Transportation Research Part B 98 (2017) 135–153

Contents lists available at ScienceDirect

Transportation Research Part B

journal homepage: www.elsevier.com/locate/trb

Optimization of vehicle and pedestrian signals at isolated

intersections
Chunhui Yu a,b, Wanjing Ma a,∗, Ke Han c, Xiaoguang Yang a
a
Key Laboratory of Road and Traffic Engineering of the Ministry of Education, Tongji University, 4800 Cao’an Road, Shanghai, PR China
b
Jiangsu Province Collaborative Innovation Center of Modern Urban Traffic Technologies, SiPaiLou #2, Nanjing, 210096, PR China
c
Department of Civil and Environmental Engineering, Imperial College London, London SW7 2BU, UK

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

Article history: In most traffic signal optimization problems, pedestrian traffic at an intersection receives
Received 18 February 2016 minor consideration compared to vehicular traffic, and usually in the form of simplis-
Revised 24 December 2016
tic and exogenous constraints (e.g., minimum green time). This could render the result-
Accepted 24 December 2016
ing signal timings sub-optimal especially in dense urban areas with significant pedestrian
traffic, or when two-stage pedestrian crosswalks are present. This paper proposes a con-
Keywords: vex (quadratic) programming approach to optimize traffic signal timings for an isolated
Traffic signal optimization intersection with one- and two-stage crosswalks, assuming undersaturated vehicular traf-
Vehicle and pedestrian delay fic condition. Both vehicle and pedestrian traffic are integrated into a unified framework,
Spatial constraint where the total weighted delay of pedestrians and vehicles at different types of crosswalks
Quadratic programing (i.e. one- or two-stage) is adopted as the objective function, and temporal and spatial con-
Isolated intersection
straints (e.g. signal phasing plan and spatial capacity of the refuge island) are explicitly
formulated. A case study demonstrates the impacts of incorporating pedestrian delay as
well as geometric and spatial constraints (e.g., available space on the refuge island) in the
signal optimization. A further analysis shows that a two-stage crosswalk may outperform a
one-stage crosswalk in terms of both vehicle and pedestrian delays in some circumstances.
© 2016 Elsevier Ltd. All rights reserved.

1. Introduction

With rapid urbanization and moterization especially in developing countries, daily commuters in dense urban areas suffer
from heavy traffic congestion that leads to delays, environmental deterioration, and economic loss (Koonce et al., 2008).
As reported in National Transportation Operations Coalition (NTOC) (2012), delays at traffic signals in road networks are
estimated to be 295 million vehicle-hours. Improving traffic signal operation has a significant impact on the efficiency of
transportation systems in dense urban areas, potentially more effective than any other operational measure in the traffic
engineering toolkit (NTOC, 2012).
Abundant studies on signal optimization have been conducted at isolated intersections, along corridors, or at network
levels. A comprehensive literature review of off-line traffic signal optimization methods is provided in Wong et al., (2005).
Conventionally, signal optimization are stage-based (Allsop, 1981; Webster, 1958) or group-based (phase-based) (Heydecker,
1992; Silcock, 1997). Compatible traffic movements are grouped and move together in the same time window, called a
stage, in the stage-based approach. Green time is then allocated to each stage. In contrast, there is no need to specify a

∗
Corresponding author.
E-mail addresses: [email protected] (C. Yu), [email protected] (W. Ma), [email protected] (K. Han), [email protected] (X. Yang).

http://dx.doi.org/10.1016/j.trb.2016.12.015
0191-2615/© 2016 Elsevier Ltd. All rights reserved.
136 C. Yu et al. / Transportation Research Part B 98 (2017) 135–153

stage structure in the group-based approach. Green time is directly allocated to each movement. Obviously, the group-based
approach is more flexible and potentially more efficient. More recently, it has been reported that signals could be optimized
together with lane markings to further improve traffic conditions (Wong and Heydecker, 2011; Wong and Wong, 2003a, b).
These various approaches adopt similar optimization objectives; that is, the minimization of vehicle delay and maximization
of intersection throughput with undersaturated and oversaturated vehicular traffic, respectively.
Based on the signal optimization for isolated intersections, signal coordination for multiple intersections along corridors
is investigated. MAXBAND (Little et al., 1981) and MULTIBAND (Stamatiadis and Gartner, 1996) are proposed to signalize in-
tersections along arterial roads. The stage-, group-, and lane-based methods were extended to the optimization of area traffic
signals (Li and Gan, 1999; Wong and Wong, 2002; Wong and Yang, 1999). In addition, relatively new methods (e.g., fuzzy
approaches and reinforcement learning algorithms) are applied to cope with complicated models and address computational
burdens in signal optimization problems (Murat and Kikuchi, 2008; Ozan et al., 2015).
Moreover, additional concerns such as traffic flow uncertainty and vehicle-derived emissions are taken into consideration
in signal optimization. The real-world traffic flows are time-varying and stochastic in nature (Cascetta et al., 2006; Szeto and
Lo, 2005), even for the same time period across days of the same category (Yin, 2008). In general, robust signal optimization
and on-line (adaptive) signal control are available to address flow uncertainty and influence traffic in an adaptive manner.
Using robust optimization, signal parameters (e.g., cycle lengths and green splits) are predetermined, for example, by a
scenario-based model using historical data in order to maintain satisfactory performance under flow fluctuations (Yin, 2008;
Zhang et al., 2010). Robust optimization has also been applied to handle uncertainties in emission estimation arising from a
macroscopic traffic modeling approach (Han et al., 2016). On the side of on-line optimization, real-time traffic measurements
and prediction of future demands are utilized to dynamically adjust signal plans (Liu et al., 2015; Tong et al., 2015). With
growing attention to the negative impacts of traffic on air quality and public health, sustainable traffic control measures have
gained much attention. Minimization of vehicle emissions is regarded as one of the objectives in the signal optimization,
mainly by means of analytical models (Khalighi and Christofa, 2015; Ma and Nakamura, 2010) and traffic microsimulations
(Mascia et al., 2016; “Brian” Park et al., 2009; Zhang et al., 2013). In addition, signal prioritization for emergency vehicles
and public transportation is also investigated (Head et al., 2006).
The numerous studies reviewed above focus on vehicles and their delays, with little or no attention given to crossing
pedestrian traffic. In fact, most of these studies only implicitly consider pedestrians by imposing constraints such as mini-
mum green time and pedestrian clearance time. Pedestrian delay is not included in the objectives of the signal optimization
problems. This seems reasonable in situations where vehicular traffic dominates pedestrian traffic, and the former receives
far greater priorities. However, this does not apply to dense metropolitan areas with significant pedestrian traffic. Indeed,
very limited studies have provided insights into this issue. Li et al., (2010b) propose a traffic signal optimization strategy
that considers both vehicular and pedestrian traffic. Vehicle and pedestrian delays are calculated by a deterministic queuing
model. Wang and Tian (2010) develop a model in an analytical form to estimate pedestrian delay at a two-stage crosswalk.
Ma et al., (2014) model pedestrian delay in the optimization of signal timing for an isolated intersection with exclusive
pedestrian phases. In these studies, however, vehicle and pedestrian delay models are nonconvex, and global optima cannot
be guaranteed. Moreover, only one-stage crosswalks have been investigated so far. Notably, two-stage crossings become in-
creasingly popular and preferred by traffic engineers when a crosswalk at an intersection is long and a safe median refuge
island is available. For example, in MOHURD (2011) a refuge island is requested when a crosswalk is longer than 16 m.
However, the limited space on the refuge island is usually ignored. Insufficient space on the island may impose restric-
tive constraints to the signal optimization problems, while raising safety concerns. Although the extended cell transmission
model in Zhang and Chang (2014) may provide a potential way to address these issues, their optimization problem is solved
with heuristic methods, which does not guarantee global optimality and may be computationally expensive. Therefore, a
rigorous and effective signal design framework for intersections with two-stage crosswalks is needed to minimize vehicle
delay as well as provide a high level of service for pedestrians.
In order to address this gap, this paper presents a convex (quadratic) programming approach to optimize fixed signals
for vehicles and pedestrians at an isolated intersection with one- and two-stage crosswalks with undersaturated vehicular
traffic.1 A unified signal timing optimization framework is developed, in which the delays of both vehicles and pedestrians
are explicitly accounted for. The impacts of spatial and temporal constraints, including space limitation of the refuge island
and the occupancy of the island, are explicitly captured and modeled through linear formulae. Next, a series of quadratic
formulae are derived to approximate vehicle and pedestrian delays at the signalized intersection with one- and two-stage
crosswalks. The feasibility of the proposed delay model is validated. The signal optimization problem is decomposed into
a series of quadratic programming problems, which guarantees global optimality and maintains computational tractability,
allowing computationally intensive tasks such as real-time signal control and signal coordination to be subsequently de-
veloped. The proposed model can be readily solved by many existing commercial solvers efficiently, which serves well the
purpose of practical applications. The effectiveness of the proposed model is validated by numerical examples.

