Advanced Gap Seeking Logic for Actuated Signal Control Using Vehicle Trajectory Data: Proof of Concept

Abstract

As detection systems improve, opportunities are emerging for using vehicle position and speed to drive signal control. This study explored how basic actuation processes might be improved by using vehicle position and speed. The position and speed of arriving vehicles were used to calculate their estimated time of arrival. Two variations on vehicle extension logic were developed that extended a green until there was a headway in the arriving traffic above a size that corresponded to a minimum flow rate, at which it was desired to terminate the phase. In the first method, the headway was measured from the point where the leading vehicle passed the stop bar, whereas the second method measured the headway from where the leading vehicle was unlikely to stop at the onset of yellow. This was compared against a control with a conventional stop bar and upstream detectors. A simulation study was carried out for an eight-phase intersection. The results suggest that at higher levels of demand, trajectory-based actuation could yield substantial reductions in delay. The trajectory-based methods were able to terminate green more efficiently, leading to reductions in delay in some cases, and reductions in emissions and fuel consumption, although there was a tradeoff in the number of split failures. These early results show promise for the development and fine-tuning of the methods.

Keywords

operations detection signal timing traffic signal

Recent innovations in sensor technology are making it increasingly possible to track vehicle position and speed in real time. In this study, these advanced sensor technology functions were referred to as advanced infrastructure sensing (AIS). AIS may be thought of as one-way communication between vehicles and infrastructure. Connected vehicles (CVs) are anticipated to facilitate two-way communication, but it is not known when CVs may form a high proportion of the vehicle fleet. Acquisition of trajectories with sensors may help fill the gap ( 1 ).

Fully actuated control is commonly used for isolated intersections. This type of control is highly efficient, capable of responding to changing traffic demands, and when properly configured can produce near optimal control. In spite of these advantages, to our knowledge, previous research has not explored how to integrate trajectory data into the actuation process, although some published methods ( 2 – 4 ) considered vehicle actuation with movement data estimated from conventional binary-output detection. One inefficiency in fully actuated control is the use of a passage timer to measure gaps, which requires that a passage time elapses, which potentially “wastes” a few seconds waiting for a second vehicle that never arrives. Direct measurement of the gap with trajectory data could avoid this wasted time. This is also possible by using a sufficiently long detector, but this requires assumptions to be made about vehicle speed. Use of trajectory data to measure both position and speed could improve gap-out methods. Consideration of a gap further upstream of the stop bar might also improve efficiency.

This study focuses on the gap detection process to evaluate potential improvements from replacing the “detection zone” concept with speed and position data. We present a proof-of-concept study carried out in simulation. The findings demonstrated that operational improvements were possible at moderate levels of traffic, indicating the potential to achieve improvements in fully actuated control.

Literature Review

Fully actuated control has been used for many years, and controller manufacturers have occasionally introduced features to improve it. Although some rather advanced features were developed before the advent of digital computers ( 5 ), these did not become commonplace until after the 1960s. These include volume density control features ( 6 ) such as added initial timing and gap reduction. Adaptive maximum green is an example of a relatively recent feature ( 7 , 8 ).

In the 1980s, inefficiencies of existing actuated control precipitated the development of a system called Microprocessor Optimised Vehicle Actuation (MOVA) ( 2 ), now widely used in the UK. This system introduces a rule-based gap seeking method. Two sets of upstream detectors are placed at 40 m (“X” detector) and 100 m (“IN” detector). After the initial minimum green period, the controller checks whether vehicles have recently crossed the X detector, using its presence for green time extension up to the “full minimum green time.” This rule is similar to conventional gap-out logic. After the full minimum green time, MOVA estimates a performance index and determines when to terminate the current phase based on anticipated arrivals.

In the United States, improvements to actuated control include changes in detection practices, such as use of lane-by-lane detection, which is more efficient than approach-based detection ( 9 , 10 ). Cesme and Furth proposed a multiheadway gap logic to address inefficiencies in phase termination across multiple lanes, achieving improvements over both approach-based and lane-by-lane detection ( 11 ).

Additional actuation processes are possible with upstream (also called “setback” or “advanced”) detection, such as Type II dilemma zone protection. Wide area detection, such as radar, which can identify the speeds and positions of individual vehicles, could be used for this purpose ( 12 ). Recently, speed and position data from radar detectors were used to directly control phase termination for both dilemma zone protection as well as to schedule the service of side-street phases to coincide with periods of low expected mainline traffic ( 13 ). In contrast to this application space, the present study focused on phase termination using such data in close range.

Control Methods

This study explored the potential uses and benefits of trajectory-based actuated control using detection near the stop bar. This section presents descriptions of the tested control methods, whereas the next section describes the test environment and presents specific settings used in the evaluation.

Stop Bar Presence Detection

In actuated signal control, after the minimum green has elapsed, the controller may extend the green until the maximum green time elapses. The controller will do so as long as detectors associated with the phase are occupied or an associated passage timer is active. The passage timer begins counting down after the last detector associated with the phase becomes unoccupied. For lane-by-lane detection, this process occurs for each lane separately, and the phase terminates after every lane has gapped out. This process checks whether queues have finished discharging by seeking a time headway greater than a certain duration. The headway (h) is equal to the reciprocal of traffic flow (q),

h = \frac{3600}{q}

(1)

where h is measured in seconds and q in vehicles per hour (vph).

Figure 1 shows the relationship between passage time and flow rate. After initial startup lost time, traffic exits the intersection at the saturation flow rate. As the queue dissipates (around 15 s in this example), the flow rate decreases. To simplify our example, let us ignore the effect of detector length and assume the gap directly corresponds to the headway. The signal operator must select a headway where it is desirable to terminate the phase. The selected headway is called the maximum allowable headway (MAH) ( 14 ).

