G Model AAP 3769 No. of Pages 11

Accident Analysis and Prevention xxx (2015) xxx–xxx

Contents lists available at ScienceDirect

Accident Analysis and Prevention journal homepage: www.elsevier.com/locate/aap

Quantifying the yellow signal driver behavior based on naturalistic data from digital enforcement cameras Bar-Gera H. a, * , Musicant O. b , Schechtman E. a , Ze’evi T a a b

Department of Industrial Engineering and Management, Ben-Gurion University of the Negev, P.O.B. 653, Beer-Sheva 84105, Israel Department of Industrial Engineering and Management, Ariel University, Ariel 40700, Israel

A R T I C L E I N F O

A B S T R A C T

Article history: Received 1 November 2014 Received in revised form 5 March 2015 Accepted 30 March 2015 Available online xxx

The yellow signal driver behavior, reflecting the dilemma zone behavior, is analyzed using naturalistic data from digital enforcement cameras. The key variable in the analysis is the entrance time after the yellow onset, and its distribution. This distribution can assist in determining two critical outcomes: the safety outcome related to red-light-running angle accidents, and the efficiency outcome. The connection to other approaches for evaluating the yellow signal driver behavior is also discussed. The dataset was obtained from 37 digital enforcement cameras at non-urban signalized intersections in Israel, over a period of nearly two years. The data contain more than 200 million vehicle entrances, of which 2.3% (5 million vehicles) entered the intersection during the yellow phase. In all non-urban signalized intersections in Israel the green phase ends with 3 s of flashing green, followed by 3 s of yellow. In most non-urban signalized roads in Israel the posted speed limit is 90 km/h. Our analysis focuses on crossings during the yellow phase and the first 1.5 s of the red phase. The analysis method consists of two stages. In the first stage we tested whether the frequency of crossings is constant at the beginning of the yellow phase. We found that the pattern was stable (i.e., the frequencies were constant) at 18 intersections, nearly stable at 13 intersections and unstable at 6 intersections. In addition to the 6 intersections with unstable patterns, two other outlying intersections were excluded from subsequent analysis. Logistic regression models were fitted for each of the remaining 29 intersection. We examined both standard (exponential) logistic regression and four parameters logistic regression. The results show a clear advantage for the former. The estimated parameters show that the time when the frequency of crossing reduces to half ranges from1.7 to 2.3 s after yellow onset. The duration of the reduction of the relative frequency from 0.9 to 0.1 ranged from 1.9 to 2.9 s. ã 2015 Elsevier Ltd. All rights reserved.

Keywords: Dilemma Zone Traffic Signals Safety Digital Enforcement Cameras Yellow Onset Crossing Probability

1. Introduction Driver behavior during the yellow signal phase influences both the safety and the efficiency of signalized intersections. The main two safety outcomes are: (i) the impact on angle accidents as a result of red-light-running1 ; and (ii) the impact on rear-end accidents as a result of conflicting decisions by consecutive drivers. Efficiency refers to traffic throughput (or capacity) of the intersection, which is influenced by the changes in traffic flow during the yellow phase. Red-light-running accidents are a major concern. In 2012, 683 people were killed and an estimated 133,000 were injured in accidents that involved red light running in the US (IIHS 2015).

* Corresponding author. E-mail address: [email protected] (H. Bar-Gera). 1 With permissive left-turns red-light-running may lead to left-turn head-on accidents as well, but in Israel all signalized left-turns are protected.

Rear-end accidents at signalized accidents require attention as well. Devlin et al. (2011, Fig. 4) report that during 2004–2009 in West Australia 20% of serious casualty accidents at signalized intersections were rear-end, compared with only 4% at signcontrolled intersections. Characterization of driver behavior during the yellow phase can be done in several related but not completely equivalent ways. One way that has been explored fairly extensively in the literature is by estimating the conditional probability that a driver will cross given the state of the vehicle at yellow onset, which is described primarily by two key factors: current speed, and distance from the stop line. Another way that has been considered less often and less thoroughly (e.g. Köll et al., 2004) is by the distribution of entrance times after the yellow onset. The main contribution of this paper is a demonstration of the possibility to model statistically the distribution of entrance time during the yellow phase as well as at the beginning of the red phase, from information collected by loop detectors at signalized

