category: literaturenote citekey: bellnovelapplicationredlight2012 title: Novel application of red-light runner proneness theory within traffic microsimulation to an actual signal junction authors: "Bell, Margaret Carol; Galatioto, Fabio; Giuffrè, Tullio; Tesoriere, Giovanni" year: 2012 date: 2012-00-00 2012 doi: 10.1016/j.aap.2011.12.010 publication: "Accident Analysis & Prevention" url: "https://doi.org/10.1016/j.aap.2011.12.010" zotero_key: 6ZA24VT4 zotero_storage: RJRD2GEN collections: transpordi_planeerimine folder: 001_artiklid firstAuthor: "Bell, Margaret Carol"

status: converted

Contents lists available at SciVerse ScienceDirect

Accident Analysis and Prevention

journal h om ep age: www.elsevier.com/locate/aap

Novel application of red-light runner proneness theory within traffic microsimulation to an actual signal junction

Margaret Carol Bell a, Fabio Galatioto^a,∗, Tullio Giuffrè b, Giovanni Tesoriere ^b

^a School of Civil Engineering and Geosciences, Newcastle University, Cassie Building, Newcastle Upon Tyne NE17RU, UK
^b Faculty of Engineering and Architecture, Kore University of Enna, Cittadella Universitaria, 94100, Enna, Italy

a r t i c l e i n f o

Article history: Received 19 July 2011 Received in revised form 7 December 2011 Accepted 22 December 2011

Keywords:

Road junction safety evaluation Potential conflict analysis Traffic microsimulation Traffic light management Road safety counter measures

a b s t r a c t

Building on previous research a conceptual framework, based on potential conflicts analysis, has provided a quantitative evaluation of 'proneness'to red-light running behaviour at urban signalised intersections of different geometric, flow and driver characteristics. The results provided evidence that commonly used violation rates could cause inappropriate evaluation of the extent of the red-light running phenomenon. Initially, an in-depth investigation of the functional form of the mathematical relationship between the potential and actual red-light runners was carried out. The application of the conceptual framework was tested on a signalised intersection in order to quantify the proneness to red-light running. For the particular junction studied proneness for daytime was found to be 0.17 north and 0.16 south for opposing main road approaches and 0.42 east and 0.59 west for the secondary approaches. Further investigations were carried out using a traffic microsimulation model, to explore those geometric features and traffic volumes (arrival patterns at the stop-line) that significantly affect red-light running. In this way the prediction capability of the proposed potential conflict model was improved. A degree of consistency in the measured and simulated red-light running was observed and the conceptual framework was tested through a sensitivity analysis applied to different stop-line positions and traffic volume variations. The microsimulation, although at its early stages of development, has shown promise in its ability to model unintentional red light running behaviour and following further work through application to other junctions, potentially provides a tool for evaluation of alternative junction designs on proneness. In brief, this paper proposes and applies a novel approach to model red-light running using a microsimulation and demonstrates consistency with the observed and theoretical results.

1. Introduction

Red light running (RLR) at signalised junctions over the recent two decades has been well researched because of the human and financial cost to individuals and Governments (Retting and Williams, 1996).Also,the effects ofthis phenomenon on road safety are documented well by several statistics (Mohamedshah et al., 2000; Milazzo et al., 2001) highlighting both a very high frequency (up to an average value of 1 violation every 3.5 min) and the respective gravity of the consequences (Wissinger, 2000), with a growing intensity in recent years. In Mohamedshah et al. (2000) a study of data from four States in the U.S. show that RLR crashes account for 16–20% of the total crashes at urban signalised intersections. Evidence presented by Milazzo et al. (2001) showed that about 5% of urban crashes are the result of red-light runners. A large database study carried out by Retting et al. (1999) to analyze the RLR phenomenon suggested that usually accidents caused by RLR drivers involve several vehicles. The trajectories of RLR crash types can be angle and side-impacts. Strictly speaking, rear-end collisions are not normally considered among the outcomes of red-light running, but nevertheless, ironically, studies highlight an increase in this type of collision as a consequence of the increase in red-light running enforcement (Persaud et al., 2005; Council et al., 2005). According to Baguley (1988), Hill and Lindley (2004) and Elmitiny et al. (2010) proneness to red-light behaviour can be considered to be governed by two main factors which influence the

However, the authors believe that the road factors category should be separated into vehicle (size, type, road worthiness) and

decision to stop, or to keep going, at the intersection. These are: - human factors, related to population characteristics;

^- road factors depending on traffic, geometric and operational characteristics present at the intersection.

^∗ Corresponding author. Tel.: +44 0191 2226329; fax: +44 0191 222 6502. E-mail addresses: margaret.bell@ncl.ac.uk (M.C. Bell), fabio.galatioto@ncl.ac.uk (F. Galatioto), tullio.giuffre@unikore.it (T. Giuffrè), giovanni.tesoriere@unikore.it (G. Tesoriere).

flow (volume, speed, etc.) related factors and infrastructure factors (topography, geometric and operational factors) characterising the intersection.

Some behavioural research has clarified the role of the two different factors on RLR in terms of "dilemma zone" (Allos and Al Haidithi, 1992) and circumstances that make red light running possible (Bonneson et al., 2001). Van der Horst and Wilmink (1986), interestingly, suggested that in drivers' stop–go decision process three fundamental criteria are present:

expectations due to previous knowledge of the intersection (especially waiting times at red lights);
assessment of the consequences of a violation;
estimation of the consequences of stopping.

In urban areas, speeds are generally constrained by legislation and/or congested condition, the authors consider speed effects on RLR mostly associated with the human factor element. Previous research results in respect of speed, are not giving consistent messages. In particular, in Kent et al. (1995) RLR rates for drivers of through-junction vehicles were essentially the same for 60 km/h as for 80 km/h speed zones. However, RLR rates were found to be significantly lower (up to 8 times) for right-turn movements (with either divided or undivided approaches – UK driving) at 80 km/h compared to 60 km/h, reflecting the heightened awareness of fatal accident risk at the higher speed. Bonneson et al. (2002a,b), on the other hand, suggested that more red-light-running (US drivers) will occur at higher speeds, but with a relationship that was not as strong as it was for approach flow rate or yellow interval duration.