Figure 1.

Relation between gap/passage time and flow rate, after Urbanik et al. ( 15 ).

To design for real-world operation, detection zone length must be taken into consideration. Conventional detectors use a “presence” mode, which reports an active input if any vehicle is present in the zone. The detector occupancy time can be calculated from

t_{o} = \frac{L_{v} + L_{d}}{1.47 v_{a}}

(2)

where

$L_{v}$ = vehicle length (ft),

$L_{d}$ = detector length (ft), and

$v_{a}$ = assumed vehicle speed (mph) and 1.47 converts vehicle speed from mph to ft/s.

The gap is related to headway, h, by

gap = h - t_{o}

(3)

Thus, if h is equal to the MAH, the gap time can be related to the cutoff flow rate, q, by the following relationship:

gap = \frac{3600}{q} - \frac{L_{v} + L_{d}}{1.47 v_{a}}

(4)

Given the length of the detector and some assumptions about vehicle length (20 ft is recommended by the Signal Timing Manual [ 15 ]) and speed, one can design a gap time to terminate the phase at a particular cutoff flow rate. As mentioned earlier, the use of any gap time introduces an inefficiency because the time must elapse even if there are no more vehicles within the gap. One alternative is to create a detection zone whose length corresponds to the gap between two vehicles traveling at v_a, which would permit termination immediately after the first vehicle exits the detection zone. The required length of detection zone is

L_{d} = 1.47 \times h \times v_{a} - L_{v}

(5)

Use of a longer detector would eliminate the need for a gap time and may provide a marginal operational improvement ( 16 ). However, in practice this may be difficult to accomplish because the lengths required are not always feasible, especially for higher speed traffic.

Trajectory-Based Actuation

If the positions and speeds of vehicles on an approach are known, it is possible to directly measure the presence of any gaps between vehicles on the approach. Further, it may be possible to measure the gap before it is close to the intersection, permitting more proactive decision making.

Let us define the position of a vehicle, $x_{i}$ , as the distance from the stop bar to the vehicle front end (ft), and the speed of a vehicle as $v_{i}$ (mph), where i is the approach order of the vehicle. Thus, the vehicle’s estimated time of arrival (ETA) $a_{i}$ , is

a_{i} = \frac{x_{i}}{1.47 v_{i}}

(6)

The headway between two vehicles is a_i_+ 1−a_i. This headway may seem like an appropriate measurement for determining phase termination. However, this ignores the variability in vehicle lengths. A particularly long vehicle necessarily introduces a longer headway, so using the time between ETAs might lead to unintended behavior in which the green terminates prematurely when a long vehicle such as a tractor-trailer passes through the intersection. Instead, it is possible to measure the gap between the leading vehicle rear end to the trailing vehicle front end. Thus the ETA of the leading vehicle front end is adjusted by the vehicle length, ℓ_i (ft),

a'_{i} = \frac{x_{i} + ℓ_{i}}{1.47 v_{i}} .

(7)

The time gap is $gs = a_{i + 1} - a'_{i}$ .

It is possible that a following vehicle will reach the intersection earlier than a leading vehicle. In this case, $gs$ will be negative. However, ordering the vehicles according to their ETA rather than the distance from the stop bar mitigates this issue.

In addition to measuring distance from the stop bar with ETAs, gaps between vehicles can be measured at upstream locations on the approach. In conventional detection, it is only possible to gap out after the leading vehicle has exited the detection zone. However, if this vehicle is within some amount of travel time to the intersection, the driver is unlikely to stop if the green ends. From this principle, it may be possible to terminate the green slightly earlier. To avoid creating a Type II dilemma zone, this travel time must be carefully selected such that virtually all vehicles at that position and speed would proceed. The gap seeking time becomes a range that starts from the time beneath which the vehicle will not stop t_n, ending at t_n + gs,

g s_{adv} = [t_{n}, t_{n} + gs]

(8)

In this paper, we refer to gap measurement at the stop bar as “gap seeking” (GS) and upstream measurement as “advanced gap seeking” (AGS).

Figure 2 demonstrates gap identification with a conventional presence detector and with trajectory-based actuated control. Each diagram shows three different time steps as two vehicles pass through the intersection. At t₁, the first vehicle is very close to the intersection, and the driver is extremely unlikely to stop if the signal were to turn yellow at this time. The second vehicle is a distance such that the gap between vehicles exceeds the termination threshold. Let us assume that the vehicles are traveling at the same speed and that the detection zone is set to the exact length needed to detect the desired gap. At t₂, the front end of the first vehicle has exited the detection zone, but its rear end is still on the detector, and t₃, the rear end of the first vehicle has exited the detection zone.

Figure 2.

Longer detector versus gap seeking methods: (a) detector-based, (b) gap seek, and (c) advanced gap seek.

With conventional presence detection (Figure 2a), the gap is unseen by the controller until the detector is unoccupied (t₃). In this example, the detection zone is long enough that a gap/passage time of zero can be used. If the detection zone were shorter, it would be necessary to use a passage time, meaning that the green would not terminate until a few seconds later. Using vehicle position and speed data, it is possible to terminate the green when a gap is detected. One option is to do so immediately after the leading vehicle has begun to exit the intersection (t₂), as shown in Figure 2b. This choice ensures that the lead vehicle will have a green indication until it crosses the stop line. It would also be possible to terminate the green earlier, as shown in Figure 2c, where the green is terminated as early as t₁. This choice assumes that the driver is close enough to proceed through the intersection.