1
The model is built assuming undersaturated vehicular traffic condition as vehicle delay is part of the objective to be minimized. In case of oversatu-
ration, maximization of intersection capacity tends to be used as the objective. In this latter case, all of the proposed linear constraints remain effective,
and the signal optimization problem can be formulated as a linear program when the reserve capacity is adopted (Wong and Heydecker, 2011; Wong and
Wong, 2003a, b).
 C. Yu et al. / Transportation Research Part B 98 (2017) 135–153 137

Fig. 1. A typical intersection with four approaches.

The remainder of this paper is structured as follows. Section 2 presents the signal optimization problem and some key
notations. In Section 3, the quadratic programming models are developed. A numerical study of the proposed model and
solution method is conducted in Section 4. Finally, some concluding remarks and recommendations are provided in Section
5.

2. Problem description and notations

2.1. Signal optimization for intersections with crosswalks

A typical intersection with four approaches is illustrated in Fig. 1. Generally, a one- or two-stage pedestrian crosswalk
is applied in each approach. As a common example, it is assumed that Approach 1 and Approach 3 are on the main road;
Approach 2 and Approach 4 are on the minor road. There are four approach lanes and three exit lanes in Approach 1
and Approach 3, two approach lanes and two exit lanes in Approach 2 and Approach 4. Two-stage crosswalks are used
in Approach 1 and Approach 3, and one-stage crosswalks are used in Approach 2 and Approach 4. A refuge island should
be installed, for example, between the 1st segment and the 2nd segment of the two-stage crosswalk in Approach 1, for
pedestrian safety (Li et al., 2010a). A two-stage crosswalk has two separate pedestrian signal phases as a general case. That
is, the signals for the 1st and 2nd segments are independent.
In practice, a typical North American dual-ring, eight-phase controller is usually used. The controller in Fig. 2 is imple-
mented in this study. Phase 1 and Phase 8 are assumed to be adjacent to reduce pedestrian clearance time due to high
pedestrian traffic. The proposed model can also be applied to other phase structures. Pedestrian movements on the 1st
and 2nd segments in Approach 1 are illustrated as an example. Obviously, vehicle and pedestrian signals are related to each
other. For example, pedestrians can cross over the 1st segment of the two-stage crosswalk in Approach 1 during Phase 1 and
Phase 8. Undersaturated vehicular traffic is assumed in which case vehicle delay is supposed to be minimized. Therefore,
the remaining problem is to optimize vehicle and pedestrian signals for the intersection for minimal vehicle and pedestrian
delays. Fixed signal timings are applied.
138 C. Yu et al. / Transportation Research Part B 98 (2017) 135–153

Fig. 2. Dual-ring, eight-phase controller.

Table 1
Parameters and variables.

General parameters:

M: A sufficiently large positive constant

Traffic parameters:

β: Weighting parameter, 0 ≤ β ≤ 1
r1 , r2 : Arrival rates of pedestrians on Side I and Side II of Approach 1 (ped/s)
rvp : Arrival rate of vehicles in Phase p (pcu/s), p = 1, …, 8
sp : Saturation flow rate of vehicles in Phase p (pcu/s), p = 1, …, 8
ρ p, max : Maximum acceptable degree of saturation of Phase p, p = 1, …, 8
C1 , C2 : Minimum and maximum signal cycle lengths (s)
gd,min : Minimum green time for drivers to avoid stop-and-go motions (s)
g1,min : Minimum Walk interval for crossing pedestrians on Side I of Approach 1 (s)
g2,min : Minimum Walk interval for crossing pedestrians on Side II of Approach 1 (s)
τ 1, τ 2: Time to cross over the 1st and 2nd segments of the crosswalk in Approach 1 if a two-stage crosswalk is applied (s)
τ: Time to cross over the crosswalk in Approach 1 if a one-stage crosswalk is applied (s)
Nmax : Maximum number of waiting pedestrians that can be accommodated on the refuge island at the two-stage crosswalk in
Approach 1

Decision variables:

C: Signal cycle length (s)

αi : 1, if Case i is optimal for a two-stage crosswalk in Approach 1; 0, otherwise. i = 1, …, 10
tp : Ending time of Phase p (s), p = 1, …, 8
gp : Green time of Phase p (s), p = 1, …, 8
tg : Starting time of the Walk interval for the crosswalk in Approach 1 if a one-stage crosswalk is applied (s)
tf : Starting time of the flashing Don’t Walk (FDW) clearance interval for the crosswalk in Approach 1 if a one-stage crosswalk is
applied (s)
tr : Starting time of the Don’t Walk or Stop interval for the crosswalk in Approach 1 if a one-stage crosswalk is applied (s)
tg1 , tg2 : Starting time of the Walk intervals for the 1st and 2nd segments of the crosswalk in Approach 1 if a two-stage crosswalk is
applied (s)
t 1f , t 2f : Starting time of the FDW clearance intervals for the 1st and 2nd segments of the crosswalk in Approach 1 if a two-stage
crosswalk is applied (s)
tr1 , tr2 : Starting times of the Don’t Walk or Stop intervals for the 1st and 2nd segments of the crosswalk in Approach 1 if a two-stage
crosswalk is applied (s)

Intermediate variables:

gp, min : Minimum green time of Phase p to clear queues (s), p = 1, …, 8

Dp : Vehicle delay in Phase p (s), p = 1, …, 8
Dmper : Total pedestrian delay at the crosswalk in Approach m (s), m = 1, 2, 3, 4
D11 , D21 : 1st stage delays of waiting pedestrians on Side I and Side II of Approach 1 (s)
D12 , D22 : 2nd stage delays of waiting pedestrians on the refuge island from Side I and Side II of Approach 1 (s)
n1 , n2 : Accumulated pedestrians on Side I and II of Approach 1 when the Walk intervals start
nt 1 , nt 2 : Introduced variables for the convenience of modeling

2.2. Notations

To facilitate the presentation of the model, key notations to be applied hereafter are summarized in Table 1.

