Cost Effective Real-Time Traffic Signal Control Using the ... - CiteSeerX

Cost Effective Real-Time Traffic Signal Control Using the TUC Strategy

© COMSTOCK

Werner Kraus Jr. and Felipe Augusto de Souza Department of Automation and Systems Engineering, Federal University of Santa Catarina, Florianópolis, SC, Brazil, E-mails: [email protected] and [email protected]

Rodrigo Castelan Carlson and Markos Papageorgiou Dynamic Systems and Simulation Laboratory, Technical University of Crete, Chania, Greece, E-mails: [email protected] and [email protected]

Luciano Dionisio Dantas Technische Universität Braunschweig, Germany, E-mail: [email protected]

Elias B. Kosmatopoulos Department of Electrical & Computer Engineering, Democritus University of Thrace, Xanthi, Greece, E-mail: [email protected]

Eduardo Camponogara Department of Automation and Systems Engineering, Federal University of Santa Catarina, Florianópolis, SC, Brazil, E-mail: [email protected]

Konstantinos Aboudolas Informatics and Telematics Institute, Centre for Research and Technology Hellas, 57001 Thessaloniki, Greece, E-mail: [email protected]

Digital Object Identifier 10.1109/MITS.2010.939916 Date of publication: 4 February 2011

IEEE INTELLIGENT TRANSPORTATION SYSTEMS MAGAZINE •

6

• WINTER 2010

1939-1390/10/$26.00©2010IEEE

I. Introduction

M

economical to implement and yet was shown in simulation any reasons have been mentioned for the less than and field implementations to perform at a level comparable expected deployment of adaptive urban traffic to other established strategies [12] such as SCOOT [11] and control systems [6], among which the high cost of BALANCE [17]. The cost savings brought by TUC are due to installation and maintenance is one of the most citthe following reasons. First, traffic data measurements are ed. This is particularly true for mid-sized cities (with less needed only once per cycle and communication latencies than 500 000 inhabitants) in developing countries which are tolerated. Second, only occupancy data is needed so that cannot afford the systems provided by the major vendors of messages sent to the central room are short, although flow urban traffic control (UTC) software. Often these cities do data is typically transmitted as well for performance assessnot have fiber optic communication cabling laid down as ment and archival. Third, the centralized calculations done part of the public infrastructure nor high quality pavement by TUC allow both savings in the number of sensors on roads. Moreover, qualified technical personplaced in the network and the interpolation nel dealing with traffic management and of missing sensor data due to breakoperations may be a scarce resource, ages and failures. Performance commonly in charge of many duAbstract– Rea l-t i me deteriorates with less sensors, ties beyond the proper setting urban traffic control systems frebut simulation studies and of time-of-day (TOD) traffic quently require precise traffic measurefield observations [12] insignal plans. The situaments and fast communications in order to achieve dicate that the benefits tion, then, is characterdesired performance levels. Such requirements may hinof real-time signal conized by a combination der the adoption of these beneficial control systems because trol are still retained of lack of resources of the installation and maintenance costs involved. The recently when loss of sensor with rapidly growing developed TUC strategy has been conceived in a way that simpliinformation is comcar numbers, leadfies measurement requirements and yet achieves performance levels pensated by the ining frequently to comparable to other well-established commercial systems. This was a terpolation of data poor performance major motivation to select TUC for a traffic control center installation from neighboring of the urban traffic in a mid-sized Brazilian city aiming at improving the traffic conditions sensors. According network and widedespite the lack of wired communication between roadside controlto a recent survey spread congestions lers and the central control room. A description of the implemented [14], other UTC systhat could otherwise system is presented, followed by field data comparing pre and post tems have to rely on be avoided or, at least, installation traffic behavior. It is found that TUC leads to a 15– historical data to subbetter managed by use 25% improvement in average network speed compared with stitute faulty detectors. of modern UTC systems. pre-existing time-of-day plans. Also, in the particular Addressing this situcase of SCOOT, [16] states ation requires low-cost Keywords– Adaptive urban traffic that performance is retained automation for traffic signal control, Feedback control, up to 15% of sensor failure, below operation. The two most relevant TUC strategy. the perceived TUC capabilities. items in the cost of UTC systems Given these features and capabilities of are the setting up of an appropriate comTUC, it has been chosen for an installation in the munication infrastructure and the installation of Brazilian city of Macaé, Rio de Janeiro state. The work was sensors to collect traffic measurements in real time. As a a close collaboration between Greek and Brazilian univerconsequence, in order to keep the related cost low, a UTC sities, local contractors and equipment vendors. Besides strategy must have low data exchange requirements in the municipality own budget resources, funding support terms of sampling rates and message size, be able to operfor the project also came from federal agencies that fosate with few sensors, and be robust to failures. Moreover, ter innovation through research & development partnerthe strategy must be simple enough to be deployed without ships between universities and industry. The result of the the need for lengthy fine-tuning, implying that few parameffort is an overall improvement of the urban traffic in the eters must be present, or else, a systematic approach for downtown area of the city in the order of 15% to 25% as tuning must be available. reflected in the corresponding increase in average traffic The TUC real-time signal control strategy [2], [3] has speed when compared to the recently adjusted TOD plans. been developed with special attention to simplicity; it is

