such as sud-
den changes in pedestrian behavior, which can be identified by sig-
nificant changes in their instantaneous speeds. Iryo-Asano et al.
(2014) indicated that some pedestrians exhibit significant changes
in their speed profiles, which could be a reaction to pedestrian sig-
nal indications, the crosswalk layout, or a combination of different
factors. This may cause pedestrians to arrive more quickly to con-
flict areas or stay longer in such areas. Either way, drivers cannot
anticipate such abrupt behavioral changes, which may lead to sev-
ere conflicts with pedestrians. The aim of this study was to inves-
tigate this phenomenon and identify influencing factors.
3. Study sites and data observation
3.1. Study sites and data collection
Empirical observation is essential to analyzing instantaneous
pedestrian speeds to verify significant behavioral changes. Thus,
several sites with low-to-medium pedestrian demand were
selected for a video survey (Table 1). Such sites were selected to
minimize the impact of interactions between pedestrians. With
high pedestrian demand, the interaction between pedestrians sig-
nificantly affects their maneuvers, which makes identifying the
impact of the crosswalk geometry and signal timing very difficult
and even impossible. When referring to turning traffic movements
and conflict area with pedestrians, it is important to recall that
Japan has a left-hand traffic system where vehicles are driven in
the left-side of the road.
Five approaches at three signalized intersections were
videotaped. All of these sites were in Nagoya City, Japan. Table 1
presents the geometric characteristics of the observed sites. The
parameters in Table 1 are defined in Fig. 1. The observed cross-
walks had significantly different geometric layouts with regard to
the length, setback distance and corner radius.
All sites had shared left turn-through phases (as given in
Table 2), while left turns during the red phase were prohibited.
Pedestrians shared the same signal phase with the through and
left-turning traffic in the same direction. Thus, left-turning traffic
Pedestrian B
(Far-side)
Far-side
Distance Ds
( )
Conflict area with
turning vehicle
Near-side
Pedestrian
(Near-side)
Fig. 1. Definitions of parameters related to crosswalk geometry and pedestrian
movement.
frequent conflicted with crossing pedestrians depending on the
demand and arrival pattern.
Table 3 lists the average demands of the pedestrians and turn-
ing traffic during the observation period. The pedestrian demand
was divided into the near and far sides in reference to the exiting
vehicular traffic, as shown in Fig. 1. Near-side pedestrians started
crossing from the side of the vehicular traffic exiting the intersec-
tion, while far-side pedestrians started crossing from the side of
the incoming vehicular traffic (Alhajyaseen et al., 2013a, 2013b).
The observed numbers of heavy vehicles were very low at all sites;
thus, they were neglected. Table 3 indicates that pedestrian and
turning traffic demands were low to medium, which made these
locations suitable for analysis of the pedestrian instantaneous
speeds and the impacts of the crosswalk geometry and signal
timing.
The analysis sites have no pedestrian countdown signals or any
means for information on remaining time. However, it is reason-
able to assume that pedestrians roughly predict the remaining
time. Firstly, the signal timing at observation sites is pre-timed
which means that the length of pedestrian green and flash green
times are fixed through the day. Furthermore, the study sites are
located either in business or in residential areas, thus most of the
pedestrians are familiar with the green and flashing green length.
3.2. Trajectory tracking
The video image processing system TrafficAnalyzer (Suzuki and
Nakamura, 2006) was used to track pedestrians from the

Table 1
Surveyed sites for pedestrian analysis.

Intersection Subject Survey date Radius of Intersection Number of Crosswalk setback Crosswalk Crosswalk
name crosswalk corner angle hl () exit lanes NO distance DS (m) length L (m) width (m)
RC (m)
Kanayama East 9:0013:00, 10/19/2012 13.4 93 1 5.0 16.2 5.8
North 9:3013:00, 10/19/2012 8.0 86 3 12.0 36.2 6.0
Ueda East 7:0010:00 and 14:0016:30, 11/2930/2012 11.5 65 2 7.5 28.7 6.3
South 14:0016:30, 11/29/2012 14.5 119 2 18.9 20.8 5.8
7:0010:00 and 14:0016:30, 11/30/2012
Fushimi South 10:0011:00 and 14:0015:00, 11/5/2012 12.2 90 3 13.3 30.4 6.9
354 W.K.M. Alhajyaseen, M. Iryo-Asano / Safety Science 91 (2017) 351360

Table 2
Signal timing plans at observed intersections.