3. Model formulation

A one- or two-stage crosswalk is usually applied in each approach of an intersection, as shown in Fig. 1. The signals of
one- and two-stage crosswalks in Approach 1 are described in detail as an example to illustrate the modeling of pedestrian
 C. Yu et al. / Transportation Research Part B 98 (2017) 135–153 139

Fig. 3. Precedence graph representation of a dual-ring controller with pedestrian signals in Approach 1 as an example.

constraints and delay. As shown in Fig. 3, a precedence graph is employed to represent the dual-ring eight phases in Fig.
2 to capture the relations between vehicle and pedestrian signals.
Pedestrians wait on Side I and Side II until the starts of the Walk intervals (i.e., tg , tg1 and tg2 ). If a one-stage crosswalk
is used in Approach 1, pedestrians can cross only during Phase 8. The pedestrian Walk interval starts when Phase 8 starts
and the Stop interval starts when Phase 8 ends to reduce pedestrian delay (i.e., tg = t7 , tr = t8 ). If a two-stage crosswalk is
used, two separate pedestrian signal phases on the 1st and 2nd segments are available. Crossing pedestrians can wait on
the refuge island. Pedestrians can cross over the 1st segment during Phase 1 and Phase 8, and the 2nd segment during
Phase 3 and Phase 4 (i.e., Phase 7 and Phase 8). To make the best use of vehicle phases, the pedestrian Walk interval on the
1st segment and Phase 8 start at the same time (tg1 = t7 ); the Stop interval on the 2nd segment starts when Phase 4 ends
(tr2 = t4 ). Note that Phases 3, 4 and Phases 1, 8 may not be fully utilized for crossing pedestrians owing to the limited space
on the refuge island (i.e., tg2 ≥ t2 , tr1 ≤ C + t1 ).

3.1. Vehicle-related constraints

In this section we present constraints pertaining to vehicular traffic at the intersection, including cycle time, phase se-
quences, and minimum green time. As the signal timings are fixed and cyclic, the following constraints are expressed for
only one cycle.

3.1.1. Cycle length

It is assumed that the minimum and maximum signal cycles, denoted as C1 and C2 , respectively, are given in advance.

C1 ≤ C ≤ C2 (1)

3.1.2. Precedence relationships of signal phases

As shown in Fig. 3, the start time t0 of the cycle is set at zero as the reference time:

t0 = 0 (2)

Phase 2 and Phase 6 terminate at the same time:

t2 = t6 (3)

Both Phase 4 and Phase 8 terminate at the time of C:

t4 = t8 = C (4)

3.1.3. Green time of each phase

For simplicity, green time of each phase is regarded as effective green duration. The green duration of Phase p is calcu-
lated as

t p − t p−1 = g p p = 1, . . . , 4, 6, . . . , 8 (5)

t5 − t0 = g5 (6)
140 C. Yu et al. / Transportation Research Part B 98 (2017) 135–153

Fig. 4. Delay models: (a) vehicle delay model; (b) pedestrian delay model.

3.1.4. Minimum green time

On the basis of the arrival rate and the saturation flow rate in Phase p, as shown in the following Fig. 4(a), the minimum
green time of each phase to clear vehicles accumulated during the red time is calculated as
rvp (C−g p )
g p,min = s p −rvp
p = 1, 2, . . . , 8 (7)

The green time gp should be no less than the minimum value gp, min :

g p,min ≤ g p p = 1, 2, . . . , 8 (8)
Substitute Eq. (7) for gp, min , and Eq. (8) can be modified as

rvpC ≤ s p g p p = 1, 2, . . . , 8 (9)
Furthermore, another minimum green time gd,min is set so that drivers can have sufficient time to prepare the changes of
signals to avoid sudden stop-and-go motions:
g p ≥ gd,min p = 1, 2, . . . , 8 (10)

3.1.5. Maximum acceptable degree of saturation

The saturation degree of Phase p is calculated as
rvp
ρp = spgp p = 1, 2, . . . , 8 (11)
C

To ensure the service level, a maximum acceptable degree of saturation ρ p,max is set:
ρ p ≤ ρ p,max (12)
Substitute the expression (11) for ρ p , and Eq. (12) can be modified as

rvpC ≤ ρ p,max s p g p p = 1, 2, . . . , 8 (13)

Note that these constraints should be set carefully since they may render the problem infeasible under high vehicle demand.

3.2. Pedestrian-related constraints of one-stage crosswalks

3.2.1. Starting time of the pedestrian Walk interval

As shown in Fig. 3, pedestrians can cross over the intersection at a one-stage crosswalk only during Phase 8 for safety
concerns. To make the best use of Phase 8, the pedestrian Walk interval starts when Phase 7 ends:
tg = t7 (14)

3.2.2. Starting time of the Do-not Walk or Stop interval

For similar purposes, the Don’t Walk or Stop interval starts when Phase 8 ends to provide pedestrians more crossing
time:
tr = t8 (15)

3.2.3. Starting time of the flashing Do-not Walk (FDW) clearance interval
Given the layout of the intersection, the FDW clearance interval τ is determined. Therefore, the starting time of the
clearance interval is calculated as
t f = tr − τ (16)
 C. Yu et al. / Transportation Research Part B 98 (2017) 135–153 141

Fig. 5. Case 1: (a) precedence graph representation, (b) accumulated pedestrians arriving at the refuge island from Side II, (c) accumulated waiting pedes-
trians on the refuge island from Side II, (d) accumulated pedestrians arriving at the refuge island from Side I, (e) accumulated waiting pedestrians on the
refuge island from Side I, and (f) total waiting pedestrian number on the refuge island. Possible maximum numbers are marked by a red square or triangle.
(For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

3.2.4. Minimum pedestrian Walk intervals

To satisfy the crossing demand on Side I and Side II, minimum pedestrian Walk intervals on both sides should be guar-
anteed:

t f − tg ≥ g1, min (17)

t f − tg ≥ g2, min (18)

The minimum values can be calibrated according to pedestrian demand as discussed in Lao et al., (2009).

3.3. Pedestrian-related constraints of two-stage crosswalks

3.3.1. Potentially optimal cases

Pedestrian signal phases on the 1st and 2nd segments are separate. As a result, it is difficult to model pedestrian delay
and the waiting pedestrian number on the refuge island. The arrival time of the pedestrians from Side II at the refuge island
is between [tg2 + τ2 , tr2 ] while the Walk interval on the 1st segment is [tg1 , t 1f ]. Similarly, the arrival time of the pedestrians
from Side I is between [tg1 + τ1 , tr1 ] while the Walk interval on the 2nd segment is [tg2 , t 2f ]. According to the relative positions
of these intervals, ten potentially optimal cases are identified for each two-stage crosswalk. The optimal pedestrian signals
at the crosswalk are produced in one of the ten cases. This is specified by


10
αi = 1 (19)
i=1

where α i = 1 if Case i is optimal; otherwise, α i = 0 (i = 1, …, 10).