Evidence presented in Chin and Quddus (2003) confirms that geometric design parameters such as visibility distance, presence of right/left (UK/Europe) queuing reservoir, dimension of central reserve, have an impact on the frequency of accidents. More specifically, Hill and Lindley (2004) investigated the effect on RLR of geometric design factors and traffic characteristics. An empirical multiple linear relationship was developed to predict the number of RLR violations per hour for an intersection as a function of the number of lanes of the primary and secondary roads and the Annual Daily Total Traffic (AADT). In particular the effect on the number of violations of an additional lane into the major road represented a larger positive compared to the smaller, and negative, effect of the addition of a lane into a minor road.

Whilst acknowledging that the human factors component of RLR is important, unlike road factors, it is a component difficult to manage other than with education and enforcement. In this study the RLR that results from human factors are referred to as the intentional RLR and those that can be influenced by traffic and demand management and control measures, and occur due to statistical probability are referred to as the unintentional RLR. Often studies qualitatively consider the influence of both human, vehicle and road factors on the decision making process but quantitative evaluation of proneness to red light violation is scarce. However, from an engineering standpoint, to establish appropriate countermeasures against the RLR phenomenon, qualitative data alone is insufficient. This paper seeks to address this knowledge gap in an attempt to provide a tool which can be used, at the design stage, to minimise proneness of RLR.

Building on the early work of Van der Horst and Wilmink criteria, a model for predicting the frequency of RLR was developed by Bonneson et al. (2002a,b). This model is based on the probability of a driver stopping (following the onset of the red indication) when that driver is at t seconds travel time from the stop line. The equation for predicting the expected number of RLR per cycle was obtained by integrating the product of the traffic flow and the probability of proceeding (the complement to the probability of stopping) during the yellow aspect. Bonneson's formulation is based on the assumption that each driver, in a given lane, decides to go (or stop)independently of anyprecedingdriver.Aquantitative evaluation of RLR proneness (or propensity) is allowed depending on factors such as approach flow, speed, amber interval duration and clearance path length.

In an attempt to extend the traditional probabilistic and empirical approaches to assess the proneness of a junction (or specific arm of a junction) to RLR depending on junction layout, this paper applies microsimulation as a tool to explore whether the effect of geometric designs (the number of lanes at a major–minor cross road junction with three and two lanes respectively), operational conditions (signal phasing, flow rate, conflicting movements, queue length, etc.) and traffic flow volumes (and their associated headway distributions) significantly affect red-light running. The motivation for the approach is twofold. Firstly, Local Authorities make investment in microsimulation models, and their validation, to design and assess the impacts of traffic management measures, therefore, extending their application to also identify potential accident hot spots brings with it added value. At times of economic uncertainty, demonstrating increasing benefits of expenditure is crucial to realising funding streams in the first instance. Secondly, the propensity of RLR is largely influenced by the signal timings usually optimised to reduce delay and stops oftraffic movements at junctions. These, with little costs, can be adjusted in response to associated RLR risk which can vary throughout the day depending on flow and capacity. On the other hand the geometric factors (number of lanes, curvature, central reserve, etc.) once fixed at the design stage became part of the infrastructure of the built environment and therefore are much more difficult and expensive to modify especially in urban areas. However, the microsimulation method proposed here would enable the geometric layout and control strategies to be explored at the design stage.

The paper builds on the earlier theoretical approach to potential conflict analysis and investigates violation proneness by applying the theory, outlined in Sections 2 and 3, to a real junction accounting for both the impact of the local (site) and general (population) characteristics of the RLR phenomenon. A description of the junction used as a case study for the research is described in Section 4 and the proneness, for each stream into the junction, is quantified based on data made available by the police and direct observation. The paper further explores, in Sections 5 and 6, the potential for a validated microsimulation of the same junction to predict the proneness of links/streams to RLR. Section 7 compares the results of the theory with the microsimulation and conclusions are drawn in Section 8.

2. Theoretical approach

Apromising way of quantifying proneness toRLRis suggested by the theory based on potential conflict analysis previously suggested by the authors (Giuffrè and Rinelli, 2004, 2005). For ease of understanding, before embarking on the development of the concepts underpinning this work, definitions of the variables are presented here:

$$N_{RLR,cri}^{I}, N{RLR,cri}^{II}, N{RLR,rti}^{II}, N{RLR,lt_i}^{II} = \text{number of potential conflicts}$$ due to RLR in streams crossing, turning right and turning left for generic arm $i$ , where $I$ and $II$ define respectively unintentional and intentional crossing conflicts

˛,ˇcr ,ˇrt,ˇlt ⁼ shares of RLR (unintentional, ˛; intentional for crossing,ˇcr ;turning right,ˇrt and turning left, ˇlt) from the generic arm i during the red-light phase

Qe RLR,i = actual number of RLR

l, ^j = gap in stream l and j

tclr ,t^∗ clr = clearance time of each lane group

t1,t² = limits of the lower and upper gap interval

tcr 1,i,tcr 2,i,trt 1,i,trt 2,i,tlt 1,i,tlt ²,i = idem to t1,t2, relating to crossing, right and left turning manoeuvre for vehicle coming from arm i

The conceptual framework to model RLRs presented here refers to three main cases: (i) straight through the intersection (Case A); (ii) right-turners (Case B); left-turners (Case C). Fig. 1 illustrates the conceptual framework applied to a simple four armapproach, for which the conventions and symbols are defined above.

Case A: As highlighted by behavioural observations of the RLR phenomenon (Giuffrè and Rinelli, 2005), two distinct scenarios are possible. The first scenario corresponds to an unintentional manoeuvre (the RLR enters the intersection blindly, without stopping at the traffic lights). In this hypothesis, the RLR coming from approach i subsequently comes into conflict with the two streams on the adjacent side arm approaches Ql and Qj. The second scenario (intentional violation) in the case examined, relates to a RLR who decides to violate the red-lights after hesitating and evaluating the gap required for performing the manoeuvre safely in relation to the positions and speeds of the vehicles in the opposing streams and the depth of the junction. Both components are influenced by the traverse-length of the junction (depends on the presence of a central reservation, number and width of lanes, vehicle length and speed) and flows on opposing streams (safe gap to pass).