Signal phasing length (sec) Cycle

Mode 1 2 3 4 length
1 2 3 4 5 6 7 8 9 10 11 12 13 14 (sec)
Vehicle
E-W Pedestrian (location S and N)
Right-turning vehicle Shared
Vehicle
S-N Pedestrian (location E and W)
Right-turning vehicle Shared
Kanayama 39 9 3 3 7 2 5 54 6 5 3 17 2 5 160
Intersection
Ueda 54 8 2 3 9 2 5 45 10 4 4 7 2 5 160
name
Fushimi 40 10 2 4 7 2 5 62 7 3 4 8 1 5 160

Signal phase plan

Green Right-turning arrow Pedestrian flashing green Amber Red

Table 3
Turning traffic and pedestrian demands at study sites.

Intersection name Subject crosswalk Turning vehicle demand (veh/h) Pedestrian demand (ped/h) Number of observed
pedestrians (ped)
Left turning Right turning Near side Far side Near side Far side
Kanayama East 148 56 103 76 189 74
North 124 36 185 153 106 267
Ueda East 46 264 49 41 39 32
South 176 52 58 56 64 71
Fushimi South 122 112 155 167 56 72

approaching sidewalk until they finished crossing. This system is
designed to support the manual tracking of vehicle/pedestrian tra-
jectories from video data (Fig. 2). Pedestrian positions were manu-
ally extracted every 0.5 s; then, their video coordinates were
converted to global coordinates by projective transformation. The
parameters of the projective transformation for each video image
were estimated with the least-squares method under the assump-
tion that the entire road surface was on a unique plane. Fig. 2
shows the inner orientation results for the camera view; the mean
absolute error of the transformation was 35 cm, and R2 of the
observed and estimated positions was 0.999, which means that
the perspective effect was negligible. Table 4 shows the absolute
coordinate transformation error when using TrafficAnalyzer. It also
summarizes the possible measurement error caused by the manual
tracking, when the observer chooses the adjacent pixel instead of
the correct one. The possible mean error caused by the manual
tracking was 5 cm.
3.3. Empirical observation of pedestrian speed change on crosswalks
Note that pedestrians waiting at the sidewalk at the beginning
of the PG phase at all sites were excluded from the analysis
because their trajectories are highly influenced by other pedestri-
ans crossing with them to form a platoon. Table 3 lists the number
of observed pedestrians who were used in the analysis. Here,
Xt xt; yt was defined as the position of a pedestrian mea-
sured at the time t. The pedestrian speed v t at the time t can be
extracted from the row data as follows:
jXt Dt  Xtj
v t 1
Dt
Although TrafficAnalyzer has a function of Kalman smoothing to
get the smoothed trajectory including speed and acceleration
information, this is not utilized in this study. The results of Kalman
smoothing method are dependent on the range of pedestrians
acceleration/deceleration and the range of observation
error. As these two has similar order in case of the pedestrian
tracking, the smoothing results remove the observation error as
well as the sudden acceleration/deceleration. Thus, it is difficult
to extract the speed change events once the smoothing method
is applied. The speed change events are extracted directly from
the location profile information following Eq. (1).
Fig. 3 demonstrates how the pedestrian speed change event
affects conflicts between pedestrians and vehicles. This is an exam-
ple of a pedestrian trajectory observed at the south crosswalk of
Fushimi Intersection. The vertical axis of Fig. 3(a) indicates the
spacing from the near-side edge of the crosswalk, while the hori-
zontal axis indicates the elapsed time since the onset of the PG
phase. The area surrounded by the red box is the conflict area with
the turning traffic. In this example, the pedestrian did not actually
meet a turning vehicle. However, consider the case as if a turning
vehicle is approaching.
This pedestrian crossed from the far side to the near side and
accelerated in the middle of the crosswalk just before entering
the conflict area. If there is a turning vehicle at the same time,
the driver would adjust his or her speed based on the expected
arrival time of the pedestrian to the conflict area according to the
 W.K.M. Alhajyaseen, M. Iryo-Asano / Safety Science 91 (2017) 351360 355

Manually tracked
pedestrian positions

Fig. 2. Image processing system TrafficAnalyzer used to track pedestrians.

Table 4
Coordinate transformation error and possible measurement error by manual tracking.