IEEE INTELLIGENT TRANSPORTATION SYSTEMS MAGAZINE •

7

• WINTER 2010

■

offsets are calculated based solely on travel times between junctions without consideration In order to improve traffic flow in the CBD, the city decided to of possible residual queues at the start of green in a downrenovate the traffic controllers and to establish a partnership with stream junction. universities to implement an adaptive real-time traffic control center. Traffic management personnel in the city were generally satisfied with the performance of the adjusted plans, but expected to achieve better results with a real-time traffic-responIn order to highlight the most important aspects of the sive strategy. work, this paper continues with a presentation of the urban For real-time traffic control deployment in the CBD, scenario and technological limitations for the installation it is not possible to use fast communication links such as in Section II. Then, Section III reviews the main aspects of fiber optic cables. Wired telephone lines are expensive the TUC strategy and how they relate to the control probto hire while installation of private cabling is also dislem at hand. Section IV describes the system installation carded due to the costs involved. Hence, a possible soluwith emphasis on the measurement and communication tion is cellular telephone technology and, in fact, GPRS approaches employed. Performance of the system is asover GSM was elected due to availability and low cost of sessed based on loop detector data that reflect the benefits monthly fees. of the control decisions taken by TUC, as shown by field Before proceeding to the details of the installation, the results of Section V. Concluding remarks and future work next section briefly reviews the TUC strategy and how it are discussed in Section VI. was configured for the control problem at hand.

II. Description of the Control Scenario The city of Macaé, Rio de Janeiro state, is located 200 km northeast of the state capital. It is an important harbor providing logistics for oil operations in the deep water exploration off the Brazilian coast. With a population of 195 000 in 2010 and a growth rate of 3.9% a year for the past decade, the city has traffic flows larger than other similarly sized urban regions in Brazil because of its economic significance. Figure 1 depicts a map of the central business district (CBD) where most of the traffic signal control is concentrated. Blue circles indicate the controlled intersections. Roads are mostly two-lane one-directional streets with side parking space, resulting in two-phase control for most intersections. Such traffic arrangement facilitates signal control and allows for relatively high flows of up to 9 000 veh/day/lane in the CBD, with most of the traffic occurring between 6:30 am and 9:00 p.m. The short lengths between adjacent intersections are a negative factor, with the shortest ones being only 100 m apart. In order to improve traffic flow in the CBD, the city decided to renovate the traffic controllers and to establish a partnership with universities to implement an adaptive real-time traffic control center. As soon as the new traffic controllers were installed, an updating of the TOD plans was conducted by a consultant with good field experience about Macaé traffic. The two main features of the resulting plans are: ■ cycle times are restricted to a maximum of 90 s in order to avoid queue spillbacks in the relatively short links of the CBD;

III. The TUC Signal Control Strategy TUC (Traffic-responsive Urban Control) [2], [3] was developed to provide coordinated real-time traffic control for large urban networks even under saturated traffic conditions. Its main feature is the functionally centralized computation of split and cycle control. This is in contrast to many other real-time control methods such as SCOOT [10,16], RHODES [9], [13], OPAC [7], [8] and PRODYN [5], where control values are functionally decentralized due to the exponential complexity of the involved algorithms in relation to the number of intersections. It is important to note that all these systems can be concentrated in a single computer with enough computational power, but only TUC carries the calculations in a functionally centralized fashion. In TUC, the control variables for splits, cycle and offsets are calculated by three corresponding independent modules. The only required measurement is the number of vehicles in the link during the cycle, which is estimated from occupancy measurements during one cycle. Based on the estimates, the cycle and offset modules compute the corresponding values and send them to the split module which, in turn, reconciles all three control actions to obtain the network-wide signal plan to be applied in the next cycle. A brief review of the calculations involved in each module is presented next.

A. Split Control Split control in TUC has the objective of minimizing the risk of oversaturation and blocking of intersections due to

IEEE INTELLIGENT TRANSPORTATION SYSTEMS MAGAZINE •

8

• WINTER 2010

FIG 1 Map of the main control area in Macaé, RJ, Brazil. Blue circles indicate traffic-light controlled intersections (available at http://transitoonline. mactran.rj.gov.br.).

queue spillover. To this end, split control is devised based on a store-and-forward model [3] that describes the network traffic dynamics as a linear, time-invariant discretetime system of the form: x 1 k 1 1 2 5 Ax 1 k 2 1 BDg 1 k 2 ,

(1)

where x is the state vector holding the numbers of vehicles xz in the links z [ Z (the set of all links under TUC control); Dg is the control vector holding Dgj,i 5 gj,i 2 gNj,i, 4i [ Fj (the set of traffic light phases), 4j [ J (the set of controlled inter-

sections) where gj,i is the green time of phase i at junction N j and gj,i is a corresponding (pre-specified) nominal value; and A 5 I and B are the state and input matrices, the latter reflecting the specific network topology, fixed staging, cycle, saturation flows, and turning rates. Based on the linear model (1) and an appropriate quadratic objective, the LQR methodology [4] for controller synthesis is used to derive an efficient gain matrix to be used in a feedback control law of the form:

IEEE INTELLIGENT TRANSPORTATION SYSTEMS MAGAZINE •

g 1 k 2 5 gN 2 Lx 1 k 2 ,

9

• WINTER 2010

(2)

also violate minimum-green constraints; a suitably designed knapsack optimization algorithm [1] modifies the green times so as to satisfy these constraints but keep the relative proportions of the green times as close to the ones produced by (2) as possible. As shown in [1], [15], algorithms with linear complexity are available to find global optimal solutions to the class of problem at hand.

