Professional Documents
Culture Documents
Pour Mehra B 2019
Pour Mehra B 2019
follower or lead, determine the trajectory computation model. dnl (t) : [tn0l , tncross
l
] → [0, dl ] the center-lane distance to
We define a lead vehicle as a vehicle whose movement stop bar profile of vehicle nl , ∀ l ∈ L, ∀ n l =
is not limited by other vehicles. Any vehicles movement 1, . . . , Nl , (in ft)
can affect its follower depending on a variety of factors The equation for AV Trajectory Optimization (AVTO):
such as speed, gap and braking capability. Considering the ⎧
cumulative effect in traffic stream in a lane, the lead vehicles ⎨
⎨ LAVO(s(t), V max , Vm(n l)
, an l , an l )
cross acc dec
⎨
⎨
have a key role in intersection total delay. We formulate a ⎨
⎨ ∀ nl = 1, ∀ l ∈ L, cnl = AV
⎨
⎨
nonlinear mathematical model (1) to address the follower–lead ⎨
⎨ FAVO(d (n−1)l (t), s(t), V
max , V cross , a dec , a dec )
⎨
⎨ m(n ) nl (n−1)l
dependencies among a set of vehicles in each lane. The AV ⎨ ∀ n = 2, . . . , N , ∀ l ∈ L, c l = AV
l l nl
Trajectory Optimization (AVTO) model (1) captures this effect dnl (t) =
⎨
⎨ LCNVE(s(t), v nl (tnl ))
0
by recursively computing the trajectories. ⎨
⎨
⎨
⎨ ∀ nl = 1, ∀ l ∈ L, cnl = C N V
Sets, Parameters, and Indices: ⎨
⎨
⎨
⎨ FCNVE(d(n−1)l (t), s(t), Vndes , anacc , andec )
L the set of incoming lanes, l ∈ L ⎨
⎨
⎩ l l l
nl the vehicle counter in lane l, nl = 1, . . . , Nl , ∀ l ∈ L ∀ nl = 2, . . . , Nl , ∀ l ∈ L, cnl = C N V
(note the vehicles are from closest to the farthest (1)
relative to the stop bar.)
Sub-models to AVTO model:
cnl the class of vehicle nl , cnl ∈ {AV, CNV}, ∀ l ∈
L, ∀ n l = 1, . . . , Nl L AV O(.) the Lead AV Optimizer
t re f the reference point to measure time (it can be set F AV O(.) the Follower AV Optimizer
to any arbitrary but constant values through analysis LC N V E(.) the Lead CNV Estimator
time horizon in seconds) FC N V E(.) the Follower CNV Estimator
t the time relative to the reference time t re f , (in sec- Eq. (1) defines the trajectory of nth vehicle in lane l
onds) to be a function of signal status at time t (i.e s(t)), char-
tn0l the detection time of vehicle nl relative to the refer- acteristics of the vehicle itself (i.e. anacc l
, andecl
), speed lim-
ence time t re f , (in seconds), nl = 1, . . . , Nl , ∀ l ∈ L its (i.e. V max , Vm(nl ) ), and trajectory of its lead vehicle
cross
cross
tnl the cross time of vehicle nl at the stop bar rela- (i.e. d(n−1)l (t)). In this study, we assume a conventional lead
tive to the reference time t re f , (in seconds), nl = vehicle, represented by LCNVE model, tends to maintain
1, . . . , Nl , ∀ l ∈ L its arrival speed when signal indicates green. Next subsec-
m(n l ) the mapping that keeps movement of AV nl , m(n l ) : tions discuss LAVO, FAVO, and LCNVE models to AVTO
1, . . . , Nl → {le f t, str aight, right}, cnl = AV problem.