The decision of when to terminate the green should balance operational efficiency with safety. Additional research is needed to better define “safe” ranges, and the decision should also consider the duration of the yellow interval and the meaning of yellow in the local jurisdiction (i.e., whether it is a permissive or restrictive yellow). In this study, we used a travel time of 2.5 s as a relatively large interval to explore the potential impact of introducing this type of operation.

Avoiding Gap-Out Errors

Use of shorter gap times results in “snappier” green termination, but there is also a risk of terminating the green prematurely (i.e., when there is still demand). In addition, at the start of green, stopped vehicles will have a speed of zero, and need time to start moving. Although the vehicles are close to the intersection, their ETAs may be very high. To resolve this issue, we assumed that vehicles with speeds under 5 mph had already “arrived” at the intersection. The advanced gap-out logic was updated to

g s_{adv} = {0, [t_{n}, t_{n} + gs]}

(9)

Mathematical Example

To offer a relatively simple example of these methods, a cell-transmission model (CTM) was used ( 17 ). In the CTM, in each time step a vehicle can travel one cell length. The layout of the CTM and the vehicle movement in each time step are presented in Figure 3. Configurations of four tested control scenarios (stop bar and trajectory-based) are summarized in Table 1. Figure 4 presents the gap-out logic for the four methods in detail.

Table 1.

Example Gap-Out Simulation Information

Control logic	Detector length (cells)	Gap time or gap seek time (time steps)	Gap out time
SB-S	2	2	t ₁₁
SB-L	4	0	t ₁₀
GS	NA	4	t ₉
AGS	NA	From 3 to 6 (inclusive)	t ₇

Note: SB-S = shorter stop bar detector; SB-L = longer stop bar detector; GS = gap seeking; AGS = advanced gap seeking; NA = not applicable.

Figure 3.

Network layout and vehicle movement in the CTM simulation.

Figure 4.

Example simulation of gap-out logic: (a) advanced gap seek, (b) gap seek, (c) longer detector, and (d) shorter detector.

In Figure 3, at time t₁, two vehicles are present in the first two cells, and three vehicles are present in the first four cells. Both the short (2-cell) detector and long (4-cell) detector will sense vehicle presence, but neither detector can identify how many vehicles exist within their zones. The GS method identifies that one vehicle has arrived at the intersection at t₁, with additional vehicles expected to arrive at t₂ and t₄. The AGS method identifies vehicle arrivals at t₄.

At t₇, the shorter and longer detectors both sense vehicle presence because of Vehicle 4. GS anticipates that Vehicles 4 and 5 will both clear the intersection by t₉. AGS ignores Vehicle 5 because it expects this vehicle to cross the intersection even if the signal turns yellow. AGS identifies a gap from t₁₀ to t₁₃ (as shown in Figure 4a). The controller terminates green at the end of t₇ and yellow begins at t₈.

At t₉, the shorter and longer detectors will both sense vehicle presence because Vehicle 5 is present at cell C₁. GS estimates that one vehicle will arrive at t₉, and that no additional vehicles will arrive in the next four time steps (as shown in Figure 4b), so the controller terminates the phase at the end of t₉ and yellow begins at t₁₀.

At time t₁₀, the shorter and longer detectors both sense that no vehicle is present. The longer detector is able to immediately measure the gap across C₁ to C₄, and can terminate current green at t₁₀, and yellow begins at t₁₁ (Figure 4c). However, with a shorter detector, the controller can see the state of cells C₁ and C₂ but C₃ and C₄ are unknown. To handle this situation, a passage timer is used. In this case a passage time of two time steps is employed. After these have elapsed without any further vehicle arrivals, the passage timer will count down to zero, the phase terminates at the end of t₁₁ and yellow starts at t₁₂ (Figure 4d).

Another way of viewing this situation is that shorter detectors rely on temporary storage of data about recent events with use of the passage timer (which effectively stores the time of the previous vehicle’s exit while trying to measure a gap of sufficient size), whereas other methods use an instantaneous measurement using only the present situation (longer detectors that can confirm that a gap is present without having to wait), or make predictions based on anticipated arrivals (the trajectory-based GS and AGS methods).

Upstream Detection

This study focused on trajectory-based methods for stop bar detection. However, because AGS starts to enter into the territory of upstream or setback detection, some control scenarios using upstream detectors were included for operational comparison. A full comparison considering dilemma zone protection will be a topic for future study.

Upstream detection can be used for actuation in lieu of or in addition to stop bar detection. This mode of operation also uses a passage time, but rather than detecting a headway, the goal is to extend the green long enough that a vehicle passing the detector can travel to a location where, if the green terminates, the vehicle is very likely to continue proceeding through the intersection. The passage time (s) would be

passage = \frac{x_{\det} - x_{go}}{1.47 v_{a}}

(10)

where

v_a = assumed vehicle speed (mph),

x _det = distance from the stop bar to the upstream detector (ft), and

x _go = distance from the stop bar below which a vehicle traveling at v_a is considered likely to proceed if the signal turns yellow.

The choice of x_go is a matter of policy. Some operators may use x_go = 0, which effectively means that they desire the signal to remain green until the driver has crossed the stop bar, whereas values >0 will end yellow at an earlier distance and time ( 18 ).

If no stop bar detection is available on the subject movement, then it is necessary to implement a mechanism to ensure that enough green time is provided to clear queues that have accumulated during red. This can be done by using a long minimum green time or with added initial timing ( 15 ).