Referring to a single stage manoeuvre (and absence of a central reservation) the RLR has to evaluate the simultaneous presence of adequate gaps both in the nearest Ql and the furthest Qj stream to be crossed. The collision likelihood is linked to the RLR's availability, or incorrect anticipation, of a safe interval to cross and potential conflict will exist if the vehicle's safe gap in the conflict stream is filled or compromised. The limits of the gap interval are t1, t2, for lower and upper time thresholds respectively. The specification of t¹ and t² is necessary purely for the purpose of computation; it does not impose any judgment on the risk level associated with each scenario considered.

These scenarios will be dealt with in turn. The number of potential conflicts associated with each of the cases can be determined as follows:

Case A – Scenario I: for unintentional RLR crossing

$$N_{RLR,cri}^I = \alpha \cdot Q{RLR,i}^e[P(\taul \le t{clr}) + P(\tauj \le t{clr}^*)]$$ (1)

$$t_{clr} = \frac{N_l w_l + l_v}{vi}; t{clr}^* = \frac{N_j w_j + l_v}{v_i}$$ (2)

where Nl and Nj are the number of lanes in arm l and j; wl, wj are the lane width in arm l and j; lv is the crossing vehicle length, and vⁱ is the crossing vehicle speed.

Case A – Scenario II: for intentional RLR crossing

$$N_{RLR,cri}^{II} = \beta^{cr} \cdot Q{RLR,i}^{e} [P(t_{1,i}^{cr} < ti^{cr} < t{2,i}^{cr}) + P(t_{1,i}^{*cr} < ti^{*cr} < t{2,i}^{*cr})]$$ (3)

Case B: In this case, it can be supposed that the RLR makes the manoeuvre intentionally from a stationary position and decides to enter the stream to his right when the gap available is deemed sufficient for performing the manoeuvre safely. Considering the cross flow i that can generate a potential conflict, as in the previous case, the overall number of conflicts can be calculated as follows:

$$N{RLR,rt{i}}^{II} = \beta^{rt} \cdot Q{RLR,i}^{e} \cdot P(t{1,i}^{rt} < t{i}^{rt} < t{2,i}^{rt})$$ (4)

Case C: In this case, it can be supposed that the RLR makes the manoeuvre intentionally from a stationary position in a single signal stage and decides to enter the path ofthe stream to his left when the gap available is deemed sufficient for performing the manoeuvre safely.

The number of potential conflicts from left RLR can be calculated with the expression:

$$N_{RLR,lti}^{ll} = \beta^{lt} \cdot Q{RLR,i}^{e} [P(t_{1,i}^{lt} < ti^{lt} < t{2,i}^{lt}) + P(t_{1,i}^{*lt} < ti^{*lt} < t{2,i}^{*lt})]$$ (5)

Referring againto the generic armi,the overallnumber of potential RLR conflicts is given by:

$$N_{RLRi} = N{RLR,cri}^{I} + N{RLR,cri}^{II} + N{RLR,rti}^{II} + N{RLR,lt_i}^{II}$$ (6)

Calculation ofthe potential conflict number is necessary for each intersection arm (i, l, k, j), in order to estimate the overall extent of RLR at the intersection:

$$N{RLR} = N{RLRi} + N{RLRj} + N{RLRk} + N{RLR_l}$$ (7)

The arrival pattern and vehicle stream gap distribution play a primary role in modelling red-light behaviour. This is reflected through the parameters tclr and t^∗ clr , as well as through the unsafe intervals governed by (t1, t2).

Given that the unsafe interval is included, it has been shown (Giuffrè and Rinelli, 2004) that the probability of a gap value for a vehicle stream, can be calculated, firstly, through the exponential negative distribution (for low volume of vehicles) and, generally, with Erlang's distribution (for high volume of vehicles). Furthermore, Giuffrè and Rinelli(2004) demonstrated thatthe time intervals t¹ and t² correspond to behaviours capable of generating potential conflicts whatever the gap distribution used. A potential advantage of the use of microsimulation in a study of RLR is that it allows this approach to be tested atjunctions where the arrival patterns are in platoons, whether signal controlled or not, and of course gap distributions potentially vary throughoutthe day depending on turning movement volumes. However, in this paper, as a first step, proneness is studied at a single junction with fixed time isolated control therefore arrival patterns are normally distributed. Also, at this feasibility stage of the research the unintentional RLR only are modelled and therefore the results are likely to be underestimated. However, in the context of the application of this approach this is not seen to be a limitation because the highway engineer potentially has control only on the unintentional RLR.

3. Quantifying proneness

In order to quantify proneness, consider the "actual" number of RLR manoeuvres to be equal to the observed violations resulting from the individual decision processes in the face of operational circumstances that make red light violation possible. Therefore, for a specific site, the "potential" RLR can be associated with the actual number of RLR manoeuvres which depends on the junction layout, traffic flow and signal operational characteristics of the intersection. The latter is consistent with the early study of Porter and England (2000) where city specific factors including ethnicity have

Fig. 1. Conceptual framework of the model.

Fig. 2. Road intersection layout.

been identified as significant predictors of red light running. In this way the actual RLR manoeuvres, as a proportion of potential RLR, which depend both on local (site) and general (population) characteristics, can be assumed to be a measure of proneness to red-light violation. In order to isolate the effect of population variables, a relationship between the potential number of RLR and the actual number can be stated simply as a linear relationship:

$$Q{RLR}^e = k \cdot Q{RLR}^p \tag{8}$$

where Q^e RLR is the actual number of RLR manoeuvres within the interval considered, Q^p RLR is the potential number of red-light violation (i.e. number of drivers that could have the opportunity to make a red-light manoeuvre), and k represents the proneness to red-light violation typically over an hour (or group of hours) in the day.

In this paper k was determined through direct observations and microsimulation for the study area.