dl the detection range of lane l ∈ L, (in ft) 1) Lead AV Optimization (LAVO) Model: In this section,
Vndes the desired speed of the follower CNV nl (in ft/s), we formulate and solve the Lead AV Optimizer (LAVO)
l
nl = 2, . . . , Nl , ∀ l ∈ L, cnl = CNV problem. The problem aims to minimize an automated lead
V max the speed limit inside the detection range, (in ft/s) vehicle’s travel time delay. The benefit of planing ahead
cross the speed limit at the stop bar for movement m(n ),
Vm(n l
at vehicle level has roots in accessible real-time arrival of
l)
(in ft/s) vehicles along with ability to harmonize movement of AV
anacc the maximum acceleration rate of vehicle nl , n l = with the signal phase and timing. IICA advises an AV to
l
1, . . . , Nl , ∀ l ∈ L (in ft/s2 ) adjust its motion to prevent lost times due to accelerating from
andec the maximum deceleration rate of vehicle nl , nl = standstill. We consider the functional form for AV trajectory—
l
1, . . . , Nl , ∀ l ∈ L (in ft/s2 ) also called the space-time function— that have several prop-
the set of phases to serve all available movements at erties: (1) provides acceleration/deceleration stages for AV
the intersection, φ ∈ trajectories; (2) produces paths that AVs can implement;
the phase-lane incidence matrix = [ηφl : ∀ l ∈ L, (3) parametrizes trajectories that are tractable to optimize.
∀ φ ∈ ], where ηφl is 1 if lane l belongs to phase φ, Hence, we formulate the lead AV trajectory considering
0 otherwise three ordered stages as:
⎧
s(t) = (tφs (t), G φ (t), Yφ (t), A R φ (t)) the vector that con- ⎨ d 1 (t) ∀ t ∈ [tn0l , tn1l ]
⎨
⎨ nl
tains signal status at time t
tφs (t) the time stamp when the green interval for phase φ dnl (t) = dn2l (t) ∀ t ∈ (tn1l , tn2l ] (2)
⎨
⎨
begins ⎩ d 3 (t) ∀ t ∈ (t 2 , t 3 ]
nl nl nl
G φ (t) the duration of green interval for phase φ
Yφ (t) the duration of yellow interval for phase φ where:
A R φ (t) the duration of all-red interval for phase φ dni l (t) is the distance to the stop bar during the i th stage,
Notice all phases that belong to except φ, receive (in ft)
red during [tφs (t), tφs (t) + G φ (t) + Yφ (t) + A Rφ (t)] tni l is the time stamp when the i th stage ends, (in sec-
onds). Note tn3l = tncross
l
since the vehicle nl crosses
Variables: at the end of the third stage.
v nl (t) : [tn0l , tncross
l
] → R+ the speed profile of vehicle nl , The stage-wise trajectory by Eq. (2) provides flexibility
∀ l ∈ L, ∀ n l = 1, . . . , Nl , (in ft/s) to adjust the speed of incoming AV—if driven too slow or
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
v nl (tn3l )2 − v nl (tn2l )2
− /v nl (tn1l ) (8)
2an3l
v nl (tn3l ) − v nl (tn2l )
tn3l − tn2l = (9)
an3l
We formulate the LAVO mathematical program to mini-
mize vehicle’s travel time delay subject to constraints from
the signal, maximum speed limit, and vehicle accelerating
capability. Defining the base travel time of a vehicle to be
Fig. 3. The three stages of a lead AV’s trajectory. the amount of time to travel the detection range at its desired
speed, we minimize the total travel time delay as of stated in
fast—then maintain the speed followed by final a stage to model (10-15). The first set of constraints, (11), bounds the
readjust the speed for a safe departure. Therefore, the first feasible area to solutions which AV will be granted green or
and third stages adjust the speed by a constant acceleration or yellow. The constraint (12) ensures following the speed limit
deceleration rate, while the vehicle maintains a constant speed starting from the second stage. The constraint (13) controls
during the second stage, see Fig. 3. The speed profile of the the crossing speed at the stop bar depending on the type of
vehicle can be formulated as: movements i.e. left turn, through, or right turn. The constraints
⎧ (14) and (15) limit all acceleration or deceleration rates to the
⎨ v nl (tn0l ) + an1l × (t − tn0l ) ∀ t ∈ [tn0l , tn1l ]
⎨
⎨ range that the vehicle can execute.