Microsimulation Environment

The traffic microsimulation software VISSIM was used to implement and evaluate alternative phase termination logic. VISSIM includes several phase-based signal controllers, such as the ring-barrier controller, virtual Econolite ASC/3, and others. These signal controllers are able to implement conventional signal control logic that is extremely similar or identical to that used in real-world controllers. However, the core decisions of conventional logics are not explicitly customizable. In addition, these controllers use only the detector presence data from the simulation, and do not possess the ability to use the vehicle speed and position data.

Traectory-Enabled Signal Controller

VISSIM also provides an application programming interface (API) for signal control consisting of an external controller dynamic link library (DLL) file in which the user can add their own custom logic using C+ +. A simpler version of this external controller was used in previous studies ( 19 ), with the addition of a status display that can show the internal states of the controller at runtime. The same platform was used to develop a fully actuated multi-ring, multiphase signal controller with all of the basic actuation functions. A view of the optional status display is presented in Figure 5. This external controller executes basic time-of-day pattern selection and can execute realistic, fully actuated eight-phase control and includes a basic coordinator. This platform was selected for the integration of vehicle position and speed data entirely within the external controller’s actuation and phase-switching processes, as opposed to placing calls to a controller from a secondary controller routine.

Figure 5.

Signal controller status display.

Extracting Data From VISSIM

The external signal controller DLL receives detector state data (presence or absence of a vehicle) from VISSIM (along with a few other items) and sends back signal state data ( 20 , 21 ). However, this API is unable to directly retrieve any vehicle (e.g., positions, speeds) or network (link, lane) data from the simulation. Instead, the data must be implemented into the control by some other means. Two ways to extract vehicle position and speed data were tested in this study:

Driver model. VISSIM provides an external driver model API in the form of a DLL written in C + +. This can replace the internal driver behavior model with a user-defined model, but also can be used to access vehicle data without changing the driver behavior. After initiating the simulation and the driver model DLL, during simulation runtime, the processes in the DLL loop through each vehicle for which the driver model is configured, extract vehicle position and speed data, and send the relevant data to the external controller. In previous work, Feng et al. used the driver model DLL to collect vehicle trajectory data of a low penetration rate of CVs through socket programming ( 22 , 23 ).

VISSIM component object model (COM). VISSIM-COM is a hierarchical model through which almost all the VISSIM attributes and methods can be manipulated ( 20 ). In this study, Python was used to write an external COM script that initiates a VISSIM object, controls the simulation, extracts information from the simulation, and sends the vehicle speed and position data to the external signal controllers. Several studies ( 24 – 26 ) have used COM to implement customized signal control logic in VISSIM.

Sending Data to External Controllers

In this study, a memory space called a named pipe was used to communicate data from either the driver model or COM to the external signal controller DLL. A named pipe is a variety of interprocess communication that establishes a memory space that multiple processes can access. Named pipes can both send large amounts of data and maintain runtime efficiency. Further detailed information can be found in Microsoft documentation on named pipes ( 27 ). For each external controller, the driver model DLL or Python-COM script creates a server-side named pipe, and each external controller creates a client-side named pipe. Through this pipe, either the driver model or Python-COM script can send relevant vehicle data as string to external controllers.

Runtime Comparison

The choice of the method of conveyance and the reporting frequency have an impact on the simulation performance, which can be nontrivial with the use of larger networks and more scenarios. To decide on the method, we compared Python-COM with the driver model, DLL. The runtime performance results are presented in Figure 6, which shows the real-time factor (i.e., the speed of simulation in multiples of real time) over the simulation runtime. The series “Only controllers” refers to the situation in which no vehicle position/speed data are transferred, and only the external controller API is running. This option has the highest real-time factor. The two options using COM have higher real-time factors than those that use the driver model for both frequencies of 1/s or 10/s. In fact, using COM with 10/s has a much higher real-time factor than the driver model with a 1/s frequency. Because Python-COM exhibited better performance, it was selected for integration into the external signal controller framework.

Figure 6.

Simulation runtime comparisons.

Vehicle Data Processing

To make use of vehicle positions, the VISSIM network data (vehicle position, lane, and link) must be mapped to the intersection location. The user defines this mapping, which is encoded in an XML file that the external controller runs at the simulation initiation. The XML file also contains the signal timing data (minimum green, yellow change interval, red clearance interval, etc.). The Python script extracts vehicle link, lane, position, speed, and length data from the VISSIM network; all of the data are brought in one step (it is possible to collect data on a per-link basis but this is much more time consuming; it is more efficient to send a larger amount of data in fewer VISSIM-COM exchanges). The script then filters through these data and sends the relevant data to each external controller.

Named pipe communication allows the data to be sent as a string. For each link, a string representing the vehicles is created. The following is an example of a typical string:

Link 1 1 1 11 20 33

1 1 12 20 32

1 1 15 20 35

Link 2 2 2 11 20 33

2 1 11 20 33

2 1 15 20 35

2 1 21 30 33

Different vehicles are separated by a line separator, while vehicle attributes are separated by spaces (the five numbers representing the link number, lane number, vehicle position, vehicle length, and vehicle speed). A long one-dimensional array for links stores this information. Later, the Python script sends the data to the relevant external signal controllers.

The external signal controller DLL receives these data and further processes them. The position and speed data are mapped to internal objects that assign the data to a particular link, lane, and approach. Thus, the controller develops an internal spatial model of the intersection layout. This feature does not exist in any current signal controller to our knowledge. For each vehicle, the controller can calculate the position distance from the stop bar, using the position of the signal head as provided by the user in the external controller’s input XML file. The assignment of the signal head is also known, permitting mapping of the arriving vehicle to the assigned signal output (phase or overlap). If there is no signal head on the link, the user can alternatively define the distance to the exit point of the link. At this point, the controllers have access to the vehicle’s position and speed data and they are associated with a particular signal output, which can be used for purposes of control.