Case 1 is shown in Fig. 5 as an example (we refer the reader to Appendix I for Cases 2–10), which only shows the traffic
pattern in one cycle. Fig. 5(a) shows the precedence graph when the following relationships of the above intervals hold:

tg2 + τ2 ≤ tg1 < t 1f ≤ tr2 (20)

tr1 > tg1 + τ1 ≥ t 2f > tg2 (21)

In Fig. 5(b), the dark line represents the number of pedestrians who arrive at the refuge island from Side II in one
cycle length; the blue line shows the pedestrian Walk interval on the 1st segment. Fig. 5(c) shows the number of waiting
pedestrians on the island who come from Side II, which is derived from Fig. 5(b). The shaded area represents the delay
of these pedestrians. Similarly, Fig. 5(d) and (e) show the information of pedestrians from Side I. Fig. 5(f) shows the total
number of pedestrians on the island from both sides which is derived by combining Fig. 5(c) and (e).
142 C. Yu et al. / Transportation Research Part B 98 (2017) 135–153

3.3.2. Starting time of the pedestrian Walk interval on the 2nd segment
In all cases, the starting time of the Walk interval on the 2nd segment should be no earlier than the ending time of
Phase 2 (t2 ):
tg2 ≥ t2 (22)

The earliest arrival time at the refuge island from Side II is tg2 + τ2 . The Walk interval on the 1st segment is [tg1 , t 1f ].
In Cases 1–4 (refer to Fig. 5 and Appendix I), tg2 + τ2 and [tg1 , t 1f ] satisfy the following constraint (23). Considering the
constraint (22), the following constraint (24) should hold as well. Eqs. (23) and (24) are effective only when α i = 1 (i = 1, 2,
3, 4).
 

4
tg2 + τ2 ≤ tg1 + 1− αi M (23)
i=1
 

4
t 2 + τ2 < tg1 + 1− αi M (24)
i=1

In Cases 5–8, tg2 is taken as t2 , which can be specified by Eqs. (22) and (25). At the same time, t2 + τ 2 (or tg2 + τ2 ) and
[tg1 , t 1f ] satisfy the following constraint (26). Eqs. (25) and (26) are effective only when α i = 1 (i = 5, 6, 7, 8).
 

8
tg2 ≤ t2 + 1− αi M (25)
i=5
   

8 
8
tg1 − 1− αi M ≤ t2 + τ2 < t 1f + 1 − αi M (26)
i=5 i=5

In Cases 9–10, tg2 + τ2 ≥ t 1f , which can be realized by Eqs. (27) and (22). Eq. (27) is effective only when α i = 1 (i = 9, 10).
 

10
t2 + τ2 ≥ t 1f − 1− αi M (27)
i=9

3.3.3. Starting time of the pedestrian Walk interval on the 1st segment
In all cases, the pedestrian Walk interval on the 1st segment and Phase 8 start at the same time to reduce pedestrian
delay:
tg1 = t7 (28)

The earliest arrival time at the refuge island from Side I is tg1 + τ1 . The Walk interval on the 2nd segment is [tg2 , t 2f ]. In
Cases 2n + 1 (n = 0, 1, 2, 3, 4) (refer to Fig. 5 and Appendix I), tg1 + τ1 and [tg2 , t 2f ] satisfy the following constraint (29). All
pedestrians from Side I have to wait on the refuge island.
 

4
tg1 + τ1 ≥ t 2f − 1− α2n+1 M (29)
n=0

In Cases 2n + 2 (n = 0, 1, 2, 3, 4), tg1 + τ1 and [tg2 , t 2f ] satisfy the following constraint (30). Pedestrians arriving at the
refuge island from Side I during [tg1 + τ1 , t 2f ] can proceed without delay while the others have to wait on the island.
   

4 
4
t 2f + 1− α2n+2 M > tg1 + τ1 ≥ tg2 − 1 − α2n+2 M (30)
n=0 n=0

3.3.4. Starting time of Don’t Walk or Stop intervals

Pedestrians can cross over the 1st segment only during Phase 8 and Phase 1. Therefore, the starting time of the Don’t
Walk or Stop interval on the 1st segment (tr1 ) should be no later than the ending time of Phase 1 for safety concerns:
tr1 ≤ t1 + C (31)
The starting time of the Don’t Walk or Stop interval on the 2nd segment is set as the ending time of Phase 8 to maximize
available pedestrian crossing time:
tr2 = t8 (32)
 C. Yu et al. / Transportation Research Part B 98 (2017) 135–153 143

3.3.5. Starting time of FDW clearance intervals

Given the layout of the intersection, the FDW clearance intervals on the 1st and 2nd segments (τ 1 and τ 2 ) are deter-
mined. Therefore, the starting time of the clearance intervals on both segments (t 1f and t 2f ) are calculated as

t 1f = tr1 − τ1 (33)

t 2f = tr2 − τ2 (34)
In all cases, tr1 (i.e., t 1f + τ1 ) is larger than t 2f to make use of available crossing time on the 2nd segment for pedestrians
from Side I:
tr1 = t 1f + τ1 ≥ t 2f (35)
Because pedestrians arriving at the refuge island from Side I earlier than t 2f can proceed without delay.
In Cases 1, 2, 5, and 6, t 1f and tr2 are related in the following Eq. (36). However, Eq. (37) holds in Cases 3, 4, 7, and 8.

t 1f ≤ tr2 + (1 − α1 − α2 − α5 − α6 )M (36)

t 1f > tr2 − (1 − α3 − α4 − α7 − α8 )M (37)

3.3.6. Minimum pedestrian Walk intervals

To satisfy the crossing demand on Side I and Side II, minimum pedestrian Walk intervals at both sides should be guar-
anteed:
t 1f − tg1 ≥ g1, min (38)

t 2f − tg2 ≥ g2, min (39)

As the pedestrian demand on Side I and Side II may differ, g1,min and g2,min can take different values.

3.3.7. Number of waiting pedestrians on the refuge island

The total number of waiting pedestrians on the refuge island can be easily identified in each case, for example, as the
points marked by a red square and a red triangle in Fig. 5(f). It is shown that there are two possible maximum values in
Case 1. These values can be easily expressed with linear formulae derived from Fig. 5(c) and (e), and are then constrained
by the capacity of the refuge island.
The following Eqs. (40) and (41) show the capacity constraints in a unified form for all cases. These equations mean that
both of the two maximum values should not exceed the island capacity Nmax . In each case with determined α i (i = 1, …,
10), the following formulae can be simplified. The calculation of the pedestrian number on the island is broken into several
parts for the convenience of illustration.