Site Absolute coordinate transformation error (m) The difference of 1 pixel on the screen (m)
Min. Max. Ave. Min. Max. Ave.
Kanayama East 0.00 0.76 0.35 0.00 0.10 0.03
North 0.02 0.62 0.18 0.01 0.26 0.05
Ueda East 0.01 0.48 0.12 0.01 0.07 0.02
South 0.01 0.57 0.26 0.01 0.12 0.05
Fushimi South 0.03 0.61 0.27 0.02 0.42 0.04

Far-side

35 6

30
Expected pedestrian 5
Spacing from near-side

25 arrival time
crosswalk edge

Actual pedestrian 4
20
Speed (m/s)

arrival time
15 3

10
2
5
1
0
The Pedestrian is on the crosswalk
30 35 40 45 50
Near-side -5 0
Elapsed time from onset of pedestrian green 30 35 40 45 50
(second) Elapsed time from onset of pedestrian green
(second)
(a) Trajectory and expected/actual arrival time to conflict area (b) Speed profile
Fig. 3. Sudden speed change of pedestrian and its impact on arrival time to conflict area.
356 W.K.M. Alhajyaseen, M. Iryo-Asano / Safety Science 91 (2017) 351360
pedestrians movement. However, the sudden sharp increase in the
pedestrian speed would lead the pedestrian to arrive early to the
conflict area by 2 s (Fig. 3), which cannot be predicted by the dri-
ver. This time is not sufficient for the driver to react and avoid
the hazardous situation. Therefore, this study focused on signifi-
cant and sudden accelerations/decelerations of pedestrians. The
next section introduces how speed change events can be extracted
from individual pedestrian speed profiles.
4. Extraction of sudden speed change events
4.1. Stepwise speed function
As shown in Fig. 3(b), the frequent minor fluctuations in pedes-
trian speeds were due to the tracking error from the video data. To
minimize the impact of these measurement errors, the following
assumptions were made to clean the data:
(1) Pedestrians do not frequently change their speed.
(2) Pedestrians instantly change their speed, so that the dura-
tion of the speed change can be neglected.
The first assumption means that, under ideal conditions, pedes-
trians tend to keep their speed with minor fluctuations while
crossing. However, if for any reason the pedestrian decides to
change his or her speed, this change was assumed to be instanta-
neous with a negligible duration of occurrence. Under these
assumptions, a stepwise function can be applied to represent
pedestrian speed profiles.
The stepwise function v s t is defined in Eq. (2).
8 According to the data obtained by Alhajyaseen et al. (2011),
>
>
v1 where t < t 1
>
< v2 where t 1 6 t < t2
v s t >
>
> ...
:
v m1 where t m 6 t
where m is the number of speed changing events, ti is the timing of
the speed change event i, and v i is the constant speed during the
time period between ti and t i1 . The value of m differs for each indi-
vidual speed profile and should be 0 if there are no speed changes.
The speed profiles of each pedestrian can be fitted to this stepwise
function by determining t i , v i , and m.
4.2. Extracting feasible speed change events
The shapes of the fitted speed profiles strongly depend on the
value of m. A larger m value means that the stepwise function bet-
ter fits the original speed profile with a smaller error. In extreme
cases, the original and fitted profiles are equivalent if m + 1 is equal
to the number of observed samples of the speed profile. However,
the focus here was to identify significant speed changes only. Con-
sidering the purpose of this analysis, only the timings when signif-
icant speed changes occur should be extracted. If the number of
speed changes m is fixed, other parameters can be uniquely esti-
mated to minimize the sum of the least-squares error between
the estimated and observed speed profiles.
Fig. 4 shows example fittings with m = 1 and 2 for observed
speed profiles of two individual pedestrians A and B. m = 2 of
Fig. 4(a) shows excessively small speed changes that may be negli-
gible, and m = 1 in Fig. 4(b) shows that the sudden acceleration
event was not followed. From these figures, Pedestrian A only
had one significant speed change event, while Pedestrian B had
two.
To determine the appropriate number of speed changes m, the
speeds before and after the extracted speed change events must
be sufficiently high. Furthermore, the fluctuation in speed during
each period between speed change events should be considered.
Even though the difference in mean speed between adjacent peri-
ods can be large, the extracted event may not be feasibly estimated
if the variation in speed is quite high during these periods. There-
fore, a statistical test is required to judge the significance of mean
speed differences.
Considering the above discussion, an algorithm to distinguish
significant speed change events is proposed, as shown in Fig. 5.
The algorithm is applied to each observed speed profile. The pro-
cess starts with fitting the observed speed profile to the maximum
number of m. Then, the output is examined. This fitted profile is
selected if all of the speed change events are identified as feasible.
However, if at least one speed change event is identified as unfea-
sible, the speed profile is fitted again with m  1 speed changes.
Then, it is examined again. This process continues until all of speed
changing events are feasible as demonstrated below:
Step 1: Set i = mmax, which is the maximum number of possible
speed change events in one speed profile.
Step 2: Estimate parameters of the stepwise function with the
given m = i that minimizes the sum of the squared error.
Step 3: The estimated speed profile is examined according to the
following two points. If both are satisfied, the speed profile of m
is selected as feasible for the subject pedestrian. If not, i is
updated to i 1, and the process goes back to Step 2.
Step 3-1: Check whether or not the difference before and
after each extracted speed change event is larger than the
threshold a. The following inequality has to be satisfied:
jv k1  v k j > a; for 8k 2 1; i 3
the difference between the 90th and 10th percentiles for the
speed distribution of each age group under the free flow con-
2
dition is approximately 0.5 m/s. This implies that a speed
change of >0.5 m/s can be considered significant. Therefore,
a was set to 0.5.
Step 3-2: Check whether the difference before and after each
extracted speed change event is statistically significant. The
Students t-test is applied to compare the difference in the
means of two speed distributions. The start and end times
of the observation are defined as t0 and ti+1, respectively.
The t-test is applied between observed speed distributions
during t k1 6 t < tk and tk 6 t < t k1 . The null hypothesis is
that the means of the two distributions are the same. If
the null hypothesis is rejected with a 95% confidence level
for 8k 2 1; i, the speed profile satisfies this condition.
5. Results and analysis
5.1. Base statistics of speed change events
Fig. 6 shows the number of observed speed profiles classified by
the number of extracted speed change events. More than 90% of
the pedestrians had a maximum of two speed changes, while more
than half did not show any significant change in speed. The per-
centage of speed change occurrence was clearly significantly
higher at longer crosswalks such as Kanayama North and Fushimi
South. This is rational because pedestrians have long travel times
at long crosswalks, which means a greater opportunity for speed
changes. Furthermore, at long crosswalks, pedestrians have a
higher probability to be on the crosswalk during the clearance time
because of the limited available green time. Pedestrians who cross
during clearance times have significantly higher crossing speeds
(Iryo-Asano et al., 2014).
Fig. 7 shows the location distribution of the speed change
events on the crosswalks. The horizontal axis shows the normal-
 W.K.M. Alhajyaseen, M. Iryo-Asano / Safety Science 91 (2017) 351360 357