TUC’s cycle control acts so as to limit the maximum observed saturation level in the network.

where g is the vector of green times gj, i, gN is the vector of nominal green times gj,Ni, and L is the state feedback matrix that minimizes the following cost function: J5

`

1 T T a 1 xk Qxk 1 Dgk RDgk 2 , 2 k50

(3)

where Q and R are non-negative definite, diagonal weighting matrices. The elements of the diagonal matrix Q are chosen as the inverse of the storage capacity of the respective link, and the elements of R are chosen by the designer. Then, L depends on system parameters given by matrices A, B and Q and essentially one single design parameter, the control weights specified by R in the quadratic cost function (3). The design methodology of the gain matrix L takes into account the storage capacity of the links as weights for the state x in order to avoid the risk of spillback of queues. Thus, matrix L lends a gating feature to the control law, that is, restricting the flow into the overloaded network links. In doing so, it protects the downstream links with high number of vehicles from oversaturation by decreasing the green times of upstream links. This feature can be accentuated by weighting the estimates of the number of vehicles on the links. Simulation studies [1] have shown that L has low sensitivity with respect to traffic parameter variations. In order to apply the control (2), online measurements of the state variables are necessary. However, the number of vehicles xz cannot be directly measured, except if video cameras are available. For this reason, local occupancy measurements oz , collected in real time by inductive loops, are transformed in estimated number of vehicles xz through appropriate non-linear functions xz 5 fz 1 oz 1 k 22 . In other words, instead of attempting to accurately estimate traffic queues, TUC uses a static nonlinear function fz 1 # 2 that maps occupancy into the number of vehicles accumulated in one cycle. fz 1 # 2 is adjusted to reflect sensor placement on the link; the nearer to the stopline, the lower the number of vehicles in the link estimated for a given occupancy. For a compromise between early detection of growing queues and better ability to measure flow (instead of halted vehicles), midblock placement of sensors is indicated [1]. The green times for the phases of each junction resulting from (2) will generally not add up to a cycle and may

B. Cycle Control TUC employs a common cycle length for a given region of the network in order to enable coordination via suitable offsets. The reasoning behind cycle control, as in all cyclic methods of traffic control, is that a longer cycle time reduces the proportion of the lost time (which is a fixed quantity) with respect to the cycle time, thus increasing the intersection capacity. On the other hand, if traffic volume is low, a longer cycle will result in wasted green time. TUC’s cycle control acts so as to limit the maximum observed saturation level in the network. More specifically, TUC applies a proportional-type feedback algorithm that uses the current maximum saturation level of a prespecified percentage of the network links as a criterion for the cycle setting. The cycle control algorithm comprises three steps: 1) The list of network links is ordered by decreasing current loads sz 1 k 2 5 xz /xzmax (xzmax is the jam capacity of link z). Then, given a user-defined p, the top p-percent links are taken and the corresponding loads are averaged to provide the average maximum load s 1 k 2 ; 2) The network cycle is calculated by the feedback control law (proportional controller) C 1 k 2 5 C N 1 K C 1 s 1 k 2 2 sN 2 ,

(4)

where CN is the nominal network cycle length; sN is a nominal average load; and KC is a control parameter. After the application of Eq. (4), the calculated cycle length is constrained within the range of permissible cycle lengths 3 Cmin, Cmax 4 , if necessary; 3) If the resulting network cycle C 1 k 2 is sufficiently high while all links approaching specific intersections have sufficiently low saturation levels, then these undersaturated intersections may be double-cycled. The steps 1 and 2 seek to adjust the cycle time to cope with the maximum levels of saturation observed in the network, while step 3 attempts to reduce delays that would occur at certain intersections with low saturation levels due to high cycle times. The justification for the proportional control of Eq. (4) is the direct proportionality between cycle time and

IEEE INTELLIGENT TRANSPORTATION SYSTEMS MAGAZINE •

10

• WINTER 2010

intersection capacity. In turn, capacity and link loads are related by an inverse first-order process whereby an increase in capacity leads to a continuous decrease in load until a new equilibrium is reached. As such, a proportional controller is well suited to regulate link loads around a desired value.

