1794

Simulated safety performance of rear-end and angled vehicle interactions at isolated intersections
Flavio Cunto and Frank F. Saccomanno

Abstract: This paper applies a calibrated microscopic simulation model to assess the safety implications of signalization at a stop-controlled isolated intersection. Safety performance is measured in terms of a crash potential index (CPI) that makes use of time-specific vehicle parameters, such as deceleration rates, spacing, and speed profiles. Four performance measures are obtained: (i) average CPI/vehicle, (ii) CPI 85th percentile, (iii) number of vehicles with CPI > 0 (defined as interacting), and (iv) number of conflicts (defined in terms of a given CPI threshold). Two types of interactions are considered, namely rear end and angled. For rear-end interactions, CPI/vehicle was found to be significantly higher following the introduction of fixed signal controls. For angled interactions, CPI/vehicle was found to decrease with signalization. For both types of interactions, the CPI 85th percentile was found to decrease nonlinearly with signalization, especially for higher assumed volumes on the major approach. Rear-end vehicle interactions increased significantly following signalization and with increasing volume, whereas no such increase was observed for angled interactions. The key observation is that the number of vehicles subject to angled interactions was found to decrease after signalization. Key words: safety performance, microscopic simulation, signalized and stop-controlled intersections. ` Resume : Cet article applique un modele de simulation microscopique etalonne pour evaluer les implications de securite ` ` dune signalisation a un carrefour isole controlee par un arret. Le rendement en matiere de securite est mesure en termes ` dindice potentiel de collision ( CPI ) qui utilise des parametres temporels des vehicules tels que les taux de decelera tion, lespacement et les profils de vitesse. Quatre mesures du rendement ont ete obtenues : (i) CPI moyen par vehicule, (ii) 85e percentile du CPI, (iii) le nombre de vehicules ayant un CPI > 0 (definis comme interdependants), et (iv) le nom ` ` bre de conflits (defini en termes dun CPI seuil donne). Deux genres dinteractions sont consideres : par larriere et a an ` ` gle. Le CPI/veh. pour les interactions par larriere sest grandement accru apres lintroduction de controles de signaux ` fixes. Le CPI/veh. pour les interactions a angle diminuait avec la signalisation. Pour les deux types dinteractions, le 85e ` percentile du CPI diminuait de maniere non lineaire avec la signalisation, surtout pour les volumes de trafic presumes plus ` ` eleves dans lapproche majeure. Les interactions entre vehicules par larriere ont augmente considerablement apres la mise en place de la signalisation et avec un volume de trafic superieur; cette augmentation na pas ete observee pour les inter` ` ` ` actions a angle. Lobservation principale est que le nombre de vehicules sujets a des interactions a angle a diminue apres la mise en place de la signalisation. Mots-cles : rendement en securite, simulation microscopique, carrefours signales et controles par un arret. [Traduit par la Redaction]

Introduction
A number of recent studies of crashes for North American urban roads suggest that over 50% of these reported crashes take place in proximity to intersections. Intersections present
Received 18 June 2008. Revision accepted 2 June 2009. Published on the NRC Research Press Web site at cjce.nrc.ca on 17 November 2009. F. Cunto.1 Department of Transportation Engineering, Universidade Federal do Ceara, Campus do Pici S/N, Bloco 703, Departamento de Engenharia de Transportes CT, CEP.: 60.455-760, Fortaleza, CE, Brazil. F.F. Saccomanno. Department of Civil and Environmental Engineering, University of Waterloo, 200, University Street West, Waterloo, ON N2L 3G1, Canada. Written discussion of this article is welcomed and will be received by the Editor until 31 March 2010.
1Corresponding

author (e-mail: flaviocunto@det.ufc.br).

special safety concerns because of unsafe driver actions and manoeuvres that result in conflicts that could lead to crashes. According to the U.S. Federal Highway Administration (FHWA 2004), approximately 3.2 million intersection-related crashes took place in 2002 in the United States, with estimated costs to society in excess of US$100 billion. The high cost of crashes at intersection locations provides strong justification for the development and implementation of cost-effective and practicable countermeasures. The microscopic model described in this paper serves as an objective scientific platform for guiding decisions on what type of countermeasure to consider, examining the effectiveness of such countermeasures under different traffic conditions, and suggesting when such countermeasure should be introduced at a given intersection. A number of traffic control strategies can be considered for stop-controlled intersections, including signalization, improvements in signal timings, signage, prioritization of movements, and changes in posted approach
Published by NRC Research Press