2 
4    
4
nt2 αi + n2 αi + r2 tg1 − tg2 + τ2 αi ≤ Nmax (40)
i=1 i=1 i=1

nt2 + nt1 ≤ Nmax (41)

where nt1 , nt2 , and n2 are calculated in Eqs. (42), (43), and (44), respectively:
  4 4
nt1 = r1 tr1 − t 2f
n=0
α2n+2 + r1C n=0
α2n+1 (42)
 
  2 
6 
10
nt2 = r2 tr2 − t 1f αi + αi + r2C αi (43)
i=1 i=5 i=9
  
n2 = r2 C − t 2f − tg2 (44)

3.4. Delay models

3.4.1. Vehicle delay

The deterministic queuing model in Li et al., (2010b) is applied. Vehicle delay Dp in each Phase p is calculated as the
shaded area in Fig. 4(a):
r p s p (C − g p )2
Dp = v   p = 1, 2, . . . , 8 (45)
2 s p − rvp
Although this delay model is simple, it will produce similar delay compared with other models such as the shock wave
delay model and the steady-stage stochastic delay model under undersaturated vehicular traffic (Dion et al., 2004).
144 C. Yu et al. / Transportation Research Part B 98 (2017) 135–153

3.4.2. Pedestrian delay at one-stage crosswalks

Take the one-stage crosswalk in Approach 1 in Fig. 3 as an example, the delays on Side I and Side II are calculated as the
shaded area in Fig. 4(b) by use of the commonly used pedestrian delay model in the Highway Capacity Manual (HCM) (TRB,
2010):
  2
r j C − t f − tg
D1j = j = 1, 2 (46)
2
Therefore, the total pedestrian delay in Approach 1 is finally determined in the following Eq. (47):


2
D1per = D1j (47)
j=1

3.4.3. Pedestrian delay at two-stage crosswalks

The calculation of pedestrian delay at a two-stage crosswalk is complicated. The delay of pedestrians from either Side I
j j
or Side II is the sum of the experienced delays at the first and second stages (i.e., D1 and D2 , respectively):


2 
2
D1per = Dkj (48)
j=1 k=1

j
The delay at the first stage D1 is calculated in a similar way to Eq. (46):
  2
r j C − t fj − tgj
D1j = j = 1, 2 (49)
2
The delay of pedestrians from Side I at the second stage D12 is calculated as the shaded area, for example, in Fig. 5(e) for
Case 1. Similarly, D12 in other cases can also be expressed by quadratic formulae. Eqs. (50)–(53) are the formulae to calculate
D12 for all cases in a unified form. The calculation of D12 is broken into several parts for the convenience of illustration.

D12 = d11 + d12 (50)

r1  1 2 4 r  2  4
d11 = t f − tg1 α2n+1 + 1 tr1 − t 2f α2n+2 (51)
2 n =0 2 n=0

  4   4 10
  8
d12 = n1 t 1f − tg1
n=0
α 2 1
2n+1 + nt1 C + tg − tr
i=1
αi +
i=9
αi + nt1 C + t2 − tr1
i=5
αi (52)

where nt1 is calculated in Eq. (42), and n1 is calculated as

  
n1 = r1 C − t 1f − tg1 (53)
Note that only certain terms in the above equations remain in each case with determined α i (i = 1, …, 10).
The delay of pedestrians from Side II at the second stage D22 is calculated as the shaded area, for example, in Fig. 5(c) for
Case 1. Similar to D12 , D22 in all cases is calculated by Eqs. (54)–(56) in a unified form:

D22 = d21 + d22 (54)

r2  2  2  2 6
r   2 4 r2  2 2 10
d21 = tr − t 1f αi + αi +
2
tg1 − tg2 + τ2 αi + t f − tg2 αi (55)
2 i=1 i=5 2 i=1 2 i=9

  10 1 2  4   2 6 10

d22 = n2 t 2f − tg2
i=9
αi + n 2 t g − t g + τ2
i=1
α 1 2
i + nt2 C + tg − tr
i=1
αi +
i=5
αi +
i=9
αi (56)

where nt2 , n2 are calculated in the above Eqs. (43) and (44), respectively. In this way, the total pedestrian delay D1per in
Approach 1 can be expressed by a quadratic formula in each case with fixed α i (i = 1, …, 10).
The formulae of vehicle delay (Eq. (45)) and pedestrian delay at a one-stage crosswalk (Eq. (46)) are convex, quadratic
while the formula of pedestrian delay at a two-stage crosswalk is nonconvex due to the terms (52) and (56). As a result, no
global optima is guaranteed although there are many existing algorithms to solve a nonconvex, quadratic model efficiently
for local optima. In this study, the following convex, quadratic formulae Eqs. (57) and (58) are proposed to replace Eqs.
(52) and (56), respectively, to calculate approximate pedestrian delay. Note that Eqs. (57) and (58) are the transformations
of Eqs. (52) and (56) in form but have no specific meanings.
 2  2
γ2 C + t2 − t 1 2 8
γ1 n1 2 + γ2 t 1f − tg1 4 γ1 nt1 2 γ2 C + tg2 − tr1 4 10
d
r
12 = α2n+1 + + αi + α i + αi (57)
2 n=0 2 2 i=1 i=9 2 i=5
 C. Yu et al. / Transportation Research Part B 98 (2017) 135–153 145

Table 2
Parameters in the numerical example.

Vehicle-related parameters Traffic demand Pedestrian-related parameters

sp 3537 pcu/h (p = 1, 5) 1
rv 300 pcu/h r1 1620 ped/h
3789 pcu/h (p = 2, 6) rv2 550 pcu/h r2 1080 ped/h
1965 pcu/h (p = 3, 7) rv3 100 pcu/h τ1 9 s (for Approaches 1 and 3)
2105 pcu/h (p = 4, 8) rv4 300 pcu/h τ2 12 s (for Approaches 1 and 3)
ρ p, max 0.9 rv5 320 pcu/h τ 12 s (for Approaches 2 and 4)
C1 60 s rv6 500 pcu/h g1, min 5s
C2 120 s rv7 200 pcu/h g2, min 5s
gd, min 5s rv8 100 pcu/h Nmax 20
β 0.6

 2   2
 γ3 n2 2 4 10
γ4 t 2f − tg2 10 γ4 tg1 − tg2 + τ2
d22 = αi + αi + αi +
2 i=1 i=9 2 i=9 2
 
2 2

4 γ3 nt2 + γ4 C + tg − tr
2 1 2 6 10
× αi + αi + αi + αi (58)
i=1 2 i=1 i=5 i=9

where γ 1 = 1.948, γ 2 = 0.149, γ 3 = 1.626, and γ 4 = 0.011 are factors, which are calibrated using signal samples randomly gen-
erated satisfying the constraints in Eqs. (1)–(13) and (19)–(39). Therefore, the delays D12 and D22 are modified in the following
Eqs. (59) and (60), whose feasibility is verified in the numerical tests.

D12 = d11 + d12 (59)


D22 = d21 + d22 (60)

3.5. Optimization models

The objective function of the optimization model is to minimize vehicle delay and pedestrian delay in each approach
with a crosswalk:
8 4
min β D p + (1 − β ) Dm
per (61)
p=1 m=1

where β (0 ≤ β ≤ 1) is a weighting parameter, and it can be calibrated by the economic losses of vehicle and pedestrian
delays. Dp is calculated in the above Eq. (45). Dm per is set at zero if there is no crosswalk in Approach m; otherwise, it can be
calculated in a similar way to D1per in the above Eq. (47) or Eq. (48).
The constraints of vehicle signals are shown in Eqs. (1)–(13). If a one-stage crosswalk is applied in Approach 1, the
constraints of pedestrian signals are specified by Eqs. (14)–(18). If a two-stage crosswalk is applied instead, Eqs. (19)–(44) are
effective. In addition, similar constraints should be added when a one- or two-stage crosswalk is used in other approaches.
As discussed above, the signal optimization problem is decomposed into several cases. In each case with determined α i
(i = 1, …, 10), all constraints are linear and the objective function is convex, quadratic, which is a convex, quadratic program-
ming model. Many algorithms such as interior-point methods are available to solve such kind of models efficiently for global
optima. Therefore, the optimal signals are produced after solving a series of convex, quadratic programming models. There
are 10n potentially optimal cases if n approaches use two-stage crosswalks at the intersection. However, the solving process
is quite efficient, because the number of intersection approaches is usually small in the real world.
It is noted that pedestrians may also cross over the 2nd segment in Approach 1 when Phase 2 and Phase 5 are both active
as shown in Fig. 2. This requires that the overlap time is long enough to satisfy the constraints of the minimum pedestrian
Walk interval and the FDW clearance interval on the 2nd segment; all waiting pedestrians on Side II can be accommodated
on the refuge island. However, these conditions are not likely to be satisfied, especially with large pedestrian traffic that is
assumed in this study.