4 4 Observed data
Observed data Onset of PFG Estimated with m = 1
3.5 3.5
Estimated with m = 2 End of PFG Estimated with m = 2
3 3

Speed (m/s)
Estimated with m = 1
Speed (m/s)

2.5 2.5
2 2
1.5 1.5
1 1
0.5 0.5
0 0
10 20 30 40 50 60 30 40 50 60 70
Elapsed time from onset of PG (s) Elapsed time from onset of PG (s)
(a) Pedestrian A (b) Pedestrian B
Fig. 4. Observed and fitted speed profiles of m = 1 and 2.

1. Set i = mmax were no clear tendencies for the near-side pedestrians, while the
far-side pedestrians decelerated either when they started crossing
2. Estimate stepwise function with m = i or when they almost completed crossing as shown in Fig. 7(b).
Fig. 8 shows the distribution of the timing when the pedestrians
changed speed. The origin of the horizontal axis represents the
Not significantly onset of the PFG phase, and the positive values along the axis mean
3. Evaluate difference of mean between different
distribution of v(t k t < t k +1 ) and that the speed change events occurred after the onset of the PFG
v(t k 1 t < t k ) for k [1, i ] phase. More than a quarter of all acceleration events occurred just
after the onset of the PFG phase. This indicates that pedestrians
Significantly different tried to hurry when they recognized the PFG indication. On the
other hand, 15% of deceleration events occurred 10 s after the
4. | vk +1 vk |> No onset of the PFG phase, which is close to or even after the onset
for k [1, i ] ? Update i  i -1 of the pedestrian red phase. Those pedestrians were probably at
Yes
No the end of the crossing, so they felt safe.
i =C0 ?