(1– σz) ⋅ lz

σz⋅ lz

j1

j2 lz

C. Offset Control Offsets should be set taking into account the existence of queues. This fact is exploited by TUC’s offset controller which is based in the following assumptions: ■ Arterials are defined here as sequences of links that do not need to correspond to physical network arterials so that any route chosen by the traffic engineer as requiring good progression can be assigned for offset calculation; ■ In case of two-directional arterials, an offset is specified for each direction, and the offset that will be finally implemented is a weighted sum of the offsets of the two directions. Alternatively, the most loaded direction may be selected in real time to determine the arterial offsets; ■ In case of arterials that do intersect, TUC considers a pre-specified priority order of the arterials according to their relative importance regarding offset specification, and offset control is implemented to each arterial sequentially, starting from the arterial that has highest priority. TUC performs offset control in a decentralized way, i.e., for successive pairs of junctions along the pre-defined arterials. Moreover, calculations are performed independently for each direction of traffic and, as stated above, the resulting offsets can be averaged to produce the final offset between successive pairs of junctions [2]. For each pair of junctions, the offset specification changes the starting time of a specific “main” stage of the upstream junction, where this main stage is uniquely determined from the arterial composition. TUC considers the possible existence of queues by means of a simple feedback control law, as follows. Consider two successive junctions j1 and j2 and link z that connects them in the j1 to j2 direction. Link z has length lz and average speed vz and receives right-of-way in the main stage of junction j2 (Fig. 2). The queue length on link z is approximately sz 1 k 2 # lz. If there are no vehicles in the link, the offset between the two intersections should be equal to the travel time under average speed for the link, that is, lz / vz. In other words, the cycle in j2 should start after the cycle in j1 (positive offset). As the number of vehicles in link z grows, the offset should decrease correspondingly in order to allow the partial discharge of the queue in j2. Then the cycle in the downstream intersection should begin earlier than in the no-queue case and, in some cases, even before the cycle in the upstream intersection (negative offset).

FIG 2 Link z with queue (gray) (adapted from [2]).

More specifically, an ideal offset would be obtained if the following two traffic waves meet exactly at the tail of the existing queue: 1) Traffic wave created due to the change to green in the upstream intersection j1; this wave moves downstream with speed vz; hence, it reaches the queue tail at time 3 1 2 sz 1 k 2 4 # lz / vz after the signal change to green; 2) Kinematic wave created by the change to green in the downstream intersection j2; this wave moves upstream (along the queue) with speed vc which is usually estimated around 15 km/h; this kinematic wave reaches the queue tail at time sz 1 k 2 # lz / vc. The ideal offset tj1,j2 can be calculated by considering both times above, which yields the following offset control law: tj1, j2 1 k 2 5

xz 1 k 2 lz 2 l Ko , vz z z xmax z

(5)

where Koz is a control parameter equal to 1 vc 1 vz 2 / 1 vc vz 2 .

D. Control Implementation The actual control implementation in this case study has been done as follows. Split control is activated at every cycle. The centralized computation is performed several seconds before the end of a reference cycle for the whole network and implemented in the following cycle. Cycle control is evaluated every n seconds, with n being a design parameter typically chosen as a multiple of the nominal cycle. Since Eq. (4) may cause large variations in two consecutive cycle times that can impact traffic due to transient effects, a modified form has been adopted in the current installation. In order to attenuate variations in cycle times, values computed by the proportional control law (4) are filtered by a nonlinear filter given by: Cf 1 k 2 5 Cf 1 k 2 1 2 1 k* 3 C 1 k 2 2 Cf 1 k 2 1 2 4 , k* 5 e

ku if c 1 k 2 $ cf 1 k21 2 , kd if c 1 k 2 , cf 1 k21 2

(6)

where Cf 1 # 2 is the filtered cycle time actually used, C 1 k 2 is the cycle given by Eq. (4), and ku and kd are the “up” and “down” control gains. The two gains make it possible to allow the cycle to change more rapidly when traffic flows are increasing than when flows are decreasing. This is

IEEE INTELLIGENT TRANSPORTATION SYSTEMS MAGAZINE •

11

• WINTER 2010

13

26

10

8 3 1

4

23

12

11

9

24

6

7

18

5

21

20

25

16

17

19

22

2

14 15

(a)

(b)

FIG 3 Schematic diagram of the two control regions comprised only of one-way streets. (a) Diagram of control region 1. (b) Diagram of control region 2.

beneficial since (i) surges in traffic are more disruptive than sudden flow drops, and (ii) shorter-than-optimal cycle times are more detrimental than longer-than-optimal ones. In the present implementation, the values chosen are ku= 0.6 and kd 5 0.4 and the cycle time is updated every two nominal traffic signal cycles. Offset control is applied every kn seconds, with k 5 1, 2, c being a design parameter. Large offset changes may occur, which create transient traffic disturbances that may affect performance. Hence, it is recommended to allow at least 10 min between consecutive offset changes.

IV. The Implemented System The CBD was divided in two regions for control purposes. Region 1, depicted in Fig. 3a, comprises the rectangular network seen next to the center of Fig. 1 plus junctions to the East and South. Region 2 corresponds to the “V” shaped network in the upper part of Fig. 1, and is shown in Fig. 3b. Each region has its own cycle time, thus coordination between intersections at the border of adjacent regions cannot be guaranteed. The main criterion for determining the regions were the distinct traffic flows in each, meaning that different cycle times are necessary. Region 1 is the busiest of the two, with traffic volumes almost twice as large as region 2. The colored paths (bold arrows) in the pictures represent the routes for which offset control is active, representing the major flows of the respective regions.