4. Numerical examples

4.1. Experimental data

In order to demonstrate the applicability of the proposed model and the effectiveness of the resulting signal timing,
we consider the intersection shown in Fig. 1 for the numerical experiment. The dual-ring, eight-phase controller is shown
in Fig. 2. Relevant parameters are summarized in Table 2. The saturation flow rates sp are calculated as the product of a
base saturation flow rate and adjustment factors as suggested in HCM. Vehicular traffic demand is set according to the
146 C. Yu et al. / Transportation Research Part B 98 (2017) 135–153

Fig. 6. Relationship between accurate delay and approximate delay: (a) β = 0.1, (b) β = 0.9.

saturation flow rates. Owing to the lack of pedestrian field data, pedestrian parameters for the four approaches are assumed
to be identical. The typical minimum pedestrian Walk interval and pedestrian speed can be found in Wang and Tian (2010).
Pedestrian-related parameters are set accordingly. The maximum acceptable pedestrian number on the refuge island Nmax
should be specified in practice, for example, at a certain level of service according to HCM. The weighting parameter β in
Eq. (61) is 0.6 in this study, as the basic unit costs of $6 per vehicle per hour and $4 per pedestrian per hour are suggested
for vehicle and pedestrian delays, respectively, in Ma et al., (2014).
The computer program for the optimization model is written in C# and solved using Gurobi 6.0.4 (Gurobi Optimization,
Inc., 2016). All the computational tests are performed on a PC equipped with an Intel 2.8 GHz CPU with 4GB memory. To
produce the optimal solution, the model is solved for one hundred (102 ) cases within 1 s, which serves well the purpose for
engineering applications.

4.2. Feasibility of approximate delay functions

In the modelling process, the convex, quadratic formulae of approximate pedestrian delays, Eqs. (57) and (58), are pro-
posed to replace the nonconvex, quadratic formulae Eqs. (52) and (56), respectively, so that the original nonconvex, quadratic
programming model turns into a convex, quadratic programming model. A simplified intersection with only one two-stage
crosswalk in Approach 1 derived from Fig. 1 is taken to verify the feasibility of the proposed approximate formulae. Three
hundred feasible signal plans are randomly generated satisfying the basic constraints (i.e., Eqs. (2)–(13) and Eqs. (19)–(39))
with a fixed cycle length of 120 s. Then the weighted pedestrian and vehicle delay can be calculated by Eq. (61). If the
original nonconvex formulae Eqs. (52) and (56) are used in Eq. (61), the calculated delay is denoted as accurate delay. If the
approximating convex formulae Eqs. (57) and (58) are used instead, the calculated delay is denoted as approximate delay.
The relationship between these two kinds of delays is depicted in Fig. 6.
Fig. 6 shows that the accurate and approximate delays are not equal due to the different analytical expressions (i.e.,
Eqs. (52), (56)–(58)). However, they have a highly positive correlation no matter a larger weight is given to pedestrian
delay (β = 0.1) or vehicle delay (β = 0.9). This means that a signal plan with a lower approximate delay is more likely to
result in a lower accurate delay. Therefore, it is feasible to use the approximate delay in the objective function in the signal
optimization model instead of the accurate delay. In addition, more tests show that the factors (γ 1 , γ 2 , γ 3 , and γ 4 ) in Eqs.
(57) and (58) are robust, and maintain stable performance when vehicle and pedestrian demand vary.

4.3. Optimization results and discussion

The optimal vehicle and pedestrian signal plan is shown in Fig. 7. The ending times of the eight vehicle phases are
illustrated as well as the starting times of the pedestrian Walk interval, the FDW interval, and the Don’t Walk or Stop
interval at a one- or two-stage crosswalk in each approach. Phases 2, 4, 6, and 8 are noticeably longer than the other phases
due to the relative large through vehicular traffic and pedestrian traffic. The optimal cycle length is the minimum value 60 s,
which is consistent with the intuition that a shorter cycle length is usually preferred when delay minimization is used as
the optimization objective. Although a cycle length of 60 s may seem short, all constraints are satisfied with current vehicle
and pedestrian demand. The two-stage crosswalks in Approach 1 and Approach 3 make it possible to accommodate crossing
pedestrians in a short cycle. A larger cycle length will be produced if pedestrian demand increases.
To validate the effectiveness of the proposed model in this paper, the pedestrian delay in the objective function and the
capacity constraints of the refuge island at a two-stage crosswalk are removed for comparison. Three levels of pedestrian
 C. Yu et al. / Transportation Research Part B 98 (2017) 135–153 147

Fig. 7. Optimal vehicle and pedestrian signals.

Table 3
Optimization results.

Pedestrian Pedestrian Limited space Vehicle Pedestrian Weighted Equivalent delay for different vehicle occupancy (s)
demand level delays in on refuge delay (s) delay (s) delay (s)
objective island 1 1.5 2 2.5 3

μped * = 0.5   684.88 1781.40 1123.49 2466.28 2808.73 3151.17 3493.61 3836.05
×  630.28 2809.39 1501.92 3439.66 3754.80 4069.94 4385.08 4700.22
Increase (%) −7.97 57.71 33.68 39.47 33.68 29.16 25.52 22.53
× × 630.28 2809.39 1501.92 3439.66 3754.80 4069.94 4385.08 4700.22
Increase (%) −7.97 57.71 33.68 39.47 33.68 29.16 25.52 22.53
μped = 1   684.88 3562.80 1836.05 4247.68 4590.13 4932.57 5275.01 5617.45
×  684.88 3760.80 1915.25 4445.68 4788.13 5130.57 5473.01 5815.45
Increase (%) 0.00 5.56 4.31 4.66 4.31 4.01 3.75 3.52
× × 630.28 5618.78 2625.68 6249.05 6564.19 6879.33 7194.47 7509.61
Increase (%) −7.97 57.71 43.01 47.12 43.01 39.47 36.39 33.68
μped = 2   684.88 7125.60 3261.17 7810.48 8152.93 8495.37 8837.81 9180.25
×  684.88 7521.60 3419.57 8206.48 8548.93 8891.37 9233.81 9576.25
Increase (%) 0.00 5.56 4.86 5.07 4.86 4.66 4.48 4.31
× × 630.28 11,237.55 4873.19 11,867.83 12,182.97 12,498.10 12,813.24 13,128.38
Increase (%) −7.97 57.71 49.43 51.95 49.43 47.12 44.98 43.01
∗
Three levels of pedestrian demand are set by μped r1 and μped r2 .