Can. J. Civ. Eng. 36: 17941803 (2009)

doi:10.1139/L09-092

Cunto and Saccomanno

1795

speeds. The focus of interest in this paper is on the introduction of fixed signal controls. A number of researchers have investigated the relationship between traffic control and crashes at intersections (Pernia et al. 2002; Abdel-Aty et al. 2005; Hadayeghi et al. 2006). Due to a myriad of methodological and empirical problems, these studies have not provided the objective platform needed for informing decisions as follows: (i) When should such controls be introduced? (ii) What form should these controls take? and (iii) What are the costs and potential safety benefits of these controls? A major FHWA report authored by Gettman and Head (2003) investigated the feasibility of adopting a microscopic simulation approach. One of the most attractive features of microscopic simulation models is their potential for investigating different safety intervention strategies in a virtual world without incurring high implementation costs or causing traffic disruptions. In general, the use of microscopic simulation for safety assessments is based on measures of safety performance, also known as proximal safety indicators or surrogate safety measures. These indicators are defined to reflect high-risk events in relation to a projected point of collision based on pairwise vehicular velocity and spacing attributes. The use of safety performance measures also constitutes in essence a proactive approach to road safety studies because it can detect safety problems whether or not they result in a subsequent crash (Barcelo et al. 2003; Archer 2005). Several measures of individual vehicle safety performance can be extracted from simulation, and these include time-tocollision (TTC), extended time-to-collision (TET), postencroachment time (PET), and initial deceleration rate (DR) (Hayward 1971; Minderhoud and Bovy 2001; Gettman and Head 2003; Huguenin et al. 2005). In this paper, a microscopic simulation model has been developed to investigate the safety implications of introducing fixed signal control at a stop-controlled intersection for different traffic conditions. The essence of this model is the use of vehicle-specific measures of safety performance obtained from microscopic traffic algorithms (car following, gap acceptance, and lane changing) to evaluate alternative countermeasures.

Microscopic simulation
The initial step in the simulation of safety performance is to choose a simulation platform capable of representing in a comprehensive manner most aspects of driver behaviour, including late reaction, erroneous judgment of the actual speed differential, spacing, and deceleration rate. Brackstone and McDonald (1999) and Mehmood (2003) presented a comprehensive review of existing car-following models. They identified four types of models: stimulusresponse, safetydistance, psychophysical or action point, and fuzzy logic based models. Psychophysical or action point car-following models are based on the assumption that a driver of the following vehicle would be able to ascertain if his (her) vehicle was approaching the lead vehicle by changes in the apparent size of this vehicle and ascertain his (her) relative speed by changes in the visual angle of the lead vehicle. The basic structure of these models was

found to be the most coherent and was best able to describe the natural everyday driver behaviour (Brackstone and McDonald 1999; Xin et al. 2008). It should be noted that, although these models were not developed explicitly to predict crashes, they can provide sound and accurate estimates of safety performance. The simulation model used in this paper is VISSIM version 4.3, distributed by PTV America. VISSIM is based on psychophysical driving algorithms developed by Wiedemann and Reiter (1992), whose car-following model considers four types of regimes where drivers adjust their desired spacing and speeds through changes in their acceleration and deceleration rates. These four driver regimes are uninfluenced driving, closing process, following process, and emergency braking (Fig. 1). The transition between these regimes has been established in VISSIM for six different distancespeed thresholds: (i) AX desired distance for standing vehicles (front-to-front distance); (ii) ABX desired minimum following distance at low speed differences; (iii) SDV perception threshold of speed difference at long distance; (iv) SDX perception threshold of growing distance in following process; (v) CLDV perception threshold of small speed differences at short, decreasing distances; and (vi) OPDV perception threshold for recognizing small speed differences at short, but increasing distances. In the uninfluenced driving regime, the following driver is trying to reach the desired speed once there is no lead vehicle in a reasonable distance (150 m) or when the distance is decreasing but the perception threshold of speed difference at long distance (SDV) has not been achieved. Speed differential (DV) and spacing (DX) for this regime should satisfy the following conditions: DV < SDV or DX > 150 m at each time interval. When both the distance between the lead and the following vehicle is less than150 m and the SDV threshold has been surpassed, the following driver enters the closing-process regime. In this regime, drivers realize that they are approaching a slower vehicle and, after a given delay, begin to decelerate. At this point, drivers intend to decelerate, matching their own desired minimum following distance (ABX). The applied deceleration rate is based on the kinematics equation for deceleration considering a moving target. An additional error term is added to represent the error of human estimation that varies for the same driver every second and a parameter to account for the learning process, also updated each second. In the following process, the following vehicle has almost the same speed as that of the lead vehicle, and the following driver does not consciously react to movements of the lead vehicle. Acceleration and deceleration rates are applied in a very low oscillating level around the average value of 0.2 m/s2. The transition from the closing process to the following process happens when DV < SDX. The following process is delimited by two perception thresholds for small speed differences at short, decreasing and increasing distances (CLDV and OPDV) and two thresholds corresponding to the minimum desired distance at low speed differences and the perception of growing distance in the following process (ABX and SDX). The emergency braking regime happens when drivers
Published by NRC Research Press