Yes 5.2. Necessary speed to complete crossing

The stepwise function with m = i is Ideally, pedestrians should complete their crossing by the end
selected as the feasible function
of the PFG phase. In this study, the minimum speed at which a
Fig. 5. Flowchart of feasible speed profile estimation. pedestrian can complete the crossing before the onset of the
pedestrian red phase was designated the necessary speed vnec
and used as an index to evaluate the actual walking speeds. vnec(t)
100% 35 17 4 2 at the time t is calculated as follows:
5 8 14
90% 4
Percentage of samples

40 49
Remaining distance to complete crossing at time t
80%
70%
10
41
41 v nec t 4
123 Remaining time till the end of PFG at time t
60%
50%
40% 215
Consider a pedestrian entering at the time t 0 . This pedestrian needs
30%
52
82 71 to walk constantly with a speed v P v nec t 0 in order to complete
184
20% the crossing before the PFG phase ends. Using this necessary speed
10%
0%
concept, the following behavior can be hypothesized. If the entering
Kanayama E Kanayama N Ueda E Ueda S Fushimi S speed v t 0 at t 0 is less than v nec t 0 , the pedestrian is expected to
m=0 m=1 m=2 m3 accelerate in order to not remain on the crosswalk after the end
of the PFG phase. If the entering speed v t0 P v nec t0 , the pedes-
Fig. 6. Number of observed speed profiles classified by number of identified speed
trian does not need to accelerate. Some pedestrians enter cross-
change events.
walks at faster than their desired walking speeds. In such cases,
they may even decelerate to their desired speeds.
Figs. 9 and 10 are observed distributions of the speed change
ized distance, where the near-side crosswalk edge was set to 0 and probabilities. Fig. 9(a) shows the relationship between v nec t 0
the far-side edge was set to 1. The vertical axis shows the percent- and the probability of the first speed change event occurrence.
age of the speed change event occurrence calculated by dividing Those with more than one speed change event (i.e., m > 1) are clas-
the number of speed change events with the number of pedestri- sified as either accelerating or decelerating based on the first speed
ans who changed speed. change event. When v nec is low, most of the pedestrians do not
The near-side pedestrians tended to accelerate at the beginning change their speeds. As v nec increases, the probabilities of both
of the crossing, where they may conflict with turning vehicles. acceleration and deceleration events become higher. Fig. 9(b)
Meanwhile, far-side pedestrians accelerated around the middle of shows a similar relationship between the probability and
the crosswalk before entering the conflict area with the turning v nec  v 1 , where v 1 is the mean speed of the time period when
traffic, which indicates that they wanted to clear the conflict area pedestrians start crossing (entering speed), as given in Eq. (2).
faster as shown in Fig. 7(a). Regarding deceleration events, there The probability of deceleration is significantly high when the value
358 W.K.M. Alhajyaseen, M. Iryo-Asano / Safety Science 91 (2017) 351360

30% 30%
Percentage of speed change events

Percentage of speed change events

25% 25%
Walking directions Walking directions
20% 20%

15% 15%

10% 10%

5% 5%

0% 0%
<0.1 <0.2 <0.3 <0.4 <0.5 <0.6 <0.7 <0.8 <0.9 <1 <0.1 <0.2 <0.3 <0.4 <0.5 <0.6 <0.7 <0.8 <0.9 <1
Normalized distance from nearside edge of crosswalk Normalized distance from nearside edge of crosswalk

Nearside (N=30) Farside (N=104) Nearside (N=130) Farside (N=194)

Far-side
Near-side

Near-side
Far-side

(a) Location of acceleration events (b) Location of deceleration events

Fig. 7. Location distribution of speed change events.

30% of v nec  v 1 is small. These results support the above hypothesis

that pedestrians are less motivated to keep walking faster if their
Percentage of occurence

Accelerate (N = 134)
25% current speed is much higher than necessary. If the value of
Decelerate (N = 324)
20%
v nec  v 1 is large, pedestrians can choose to either accelerate or
decelerate. The acceleration behavior suits the hypothesis, while
15% the deceleration behavior is contradictory because the pedestrians
decelerate even though they cannot complete the crossing before
10% the end of the PFG phase.
To further understand these behaviors, Fig. 10 shows the posi-
5%
tion distributions of deceleration events with different ranges of
0% v nec  v 1 . Pedestrians with positive v nec  v 1 (i.e., those who could
-50 -40 -30 -20 -10 0 10 20 30 not complete the crossing with their entering speed) clearly decel-
Elapsed time from the onset of PFG (s) erated on the latter half of crosswalks because they may have felt
safe and confident in finishing the crossing as they approached the
Fig. 8. Distribution of the timing of speed change events. end of the crosswalk. On the other hand, when v nec  v 1 was neg-
ative, most of the deceleration events occurred at the beginning of

100%
100%
90%
90%
80%
80%
Choice probability
Choice probability

70%
70%
60%
60%
50%
50%
40%
40%
30%
30%
20%
20%
10%
10%
0%
0%

Vnec - V1 (m/s)
Vnec (m/s)
Decelerate No speed changes Accelerate Decelerate No speed changes Accelerate

(a) v nec (b) difference between v nec and v1

Fig. 9. Impact of v nec on probability of speed change events.