A. TUC Control Parameters The entries bij of B in (1) are determined as follows. In the control regions, all intersections have two approaches. In this case, the dynamics of link z is given by [3]: xz 1 k 1 1 2 5 xz 1 k 2 1 ta,z Sa Dga 1 k 2 1 tb,z Sb D gb 1 k 2 2 Sz Dgz 1 k 2 ,

(7)

where ta,z and tb, z are the turning rates into z from approaches a and b, Sa and Sb are the saturation flows of the approaches, Dga and Dgb are the respective incremental green times, Sz is the saturation flow that leaves link z, and Dgz is its incremental green time. Collecting the appropriate terms, (7) can be written in vector form as: Dga 1 k 2 xz 1 k 1 1 2 5 xz 1 k 2 1 3 baz bbz bzz 4 £ Dgb 1 k 2 § , Dgc 1 k 2

(8)

where baz 5 ta,z Sa, bbz 5 tb, z Sb and bz z 5 2 Sz are the elements of the line of the B matrix associated with the dynamics of link z. For the whole control region, the entries of B are arranged such that the correct green times multiply the respective coefficients for each link. Traffic parameters necessary for the entries of the input matrix B, for nominal control times, and for the LQR calculation of matrix L were conured as follows: ■ Saturation flows: a field survey was conducted to estimate saturation flows per traffic lane according to the procedures described in [18]. The values found are in the range of [1200,1650] veh/h, lower than the expected values of above 1800 veh/h. The reason for the lower values is the existence of side street parking and the relatively high interference of pedestrians and bicycles in the streets; ■ Turning rates: although field surveys were available, the data were not adopted because of the time of the year when the surveys were conducted, mostly during school holidays. Traffic in the city changes considerably when there are no school classes. Rough empirical estimations were conducted in order to assign turning rates in one out of four possibilities: 80/20; 70/30; 60/40; 50/50; N ■ Nominal green times g : the existing green splits of the TOD plans were adopted;

IEEE INTELLIGENT TRANSPORTATION SYSTEMS MAGAZINE •

12

• WINTER 2010

Loop Sensor

Controller GPRS Modem

Firewall Internet

Megabit Switch

Megabit Switch

Operator Workstation

Controller Loop Sensor

GPRS Modem

Application Server

GPRS/Web Server

Database Server

FIG 4 Simplified physical architecture of the control system.

■

Link storage capacities: used in the LQR calculation as weights for the quadratic cost of the states xz, the storage capacities where based on average vehicle lengths of 6 m for the control regions.

B. Communication and Computational Infrastructure The overall physical architecture of the control system is presented in Fig. 4. Based on the requirements and restrictions posed by the scenario at hand, the following technical solutions were adopted: ■ Communications: GSM/GPRS modems were installed in the traffic signal controllers. The bit rate between modem and controller is 9600 bps due to controller characteristics. On the control room side, an Internet connection by radio with 1 Mbps download rate and 250 kbps upload is the medium for the GSM/GPRS link. In terms of data traffic, although only occupancy measurements are required for control purposes, vehicle counts are also transmitted for evaluation and archival reasons, resulting in messages of 20 bytes for measurements (download) and 18 + [number of stages] bytes for actuation (upload); ■ Traffic sensors: inductive loop detectors are positioned in the middle of the respective blocks, one per lane. The average distance between detectors and the controller cabinet is 60 m, with cabling laid down under the sidewalks; ■ Traffic controllers: existing controllers in Macaé were not modular. In close cooperation with the controller vendor, the CPU card was redesigned, both in hardware

and software, for real-time operation. The time base of the controller is obtained periodically from a GPS module installed in each cabinet, thus assuring networkwide synchronization of controller clocks with the control room; ■ Control room: servers for database, control algorithm and operator workstations with multi-display graphiccards are configured as an intranet with restricted access from the outside world. For the necessary internet services, one computer without protection of the firewall security rules hosts the GPRS/GSM and web servers. The web server provides the public with internet access to the information about the current traffic state (see Fig. 1). For the control loop, the basic data traffic occurs once per traffic cycle. It consists of incoming traffic measurements and outgoing control commands that communicate over the GPRS/GSM service. GPRS/GSM modems send messages to the IP address of the server located in the control room. The server, on its turn, must keep a list of the dynamic IP addresses assigned by the wireless telephone provider, which may change from time to time. One-way latency in the available GPRS network in Macaé can be as high as 5 s. Considering the round trip of the request/response messaging, total time for the communication can reach 10 s. Clearly, the control algorithm must be robust to message delay and inaccuracy in sampling time. Ideally, traffic data would be requested at the last second of the cycle, but in reality requests are originated 15 to 20 s before the end of the cycle. Note that even though it is possible to initiate the sampling process at a precise instant, latencies

IEEE INTELLIGENT TRANSPORTATION SYSTEMS MAGAZINE •

13

• WINTER 2010

TUC achieves a better balance of speeds in the intersections in such a way that the overall intersection speeds are increased by 10% for intersection 6 and by 36% for intersection 7.

make it impossible to guarantee that the obtained samples have the same precise interval. Still, as shown in the field results presented below, TUC is robust to such errors in the sampling period. An alternative to GPRS/GSM that can be used in other scenarios is the 3G (UMTS) wireless telephone service, which presents increased bandwidth and low latencies. However, the lack of availability of 3G networks in developing countries like Brazil should not mean that cities in these countries must

wait to have workable and effective real-time adaptive systems in place to better manage their traffic networks.