1796

Can. J. Civ. Eng. Vol. 36, 2009

need to react to avoid a crash and to come back to a distance greater than the minimum desired distance for standing vehicles (AX). The transition to the emergency braking situation can happen either from the closing process or from the following process. If the vehicle is still in the closing process (i.e., DV > SDX) and the spacing between the lead and following vehicles becomes smaller than ABX, then the emergency braking regime begins. If the vehicle is in the following regime and the lead vehicle brakes suddenly, leading to DV > CLDV and DX < ABX, then the emergency regime is also initiated. The gap-acceptance behaviour for left-turn, right-turn, and crossing manoeuvres was modeled using the conflict area feature in VISSIM. Based on intersection geometric characteristics as defined by the user, VISSIM automatically detects overlapping areas (conflict areas) and the user establishes which movement should yield the right-of-way. The driver in a lower priority approach evaluates gaps in the main stream, the situation behind the conflict area, and the current speedacceleration profile of the main stream before deciding to proceed or stop at the intersection. Unfortunately, there is little information about how the software performs these tests.

The CPI expression used in this analysis is of the following form:

tf i  X  P MADRa1 ;a2 ;:::an < DRACi;t Dtb

CPIi

ttii

Ti

Estimating safety performance

Safety-performance measures attempt to capture real-time vehicle interactions in the traffic stream and hence explain their potential for crashes. Safety performance is a more inclusive expression of high-risk driving behaviour than reported crash history, since it also accounts for near misses, which are indicative of lack of safety but do not result in actual crashes. A number of measures of safety performance can be established, including time-to-collision (TTC), deceleration rate to avoid the crash (DRAC), and postencroachment time (PET), among others (Gettman and Head 2003). Gettman and Head (2003) and Archer (2005) have explicitly recognized the relevance of DRAC as a safetyperformance measure that considers the role of speed differentials and decelerations in potential crash occurrence. The conventional DRAC measure, however, fails to accurately reflect traffic conflicts because it does not consider the vehicles braking capability over time for prevailing road and traffic conditions. Hence, a desirable measure of safety performance is one that includes both required deceleration rates and braking capabilities for individual vehicles. A crash-potential index (CPI) is established in this paper from the perspective of a driver in the traffic stream (response vehicle, RV) responding to a given stimulus provided by another vehicle (stimulus vehicle, SV). More precisely, CPI is defined in terms of the probability that the DRAC experienced by a given RV exceeds its maximum available braking capability (MADR) for every 0.1 s of simulated time. DRAC is estimated for each time interval as a function of the differential speed and spacing between SV and RV using Newtonian physics. The SV is the vehicle responsible for the initial action (braking for a traffic light stop sign, changing lanes, and (or) accepting a gap), and the RV is the vehicle immediately affected by SV action and must respond to avoid dangerous interactions. It is worth noting that an RV response, whether braking or changing lanes, can also work as a stimulus for the subsequent vehicle in the simulation.

where CPIi is the crash-potential index for RV i; MADRa1 ;a2 ;:::an is a random variable following a normal distribution according to the vector of traffic and environmental attributes a1, a2, . . ., an; DRACi,t is the deceleration rate to avoid the crash for RV i during time interval t; tii is the initial simulated time interval for RV i; tfi is the final simulated time interval for RV i; Dt is the simulation time interval (s); b is a state variable (b = 1 if RV is approaching SV (interaction), otherwise b = 0); Ti is the total simulated time for vehicle i (s). A number of other safety-performance measures can be established using the CPI expression given by eq. [1]. One such measure is vehicle interaction, defined as the situation where a stimulus is presented to the RV such that braking is required to avoid a future collision (CPI > 0). In this case, CPI is estimated assuming MADR to be normally distributed, with an average value of 8.45 m/s2 and a standard deviation of 1.40 m/s2. These values were obtained from field tests for different vehicles with initial speeds from 80 to 100 km/h coming to a full stop (MOVIT 2006; Neilsen 2007). Vehicle conflicts, on the other hand, are a subset of interactions such that DRAC exceeds MADR. For conflicts, individual values of MADR are assigned to individual vehicles entering the simulation, and a conflict is ascribed every time interval DRAC exceeds MADR. To avoid negative and unrealistic values of MADR, the normal distribution was truncated to minimum and maximum values of MADR of 4.2 and 12.7 m/s2, respectively. Since safety performance in this paper is measured in terms of rear-end and angled interactions, separate expressions for DRAC need to be defined. For rear-end vehicle interactions, DRAC can be expressed as 2 DRACREAR RV;t1 VRV;t VSV;t 2 2XSV;t XRV;t LSV;t