3.1. Determining the k parameter

As was shown above, in order to estimate proneness to red-light violation, the k parameter can be determined by the ratio between actual and the potential RLR. The number of potential RLR, Q^p RLR, can be empirically determined and considered to be:

$$Q_{RLR}^p = Q_y^p + Q_r^p \tag{9}$$

where Q^p ^y is the number of vehicles that reach the intersection whilst the lights are amber and Q^p ^r is the number of vehicles that stand at the front of the queue whilst the lights are red.

Itis importanttonote thatthe actualunintentionalRLRare those that reach the intersection during the extension of the effective green time, e, with a duration equal to about 2 s1 at the junction studied.

Hence, altogether, considering the vehicles entering from the generic arm i, in order to evaluate violation proneness at an assigned site we can refer to the k-factor as follows:

$$k = \frac{Q{RLR,i}^{e}}{Q{RLR,i}^{p}} = \frac{Q{RLR,i}^{e}}{(Q{i}^{p} \cdot \delta{1} + Q{y{i}}^{p} \cdot \delta{2})}$$

$$= \frac{Q{RLR,i}^{e}}{[1/c \cdot (3600 \cdot N{i} \cdot \delta{1} + Q{i} \cdot \Delta t \cdot \delta_{2})]}$$ (10)

where ı1, depends on the flow rate Qi and number of lanes Ni in arm i, and represents the proportion of signal cycles in the hour during which a sufficient probability of a vehicle arriving atthe stop-line in the red time interval, ri in seconds, exists; ı2, depends on flow rate Qi and number of lanes Ni in arm i, and represents the proportion of signal cycles in the hour during which a sufficient probability of a vehicle arriving at the stop line in the time interval t = yi − c exists, where yi is the duration of the amber on arm i, and c is the duration of the traffic signal cycle in seconds.

Clearly, the values of the coefficients ı¹ and ı² depend on the appropriate arrival pattern P(x) and gap distribution P(-) for the stream examined; to account for the number of lanes at the approach one can refer to the following expressions:

a) one lane per approach: b) two lanes per approach: $$-\delta_1 = 1 - P(x = 0; \tau = r_i) = 1 - P(\tau > r_i)$$

$$-\delta_2 = 1 - P(x = 0; \tau = \Delta t) = 1 - P(\tau > \Delta t)$$

$$\tau = \lambda_1 + P(x = 0; \tau = \lambda_1) + P(x = 1; \tau = \lambda_1)$$

$$\tau = \lambda_1 + P(x = 0; \tau = \Delta t) + P(x = 1; \tau = \lambda_1)$$

¹ The HCM (2000) suggests that e can take a value greater than 2 s when there is traffic congestion.

Fig. 3. Regression between VR and MR.

The theory of RLR proposed above was applied to a real junction in Milan, Italy, using direct measurement as well as microsimulation. These will be dealt with in turn in the next sections.

4. Case study

A road intersection in Milan, Italy, was studied in detail over a period of three years. Accident and red-light running data were made available by the police and on-street measurements of traffic flows were carried out. This junction was a typical four arm intersection with three lanes per direction on the main and two lanes each way on the secondary road. A wide central reservation (17 m) divided the opposing lanes on the main road (see Fig. 2). The contrasting geometry width of conflicting arms (36 against 17 m oftraverse-length) ofthe junction meantthis intersection was particularly suitable for this study. The traffic signal control was two-stage fixed-time at this junction with cycle time of 65 s during the daytime and 50 s at night. AADT (Annual Average Daily Traffic) weekday values on the approaches – both directions – were: 42,649 veh/day and 18,759 veh/day on the main and secondary road respectively.

4.1. On-site measurements

For the case study presented here, the number of actual RLRs was assumed to be equal to the observed violations. The junction selected for the microsimulation application was the one with the highest rate of RLR in a wider study of 12 junctions, in Milan (Italy), each operating at different fixed cycle time.

Violation records were collected by the police through field measurements in three different time intervals (9:00–12:00 h, 14:00–17:00 h, 23:00–02:00 h) during the day. Police records are based on red light cameras set to trigger at the instant the presence of a vehicle is detected as crossing the stop line in the red. These records have been used to validate the theoretical approach. However, because the police records were limited to fixed periods and RLR data was needed for the entire day,this was generated by direct measurement from continuous video recording. In order to check for measurement bias a comparison of the RLR rates captured with the video were compared with the similar data from the police, for the three periods of the day. The results are presented in Table 1 and plotted in Fig. 3. A regression analysis with R² = 0.994 revealed a slope of 0.99 and an intercept of 0.93 which were shown not to be statistically significantly different from 1, and 0 respectively, therefore eliminating concerns regarding bias.

Confining the analysis to a day time period (9:00–11:00 h and 14:00–17:00 h) and night-time period (23:00–02:00 h) the aggregated violation rates were quantified for the case study junction separate from the original 12 junctions.

With reference to Table 2, the average RLR is presented for day time and night time respectively, for the selected intersection and the average over all 12 junctions studied is given in brackets. The metrics are normalised per hour, per 1000 vehicles and per 10,000 vehicles-cycle to accommodate the operation of junctions at different cycle times. For the police recorded data, in addition to the number of violations, a description of the type of collision (in the context of a crossing or turning manoeuvre to the right or left) is known. The violation records for the case study were based on reports for the period 2002–2004, for two weeks per year.

The results in Table 2 clearly show the degree of variation of violation per 1000 vehicles (the highest compared to the average RLR rate over 12 junctions) is a more realistic parameter to compare the performance of the different arms of the junction due to the fluctuation in the traffic flow from hour to hour within the observation period. The secondary road violations are consistently 2–3 times higher than the main road for both daytime and night time periods. This is due to the fact that the journey time to traverse the longer distance across the junction which attracts longer (+2 s) secondary road amber durations. The suggestion is that this may give a false sense of security and lowers the perceived risk for secondary road flows leading to higher RLR rates. On the other hand it may simply reflect the "higher gap lengths" presented by the conflicting flows.