System Architecture

Figure 7 presents a block diagram of the testing platform. Figure 7a shows the steps involved for initialization of the simulation, whereas Figure 7b exhibits how the system operates throughout subsequent simulation time steps.

Figure 7.

System architecture: (a) initiate program and (b) simulation steps.

The program initiation procedure is as follows:

The Python-COM script initiates VISSIM, accesses VISSIM’s objects and attributes, and controls the simulation.

In the Python-COM script, a loop goes through all VISSIM signal controller objects and identifies external controllers.

For each external controller, a server-side named pipe is opened.

The Python-COM script parses an XML file that maps the VISSIM network to relevant external controllers.

VISSIM initiates all the conventional and external controllers.

External controllers create client-side named pipes.

External controllers import user-defined signal-related data (e.g., signal timing plan, phase configurations, detector configurations) from XML files.

During simulation runtime, the following procedure is executed on each time step:

The Python-COM script runs the VISSIM simulation for a single time step.

The VISSIM model sends vehicle speed, location, and length data to Python via COM.

VISSIM sends detector data (if any) to external controllers via the external signal controller API.

For each external signal controller, the Python-COM script creates a string of vehicle data and sends vehicle data to the controller via named pipes.

Each external signal controller processes its received string to extract vehicle position and speed data, and associate them with a movement and phase.

Each external signal controller determines the signal state.

VISSIM retrieves the signal states via the external signal controller API.

Study Network

A simulation model was developed of an isolated eight-phase intersection using the geometry of the intersection of Austin Bluffs Parkway and North Academy Boulevard in Colorado Springs, CO (Figure 8). This intersection has two protected left-turn lanes and one right-turn lane on each approach. Simulations were run for 4,500 s, with the first 900 s for simulation startup. Five iterations with different random seed numbers were run for each scenario.

Figure 8.

Geometry of the test intersection.

A hypothetical volume scenario was created that was intended to yield a moderate amount of traffic. Through-, left-turn-, and right-turn volumes were 1,125 vph, 300 vph, and 75 vph, respectively, for all for approaches. Table 2 lists the settings used for the six tested control scenarios.

Table 2.

Types of Signal Control Tested

Type of control		Detector/control parameter	Value
Abbreviation	Description of control	Detector/control parameter	Value
SB-S	Shorter stop bar detector; nonzero passage time	Stop bar detector length	40 ft
		Passage time, through	2.2 s
		Passage time, left turn	2.0 s
SB-L	Longer stop bar detector; passage time of zero	Stop bar detector length, through	168 ft
SB-L	Longer stop bar detector; passage time of zero	Stop bar detector length, left turn	156 ft
U-S	Upstream detector with short left-turn stop bar detector	Upstream detector location, through	300 ft
		Upstream detector length, through	6 ft
		Passage time, through	3.0 s
		Stop bar detector length, left turn	40 ft
		Passage time, left turn	2.0 s
U-L	Upstream detector with long left-turn stop bar detector	Upstream detector location, through	300 ft
		Upstream detector length, through	6 ft
		Passage time, through	3.0 s
		Stop bar detector length, left turn	156 ft
		Passage time, left turn	0 s
GS	Gap seeking (trajectory-based)	Gap seek time, through	2.9 s
GS	Gap seeking (trajectory-based)	Gap seek time, left turn	2.7 s
AGS	Advanced gap seeking (trajectory-based)	Advanced gap seek time, through	{0, [2.5, 5.4]} s
AGS	Advanced gap seeking (trajectory-based)	Advanced gap seek time, left turn	{0, [2.5, 5.2]} s

For the two scenarios with upstream detection, added initial timing was used for the through phase on the approach. The added initial timing settings were 1.5 s per actuation, maximum initial green of 40 s, and zero “actuations before.” The upstream detectors were located 300 ft or 5 s travel time from the stop bar. A 3-s gap was used such that the green would terminate when a vehicle was 2 s travel time from the stop bar. This was intended to mimic conventional dilemma zone protection.

Results

Aggregate Performance Measures

Table 3 presents the performance comparisons for all four control methods. This table summarizes the total delay by approach, overall intersection total delay, average cycle length, split failure, emissions results, and fuel consumption. Emissions and fuel consumption were obtained using the default models employed by VISSIM to yield results through its “Node” data collection object. The values are the average of the five simulation random seeds.

Table 3.

Signal Performance Comparisons

		SB-S	SB-L	GS	AGS	U-S	U-L
Delay (vehicle hours)	EB Left	3.28	3.33	3.15	2.74	3	2.93
	EB Through	10.00	9.92	8.87	7.94	8.91	8.77
	EB Right	0.01	0.01	0.01	0.01	0.01	0.01
	SB Left	3.34	3.33	3.14	2.83	3.06	2.98
	SB Through	9.12	8.32	8.70	8.00	9.39	8.60
	SB Right	0	0	0	0	0	0
	WB Left	3.24	3.19	3.06	2.68	2.94	2.83
	WB Through	10.47	10.29	9.06	8.31	9.38	8.71
	WB Right	0.01	0.01	0.01	0.01	0.01	0.01
	NB Left	3.48	3.50	3.15	3.01	3.05	3.01
	NB Through	9.21	8.33	8.75	8.03	9.32	8.70
	NB Right	0.01	0.01	0.01	0.01	0.01	0.01
	Total	52.17	50.24	47.91	43.57	49.09	46.55
Average cycle length (s)		87.8	83.1	77.7	67.6	79.1	75.2
Split failures		89.2	76.7	66.8	76.8	89.4	68.0
Number of stops		4,825	4,797	4,770.8	4,807.4	4,817.4	4,739.6
Throughput (veh)		5,993	6,009	6,001	6,003	5,996	5,999
Emissions, CO (g)		8,892	8,795	8,674	8,458	8,961	8,581
Emissions, NOx (g)		1,730	1,711	1,687	1,645	1,743	1,669
Fuel consumption (U.S. gal)		127.2	125.8	124.1	121.0	128.2	122.8

Note: EB = eastbound, SB = southbound, WB = westbound, NB = northbound.