where t is the time interval, X is the position of the vehicles, L is the vehicle length, and V is the velocity. Angled interactions emerge after vehicles in lower priority manoeuvres accept specific gaps and therefore become the SV. Since SV and RV trajectories are not parallel for angled interactions, conflict areas need to be defined for all combinations of approaching movements. Vehicle speeds and distances to crash zones are used to verify the existence of a collision course between SV and RV. Figure 2 illustrates parameter definitions for angled interactions resulting from right-turn manoeuvres, with the corresponding expression for DRAC of the form 3 DRACANGLED RV;t1 VRV;t VSV;t 2 2XCAm;n XRV;t

where XCAm;n is the position of the closest boundary of the conflict area for SV movement m and RV movement n.
Published by NRC Research Press

Cunto and Saccomanno Fig. 1. Perception and reaction thresholds and distances in the carfollowing model (CFM) of Wiedemann and Reiter (1992) (source: PTV 2008). Fig. 2. Framework to establish CPI for a given time interval.

1797

Vehicles are supposed to apply DRAC to avoid high-risk situations. In reality, combinations of differential speed and spacing are perceived by the RV driver who, after a given delay or reaction time, will respond in a particular fashion. The simulation algorithm updates DRAC for every time interval (0.1 s) based on driver reaction in the previous interval. For example, in a given interval (ti) the RV driver applies a deceleration rate greater than DRAC. In the next time interval it is expected that DRAC will be lower, assuming that the SV driver did not decelerate, and vice versa.

Intersection and model specification

The intersection being simulated in this paper consists of four legs with two lanes in each major approach (westbound and eastbound) and one lane in each minor approach. All lanes have a fixed width of 3.5 m and an angle between major and minor approaches of 908. The simulation experiment was designed to investigate differences in CPI-based measures of safety performance caused by three factors: (i) introduction of fixed signal control to replace a previous stop sign in the minor approach, (ii) changes in traffic volume in the major approach, and (iii) type of vehicle interaction (rear end or angled). In total, 10 different traffic volumes were investigated for both stopcontrolled and signalized cases. Table 1 summarizes the assumed approach volumes and relevant traffic parameters for the simulation experiment. It is worth noting that right turns on red have been permitted for the signalized intersection. However, the traffic signal option does not allow for advanced-green left-turn manoeuvres. Signals are assumed to operate independently

of timings at other adjacent intersections. Semi-actuated phases have been optimized using Synchro 7. The traffic directional split is assumed to be constant for all approaches during the simulation period for both signalized and unsignalized cases (i.e., 5% for left turns, 5% for right turns, 90% through movements). One of the major steps in applying microscopic simulation is to ensure that traffic model inputs (in this case, VISSIM input parameters) are accurately calibrated based on observational vehicle tracking data. Only in this way can we ensure that simulated measures of safety performance closely reflect what can be observed in the real world. Cunto and Saccomanno (2008) describe an in-depth heuristic procedure for calibrating and validating VISSIM carfollowing, gap-acceptance, and lane-change parameters based on observed vehicle tracking data. The data were obtained from the Next Generation SIMulation (NGSIM) program administered by the FHWA (FHWA 2006). The microscopic behaviour of road users has been calibrated and validated in terms of the CPI for every vehicle in the data sample. This differs considerably from conventional calibration procedures based on macroscopic traffic attributes such as average speed, volume, delay, and density. The primary aim of the calibrationvalidation procedure was to obtain values of simulated safety performance based on specified model inputs that closely matched observed safety performance from NGSIM. The results of this calibration yielded best estimate values for those inputs that were found to be statistically significant in explaining safety performance measures as obtained from the simulation. These values are summarized in Table 2. VISSIM default values were used for those inputs not found to be statistically significant for simulating safety performance. Each simulation has a 15 min duration plus a 5 min warm-up interval. To account for variations in different simulation runs for the same traffic volumes, 15 replicates were carried out for each run using different random seeds. In total, the analysis in this paper involved 300 simulation runs.
Published by NRC Research Press

1798 Table 1. Intersection approach volumes and traffic attributes. Volume (vph) Stop-controlled Major approach LOSb A A A A A A A B B B Minor approach LOS C C D F F F F F F F

Can. J. Civ. Eng. Vol. 36, 2009

Signalized Major approach LOS A A A A A B B B B B Minor approach LOS A A A A A A A A B B

Levela 1 2 3 4 5 6 7 8 9 10
a b

Eastbound 300 400 500 600 700 800 900 1000 1100 1200

Westbound 300 400 500 600 700 800 900 1000 1100 1200

Northbound 100 100 100 100 100 100 100 100 100 100

Southbound 100 100 100 100 100 100 100 100 100 100

Cycle (s) 40 40 40 40 40 40 45 45 50 60

Levels 710 satisfy the FHWA Manual of Uniform Traffic Control Devices (MUTCD) warrant 3 for signalization. LOS, level of service.