Indeed, for each red-light manoeuvre, considered separately by direct observation, showed a prevalence of RLR straight ahead movement both on the main and on the secondary roads (respectively over 75% and 50% of all RLR for that stream irrespective of turning movement) whilst a considerable share of left turn (about 38%) was also registered on the secondary road from arm 4 to 3. The extent of the RLR phenomenon (from both the absolute and the relative point of view) was marked, especially in the daytime and on the secondary road. This circumstance has been generated by specific operational conditions of this intersection when considered in more detail. Regarding the secondary road, inappropriate day time,traffic light phasing and traffic volume peaks has been the primary contributing factors to high RLR rates. Therefore, the large differences in the violation rates between daytime/night time and main/secondary road operations suggested this as a suitable junction to further investigate proneness and carry out the calculation of k values separately for these operational conditions at the intersection. Furthermore it was possible to explore whether k varied for different arms, turning movement, between day/night time or perhaps an unique k value exists for the whole junction.

4.2. Calculation of k factors by direct measurement

The number of potential RLR for the isolated junction was calculated assuming the hourly flows at each approach; ı¹ and ı² conform to a Poisson distribution for flows less than 400 veh/h and a Poisson generalized distribution in other cases. Note that the calculated Q^p RLR values, compared with the actual traffic volume, vary perceptibly during the day, depending on operational conditions (primarily number of lanes, cycle length and turning traffic volumes) and ranged from a minimum to a maximum of 0.8% to 17.1% and from 8.6% to 48.3% for the main and secondary roads respectively. Finally, given that pooling together the daytime and night time hours, during which violations were known to have occurred, the k parameter values based on direct measurement were determined for each link (north, south, east and west), as below:

a) main approach: b) secondary approach: -daytime: 0.17 (north); 0.16 (south); -daytime: 0.42 (east); 0.59 (west); -night time: 0.12 (north); 0.06 (south); -night time: 0.16 (east); 0.20 (west).

5. Microsimulation tool description

Microscopic traffic simulators are probably the most powerful and versatile traffic analysis tools for individual vehicle movements

Table 1 RLRs captured with the video (VR) and measured by the police (MR).

Hours	North–South				South–North					West–East					East–West
	CR		LT		RT		CR		LT		RT		CR		LT		RT		CR		LT		RT
	MR	VR	MR	VR	MR	VR	MR	VR	MR	VR	MR	VR	MR	VR	MR	VR	MR	VR	MR	VR	MR	VR	MR	VR
9–12	7	8	33	31	7	7	29	30	3	3	10	15	16	16	38	37	11	10	42	44	41	41	7	7
14–17	13	13	53	56	8	11	57	57	0	2	11	14	47	47	37	39	11	11	66	66	35	35	7	7
23–2	23	25	4	4	3	4	13	13	0	0	3	4	13	16	6	6	2	3	27	27	0	2	4	4

Table 2 RLR rates for the case study junction.

| | Main flow arms | | | Secondary flow arms | | | | | | |-----------------------|----------------------------|------------------------------|-------------------------------------|----------------------------|------------------------------|-------------------------------------|--|--|--| | | Violations per h | Violations per 1000 vehs | Violations per
10,000 vehs-cycle | Violations per h | Violations per 1000 vehs | Violations per
10,000 vehs-cycle | | | | | Daytime
Night time | 38.50(1.68)
15.33(0.66) | 15.58 (1.79)
17.12 (0.71) | 2.81 (1.68)
2.38 (0.63) | 61.33(1.81)
17.33(0.81) | 56.44 (2.53)
44.00 (1.07) | 10.19 (2.44)
6.11 (0.97) | | | |

Note: the average values measured over all 12 junctions are given for comparison in brackets.

along roads and through junctions in a network. In this and the next section a novel approach to modelling RLR with traffic microsimulation is proposed and subsequently tested to evaluate the ability of microsimulation to reproduce the RLR phenomena.

The AIMSUN (Advanced Interactive Microscopic Simulator for Urban and Non-Urban Networks, http://vvww.aimsun.com) model was used because of its ability to appropriately represent road network geometry, model in detailthe behaviour ofindividual vehicles to reproduce explicitly traffic signal control plans, pre-timed and actuated and has an easy-to-use graphical user interface. Also, a particular feature of AIMSUN is its ability to capture the empirical evidence that driver behaviour often depends on local circumstances (i.e. the acceptance of speed limits on road sections, the influence of gradients, driver interaction whilst travelling in adjacent lanes, etc.).

5.1. Validation and discussion of the microsimulation model of RLR

The first step was to set up and validate the microsimulation model against measured data. This was achieved by generating flows over the range 200–2000 veh/h. The variability within hours for each flow was simulated assuming a range of flow variation ±50% in steps of 10% and selecting 10 different random seeds for each scenario modelled. This generated a modelled sample set of 495 data. Also, this exercise extended the operational range of this junction into congestion at flow levels well beyond the range measured at this junction which was between 300 and 1300 veh/h as shown in Fig. 4.

Fig. 4. Observed compared with simulated flow for the arms of the junction.

As can be seen from the graph the overall correlation is very good with R² = 0.99, if confined to the operational flows (below 1300 veh/h) the correlation is R² = 0.98, for higher flows (above 1300 veh/h) the correlation reduces, but still high with an R² = 0.91. The ability of the model to reproduce the reality to such a high level ofperformance atthis junction was due to the richness of data available from GIS and by direct observation from video recordings; the relatively straightforwardjunctiongeometry, knownhourly arrival patterns and turning movements throughout the day, speeds and flows at what is essentially an isolated junction with known fixed time plans.

In order to simulate the RLR events, the microsimulation was set up to have a double traffic light at each entry link, one representing the existing traffic light and the second used as a virtual one. The two sets of traffic lights per link were set up with different offsets (0–4 s in steps of 1 s) to ensure that as the vehicle crosses the first stop-line, it is presented with a green with 0, 1, 2, 3 and 4 s delay to the start of the red to stop at the following (downstream) stop-line. In reality, the delay of 1 s to amber after the green at the first stop-line 'allows' the 'as if' RLR to be measured as the vehicle responds to the downstream dummy stop-line which forces the stop. Stepping the offset by a further 2, 3 and 4 s allows the additionalRLRcrossing during later elapsed periods.In this way,the two possible behaviours, decelerating to stop and accelerating to pass through the junction (in reality during the amber) were modelled. This approach was adopted because the micro-simulator is set up notto allow vehicles to cross a stop-line during the red. In the model as drivers approach signal aspects they are presented with the same aspect except at the instant of change from green to amber. It is at this time, when the drivers respond to the virtual downstream lights thatRLRis being successfullymodelled (stopping/or not stopping) at the actual stop-line. The data for the post processing and analysis of the violations was made possible by locating two sets of detectors to count the number of vehicles crossing every second (see Fig. 5). In this preliminary study, the vehicles crossing the actual stop-line during the amber to red light phases were extracted using a post-processing filter algorithm and subjected to analysis.