V. Field Results

Presentation of the field results is based on the three most important junctions of control region 1, each with two approaches. Being the busiest of the two regions, region 1 benefited most from TUC operation. Comparison is made between day long operation of TOD control and TUC. For the sake of performance evaluation, traffic counts from loop detectors are taken as well as occupancy. Since daily traffic flows are not the same, occupancy is not used directly for evaluation purposes. Rather, an equivalent spot speed Veq is computed in order to combine flow and occupancy in one indicator according to Veq 5

6Q ot1000

km/h,

(9)

where Q is the hourly flow, ot is the time occupancy of the detector, and vehicle lengths are assumed constant and 110 equal to 6 m. For the purposes of the comparative analysis, it is sufficient to assume that traffic composition is similar 100 on all days so that the vehicle length of 6 m represents a 90 normalized quantity rather than the actual effective ve80 hicle length. The daily averages presented are harmonic means of the hourly speeds. 70 Due to roadwork on pavements right after the comple60 tion of the installation, there was a loss of around 30% in 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 the number of sensors in region 1 (17 out of 60). Since the Time-of-day (h) broken sensors were spread over a large area, it was possible to use interpolation of information from neighboring FIG 5 Day trend of cycle time comparing time-of-day and TUC (note the maximum TOD cycle set to 90 s). sensors to estimate the missing occupancy data. Table I summarizes the results for the intersections considered. Numbering follows the convention in Fig. 3a. As can be Table I. Comparison of time-of-day and TUC for three selected intersections seen, intersection 4, the busiest of the city, of region 1. Data is from 6:00 to 21:00; N is the vehicle count and Veq is shows improvements in speed even in prescomputed by (9). ence of greater daily flows, with 21% increase in speed for approach 1 and 15% for Approach 1 Approach 2 Total approach 2. Results for intersections 6 and 7 Intersection N Veq (km/h) N Veq (km/h) N Veq (km/h) show a different trend. Both have approaches 1 worse off in speed, but that is contrast4 TOD 6458 5.1 7617 9.8 14075 7.8 ed with more significant improvements TUC 7628 6.5 8676 11.3 16304 9.1 % 18.1 20.7 13.9 15.1 15.8 16.3 in speeds for the respective approaches 2 (these are two segments along the same 6 TOD 6311 14.7 6151 8.7 12462 11.7 road that runs East-West in Fig. 3a). It can TUC 6509 11.1 6794 14.7 13303 12.9 be seen that TUC achieves a better balance % 3.1 -24.5 10.5 69.0 6.7 10.2 of speeds in the intersections in such a way that the overall intersection speeds are in7 TOD 5480 14.2 6160 13.1 11640 13.6 TUC 5256 11.7 7083 23.7 12339 18.6 creased by 10% for intersection 6 and by % -4.1 -17.8 15.0 81.1 6.0 36.5 36% for intersection 7. 120

Cycle (s)

TUC TOD

IEEE INTELLIGENT TRANSPORTATION SYSTEMS MAGAZINE •

14

• WINTER 2010

60

500

50

400 300

Total Flow: TOD: 6,311 TUC: 6,505

200 TOD TUC

100

Mean Speed (km/h)

Flow (vch/h)

600

40 30 20 10

6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Time-of-day (h) (b)

600

60

500

50

400

100

Total Flow: TOD: 6,151 TUC: 6,794 TOD TUC

0

Mean Speed (km/h)

Flow (vch/h)

6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Time-of-day (h) (a)

200

TOD TUC

0

0

300

Day Average: TOD: 14.7 km/h TUC: 11.1 km/h

Day Average: TOD: 8.7 km/h TUC: 14.7 km/h

40

TOD TUC

30 20 10 0

6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Time-of-day (h) (c)

6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Time-of-day (h) (d)

FIG 6 Comparison of time-of-day (TOD) and TUC, intersection 6. (a) Traffic flow, approach 1. (b) Equivalent speed, approach 1. (c) Traffic flow, approach 2. (d) Equivalent speed, approach 2.

One reason for such marked improvement is the inadequacy of TOD offsets under certain traffic regimes. Although not shown in this paper, field data indicate that TOD offsets assume that no queues are present at downstream links even in peak hours, while TUC takes queues into account so that large negative offsets are in place when needed. Another improvement comes from cycle control. TOD plans are restricted to 90 s maximum cycle time due to concerns about spillbacks on the rather short link lengths between intersections. TUC, by virtue of its restriction of flow release when downstream occupancies are high, is allowed to apply cycles up to 120 s. Figure 5 shows the cycle time excursions during one day. Clearly the trend of TOD is followed by TUC in early and later hours, but with different behavior during busier daytime traffic. Finally, field observations also led to the conclusion that split control works well in distributing the available green times so as to avoid green starvation or unduly short green times. More detailed traffic behavior can be seen in Fig. 6, which depicts hourly data for approaches 1 and 2 of intersection 6. As can be seen, traffic flow is higher on both approaches under TUC control. In terms of speeds, how-

ever, it is seen that TUC degrades approach 1 in order to better distribute flows, mainly during the busier hours of the day. Such balancing effects has been noted in many other intersections of the network, along with an overall improvement of traffic.

Conclusion The paper reports on practical results from a field deployment of the TUC strategy in the city of Macaé, Rio de Janeiro state, Brazil. Aspects of the installation were described, highlighting the economic data communication structure and the expedited nature of the selection of input parameters used for TUC configuration. Analysis of performance for three major intersections show a marked improvement brought by TUC over TOD plans. Results are particularly encouraging given the conditions of the installation. GPRS/GSM communication in the city can reach round-trip latencies of up to 10 s, thus introducing uncertainties in the measurements. Moreover, some loops have been already lost due to inadvertent works in the pavement. Such shortcomings highlight TUC’s robustness that arises from region-wide consideration of

IEEE INTELLIGENT TRANSPORTATION SYSTEMS MAGAZINE •

15

• WINTER 2010

measurements for local control decision and sampling period of one cycle. In conclusion, the preliminary data shown in the paper indicate that it is possible to improve traffic management without necessarily using high-capacity data links and accurate traffic modeling for control design. TUC has proven to be relatively cheap to install when compared to established commercial systems which, in other studies, have performed similarly well or slightly worse [12]. Overall, the TUC strategy is well appropriate for cities that would like to deploy highperformance real-time traffic control systems without the more stringent installation requirements of more traditional systems.