Table 2. VISSIM calibrated input parameters. Input parameter Desired deceleration CC0 CC1 Calibrated 2.6 3.0 1.5 Description Maximum deceleration (m/s2) drivers are willing to apply in normal (not emergency) situation Standstill distance (m); defines the desired distance between stopped cars Headway time (s); defined as the minimum time a driver wants to keep from the lead vehicle; the higher the value, the more cautious the driver; CC0 and CC1 are combined to express the safety distance

Detailed vehicle information for every 0.1 s in the simulation was obtained from VISSIM and stored in *.fzp output files. These files were scanned by a search algorithm coded in VisualBasic.net to determine the RV and its corresponding SV, as well as DRAC and CPI, in 0.1 s increments for every vehicle in the simulation. It has been recognized that microscopic simulation users should perform visual inspections during simulation run time to search for abnormal vehicle behaviour such as sudden vehicle stoppage and abrupt lane change. Two of these situations have been observed during VISSIM simulations that would have impacted safety performance measurements: (i) crashes during gap-acceptance situations, and (ii) vehicles too close during lane-change manoeuvres. Crashes have been observed in VISSIM for complex intersection environments where a considerably high number of vehicles and turning manoeuvres exist. These situations must be excluded for the analysis because it is considered that VISSIM cannot realistically simulate crashes. The algorithm is based on a crash-avoidance logic. In this paper, VISSIM output files (*.fzp files) were filtered for vehicles occupying the same space at the same time throughout a VisualBasic.net application developed for estimating CPI, named CPI calculator. Such occurrences were identified by unreasonably high DRAC values (>15 m/s2) and (or) very low (or negative) spacing between lead and following vehicles. In situations where the following vehicle has started to change the lane to improve its travel speed (discretionary lane change), the following driver starts to look at the target lane, although for a brief period, say 0.20.5 s, the vehicle is still occupying its original lane. In some occasions, especially when the lead vehicle in the original lane is too close, this transition yields very low spacing between vehicles with considerably high speed differentials. In the CPI calcula-

tor VisualBasic.net application, these situations have also been filtered by comparing the status of the following vehicles VISSIM variable called Lch (direction of current lane change). The signs < or > in the Lch variable indicate that the vehicle has started the lane-change manoeuvre to the left or right lane, even though the vehicle is still in the original lane.

Simulation results
This section presents the major findings from simulation of safety performance for the two intersection traffic control strategies, namely stop control and fixed signal control. As noted previously, four parameters based on CPI were used to reflect changes in intersection safety that result from signalization: (i) CPI/vehicle, (ii) CPI 85th percentile (CPI85), (iii) percentage of vehicles interacting (CPI > 0), and (iv) percentage of vehicles in conflict (DRAC > MADR). CPI/vehicle is obtained by summing the CPI for all interacting vehicles and dividing this by the number of simulated interacting vehicles. This measure reflects the average individual safety performance associated with each traffic scenario. CPI85 establishes a threshold for comparing safety performance of individual traffic control (value of CPI/vehicle that is exceeded 15% of the time). The percentage of vehicles interacting represents the total number of vehicles with CPI greater than zero divided by the total of simulated vehicles. The percentage of vehicles in conflict captures extreme traffic interactions where the maximum deceleration required to avoid the crash (DRAC) exceeded the braking capacity of the vehicle (MADR). This value is also divided by the total number of vehicles in the simulation. Table 3 summarizes the simulated average measures of safety performance for both stop-controlled and signalized scenarios. These are based on the average of 15 simulations
Published by NRC Research Press

Cunto and Saccomanno

Table 3. Safety performance indicators for stop-controlled and signalized intersections. Rear-end interactions Volume level Stop-controlled 1 2 3 4 5 6 7 8 9 10 Signalized 1 2 3 4 5 6 7 8 9 10
Published by NRC Research Press