6. Analysis of the microsimulation model of RLR

The data generated by the microsimulation model was based on 15 runs with different seeds. This was greater than the minimum 6–9 runs suggested by the manual and the 10 used in the initial model validation, but was adopted to ensure integrity of the modelled data. The initial run of the microsimulation was for the

Fig. 5. RLR model layout.

base case (see Norm in Fig. 6) stop-line position, 0 m, and green offset between the actual and virtual stop-line of 0 s. The first step in the analysis of the model output was to count the number of RLR across the stop-line in the first 2 s for each cycle throughout the peak period. Next the stop-line was kept constant (0 m for the Norm) and the offset was systematically changed by 1, 2, 3 and 4 s and the number of vehicles crossing the stop-line were counted in the same way. This count represents the vehicle 'as if' crossing in the

Fig. 6. RLR rates and sensitivity analysis results (0–4 s lapsed time) for primary arms (1 and 3) and secondary arms (2 and 4).

first, second, third and fourth second of the amber duration. Therefore, Fig. 6 illustrates how the number of RLR changes as the elapse time to red start (in s) changes from scenario A, with elapse time of 4 s, to scenario E, with elapse time of 0 s. In essence the number of RLR is a measure of the vehicles that, by virtue of the distribution of gaps between vehicles in conflicting streams, are simulated in the model. This depends on the arrival and opposing flows, vehicle speeds, the traffic signal timing plan (greens, inter-greens, cycle times) for the junction modelled, saturation flows and geometric layout of the junction (the width and depth of junction, number of lanes, etc.), all of which govern the relationship between the RLR and the lapsed time.

With reference to Fig. 6 the very different performance of the four arms of the junction is evident. It is clear that the major road stop-lines in general are more prone to RLR particularly in the short lapse time to red start and with much less variability compared to the minor roads. The larger variation in the side road is due to the randomness in the lower flows and the relatively higher levels of conflicting flows, in particular the significant left turn movements. Arm 1 illustrates how the high level of dominant straight ahead with low levels of conflicting left-turn flows results in RLR rate which is fairly stable and varies little across lapsed time to red. On the other hand the opposing Arm 3 flow is characterised by a dominant conflicting left turn flow from Arm 1, resulting in about 25% higher RLR rate.

It is clear that the rate of change of RLR certainty rises more dramatically with increasing time to red start for side road compared to main road traffic. This predominantly is due to the much longer distance to cross the junction for side compared to main road (3 streams with central reserve rather than 2 streams without) and therefore more sensitive to longer intergreen and greater chance to be trapped within the junction and conflict potential with left turning traffic.

A further test of how the RLR occurrence varied with the stopline position for each of the 4 arms of the simulated junction was carried out. Two positions relative to the current stop-line (0) were defined (−2) and (−4) m set back upstream of the link. This simulates the situation where vehicles are given a longer distance to stopbetweenthe actual andthe virtual stop-line (synonymous with changing geometric design or signal control timings) and therefore less risk to red-light running. In essence, advancing the stop-line means that drivers are crossing an 'advanced' stop-line during the green duration (equivalent to the time taken to travel the 2 m or 4 m at the vehicle speed) and therefore given more time to stop at the downstream dummy stop-line. These are considered to be 'as-if' RLR, but in fact they have more time to stop at the actual stop-line. The results in Fig. 6 show that due to the high dominance of straight ahead flows of link 1 there is little effect if the junction has widened. On the other hand the influence of the relatively high left turn affecting link 3 implies, depending on the conflicting flow arrival pattern, an increase (or decrease) in RLR is simulated. As for the side roads there is a much more unstable response which is mostly increasing RLR rate. This reflects the different distribution of arrival patterns associated with the time equivalent to travelling −2 m and −4 m. The graphs clearly show the interaction between the integreen and the arrival pattern of(or gaps in)the traffic across the stop-line.

6.1. Sensitivity testing

In order to clarify the interpretation of the results further, the outputs of the sensitivity analysis are presented by dividing the RLR count at the dummy stop-line, during four seconds amber, by the total flow crossing the actual stop-line during the amber, throughout the simulated peak period, to provide a measure of certainty. The variation ofthe certainty of red-light running (expressed as a %), where the amber lapsed time to red varies from 0 to 4 s is shown in Fig. 7. It is clear that with zero offset any vehicle crossing the dummy stop-line, Q*, has also crossed the actual stop-line, Q, therefore certainty Q*/Q = 100%.

With reference to Fig. 7, the results show that for the main road links (1 and 3), as expected, because the straight ahead flows are dominant, the certainty in red-light running increases to 100% as the elapse time to red decreases to 0 s. As link 1, relatively, has fewer opposing left turners (i.e. the turning movement that conflicts is from arm 3 to 2) than link 3 (left turn arm 1 to arm 4 the certainty of RLR decreases faster with the more time to stop. The higher certainty for link 3 is due to a higher proportion of left turners from 3 to 2 relative to 1 to 4 because although presenting much longer gaps between vehicles has more vehicles requiring gaps to cross.

For arms 2 and 4 the certainty of RLR drops fairly rapidly with lag time to red because flows are much lower as well as there being a much longer traverse distances across the junction with higher intergreen. The much lower flow level on link 2 probably explains the dramatic fall in certainty on link 2 (E to D for norm, −2 m and −4 m) which, at this link, results from the arrival pattern of flow at the stop line presented by the microsimulation model as opposing left turn from arm 4 to 3 is low. This result is consistent with the mathematical approach (see Section 3) which demonstrated that the red light running behaviour is primarily affected by the distribution of gaps which are much larger for the low flow. Turning now to link 4 (601 veh/h) with more than twice the flow compared to link 2 (290 veh/h), the certainty in red-light running remains high to C and as the distance given for stopping increases. This is due, in part, to the dominating effect of the left turning flow (arm 4 to 3 at a level of 194 veh/h) which is almost 1/3 of the total exiting flow of arm 4 and of a similar level to the straight ahead flow (arm 2 to 4 at a level of 240 veh/h), and because traffic crosses the stop-line to wait in the junction.