Acknowledgment The authors acknowledge the contribution of technical personnel at Macaé Trânsito e Transportes (MACTRAN), Mrs. Lais Meirelles and Mr. Fabiano Lima and consultant Mr. Pedro Bortolotto. Werner Kraus Junior thanks CNPq and FINEP/SEBRAE of Brazil for partially supporting this work through grants 310374/07-3 and 024/07 respectively.

About the Authors Werner Kraus Jr. was born in Blumenau, Brazil, in 1964. He holds B.El. Eng. (1986) and M.El.Eng. (1991) degrees from the Federal University of Santa Catarina, Brazil, and a Ph.D. (1997) from the Autralian National University. Since 2000 he has been with the Department of Automation and Systems Engineering at the Federal University of Santa Catarina. His main interests are control of urban mobility systems and cooperative systems for traffic management and control, with emphasys in the implementation of prototype systems in real scenarios. Felipe Augusto de Souza graduated in Control and Automation Engineering from the Federal University of Santa Catarina in 2007. He is currently a masters student at the same university, with main interests in the area of urban traffic control and software engineering applied to urban traffic management and control systems. He is a recipient with co-authors of the best paper award in 6th IEEE Conference on Automation Science and Engineering (CASE 2010).

Federal University of Santa Catarina, Brazil; and the Bachelor degree in Business Administration (2006) from the University of the State of Santa Catarina, Brazil. Since 2007 he is a PhD candidate and a research assistant at the Technical University of Crete, Greece. His main research interests include automatic control and optimization theory and applications to traffic and transportation systems. Markos Papageorgiou received the Diplom-Ingenieur and Doktor-Ingenieur (honors) degrees in Electrical Engineering from the Technical University of Munich, Germany, in 1976 and 1981, respectively. He was a Free Associate with Dorsch Consult, Munich (1982–1988), and with Institute National de Recherche sur les Transports et leur Scurit (INRETS), Arcueil, France (1986–1988). From 1988 to 1994 he was a Professor of Automation at the Technical University of Munich. Since 1994 he has been a Professor at the Technical University of Crete, Chania, Greece. He was a Visiting Professor at the Politecnico di Milano, Italy (1982), at the Ecole Nationale des Ponts et Chausses, Paris (1985–1987), and at MIT, Cambridge (1997, 2000); and a Visiting Scholar at the University of California, Berkeley (1993, 1997, 2001) and other universities. Dr. Papageorgiou is author or editor of 4 books and of over 350 technical papers. His research interests include automatic control and optimization theory and applications to traffic and transportation systems, water systems and further areas. He is the Editor-in-Chief of Transportation Research—Part C. He also served as an Associate Editor of IEEE Control Systems Society—Conference Editorial Board, of IEEE Transactions on Intelligent Transportation Systems and other journals. He is a Fellow of IEEE. He received a DAAD scholarship (1971–1976), the 1983 Eugen-Hartmann award from the Union of German Engineers (VDI), and a Fulbright Lecturing/Research Award (1997). He was a recipient of the IEEE Intelligent Transportation Systems Society Outstanding Research Award (2007) and of the IEEE Control Systems Society Transition to Practice Award (2010). He was presented the title of Visiting Professor of the University of Belgrade, Serbia (2010). Luciano Dionisio Dantas received the Diploma of Control and Automation Engineer and M.Sc. in Electrical Engineering degree from the Universidade Federal de Santa Catarina, Brazil in 2003 and 2005 respectively. He was enrolled as project engineer in the pilot

Rodrigo Castelan Carlson received the Bachelor degree in Control and Automation Engineering (2004) and the Master degree in Electrical Engineering (2006) from the