Angled interactions Vehicles interacting (%) 19.8 22.4 27.7 31.6 37.1 42.7 47.2 51.1 56.5 60.5 27.2 31.0 36.5 42.1 47.8 54.1 59.6 64.3 70.4 76.1 Vehicles in conflict (%) 0.00 0.00 0.00 0.02 0.00 0.02 0.00 0.02 0.01 0.02 0.00 0.00 0.00 0.06 0.03 0.10 0.09 0.04 0.27 0.33 CPI/vehicle 1.45105 1.98105 8.77105 1.36105 3.14108 6.30105 8.53105 1.82105 1.02104 1.04104 1.59109 1.68109 1.18108 6.74106 9.15104 6.40104 1.00103 5.27104 3.28104 5.24104 CPI85 9.98109 2.63108 2.36108 1.42108 1.83108 2.72108 3.16108 3.00108 3.03108 3.65108 1.59109 3.34109 3.17108 1.35105 1.61103 1.28103 2.20104 1.15103 2.12105 7.19104 Vehicles interacting (%) 4.8 4.5 6.2 6.7 6.9 7.1 7.0 5.9 5.1 4.2 0.2 0.1 0.4 0.3 0.4 0.5 0.9 0.7 1.0 1.1 Vehicles in conflict (%) 0.03 0.00 0.02 0.02 0.00 0.06 0.04 0.02 0.06 0.04 0.00 0.00 0.00 0.00 0.03 0.03 0.06 0.02 0.04 0.08

No. of simulated vehicles 198 247 300 347 396 454 506 555 610 659 199 250 301 349 398 454 506 557 614 666

CPI/vehicle 1.27108 5.87109 1.17108 1.85106 7.48108 9.90106 2.58108 2.64106 1.88106 1.67106 4.08108 6.22108 2.09107 9.45106 1.32106 6.03106 4.61106 4.84106 1.13105 1.35105

CPI85 1.48108 6.60109 9.80109 7.44109 5.12109 7.04109 4.51109 4.24109 3.06109 2.65109 6.35108 5.67108 5.21108 4.29108 4.35108 4.07108 3.65108 3.20108 3.23108 3.35108

1799

1800 Fig. 3. CPI/vehicle versus volume for rear-end interactions.

Can. J. Civ. Eng. Vol. 36, 2009

Fig. 4. CPI85 versus volume for rear-end interactions.

carried out using a different number of seeds. Table 3 also provides results for different major approach volumes and the two types of interactions of interest (rear end and angled).

Rear-end interactions
When comparing the two traffic control strategies for rear-end interactions, the introduction of signalization consistently results in an increase in the average CPI/vehicle values, suggesting that signalization may be producing negative safety dividends, and this could be due to disruptions in traffic flow along the major approach that did not exist under stop control. It should be noted that the introduction of the traffic signal results in an increase in the percentage of vehicles interacting of from 7% to 15%, conditional on increasing volume. The relationship between volume and CPI/vehicle and between volume and percentage of vehicles in conflict is not as consistent as the relationship between volume and percentage of vehicles interacting, and this could be explained by the high variability in these measures. The influence of volume on the percentage of vehicles in conflict seems to

become more apparent for the signalized scenario at higher volumes. A visual analysis of the results for CPI/vehicle and CPI85 (Figs. 3 and 4) seems to support the findings from Table 3. From Fig. 3, a high degree of variability in CPI/vehicle was observed with increasing volume (heteroscedasticity). This suggests that experiments to detect differences in simulated CPI/vehicle should be carefully designed to account for more subtle differences in the measure of safety performance and that a large number of simulations may be required to enhance confidence in the results. Figure 4 indicates that, for rear-end interactions, CPI85 decreases with an increase in volume regardless of traffic control, stop or fixed signal. This suggests that high-risk drivers are restricted by volume in achieving their desired speeds. Volume in this instance acts as a kind of speed-dampening effect to discourage high-risk behaviour as measured by the 85th percentile. Figures 5 and 6 provide additional evidence concerning the influence of volume on number of vehicles in conflict for stop and fixed signal control, respectively. The previous figures show that as volume increases, the percentage of vehicles not interacting (i.e., CPI = 0) decreases for both stopcontrolled and signalized scenarios, and the percentage of
Published by NRC Research Press

Cunto and Saccomanno Fig. 5. Distributions of CPI values for rear-end interactions in stopcontrolled scenario.

1801 Fig. 7. Distributions of CPI values for angled interactions in stopcontrolled scenario.

Fig. 6. Distributions of CPI values for rear-end interactions in signalized scenario.