 W.K.M. Alhajyaseen, M. Iryo-Asano / Safety Science 91 (2017) 351360 359

35% Table 5
Multinomial logit model for occurrence of speed change events.
30% Vnec - V1 > 0 (N=98)
Percentage of occurence

Alternatives Variables Estimate t-value

Vnec - V1 <= 0 (N=226)
25%
Acceleration Intercept 3.20 7.17a
Entering speed v1 (m/s) 0.213 0.862
20%
vnec  v1 (positive only) (m/s) 0.289 2.93b
15% Crosswalk length (m) 0.0594 6.35a
Deceleration Intercept 9.15 16.97a
10%
Entering speed v1 (m/s) 4.05 15.89a
5% vnec  v1 (positive only) (m/s) 0.262 2.65b
Crosswalk length (m) 0.0350 3.37a
0% Number of samples 1192
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Log likelihood 790.9
Normalized distance from starting edge of the crosswalk Initial log likelihood 1200.8
McFadden R2 0.341
Fig. 10. Position distribution of deceleration events with different ranges of a
Significance level: 0.1%.
v nec  v 1 . b
Significance level: 1%.

the crosswalk. This is rational because the available time to the end
of the PFG phase was more than enough to finish crossing, which
may encourage pedestrians to slow down.
6. Modeling speed change probability
Based on the above discussion, a multinomial logit model is
proposed to estimate the probabilities that a pedestrian will accel-
erate, decelerate, or not change speed during a crossing. In the
multinomial logit model, the probability of each alternative is
given by the following functions:
expF acc
Pracc
expF acc expF dec expF noc
expF dec
Prdec
expF acc expF dec expF noc
expF noc
Prnoc
expF acc expF dec expF noc
where Pr acc , Pr dec , and Pr noc are the probabilities to choose the alter-
natives of acceleration, deceleration, and no speed change, respec-
tively, and F acc , F dec , and F noc are deterministic terms of utility of
the alternatives. The deterministic utility terms of the acceleration
and deceleration alternatives were assumed to be linear functions
of the explanatory variables given in Eqs. (8) and (9):
F acc b0;acc b1;acc v 1 b2;acc d2 v nec t 0  v 1 b3;acc L
F dec b0;dec b1;dec v 1 b2;dec d2 v nec t 0  v 1 b3;dec L
F noc 0
where bn;acc and bn;dec are the parameters to be estimated. v 1 is the
mean speed until t 1 , as defined in Eq. (2), and v nec t 0  v 1 is the
same value as presented in Fig. 10. d2 is set to 1 if
v nec t0  v 1 > 0; otherwise, it is set to 0. L is the crosswalk length.
The deterministic utility of the no speed change alternative is set to
the base utility with a value of 0, because these functions simply
evaluate the relative difference between the utilities of alternatives.
Table 5 gives the modeling results using the collected empirical
data. As the signs for all of the coefficients of the acceleration and
the deceleration alternatives were similar, the relative differences
in the coefficient values needed to be examined. The entering
speed was found to have a fairly strong impact upon the selection
of deceleration. A positive v nec t 0  v 1 value was found to increase
the probabilities for both acceleration and deceleration. The cross-
walk length was found to increase the possibility of either speed
change, though it has a larger impact on the acceleration alterna-
tive. All of these results were feasible and did not contradict the
empirical analysis results. Other intersection layout parameters,
such as curb radii, intersection angles, setback distances and cross-
ing directions (farside/nearside), are tested but they turned to be
insignificant. One of the reasons could be that observation sites
are fully equipped with physical barriers. Even at intersection cor-
ners, fences are installed so that pedestrians cannot freely start
crossing from the curbside. Thus, pedestrians are forced to move
to crosswalk edges without making shortcuts. This may signifi-
cantly limit the impact of crosswalk setback distance and curb radii
on pedestrian speed behavior.
5
7. Conclusions
6 This study provides an empirical evidence that pedestrians do
change their behavior while crossing which may leave severe
safety consequences considering the conflicts with turning vehicu-
7 lar traffic. The characteristics of pedestrians sudden speed change