For the right-turn movements, because their volumes were relatively low and right turn on red was possible, there was almost no difference between the control methods. For the other movements, the performance consistently improved with implementation of trajectory-based methods. SB-S consistently had the highest delay across all left-turn and through movements, whereas AGS had the lowest delay for these movements. AGS was tied with U-L for some movements. The performance of GS was better than both SB-S and SB-L, and was comparable with U-S, but not as good as U-L. The total delay results tended to follow the same trends. AGS had the lowest total delay, followed by U-L, then GS, then the other methods, with SB-S having the highest delay.

In relation to split failures, SB-S and U-S performed worse than the other methods. GS performed the best for minimizing total split failures, whereas AGS had a higher number than GS and both upstream detector methods. Concerning the number of stops, there was only a 2% difference between the control scenarios with the most and least stops. Total throughput was similar, suggesting there were no sustained problems with unmet demand in any of the scenarios. The emissions results tended to follow the same trends as total delay.

The results indicated that AGS may be too aggressive at terminating green, since it had a higher number of split failures than other methods. However, this aggressive phase termination also reduced the cycle length and consequently the total delay. There is likely to be a tradeoff between these two aspects of operation. Additional analysis of the impact of the headway measurement location may help balance the performance of AGS.

Statistical Comparison of Delay

Table 4 presents the difference in average delays for each approach for pairwise comparisons between the tested methods. A positive value means the method for that row experienced lower delay and the column experienced higher delay. To test for statistical significance of the results, two-tailed independent sample t-tests were conducted. Most of the comparisons saw significant differences in delay, with the exception of comparisons between SB-S and SB-L. The results showed that AGS had significantly lower delay than most of the other methods under consideration for most movements, whereas GS had significant differences in most cases, although it did not always produce lower delay.

Table 4.

Pairwise Difference of Average Delays (Seconds per Vehicle)

	EBL	EBT	SBL	SBT	WBL	WBT	NBL	NBT
	SB-S
SB-L	−0.62	0.45	−.07	2.72**	0.37	0.70	−0.44	2.94**
GS	1.69	3.80**	2.22*	1.46**	2.07*	4.51**	3.33**	1.55**
AGS	6.41**	6.87**	6.06**	3.61**	6.74**	6.95**	5.36**	3.79**
U-S	3.21**	3.64**	3.11**	−0.82*	3.62**	3.53**	4.62**	−0.38
U-L	4.19**	4.14**	4.21**	1.67**	4.94**	5.69**	5.25**	1.62**
	SB-L
GS	2.31**	3.34**	2.29**	−1.26**	1.70*	3.81**	3.77**	−1.38**
AGS	7.04**	6.42**	6.13**	0.89**	6.37**	6.25**	5.81**	0.85**
U-S	3.84**	3.19**	3.18**	−3.53**	3.25**	2.83**	5.06**	−3.32**
U-L	4.82**	3.69**	4.28**	−1.04**	4.57**	5.00**	5.69**	−1.32**
	GS
AGS	4.72**	3.08**	3.83**	2.15**	4.67**	2.44**	2.03**	2.23**
U-S	1.53	−0.15	0.89	−2.28**	1.55	−0.98**	1.29	−1.93**
U-L	2.50**	0.34	1.99**	0.21	2.87**	1.19**	1.92**	0.06
	AGS
U-S	−3.20**	−3.23**	−2.95**	−4.42**	−3.13**	−3.42**	−0.74	−4.17**
U-L	−2.22**	−2.73**	−1.84**	−1.93**	−1.80*	−1.25**	−0.11	−2.17**
	U-S
U-L	0.98	0.50	1.10	2.49**	1.32	2.17**	0.63	1.99**

Value is significantly different at 99% confidence interval.

Value is significantly different at 95% confidence interval.Note: SB-L = longer stop bar detector; GS = gap seeking; AGS = advanced gap seeking; U-S = upstream detector with short left-turn stop bar detector; U-L = upstream detector with short left-turn stop bar detector.

Considerations for Implementation

Some detection systems can identify vehicle positions and speeds, but the data are converted into detection zone occupancy because controllers cannot currently receive position and speed data from detectors. Radar detection systems have had this ability for several years, and advances in analysis of video, infrared, and LiDAR are likely to extend such capabilities to other systems in the future ( 28 – 30 ). If AIS technologies are employed for trajectory acquisition, they might achieve high levels of penetration more quickly than vehicle-to-infrastructure communication.

One challenge for video camera or radar detection is that the quality of detection may drop substantially depending on site conditions. Variations in weather, lighting, the possibility of occlusion from roadside vegetation, and other issues may degrade detector performance. The positions of vehicles that are identified on the approach but temporarily disappear from the sensor could be estimated until they reappear using certain error correction methods. Indeed, radar detection systems have been using such tools for some time.