Fig. 8. Distributions of CPI values for angled interactions in signalized scenario.

vehicles interacting (CPI > 0) increases for both control strategies. This result appears to be reasonable, considering that at higher volumes the average spacing between vehicles is reduced with increased interactions requiring braking. The introduction of the signalization significantly increases the percentage of vehicles with a high probability of DRAC exceeding MADR. For example, for a volume of 2600 vehicles per hour (vph), the percentage of vehicles with CPI between 108 and 107 increases from 4% to 14% after signalization. This suggests that for rear-end interactions there is a shift in the CPI/vehicle distribution to the right, and hence a reduction in safety following the introduction of fixed signal controls. Angled interactions For angled interactions, signalization yields lower CPI/vehicle values at lower traffic volumes. However, no such trend was observed at higher volumes. At lower volumes we would expect greater opportunities for left-turn manoeuvres from the major approach, whereas at higher volumes the

presence of a traffic signal creates a kind of platooning effect along the major approach, reducing the number of available safe gaps for left-turn vehicles (and hence higher average CPI/vehicle). For example, for the stop-control strategy at volume level 8, Table 3 suggests that 5.9% of vehicles are interacting compared with only 0.7% for the signalized case. Nevertheless, the 0.7% of vehicles yield CPI values for the signalized case that are considerably higher than those for the stop-control case. This suggests that in the absence of an advanced green for left-turn manoeuvres, a higher proportion of shorter gaps is accepted. For angled interactions and stop-controlled intersections, the percentage of vehicles interacting increases at low volumes until a maximum value of 2000 vph (volume level 7) is reached and decreases thereafter. For signalized intersections, the percentage of vehicles interacting was found to increase consistently with an increase in volume. Figures 7 and 8 illustrate the frequency of simulated angled interactions for different CPI values and volume for stop-controlled and signalized scenarios, respectively.
Published by NRC Research Press

1802 Table 4. Two-way ANOVA of CPI/vehicle for rear-end interactions. Source Volume Control Volume control Error Total Type III sum of squares 15.74 7.03 6.70 62.66 108.47 df 1 1 1 176 180 MS 15.74 7.03 6.70 0.36 F 44.20 19.74 18.81

Can. J. Civ. Eng. Vol. 36, 2009

Significance 0.00 0.00 0.00

Note: df, degrees of freedom; F, F statistic; MS, means squared.

Table 5. Two-way ANOVA of CPI/vehicle for angled interactions. Source Volume Control Volume control Error Total Type III sum of squares 21 326.5 18 432.12 15 209.71 630 358.08 743 035.21 df 1 1 1 176 180 MS 21 326.50 18 432.12 15 209.71 3 581.58 F 5.95 5.15 4.25 Significance 0.02 0.02 0.04

Note: df, degrees of freedom; F, F statistic; MS, means squared.

For angled interactions, a key finding of the simulation runs is that the number of vehicles interacting is reduced after signalization. However, a small number of vehicles with very high CPI values continue to be present after signalization, especially under high volumes. The signalization scenario considered in this analysis does not permit advanced-left-turn control, even at high volumes. In practice, an advanced green for left-turn movements would be considered at such volumes, reducing the problem of unsafe gap acceptance and resultant angled interactions. Analysis of variance To statistically verify the influence of volume and type of control in the average CPI/vehicle, the results in Table 3 were grouped into two major categories: (i) low volume (levels 13), and (ii) high volume (levels 810). A two-way analysis of variance (ANOVA) was carried out for rear-end and angled interactions, and the results are summarized in Tables 4 and 5, respectively. The ANOVA results indicate that all main factors (volume and type of control) and their interactions have a significant effect on average CPI/vehicle for both rear-end and angled interactions. This confirms the findings from visual inspection that the introduction of fixed signal controls at the stop-controlled intersection can compromise safety by increasing the potential for rear-end traffic conflicts.

Conclusions
A microscopic simulation model has been presented for investigating the safety implications of introducing fixed signal control at a stop-controlled intersection for different traffic conditions. The model produces a number of meaningful measures of safety performance such as crash potential index per vehicle (CPI/vehicle), value of CPI that is exceeded 15% of the time (CPI 85th percentile, CPI85), percentage of vehicles interacting, and percentage of vehicles in conflict.

A simulation experiment was carried out by applying the model to a stop-controlled, four-legged intersection with major and minor approaches. The analysis investigated differences in CPI-based measures of safety performance resulting from changes in volume for two types of vehicle interactions, namely rear end and angled. For rear-end interactions, the introduction of fixed signal control increased the percentage of vehicles interacting and the average CPI/vehicle and CPI85. This suggests that signalization could result in an increase in rear-end crash risk when compared to a stop-signal control. An increase in volume on the major approach was found to yield higher levels of vehicles interacting and in conflict with associated higher crash risks. For angled interactions, the introduction of fixed signal control resulted in a reduction in the percentage of vehicles interacting. The effect on safety performance, however, was not consistent for all assumed volumes. At low volumes, CPI/vehicle was found to be lower after signalization (i.e., safety was enhanced), but at high volumes CPI/vehicle was found to increase slightly. This inconsistency was due to the absence of an advanced-green signal for left-turn manoeuvres from the major approach to the minor approach. The relationship between CPI/vehicle and signal control and volume was found to be statistically significant at the 5% level for both rear-end and angled interactions. The application of microscopic simulation presented in this paper provides an alternative behavioural platform for investigating the safety implications resulting from the introduction of different countermeasures for varying traffic conditions.