v nl (t) = v nl (tn1l ) ∀ t ∈ (tn1l , tn2l ] (3)
⎨
⎨
3
dnl (tn0l )
⎩ LAVO: min (tni l − tni−1 ) − (10)
v nl (tnl ) + anl × (t − tnl ) ∀ t ∈ (tnl , tnl ]
2 3 2 2 3
v,a l
Vndes
i=1 l
where: subject to
v nl (tn0l ) is the arrival speed of vehicle nl at detection time, ηφl × tφs (t) ≤ ηφl × (tn3l − tn0l ) ≤ tφs (t)
(in ft/s)
+ G φ (t) + Yφ (t) ∀ φ ∈
v nl (tni l ) is the speed of vehicle nl at the end of the i th
stage, (in ft/s) (11)
ani l is the acceleration/deceleration rate of vehicle nl 0≤ v nl (tn1l ) ≤V max
(12)
within the i th stage, (in ft/s2 )
0 ≤ v nl (tn3l ) ≤ Vm(n
cross
l)
(13)
Fig. 3 demonstrates a sample trajectory indexed by time and
distance. andec
l
≤ an1l ≤ anacc
l
(14)
Using the instantaneous velocity differential equation dec
an l ≤ an3l ≤ acc
an l (15)
−ddnl (t) = dv nl (t) × dt, the trajectory function can be
derived as: The gradient of the travel time delay, function 10, indicates
an1l monotonic variation with respect to the set of variables. This
dn1l (t) = dnl (tn0l ) − v nl (tn0l ) × (t − tn0l ) − × (t − tn0l )2 suggests the optimal solution to belong to the boundaries of
2 the feasible region; otherwise, the objective function could
∀ t ∈ [tn0l , tn1l ] (4) be decreased along the negative of the gradient. Therefore,
dn2l (t) = dnl (tn1l ) − v nl (tn1l ) × (t − tn1l ) model (10-15) can be solved by devising a constant time
algorithm to probe the boundaries of the feasible region for
∀ t ∈ (tn1l , tn2l ] (5) the minimizer, see Fig. 4. It first starts with assigning an
an3l obvious non-optimal value, M, to deln∗l which holds the
dn3l (t) = dnl (tn2l ) − v nl (tn2l ) × (t − tn2l ) − × (t − tn2l )2 global minimum value of the objective by the end of the
2
procedure. The indicator f lag is used to detect infeasibility
∀ t ∈ (tn2l , tn3l ] (6) due to an empty feasible region. The set A includes all LAVO
decision variables, and every time a variable is selected to
Note the trajectory function has to comply with first degree
determine the edge to explore. Then the for-loop fixes the
of continuity— meaning not only the distance but also the
rest of variables to the bounds; note there may be multiple
speed of vehicle remains differentiable. This is to provide a
combinations based on the choice of either the lower bound
smooth speed transition between stages. Finally, using fun-
or the upper bound. When all variables except the selected
damental motion equations, the following equations gives the
one are determined, a single-variable constrained minimization
travel time of stages:
problem can be easily solved to find the minimum point on
v nl (tn1l ) − v nl (tn0l ) that edge. Repeating this procedure for all edges, the algorithm
tn1l − tn0l = (7) outputs the global minimum, if any exists.
an1l
2) Follower AV Optimization (FAVO) Model: The Follower
v nl (tn1l )2 − v nl (tn0l )2 AV Optimizer (FAVO) model computes the trajectory of an
tn2l − tn1l = dnl (tn0l ) −
2an1l automated follower. If the vehicle in front is going to cross
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
Eq. (16), shown at the bottom of this page, yields the speed
Fig. 8. Enhanced adaptive signal control.
profile of follower CNV, where:
t is the time steps to compute trajectory points and the
reaction time of CNVs
– The unserved vehicle is in a lane that belongs to the
v nl (t + t) is the speed of follower vehicle t seconds
last scheduled phase. Based on the trajectory of this
after t
vehicle, extend the green interval to ensure safe cross
L nl is the length of nth vehicle in lane l
at the stop bar.
In the context of the AVTO model, the algorithm in Fig. 7
– The unserved vehicle is in a lane that does not belong
implements the Gipps car-following model to compute a
to the last scheduled phase. Schedule the phase that
CNV’s trajectory.
serves this vehicle for a green interval long enough to
guarantee safe cross at the stop bar.