IEEE INTELLIGENT TRANSPORTATION SYSTEMS MAGAZINE •

16

• WINTER 2010

implementation of TUC in Maca, Brazil from 2005 to 2008. He is currently, since 2009, seeking the PhD degree at the Institut fr Verkehr und Stadtbauwesen—Technische Universitt Braunschweig, Germany in the area of Traffic Control Systems. Elias B. Kosmatopoulos received the Diploma, M.Sc. and Ph.D. degrees from the Technical University of Crete, Greece, in 1990, 1992, and 1995, respectively. He is currently an Associate Professor with the Department of Electrical and Computer Engineering, Democritus University of Thrace, Xanthi 67100, Greece and a Senior Researcher with the Informatics & Telematics Institute, Center for Research and Technology–Hellas (ITI-CERTH), Greece. He was a faculty member of the Department of Production Engineering and Management, Technical University of Crete (TUC), Greece, a Research Assistant/Associate Professor with the Department of Electrical Engineering-Systems, University of Southern California, CA, USA, and a Postdoctoral Fellow with the Department of Electrical & Computer Engineering, University of Victoria, B.C., Canada. Dr. Kosmatopoulos’ research interests are in the areas of adaptive optimization and control, energy efficient buildings, robotics swarms and intelligent transportation systems. He is the author of over 40 journal papers. Currently he is leading the intelligent control developments in 4 European Commission-funded projects involving swarms of flying robots, swarms of underwater robots, positive-energy buildings and traffic control systems. Eduardo Camponogara received the Ph.D. degree in Electrical and Computer Engineering from Carnegie Mellon University in 2000. After being a postdoctoral fellow at the Institute for Complex Engineered Systems, he joined the faculty of the Department of Automation and Systems Engineering at the Federal University of Santa Catarina, Brazil. His research interests include systems optimization, distributed decision making, and traffic control engineering. Konstantinos Aboudolas received the Diploma, M.Sc. and Ph.D. degrees from the Technical University of Crete, Greece, in 1999, 2003, and 2009, respectively. He is currently a Postdoctoral Fellow in the Informatics and Telematics Institute at the Centre

for Research and Technology Hellas, Greece. From 1999 to 2009, he was a Research and Teaching Assistant at the Technical University of Crete. His research interests lie in the application of control and optimization techniques to transportation systems, multi-robot systems, and further areas. Dr. Aboudolas is the co-author of some 20 journal and conference papers. He is a member of the IEEE and the Technical Chamber of Greece.

References [1] C. Diakaki, “Integrated control of traffic flow in corridor networks,” Ph.D. thesis, Dept. Prod. Eng. Manage., Tech. Univ. Crete, Chania, Greece, 1999. [2] C. M. Diakaki, V. Dinopoulou, K. Aboudolas, M. Papageorgiou, E. BenShabat, E. Seider, and A. Leibov, “Extensions and new applications of the traffic signal control strategy TUC,” Transport. Res. Rec., vol. 1856, pp. 202–216, 2003. [3] C. M. Diakaki, M. Papageorgiou, and K. Aboudolas, “A multivariable regulator approach to traffic-responsive network-wide signal control,” Contr. Eng. Prac., vol. 10, pp. 183–195, 2002. [4] P. Dorato, C. Abdallah, and V. Cerone, Linear-Quadratic Control: An Introduction. New Jersey: Prentice-Hall, 1995. [5] J. L. Farges, L. Khoudour, and J. B. Lesort, “PRODYN: on site evaluation,” in Proc. 3rd IEE Conf. Road Traffic Control. London, U.K.: IEE, 1990, pp. 62–66. [6] K. Fehon, “Adaptive traffic signals: are we missing the boat?” in Proc. ITE District 6 2004 Annu. Meeting, 2004, pp. 1–12. [7] N. H. Gartner, “OPAC: a demand-responsive strategy for traffic signal control,” Transport. Res. Rec., vol. 906, pp. 75–81, 1983. [8] N. H. Gartner, F. J. Pooran, and C. M. Andrews, “Implementation of the OPAC adaptive control strategy in a traffic signal network,” in Proc. 2001 IEEE Intelligent Transportation Systems, 2001, pp. 195–200. [9] K. L. Head, P. B. Mirchandi, and S. Shelby, “The RHODES prototype: a description and some results,” in Proc. 77th Transportation Research Board Annual Meeting, Tucson, 1998, pp. 1–12. [10] P. B. Hunt, D. I. Robertson, R. D. Bretherton, and M. C. Royle, “The SCOOT on-line traffic signal optimisation technique,” Traffic Eng. Contr., vol. 23, pp. 190–199, 1982. [11] P. B. Hunt, D. I. Robertson, R. D. Bretherton, and R. I. Winton, “SCOOT – a traffic responsive method of coordinating signals,” Transport Road Res. Lab., Tech. Rep. 1014, 1981. [12] E. Kosmatopoulos, M. Papageorgiou, C. Bielefeldt, V. Dinopoulou, R. Morris, J. Mueck, A. Richards, and F. Weichenmeier, “International comparative field evaluation of a traffic-responsive signal control strategy in three cities,” Transport. Res. A, vol. 40, no. 5, pp. 399–413, 2006. [13] P. Mirchandani and F.-Y. Wang, “RHODES to intelligent transportation systems,” IEEE Intell. Syst., vol. 20, no. 1, pp. 10–15, 2005. [14] NCHRP, “Adaptive traffic control systems: domestic and foreign state of practice,” in National Cooperative Highway Research Program (NCHRP) Synthesis 403. Washington, DC: Transportation Research Board, 2010. [15] P. M. Pardalos and N. Kovoor, “An algorithm for a singly constrained class of quadratic programs subject to upper and lower bounds,” Math. Program., vol. 46, no. 1, pp. 321–328, 1990. [16] D. I. Robertson and R. D. Bretherton, “Optimizing networks of traffic signals in real time – the SCOOT method,” IEEE Trans. Veh. Technol., vol. 40, no. 1 part 2, pp. 11–15, 1991. [17] GEVAS Software (2009). BALANCE: Traffic-Adaptive Network Control. GEVAS Software Gmbh [Online]. Available: www.gevas.eu [18] TRB, Highway Capacity Manual, 3rd Ed. Washington, DC: Transportation Research Board, 1994.

IEEE INTELLIGENT TRANSPORTATION SYSTEMS MAGAZINE •

17

• WINTER 2010