References
Abdel-Aty, M.A., Keller, J., and Brady, P.A. 2005. Analysis of the types of crashes at signalized intersections using complete crash data and tree-based regression. Transportation Research Record 1908, pp. 3745. doi:10.3141/1908-05. Archer, J. 2005. Methods for the assessment and prediction of trafPublished by NRC Research Press

Cunto and Saccomanno fic safety at urban intersection and their application in micro-simulation modelling. Ph.D. thesis, Department of Infrastructure, Royal Institute of Technology, Stockholm, Sweden. Barcelo, J., Dumont, A., Montero, L., Perarnau, J., and Torday, A. 2003. Safety indicators for microsimulationbased assessments. In Proceedings of the 82nd Annual Meeting of the Transportation Research Board, Washington, D.C., 1213 January 2003. Transportation Research Board, Washington, D.C. Brackstone, M., and McDonald, M. 1999. Car-following: a historical review. Transportation Research Part F: Traffic Psychology and Behaviour, 2(4): 181196. doi:10.1016/S1369-8478(00) 00005-X. Cunto, F., and Saccomanno, F.F. 2008. Calibration and validation of simulated vehicle safety performance at signalized intersections. Accident Analysis and Prevention, 40: 11711179. doi:10.1016/j.aap.2008.01.003. PMID:18460386. FHWA. 2004. Intersection safety briefing sheets. Federal Highway Administration (FHWA), US Department of Transportation, Washington, D.C. Available from www.ite.org/library/ IntersectionSafety/BriefingSheets.pdf. [accessed October 2008]. FHWA. 2006. NGSIM lankershim data analysis: summary report. Federal Highway Administration (FHWA), US Department of Transportation, Washington, D.C.. Available from www.ngsim. fhwa.dot.gov. [accessed August 2009]. Gettman, D., and Head, L. 2003. Surrogate safety measures from traffic simulation models. Report FHWA-RD-03-050, Federal Highway Administration (FHWA), US Department of Transportation, McLean, Va. Hadayeghi, A., Malone, B., and De Gannes, R. 2006. Development of new crash experience warrants for traffic signals in Ontario, Canada. Transportation Research Record 1953, pp. 120127. doi:10.3141/1953-14. Hayward, J. 1971. Near misses as a measure of safety at urban intersections. Ph.D. thesis, Department of Civil Engineering, Pennsylvania State University, University Park, Pa.

1803 Huguenin, F., Torday, A., and Dumont, A. 2005. Evaluation of traffic safety using microsimulation. In STRC05: Proceedings of the 5th Swiss Transport Research Conference, Ascona, Switzerland, 911 March 2005. ETH, Zurich, Switzerland. Mehmood, A. 2003. Integrated simulation model for driver behaviour using system dynamics. Ph.D. thesis, Department of Civil and Environmental Engineering, University of Waterloo, Waterloo, Ont. Minderhoud, M.M., and Bovy, P.H. 2001. Extended time-to-collision measures for road traffic safety assessment. Accident Analysis and Prevention, 33: 8997. doi:10.1016/S0001-4575(00) 00019-1. PMID:11189125. MOVIT. 2006. Braking distances [online]. Available from www. movit.de/rahmen/stoptbl.htm [accessed 3 March 2007]. Neilsen, J. 2007. Stopping distances. Defensive driving safe driving training SDT [online]. Available from www.sdt.com.au/ [cited 3 March 2007]. Pernia, J.C., Lu, J.J., Xie, X., Weng, M., and Snyder, D. 2002. Development of models to quantify the impacts of signalization on intersections crashes. In Proceedings of the 81st Annual Meeting of the Transportation Research Board, Washington, D.C., 1317 January 2002. Transportation Research Board, Washington, D.C. PTV. 2008. VISSIM version 4.30 user manual. Planung Transport Verkehr AG (PTV), Karlsruhe, Germany. Wiedemann, R., and Reiter, U. 1992. Microscopic traffic simulation: the simulation system MISSION, background and actual state. CEC Project ICARUS (V1052), Final Report, Vol. 2, Appendix A. Commission of European Communities (CEC), Brussels. Xin, W., Houdos, J., Michalopoulos, P., and Davis, G. 2008. The less-than-perfect driver: a model of collision-inclusive car-following behavior. In Proceedings of the 87th Annual Meeting of the Transportation Research Board, Washington, D.C., 1317 January 2008. Transportation Research Board, Washington, D.C.

Published by NRC Research Press