events at crosswalks were analyzed. Extracted data from video-
tapes collected at five signalized crosswalks in Nagoya city, Japan,
is used. Based on the results, the following can be concluded:
A simple demonstration based on the empirical data showed
that sudden speed changes may have a significant impact upon
the decision-making of approaching vehicles in conflict.
8 A methodology to extract instant and significant speed change
events is proposed. This method is useful for clearly and statis-
9 tically identifying potentially dangerous locations and timings.
According to the empirical analysis, the location distribution
10 of acceleration events is concentrated at the entrance points
to the pedestrianvehicle conflict area for both far-side and
near-side pedestrians.
The necessary speed to complete the crossing before the end of
the PFG phase is proposed as an index for the acceleration and
deceleration behavior choice. If v nec t0  v 1 < 0, pedestrians
tend to decelerate; if, v nec t0  v 1 > 0, they chose to either
accelerate or decelerate. A multinomial logit model was esti-
mated for the speed change choice. The results implied that
the entering speed, necessary speed vnec, and crosswalk length
have a significant impact on the speed change choice.
Although the pedestrian speed has been discussed in the litera-
ture, previous studies mainly focused on the average walking
speed. This paper presents important suggestions for understand-
ing pedestrian maneuvers in detail from a safety viewpoint. Sud-
den pedestrian speed changes are important events that may
significantly contribute to the severity of pedestrianvehicle con-
360 W.K.M. Alhajyaseen, M. Iryo-Asano / Safety Science 91 (2017) 351360
flicts since drivers cannot easily expect them. The developed
pedestrian speed profile model can contribute to the representa-
tion of realistic conflicts with vehicles. This can be utilized to esti-
mate the pedestrian-vehicle conflict risk as a part of traffic
simulation for safety assessment, by integrating with other maneu-
ver models of vehicles and pedestrians. Another application can be
a real-time information provision to vehicles to alert the risk of
hazardous conflicts. Such system is expected to be useful not only
for drivers but also for the development of avoidance maneuver
modeling of autonomous vehicles.
It is also important to mention that the proposed methodology
can be used to identify proper crosswalk layout that minimizes
acceleration/deceleration events. Moreover, it can be used to
adjust pedestrian clearance interval (adaptive signal setting) by
detecting pedestrians and estimating the needed crossing time
considering the possibility of behavioral changes.
For future work, quantitative analysis of the amount of speed
change and modeling of the timing and location distributions of
speed change events are required. It should be noted that these
analyses are based on the empirical data taken in Japan, where
the shorter clearance time is set as compared to other countries
(Iryo-Asano and Alhajyaseen, 2014). The occurrence of the sudden
speed change may also be dependent on the PFG length. Since
intersections with longer PFG is not common in Japan, an interna-
tional comparison to confirm the sudden behavioral changes is
considered as one of the future works.
Acknowledgements
This work was supported by JSPS KAKENHI Grant No.
15H05534. Furthermore, authors are grateful to Prof. Hideki Naka-
mura and Mrs. Zhang Xin for their efforts in data collection.
References
Alhajyaseen, W., 2014. Pedestrian speed at signalized crosswalks; analysis and
influencing factors. Int. J. Eng. Manage. Econ. 5 (3/4), 258272.
Alhajyaseen, W., Asano, M., Nakamura, H., 2012a. Estimation of left-turning vehicle
maneuvers for the assessment of pedestrian safety at intersections. J. Int. Assoc.
Traffic Safety Sci.. IATSS Res. 36 (1), 6674.
Alhajyaseen, W., Asano, M., Nakamura, H., 2012b. An integrated model for
reproducing the maneuver of left turners considering dynamic reaction to
crossing pedestrians. In: Transportation Research Board (TRB) 91st Annual
Meeting, CD-ROM, Washington DC, USA (16 pages).