The sensitivity analysis was extended to test how the RLR occurrence, for a given stop-line position, responds to the traffic flow variation on each of the 4 links of the junction. Ten different scenarios (apart from the base case), have been simulated from −50% up to +50% flow variation in steps of 10%. The signal settings for the stages remained the same for all simulated flows and the junction maintained degrees of saturation level below 95%. The sensitivity test was carried out by increasing and decreasing the flows on adjacent arms for example, between the link pairs 1 and 2 (leaving the other two links 3 and 4 at the base case flows), similarly for all the other pair combinations at the junction. In this way the independent influence of the flow variation on each link pair in turn was tested.

Fig. 8a and b present the results of the sensitivity analysis considering the flow changes of −50% to +50% of the base case. The occurrence of the RLR per unit flow (Fig. 8a) for arm 1, 2 and 4, relative to 3, appear not to be sensitive to the changes in flow (−50% to +50%), however when considering the change in the RLR occurrence relative to the base case, as shown in Fig. 8b, it is clear that the RLR occurrence is sensitive to changes in flow, which is consistent with earlier studies (Porter and England, 2000; Elmitiny et al., 2010). However, the results of link 3 exhibit a rather different characteristic and have values an order of magnitude higher than those on the other 3 arms. Also, it can be seen from Fig. 8a that at the lowest flow levels (−50%) the occurrence of RLR is 0.113 and falls monotonically to the lowest level of 0.059 at higher level of flows (+50%). This at first sight seems counter intuitive however, given the approach taken for this particular sensitivity test, the conflict of the left turn flow from link 1 (with base case flows) dominates the much lower flows (−50%) on link 3 and the reverse is true at the higher flows (+50%). With reference to Fig. 8b, for link 3 the change in occurrence relative to the base case is of the similar levels to

Fig. 7. RLR certainty and sensitivity analysis results.

the other three links. It is interesting to note that Fig. 8b shows an approximate mirroring of behaviour about the base case of positive and negative % variation in flow for conflicting arms 1 and 3, and 2 and 4 with the latter pair less stable than the former. This illustrates the important role that flow levels as well as the dominant left turning movements have on RLR.

The possibility of using the performance measures (RLR rates and certainties) developed in this work to identify parameters that lead to safer designs was explored. A scrutiny of the performance measures reveals evidence that the safer position is different from link to link,therefore atthis initial stage ofthe investigation,theRLR rates and certainties were aggregated over all streams and 4 arms of the junction for each of stop-line distance norm, −2 m and −4 m and the results are presented in Table 3. The results clearly show that −2 m presents the lower values for both certainty and rates

Table 3 Junction RLR rates and certainty by stop-line position.

| | | Stop-line position | | | | | | | |-----|--------------------|--------------------|---------------|---------------|--|--|--|--| | | | Norm | −2 | −4 | | | | | | RLR | Certainty
Rates | 370%
0.455 | 347%
0.409 | 363%
0.465 | | | | |

demonstrating the potential benefit of this approach to identify the position of the stop-lines which reduce the likelihood of RLR.

7. Comparison of theory with microsimulation

In order to evaluate the performance of this novel approach to quantify the RLR using microsimulation and the signal junction setup, described in Section 5, against the theory, the values of the proneness k from the microsimulation outputs available for day time only, have been estimated. With reference to Eq. (8), proneness is expressed as a ratio ofthe actual unintentional RLR, Q^e RLR, and the potential RLR, Q^p RLR. The actual unintentional RLR is number of RLRs related to the 3 lapse times E, D, C (0, 1 and 2 s) associated with the specific time period (amber − e, where e is equal to 2 s) for this junction. The potential RLR is the number of RLRs passing across the actual stop-line in the 4 s amber plus the number of vehicles standing at the front of the queue in the red.

The results are presented below:

main approach: secondary approach: -daytime: 0.20 (north); 0.15 (south); -daytime: 0.55 (east); 0.42 (west);

Comparing the empirical values, 0.17 (north) and 0.16 (south) for the main approaches and 0.42 (east) and 0.59 (west) for the secondary approaches, with the modelled, the results are most encouraging. The proneness for the primary compared to the secondary were consistently lower as was the case for the direct

Fig. 8. For each arm of the junction: (a) occurrence of RLR per unit flow and (b) the change in RLR occurrence relative to the base case.

measurement and by a similar magnitude. For the main approaches compared to the measured there is a consistency of magnitude and order, for the secondary approaches whilst the magnitudes are similar, the values are interchanged for east and west approaches. Work is planned to explore this further.

The microsimulation has enabled parameters of violation other than proneness namely RLR rates and certainty to be quantified. It is clear that the different features become apparent in the data. Whilstfurther research will improve understanding,this work provides evidence that commonly used violation rates could cause inappropriate evaluation of the extent of RLR phenomenon.

8. Conclusions

This paper summarised themathematicalframework developed by the authors in previous research, to describe, in an analytically generic way, the phenomenon of red-light running for each arm of a signal controlled junction treating each turning movement (straight on, left and right turners) separately and added together. Thepaper suggests two components of cause of risk. Thefirst relates to the geometric design and the nature of the flow characteristic of the traffic arriving and moving through the junction and the second human's natural tendency to take risks, calculated or otherwise. Furthermore, a key finding was that the changes in opposing flow levels have the largest influence on the magnitude of the risk.

Based on earlier research the parameter k (proneness) was formulated to assess, in this work, the performance of a link (or lane, turning movement) for different junctions in a standard way. Proneness is expressed as a ratio of the actual to the potential redlight violations. Using a primary/secondary cross roads at a junction in Milan, as a case study, data on red light runners was collected by direct observation. Based on direct observation the results showed proneness k values statistically significantly lower, at the 95% confidence, for the primary compared to the secondary road stop-lines, with night time values slightly higher than during the day time.