demand are tested, which is the product of the basic pedestrian demand in Table 2 and a factor μped . The optimization
results are shown in Table 3. Under low pedestrian demand (μped = 0.5), vehicle delay decreases by ∼8% after pedestrian de-
lay is removed from the objective function while pedestrian delay increases by ∼58%. Therefore, the inclusion of pedestrian
delay in the objective makes a significant difference, especially from the perspective of pedestrian traffic. Actually, in that
case, the refuge islands at the two-stage crosswalks in Approach 1 and Approach 3 suffice to accommodate crossing pedes-
trians. The capacity constraints of the refuge islands are loose. Therefore, no differences are observed when the capacity
constraints are further relaxed. As pedestrian demand increases (μped = 1), the capacity constraints become tight. Feasible
solutions are reduced greatly. As a result, the increase of pedestrian delay is only 5.56% while no decrease of vehicle delay
occurs after the removal of pedestrian delay from the objective. As expected, however, pedestrian delay increases and vehi-
cle delay decreases in the same way as the case with μped = 0.5 after the relaxation of the island capacity constraints. With
further increasing pedestrian demand (μped = 2), the optimization results remain the same as the case with μped = 1 due to
the capacity constraints. Note that the proposed model may be infeasible with extremely high pedestrian demand due to
the limited space on the refuge island in Approach 1 and Approach 3.
It is also observed that same results are produced with different levels of pedestrian demand when neither the pedestrian
delay nor the capacity constraints of the refuge islands is considered. That is, the total vehicle delay, the relative decrease
of vehicle delay, and the relative increase of pedestrian delay are the same although the total pedestrian delay increases
with increasing pedestrian demand. This is intuitive because pedestrian demand is supposed to have no impacts on the
optimization results of the model that does not consider the pedestrian delay in the objective and the limited space on the
refuge islands.

4.4. Comparison between two- and one-stage crosswalks

It is widely believed that the refuge island of a two-stage crosswalk benefits pedestrian traffic at the expense of traffic
operational efficiency. Because one extra lane might be available for vehicular traffic otherwise. Nevertheless, it should be
148 C. Yu et al. / Transportation Research Part B 98 (2017) 135–153

Fig. 8. Relative increase of delay (%) when two-stage crosswalks in Approach 1 and Approach 3 are replaced by one-stage crosswalks: (a) total pedestrian
delay, and (b) total vehicle delay.

recognized that more green time can be allocated to vehicle flows as a result of the reduced FDW clearance interval when
a one-stage crosswalk is replaced by a two-stage one. Therefore, it is a trade-off between spatial and temporal resources.
A further analysis is conducted to investigate the impact of a two-stage crosswalk on vehicular traffic compared with a
one-stage crosswalk.
The two-stage crosswalks in Approach 1 and Approach 3 in Fig. 1 are replaced by one-stage crosswalks so that one extra
lane is added for through traffic in both approaches. Because, in the real world, the main traffic flows at an intersection of a
main road and a secondary road are usually through flows. The FDW clearance intervals in Arm 1 and Arm 3 are prolonged
to 24 s accordingly. For better illustration, the pedestrian volumes in Table 2 are halved and the cycle length is taken as
100 s. Flow factors μ13 and μ24 are introduced to represent the varying through flows in Approaches 1, 3 and Approaches
2, 4, respectively. The vehicle arrival rates in Phases 2, 4, 6, and 8 in Table 2 are modified as rv2 = 550μ13 , rv4 = 300μ24 ,
rv6 = 500μ13 , and rv8 = 100μ24 , respectively.
The relative increase of pedestrian and vehicle delays after one-stage crosswalks are installed in Approach 1 and Ap-
proach 3 are portrayed in Fig. 8(a) and (b), respectively. Obviously, a two-stage crosswalk benefits pedestrian traffic (i.e.,
positive impacts). The increase of pedestrian delay exceeds 7.5% with varying traffic flows. In contrast, the impact of a two-
stage crosswalk can be either positive or negative in terms of vehicle delay. In most cases, the impact is negative (i.e.,
increasing vehicle delay), which means that the benefits got by setting a central refuge island is relatively smaller than the
capacity drop resulted by reducing a lane. However, the impact is positive (i.e., decreasing vehicle delay) with relatively
low flows of through traffic in the four Approaches (e.g., μ13 = 0.6, μ24 = 0.6). This means that the reduced vehicle delay,
owing to the shortened FDW clearance time by the installed refuge island at a two-stage crosswalk, dominates. Therefore, a
two-stage crosswalk at an intersection may benefit both vehicular traffic and pedestrian traffic in some circumstances.

5. Conclusions and recommendations

This paper presents a convex (quadratic) programming approach to optimize signal timing for both vehicular and pedes-
trian traffic at an isolated intersection with one- and two-stage crosswalks. The optimization model assumes undersaturated
condition for vehicular traffic. The objective is to minimize vehicle and pedestrian delays. Quadratic approximations of the
delays are proposed in the objective function. The limited space on the refuge island at a two-stage crosswalk is explicitly
taken into consideration, and is captured in linear forms. The proposed model can be solved efficiently for global optima
by existing solvers, which serves well the purpose of further engineering applications, such as real-time signal control and
signal coordination.
Several numerical studies have shown the effectiveness of the proposed signal optimization approach. A correlation anal-
ysis is conducted to verify the quadratic approximation of the delays. In particular, a highly positive correlation is found
between the exact delay (calculated by the original nonconvex formulae) and approximate delay (by the proposed quadratic
formulae). The optimization results account for the limited space on the refuge island at a two-stage crosswalk, and the
pedestrian delay to be minimized. Further analyses show that a two-stage crosswalk at an intersection can benefit both
vehicular and pedestrian traffic in some circumstances.
The model relies on the undersaturated vehicular traffic condition, as the primary objective is to minimize delays. Under
oversaturation condition, the maximization of intersection capacity may be considered as the objective instead. In this case,
all of the proposed linear constraints remain effective, and the signal optimization problem can be formulated as a linear
program when reserve capacity is adopted (Wong and Heydecker, 2011; Wong and Wong, 2003a,b). The proposed framework
 C. Yu et al. / Transportation Research Part B 98 (2017) 135–153 149

will be further validated using field data in a future study. In order to cope with significant traffic flow variations, adaptive
signal plans may be devised and optimized for different time periods (e.g., peak and off-peak) or day categories (e.g. week-
days and holidays). The extension of the proposed model to actuated control and signal coordination has been conceived.
Other topics such as the interactions among pedestrian platoons and the integration of long and short cycle lengths will
also be explored.

Acknowledgements

The research is supported by the National Natural Science Foundation of China (Nos. 51238008 and 51178345), the Fok
Ying-Tong Education Foundation (No. 151076), and the Science and Technology Commission of Shanghai Municipality (No.
15DZ110 080 0). The authors also would like to thank Professor Yanfeng Ouyang for his constructive comments, which lead
to a much improved paper.

Appendix I

Figs. 9–17

Fig. 9. Case 2.

Fig. 10. Case 3.

150 C. Yu et al. / Transportation Research Part B 98 (2017) 135–153

Fig. 11. Case 4.

Fig. 12. Case 5.

Fig. 13. Case 6.

C. Yu et al. / Transportation Research Part B 98 (2017) 135–153 151

Fig. 14. Case 7.

Fig. 15. Case 8.