Alhajyaseen, W., Asano, M., Nakamura, H., 2013a. Stochastic approach for modeling
the effects of intersection geometry on turning vehicle paths. Transport. Res.
Part C: Emerg. Technol. 32, 179192.
Alhajyaseen, W., Asano, M., Nakamura, H., 2013b. Left-turn gap acceptance models
considering pedestrian movement characteristics. Acc. Anal. Prevent. 50, 175
185.
Alhajyaseen, W., Nakamura, H., Asano, M., 2011. Effects of bi-directional pedestrian
flow characteristics upon the capacity of signalized crosswalks. Proc. Soc.
Behav. Sci. 16, 526535.
Avineri, E., Shinar, D., Susilo, Y.O., 2012. Pedestrians behaviour in cross walks: the
effects of fear of falling and age. Acc. Anal. Prev. 40, 3034.
Dollar, P., Wojek, C., Schiele, B., Perona, P., 2012. Pedestrian detection: an evaluation
of the state of the art. IEEE Trans. Pattern Anal. Mach. Intell. 34 (4), 743761.
Fruin, J.J., 1971. Designing for pedestrians: a level of service concept. High. Res. Rec.
(355), 115
Hashimoto, Y., Gu, Y., Hsu, L., Iryo-Asano, M., Kamijo, S., 2016. A probabilistic model
of pedestrian crossing behavior at signalized intersections for connected
vehicles. Transport. Res. Part C 71, 164181.
Iryo-Asano, M., Alhajyaseen, W., Nakamura, H., 2014. Analysis and modeling of
pedestrian crossing behavior during the pedestrian flashing green interval. IEEE
Trans. Intell. Transp. Syst. 99, 112.
Iryo-Asano, M., Alhajyaseen, W., 2014. Analysis of pedestrian clearance time at
signalized crosswalks in Japan. Proc. Comp. Sci. 32, 301308.
Koh, P.P., Wong, Y.D., 2014. Gap acceptance of violators at signalized pedestrian
crossings. Acc. Anal. Prevent. 62, 178185.
Lam, W.H.K., Cheung, C.-Y., 2000. Pedestrian speed/flow relationships for walking
facilities in Hong Kong. J. Transport. Eng. 126 (Part 4), 343349.
LaPlante, J., Kaeser, T.P., 2007. A history of pedestrian signal walking speed
assumptions. In: Paper presented at the 3rd Urban Street Symposium, 2427
June 2007, Seattle, Washington.
Lee, J.Y.S., Lam, W.H.K., 2008. Simulating pedestrian movements at signalized
crosswalks in Hong Kong. Transport. Res. Part A 42, 13141325.
Montufar, J., Arango, J., Porter, M., Nakagawa, S., 2007. The normal walking speed of
pedestrians and how fast they walk when crossing the street. In: Proceedings of
the 86th Annual Meeting of the Transportation Research Board, Washington,
DC, USA.
Navin, F.D., Wheeler, R.J., 1969. Pedestrian flow characteristics. Traffic Eng. Control
39, 3036.
National Police Agency, Japan, 2015. Fatal Traffic Accidents in 2014 <http://www.e-
stat.go.jp/SG1/estat/List.do?lid=000001129473> (in Japanese).
Schmitz, J.N., 2011. The Effects of Pedestrian Countdown Timers on Safety and
Efficiency of Operations at Signalized Intersections. In: University of Nebraska
Lincoln, Civil Engineering Thesis, Dissertations, and Student Research, Paper 28.
Supernak, J., Verma, V., Supernak, I., 2013. Pedestrian countdown signals: what
impact on safe crossing. Open J. Civil Eng. 3, 3945.
Suzuki, K., Nakamura, H., 2006. TrafficAnalyzer the integrated video image
processing system for traffic flow analysis. In: Proceedings of the 13th World
Congress on Intelligent Transportation Systems, London (2006) (8 pp. in CD-
ROM).
Tarawneh, M.S., 2001. Evaluation of pedestrian speed in Jordan with investigation of
some contributing factors. J. Safety Res. 32, 229236.
Wang, W., Guo, H., Gao, Z., Bubb, H., 2011. Individual differences of pedestrian
behaviour in midblock crosswalk and intersection. Int. J. Crashworth. 16 (1), 1
9.
Xu, Y., Li, Y., Zhang, F., 2013. Pedestrians intention to jaywalk: automatic or
planned? A study based on a dual-process model in China. Acc. Anal. Prevent.
50, 811819.
Yang, Y., Sun, J., 2013. Study on pedestrian red-time crossing behaviors: integrated
field observation and questionnaire data. In: Transportation Research Board
(TRB) 92nd Annual Meeting, Washington DC, USA.
Zhang, X., Chen, P., Nakamura, H., Asano, M., 2013. Modeling pedestrian walking
speed at signalized crosswalks considering crosswalk length and signal timing.
In: Proceedings of the 10th International Conference of the Eastern Asia Society
for Transportation Studies (15 pages, USB).