A microsimulation model was set up for the same junction where actual measurements of violations were conducted. Red light running behaviour was simulated by setting up a closely spaced dummy signal on each arm of the junction and using the offset and a virtual stop-line as a mechanism to simulate red-light behaviour. On the whole the method developed to model RLR in microsimulation was found to be satisfactory. The results from the microsimulation were consistent with direct observation which was based on the theory proposed in early research, and showed much promise. Higher risk was measured on secondary arms and for left turners from primary into secondary roads, particularly when traffic conflict on opposing flows were high. Consistent with the empirical approach the arrival flow level was shown to have a first order effect.

Whilst this study has shown that the proposed theory has application irrespective of the distribution of approach flows, having considered a single isolated junction with a normal distribution, further research will explore the effect of arrivals in platoons created by co-ordination of signal control across a network and the change in flows over the day. Building on the initial work in this study the effect of systematic changes (−50% to +50%) on the base case flows, the impact on RLRs of changes in average annual daily flows will be considered, also by collecting RLR data from a disparate set of contrasting junctions. However, before substantial further work can be undertaken, a python code, to extract automatically from the simulation the number of vehicles related to potential RLR behaviours will be needed in order to reduce the post-processing time of the large amounts of data produced by microsimulation required to estimate proneness.

In a quest to deliver safer junctions, future work will delve deeper into the geometric characteristics of the junction, traffic conditions and signal control which were shown to significantly influence the red-light violation in the research reported here. It is anticipated that the resulting tool will allow the practitioner to identify vulnerable links and junctions in their network and test out their junction design (and re-design) to evaluate the proneness to RLR before construction.

Finally, microsimulation has been successfully used to explore the interactions between straight on and turning traffic during the period following the start of red and thus to estimate RLR proneness. This added value offers local authorities additional benefits justifying continued investment in validation of microsimulation at a time of budget constraints.

References

Allos, A.E., Al Haidithi, M.I., 1992. Driver behavior during onset of amber at signalized junctions. Traffic Engineering and Control 33 (5), 312–317.
Baguley, C.J., 1988. Running the red at signal on high speed roads. Traffic Engineering and Control 29 (7), 415–420.
Bonneson, J., Brewer, M., Zimmerman, K., 2001. Review and Evaluation of Factors that Affect the Frequency of Red Light-Running. FHWA/TX-02/4027-1. Federal Highway Administration, Washington, DC.
Bonneson, J., Zimmerman, K., Brewer, M., 2002a. Engineering Countermeasures to Reduce Red Light Running. FHWA-TX-03/4027-2. FHWA, U.S. Department of Transportation.
Bonneson, J., Zimmerman, K., Brewer, M., 2002b. Engineering Countermeasures to Reduce Red-Lightrunning. Texas Transportation Institute, The Texas A&M University System, Report 4027-2.
Chin, H.C., Quddus, M.A., 2003. Applying the random effect negative binomial model to examine traffic accident occurrence at signalizedintersections.AccidentAnalysis and Prevention, 253–259.
Council, F.M., Persaud, B.N., Lyon, C., Eccles, K., Griffith, M.S., 2005. Safety Evaluation of Red-light Camera. Publication FHWA-HRT-05-48. FHWA, U.S. Department of Transportation.
Elmitiny, N., Yan, X., Radwan, E., Russo, C., Nashara, D., 2010. Classification analysis of driver's stop/go decision and red-light running violation. Accident Analysis and Prevention, Elsevier 42, 101–111.
Giuffrè, T., Rinelli, S., 2004. Operational conditions effects on safety of urban signalized intersections – main experimental evidences. In: 2nd International SIIV Congress, Florence, Italy, p. 21.
Giuffrè, T., Rinelli, S., 2005. Illegal behaviors at signalized intersection. A model to evaluate potential conflicts. Le Strade (1), 105–109 (in italian).
HCM (Highway Capacity Manual), 2000. Transportation Research Board. National Research Council, Washington, DC.
Hill, S.E., Lindley, J.K., 2004. Red Light Running Prediction and Analysis. Transport Research Board, Annual Meeting, Washington, DC.
http://vvww.aimsun.com (01/06/2011).
Kent, S., Corben, B., Fildes, B., Dyte, D., 1995. Red Light Running Behaviour at Red Light Camera and Control Intersections. Monash University Accident Research Centre, Report #73.
Milazzo, J.S., Hummer, J.E., Prothe, L.M., 2001. A Recommended Policy of Automated Electronic Traffic Enforcement of Red Light RunningViolations in North Carolina. FinalReport.Institute for TransportationResearchandEducation,NorthCarolina State University, Raleigh, NC.
Mohamedshah, Y.M., Chen, L.W., Council, F.M., 2000. Association of Selected Intersection Factors with Red-Light Running Crashes. Publication FHWA RD-00-112. FHWA, U.S. Department of Transportation.
Persaud, B.N., Council, F.M., Lyon, C., Eccles, K., Griffith, M.S., 2005. A multijurisdictional safety evaluationof red light cameras.In: Presented at 83rdAnnual Meeting of Transportation Research Board, Washington, DC.
Porter, B.E., England, K.J., 2000. Predicting red-light running behavior: a traffic safety study in three urban settings. Journal of Safety Research 31 (1), 1–8.
Retting, R.A., Williams, A.F., 1996. Characteristics of red light violators: results of a field investigation. Journal of Safety Research 27.1, 9–15.
Retting, R.A., Ulmer, R.G., Williams, A.F., 1999. Prevalence and characteristics of red light running crashes in the United States. Accident Analysis and Prevention 31, 687–694.
Van der Horst, R., Wilmink, A., 1986. Drivers' decision-making at signalized intersections: an optimization of the yellow timing. Traffic Engineering and Control 27 (12), 615–622.
Wissinger, L.M., 2000. Issues surrounding the operation and installation of red light cameras. Master's thesis. N.C. State University.

bellnovelapplicationredlight2012.md 60 KB Cronologia Originale