Fig. 16. Case 9.

152 C. Yu et al. / Transportation Research Part B 98 (2017) 135–153

Fig. 17. Case 10.

References

“Brian” Park, B., Yun, I., Ahn, K., 2009. Stochastic optimization for sustainable traffic signal control. Int. J. Sustainable Transp. 3, 263–284.
Allsop, R.E., 1981. Computer Program Sigset for Calculating Delay-Minimising Traffic Signal Timings – Description and Manual for Users. University College,
London Transport Studies Group, Research Report.
Cascetta, E., Gallo, M., Montella, B., 2006. Models and algorithms for the optimization of signal settings on urban networks with stochastic assignment
models. Ann. Oper. Res. 144, 301–328.
Dion, F., Rakha, H., Kang, Y.-S., 2004. Comparison of delay estimates at under-saturated and over-saturated pre-timed signalized intersections. Transp. Res.
Part B 38, 99–122.
Gurobi Optimization, Inc., 2016. Gurobi optimizer reference manual.
Han, K., Liu, H., Gayah, V.V., Friesz, T.L., Yao, T., 2016. A robust optimization approach for dynamic traffic signal control with emission considerations. Transp.
Res. Part C. 70, 3–26.
Head, L., Gettman, D., Wei, Z., 2006. Decision model for priority control of traffic signals. Transp. Res. Rec. 1978, 169–177.
Heydecker, B.G., 1992. Sequencing of traffic signals. In: Griffiths, J.D. (Ed.), Mathematics in Transport Planning and Control. Clarendon Press, Oxford,
pp. 57–67.
Khalighi, F., Christofa, E., 2015. Emission-based signal timing optimization for isolated intersections. 94th Annual Meeting of the Transportation Research
Board.
Koonce, P., Rodegerdts, L., Lee, K., Quayle, S., Beaird, S., Braud, C., Bonneson, J., Tarnoff, P., Urbanik, T., 2008. Traffic Signal Timing Manual. Federal Highway
Administration.
Lao, Y., Wu, Y.-J., Wang, Y., Yang, X., 2009. Applicability of signalized single and double phase midblock crossings in highly populated areas. 88th Annual
Meeting of the Transportation Research Board, Washington, D.C.
Li, L., Yang, X., Yin, L., 2010a. Exploration of pedestrian refuge effect on safety crossing at signalized intersection. Transp. Res. Rec. 2193, 44–50.
Li, M.-T., Gan, A., 1999. Signal timing optimization for oversaturated networks using Transyt-7f. Transp. Res. Rec. 1683, 118–126.
Li, M., Alhajyaseen, W.K., Nakamura, H., 2010b. A traffic signal optimization strategy considering both vehicular and pedestrian flows. 89th Annual Meeting
of the Transportation Research Board, Washington, D.C.
Little, J.D., Kelson, M.D., Gartner, N.M., 1981. Maxband: a program for setting signals on arteries and triangular networks. Transp. Res. Rec. 795, 40–46.
Liu, H., Han, K., Gayah, V.V., Friesz, T.L., Yao, T., 2015. Data-driven linear decision rule approach for distributionally robust optimization of on-line signal
control. Transp. Res. Part C 59, 260–277.
Ma, D., Nakamura, H., 2010. Cycle length optimization at isolated signalized intersections from the viewpoint of emission. In: 7th International Conference
on Traffic and Transportation Studies. Kunming, China. ASCE, pp. 275–284.
Ma, W., Liao, D., Liu, Y., Lo, H.K., 2014. Optimization of pedestrian phase patterns and signal timings for isolated intersection. Transp. Res. Part C. 58,
502–514.
Mascia, M., Hu, J., Han, K., North, R., Van Poppel, M., Theunis, J., Litzenberger, M., 2016. Impact of traffic management on black carbon emissions: a mi-
crosimulation study. Netw. Spatial Econ. doi:10.1007/s11067- 016- 9326- x.
MOHURD, 2011. Code for Planning of Intersections on Urban Roads. China Planning Press, Beijing.
Murat, Y., Kikuchi, S., 2008. Fuzzy optimization approach: comparison with the classical optimization method using the problem of timing a traffic signal.
Transp. Res. Rec. 2024, 82–91.
National Transportation Operations Coalition (NTOC), 2012. 2012 National Traffic Signal Report Card. Washington, DC Technical Report.
Ozan, C., Baskan, O., Haldenbilen, S., Ceylan, H., 2015. A modified reinforcement learning algorithm for solving coordinated signalized networks. Transp. Res.
Part C 54, 40–55.
Silcock, J.P., 1997. Designing signal-controlled junctions for group-based operation. Transp. Res. Part A 31, 157–173.
Stamatiadis, C., Gartner, N.H., 1996. Multiband-96: a program for variable-bandwidth progression optimization of multiarterial traffic networks. Transp. Res.
Rec. 1554, 9–17.
Szeto, W.Y., Lo, H.K., 2005. Strategies for road network design over time: robustness under uncertainty. Transportmetrica 1, 47–63.
Tong, Y., Zhao, L., Li, L., Zhang, Y., 2015. Stochastic programming model for oversaturated intersection signal timing. Transp. Res. Part C 58, 474–486.
Transportation Research Board, 2010. Highway Capacity Manual. Transportation Research Board, National Research Council, Washington, D.C.
Wang, X., Tian, Z., 2010. Pedestrian delay at signalized intersections with a two-stage crossing design. Transp. Res. Rec. 2173, 133–138.
Webster, F.V., 1958. Traffic Signal Settings. Her Majesty’s Stationery Office, London, England.
Wong, C.K., Heydecker, B.G., 2011. Optimal allocation of turns to lanes at an isolated signal-controlled junction. Transp. Res. Part B 45, 667–681.
Wong, C.K., Wong, S.C., 2002. Lane-based optimization of traffic equilibrium settings for area traffic control. J. Adv. Transp. 36, 349–386.
 C. Yu et al. / Transportation Research Part B 98 (2017) 135–153 153

Wong, C.K., Wong, S.C., 2003a. A lane-based optimization method for minimizing delay at isolated signal-controlled junctions. J. Math. Modell. Algorithms
2, 379–406.
Wong, C.K., Wong, S.C., 2003b. Lane-based optimization of signal timings for isolated junctions. Transp. Res. Part B 37, 63–84.
Wong, C.K., Wong, S.C., Tong, C.O., 2005. Optimization methods for off-line traffic signal settings: recent advances and prospective future research. J. Transp.
Syst. Eng. Inf. Technol. 5, 36–54.
Wong, S.C., Yang, C., 1999. An iterative group-based signal optimization scheme for traffic equilibrium networks. J. Adv. Transp. 33, 201–217.
Yin, Y., 2008. Robust optimal traffic signal timing. Transp. Res. Part B 42, 911–924.
Zhang, L., Yin, Y., Chen, S., 2013. Robust signal timing optimization with environmental concerns. Transp. Res. Part C 29, 55–71.
Zhang, L., Yin, Y., Lou, Y., 2010. Robust signal timing for arterials under day-to-day demand variations. Transp. Res. Rec. 2192, 156–166.
Zhang, X., Chang, G.-l., 2014. Optimal control strategies with an extended cell transmission model for massive vehicular-pedestrian mixed flows in the
evacuation zone. J. Adv. Transp. 48, 1030–1050.