It is possible that AIS may not be able to detect a high percentage of vehicles, or may experience a high amount of error in vehicle position and speed measurement under certain conditions. Depending on the nature of the errors, algorithm performance may degrade. “Fail-safe” modes are likely to be needed to deal with these situations. One option may be falling back to a conventional actuated control plan. More reliable detection methods could most likely result from the development of hybrid detection systems that make use of more than one type of sensing technology. Chavez-Garcia and Aycard used a fusion of sensors to get 100% object tracking accuracy on a test track ( 31 ). However, fusion could add redundancy to the detection system.

Another potential solution may be to use a trajectory prediction algorithm. Yao et al. proposed a dynamic platoon dispersion model to predict vehicle arrival time using connected vehicle data and optimized signal timing with genetic algorithms ( 32 ). Feng et al. developed an algorithm called “estimation of location and speed” to develop an arrival table using connected vehicle data ( 23 ). Both studies were conducted in microsimulation, so real-world study is needed to investigate applications to real sensor data. Further exploration of these practical issues and the accuracy of different types of existing detection systems are subjects for future study.

Conclusion

This study proposed two trajectory-based GS methods. The first, GS, looked for gaps in arrival time within the next few seconds after a vehicle’s front end crosses the stop bar. The second, AGS, searched for gaps further away from the intersection. To compare the performance of these methods, two conventional detector-based methods were also developed, one with a shorter detector (SB-S) and another with a longer detector (SB-L). SB-S required the use of a passage time whereas SB-L could use a passage time of zero. To offer a fair comparison of AGS, two methods with upstream detection were tested, with variants on long (U-L) or short (U-S) detectors on the left-turn phases. Table 2 provides details of the control programming.

The preliminary results of this proof-of-concept study demonstrated that use of trajectory data did facilitate improvements in operational performance. Conventional methods like SB-L, SB-L, or U-S were outperformed by both GS and AGS. SB-L performed slightly better than GS. AGS performed better than the other tested methods in relation to delay, although there were tradeoffs in split failures because of the more aggressive phase termination under AGS. Altogether, these results demonstrated promise for the concepts employed under the trajectory-based methods. Further research is needed to refine and field test these methods, as well as to address the practical issues discussed in this paper.

The present study focused on stop bar detection and impacts on operational efficiency. Many other applications are possible using trajectory data. Examples include platoon-based extension of green, enhanced decision-making for serving protected-permitted left-turn phases, adjustments to pedestrian clearance intervals, and red clearance extension in the event of red light running. For each application, operational data will also need to be combined with their potential impacts on intersection safety. Future research will investigate these concepts, among others, in greater detail. The introduction of these types of control concepts into the existing cabinet-controller ecosystem could potentially lay the foundation for future advances that would be enabled by a proliferation of vehicle-to-infrastructure connectivity.

Footnotes

Author Contributions

The authors confirm contribution to the paper as follows: study conception and design: C. M. Day; data collection: A. Shams; analysis and interpretation of results: A. Shams; draft manuscript preparation: A. Shams, C. M. Day. All authors reviewed the results and approved the final version of the manuscript.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was supported in part by an internal grant from Iowa State University.

ORCID iDs

Andalib Shams

Christopher M. Day

Data Accessibility Statement

Some or all data, models, or code that support the findings of this study are available from the corresponding author on reasonable request.

References

Talukder

M. A. S.

Lidbe

A. D.

Tedla

E. G.

Hainen

A. M.

Atkinson

Trajectory-Based Signal Control in Mixed Connected Vehicle Environments. Journal of Transportation Engineering, Part A: Systems, Vol. 147, 2021, p. 04021016.

Vincent

R. A.

Peirce

J. R.

MOVA: Traffic Responsive, Self-Optimizing Signal Control for Isolated Intersection. Transport and Road Research Laboratory, Crowthorne, Berkshire, 1988.

Gershenson

Rosenblueth

Self-Organizing Traffic Lights at Multiple-Street Intersections. Complexity, Vol. 17, No. 4, 2011, pp. 23–39. https://doi.org/10.1002/cplx.

Cesme

Furth

P. G.

Self-Organizing Traffic Signals Using Secondary Extension and Dynamic Coordination. Transportation Research Part C: Emerging Technologies, Vol. 48, 2014, pp. 1–15. https://doi.org/10.1016/j.trc.2014.08.006.

Tyack

F. G.

Street Traffic Signals, With Particular Reference to Vehicle Actuation. IEE Journal, Vol. 82, 1938, pp. 125–154.

Duemmel

R. A.

Operation and Theory of Timing Two-Phase Volume Density Controllers. Traffic Engineering, Vol. 36, No. 12, 1966, pp. 41–45.

Yun

Best

(Brian) Park

Evaluation of Adaptive Maximum Feature in Actuated Traffic Controller. Transportation Research Record: Journal of the Transportation Research Board, 2007. 2035: 134–140.

Click

S. M.

Rajagopalan

Evaluation of Variable Maximum Green Time to Improve Rural Traffic Signal Operations. Transportation Research Record: Journal of the Transportation Research Board, 2008. 2080: 75–83.

Smaglik

E. J.

Bullock

D. M.

Urbanik

Evaluation of Lane-by-Lane Vehicle Detection for Actuated Controllers Serving Multilane Approaches. Transportation Research Record: Journal of the Transportation Research Board, 2005. 1925: 123–133.

10.