B. Enhanced Adaptive Signal Control
The timing for each phase should meet several practical
With Trajectory Optimization
criteria. Any green interval lower than a minimum green or
The IICA framework makes SPaT decisions that are in higher than a maximum green causes too frequent or late phase
coordination with the planned trajectories to avoid lost times switches which forces excessive delay to vehicles and should
due to low green time utilization. This section devises an be avoided.
adaptive signal logic which controls whether to extend or Fig. 9 illustrates the association of each arrival inter-
switch an ongoing phase. The proposed algorithm re-evaluates val, tφarr , and the corresponding SPaT decision for a few
the signal control status every time a new vehicle is detected. consecutive phases. The minimum time to travel the detection
As shown in Fig. 8, a traffic generation module emulates the lag
range necessitates a lag time, tφ , between the end of the
arrival of the vehicles at the detection distance of the lanes. green and the end of the associated arrival interval. In other
Before solving AVTO model (1), the position of the arrived words, the lag time represents the time between when the
vehicle is updated assuming it traveled at the detected speed algorithm makes signal decision and when the corresponding
during the message transmission time. Given the ongoing traffic will depart at the stop bar.
signal phase and timing, the solution to AVTO model (1)
for the entered vehicles determines if the upcoming green
IV. A LGORITHM I MPLEMENTATION
intervals serve all the vehicles. In case the ongoing signal
AND N UMERICAL R ESULTS
plan is insufficient, the algorithm asks the signal controller
for either enough green extension or a switch of the phase. We programmed the proposed IICA process in [19] and was
Therefore, the decision on signal phase and timing falls into run on an Ubuntu machine with Intel Core i7-8550U CPU
one of the following categories: and 8 GB RAM with no noticeable delay per iteration. A test
• The ongoing signal plan provides all the vehicles enough four-leg intersection as shown in Fig. 10 was considered. The
green time to safely cross the intersection. intersection includes six incoming lanes, four discharge lanes,
• The ongoing signal plan is insufficient, and: and two approaches with exclusive left turn lanes.
v nl (t) v n (t)
v nl (t + t) = min v nl (t) + 2.5anl × t × (1 − des ) × 0.025 + ldes , andec
acc
l
× t
Vnl Vnl
v
(n−1)l (t)
2
+ anl × 2 × d(n−1)l (t) − dnl (t) + L nl + t × anl × t + v nl (t) +
dec dec (16)
andec
l
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
Fig. 9. Schematic SPaT (G φ , Yφ and ARφ are the green, yellow and all-
red intervals for phase φ; tφarr indicates the arrival interval in phase φ;
lag
tφ denotes the time before the end of yellow interval in phase φ).
Fig. 11. Cumulative arrival and departure curves after 15 minutes of
simulation. (Notice departures occur at the stop bar while arrivals occur at
the detection distance, in this case 1000 feet far from the stop bar.)
TABLE I
W IEDEMANN 99 PARAMETERS VALUE
Fig. 12. Sensitivity analysis for 3000 scenarios (panels in same column
associate with same saturation headway; panels in same row associate with
same measure of effectiveness on the vertical axis).
actuated signal control— the state-of-the-art control system — based on the exact definition each simulator uses to
in VISSIM [20]. In order to produce quantitative outputs collect them. This criterion makes the comparison on
comparable to the proposed IICA framework, we take three the basis of measured travel times more reliable than
primary considerations into account: a judgment based on obtained travel time delays which
• Both simulators are set up to hold identical values for VISSIM collects in a specific way.
those control parameters that are shared. This category Fig. 13 and 14 compare the Average Travel Time (ATT) per
of parameters includes simulation period, intersection mile and average effective green duration under IICA and fully
geometry, sets of phases, minimum and maximum green actuated control logic. The 45-degree line passing through the
duration, inter-arrival distribution, arrival speed distribu- origin divides the quadrant into two areas. For each scenario,
tion, length of incoming lane to collect travel time data, the logic with lower ATT per mile — equivalent to a point
vehicles acceleration/deceleration capabilities. that resides in the side of 45-degree line closer to the logic
• Simulators are calibrated to behave likely under the with higher ATT per mile—is preferable. The first row of
boundary conditions. To rule out the possibility that panels in Fig. 13 shows both IICA and actuated signal control
different car-following models for CNVs affect the travel implemented in VISSIM both operate at almost identical ATT
times, we calibrate the VISSIM’s Wiedemann 99 car- per mile. Hence, for near free-flow condition, neither presence
following model, with values reported in Table. I, to have of programmable AVs nor the optimization framework to plan
same travel time as IICA’s Gipps car-following model AV trajectories have a significant effect on ATT per mile.
under different flows when the light is green. In other This trend starts to change as flow increases to the range
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.