Smaglik

E. J.

Bullock

D. M.

Sturdevant

J. R.

Urbanik

Implementation of Lane-by-Lane Detection at Actuated Controlled Intersections. Transportation Research Record: Journal of the Transportation Research Board, 2007. 2035: 81–87.

11.

Cesme

Furth

P. G.

Multiheadway Gap-Out Logic for Actuated Control on Multilane Approaches. Transportation Research Record: Journal of the Transportation Research Board, 2012. 2311: 117–123.

12.

Sharma

Bullock

Peeta

Estimating Dilemma Zone Hazard Function at High Speed Isolated Intersection. Transportation Research Part C: Emerging Technologies, Vol. 19, No. 3, 2011, pp. 400–412. https://doi.org/10.1016/j.trc.2010.05.002.

13.

Day

C. M.

Helt

Sines

Emtenan

A. M. T.

Leveraging Sensor-Based Vehicle Position and Speed Information in Traffic Signal Control With Existing Infrastructure. Proc., 2019 IEEE Intelligent Transportation Systems Conference, Auckland, NZ, IEEE, New York, 2019, pp. 4049–4054. https://doi.org/10.1109/ITSC.2019.8917351.

14.

Koonce

Rodegerdts

Lee

Quayle

Beaird

Braud

Bonneson

Tarnoff

Urbanik

Traffic Signal Timing Manual. Publication FHWA-HOP-08-024. Federal Highway Administration, Washington, D.C., 2008.

15.

Urbanik

Tanaka

Lozner

Lindstrom

Lee

Quayle

Beaird

, et al. Signal Timing Manual, 2nd ed. Transportation Research Board, Washington, D.C., 2015.

16.

Yang

Evaluation of Traffic Signal Timing Policies for Multiple Objectives. Doctoral dissertation. Iowa State University, Ames, 2021.

17.

Daganzo

C. F.

The Cell Transmission Model: A Dynamic Representation of Highway Traffic Consistent With the Hydrodynamic Theory. Transportation Research Part B: Methodological, Vol. 28, No. 4, 1994, pp. 269–287. https://doi.org/10.1016/0191-2615(94)90002-7.

18.

Urbanik

Koonce

The Dilemma With Dilemma Zones. Proc., ITE District, Pittsburgh, PA, 2007.

19.

Day

C. M.

Bullock

D. M.

Investigation of Self-Organizing Traffic Signal Control With Graphical Signal Performance Measures. Transportation Research Record: Journal of the Transportation Research Board, 2017. 2620: 69–82.

20.

PTV Vissim. PTV Vissim 2020 Introduction to the COM API. Planung Transport Verkehr AG, Karlsruhe, Germany, 2020.

21.

PTV. PTV Vissim 10 User Manual. Planung Transport Verkehr AG, Karlsruhe, Germany, 2021, pp. 265–297.

22.

Signal Simulation Subcommittee, and Transportation Research Board Traffic Signal System Committee. Traffic Signal Simulation Cookbook, Version 1.0. Scribd, Inc., San Francisco, CA, 2020.

23.

Feng

Head

K. L.

Khoshmagham

Zamanipour

A Real-Time Adaptive Signal Control in a Connected Vehicle Environment. Transportation Research Part C: Emerging Technologies, Vol. 55, 2015, pp. 460–473. https://doi.org/10.1016/j.trc.2015.01.007.

24.

Shams

Day

C. M.

Impact of Sensing Range on Real-Time Adaptive Control of Signalized Intersections Using Vehicle Trajectory Information. Presented at 100th Annual Meeting of the Transportation Research Board, Washington, D.C., 2021.

25.

Tan

K. L.

Sarkar

Poddar

Sharma

Deep Reinforcement Learning for Adaptive Traffic Signal Control. arXiv Preprint arXiv:1911.06294, 2019, pp. 286–293.

26.

Cabrejas-Egea

Zhang

Walton

Reinforcement Learning for Traffic Signal Control: Comparison With Commercial Systems. Transportation Research Procedia, Vol. 58, 2021, pp. 638–645. https://doi.org/10.1016/j.trpro.2021.11.084.

27.

Microsoft. Named Pipes. Windows App Development, 2021. https://docs.microsoft.com/en-us/windows/win32/ipc/named-pipes.

28.

Grossman

Hainen

Remias

Bullock

Evaluation of Thermal Image Video Sensors for Stop Bar Detection at Signalized Intersections. Transportation Research Record: Journal of the Transportation Research Board, 2012. 2308: 184–198.

29.

Zhang

Xiao

Coifman

Mills

J. P.

Vehicle Tracking and Speed Estimation From Roadside Lidar. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, Vol. 13, 2020, pp. 5597–5608. https://doi.org/10.1109/JSTARS.2020.3024921.

30.

Zhao

Liu

Zheng

Detection and Tracking of Pedestrians and Vehicles Using Roadside LiDAR Sensors. Transportation Research Part C: Emerging Technologies, Vol. 100, 2019, pp. 68–87. https://doi.org/10.1016/j.trc.2019.01.007.

31.

Chavez-Garcia

R. O.

Aycard

Multiple Sensor Fusion and Classification for Moving Object Detection and Tracking. IEEE Transactions on Intelligent Transportation Systems, Vol. 17, No. 2, 2016, pp. 525–534. https://doi.org/10.1109/TITS.2015.2479925.

32.

Yao

Jiang

Zhao

Luo

Peng

A Dynamic Optimization Method for Adaptive Signal Control in a Connected Vehicle Environment. Journal of Intelligent Transportation Systems: Technology, Planning, and Operations, Vol. 24, No. 2, 2020, pp. 184–200. https://doi.org/10.1080/15472450.2019.1643723.