G Model AAP 3769 No. of Pages 11

2

H. Bar-Gera et al. / Accident Analysis and Prevention xxx (2015) xxx–xxx

intersections. In this particular case the loop detectors are part of a digital enforcement system. This system also includes on-site digital camera and a link to a control center that maintains a log of the intersection entrance data. In principle similar data can be obtained from loop detectors installed for operational (or other) purposes, an opportunity that will be further discussed in Section 7. A secondary contribution of this paper (see Section 6) is a conceptual discussion on the conditions under which both characterizations of driver behavior during the yellow phase (conditional crossing probability and entrance time distribution) contain equivalent information. When such conditions are satisfied, the advantage of the proposed methodology is mainly the ability to collect large sample sizes at lower costs. Otherwise, the two approaches are complementary, and the importance of the proposed method is even more acute. Red light cameras have been in use for nearly 50 years, traditionally using wet film technologies, while the first digital enforcement camera was introduced in 1997 (GATSO, 2014). Previous studies of red light cameras (e.g. Miller et al., 2006; Shin and Washington, 2007; Retting et al., 2008; Erke, 2009) focused primarily on the safety benefits of camera installations. To the best of our knowledge, the potential of datasets generated by enforcement cameras, and especially by digital systems, was not thoroughly explored. Digital red-light cameras typically have direct communication to a control center, thus allowing storage of data about each passing vehicle. One such data element, stored by some digital red light camera systems, is the exact stop line crossing time after the yellow onset. Our proposed methodology for quantifying yellow behavior relies on this data element. The remainder of the paper is organized as follows. Section 2 presents background on yellow signal driver behavior studies. Section 3 describes the methodology and defines the notation used in our analyses. Section 4 presents the dataset. Section 5 presents the results. Section 6 describes a conceptual framework for evaluating the conditions under which the distribution of entrance times contains equivalent information to the conditional probability of crossing. Conclusions and opportunities for future research are discussed in Section 7. 2. Background Studies of yellow behavior commonly focus on three main variables of analysis: two state variables – speed and distance from the stop line at yellow onset, and a binary variable representing the decision of the driver to cross or to stop. It is also quite common to consider the ratio between distance and speed at yellow onset (with appropriate conversion of units). This ratio can be interpreted as extrapolated entrance time (EET), especially if drivers maintain (nearly) constant speed, as it was at yellow onset. (Previous studies used various alternative terminologies for EET, as indicated hereon.) Because EET is determined at yellow onset (it is derived from the two original state variables), it is not affected by driver’s response to the signal transition which comes later. In that sense its use as explanatory variable for driver decisions is as appropriate as the use of speed and distance. A brief overview of the literature on yellow behavior is presented next, covering three topics: Examples of simulator and other lab-type studies (Subsection 2.1); Examples of naturalistic studies (Subsection 2.2); and a discussion about the need for large sample sizes (Subsection 2.3), which is one of the key advantages of the methodology proposed here. 2.1. Simulator and experimental studies Driver behaviors at yellow onset have been studied extensively in simulator and lab-type studies. Haque et al. (2013) used a

driving simulator to study the effect of cell-phone usage on yellow behavior of 58 drivers at urban intersections with 72 km/h (45 mph) speed limit. The information collected for each participant involved 5–6 scenarios with different EETs (denoted there as TSL) and phone usage (Yes/No). Yellow duration was a continuous variable between 2.46 and 3.85 s. The simulator program was designed so that the switch from green to yellow will occur when the EET is equal to the yellow duration, thus requiring drivers to cope with a situation of dilemma. The researchers analyzed the data with classification trees and with a repeated measures logistic regression model. The explanatory variables in the latter model were EET, age, gender and cell-phone usage. All variables had first order significant effect, some of their interactions were also significant. The odds ratio for EET was 0.1037 (CI = [0.0145, 0.7403]), implying that larger EET results in smaller likelihood to cross. Drivers using the phone were much less likely to cross (odds ratio = 0.0005, CI = [1.1E-6, 0.2372]). Yet this effect was relaxed (interaction) as the EET was larger (odds ratio = 10.616, CI = [1.55, 72.88]). Bar-Gera et al. (2013) studied the effect of an in-vehicle stopping decisions advisory system. Twenty drivers participated in the driving simulator experiment. Each participant drove through 28 interurban traffic lights, 14 with and 14 without the evaluated system. The key parameter in the experiment was EET (denoted there as AOT), set at seven values, from 2 to 5 s at 0.5 s increments. Designated speed limit was 90 km/h, and yellow duration was 4 s. The probability of crossing was evaluated as a function of EET by logistic regression. The time range between 0.9 crossing probability and 0.1 crossing probability was 1.03 s in the base scenario, and 0.31 s with the advisory system. This means that the proposed advisory system is effective in reducing driver behavior uncertainty, and thus can potentially improve safety. Rakha et al. (2007) studied perception-reaction times and the probability to stop following yellow onset using an empty 2.1 km test track. Sixty drivers were instructed to drive an instrumented car at 72 km/h (45 mph), passing through traffic lights 24 times. 4 s yellow was initiated when the car was at 32 m, 55 m, 66 m, 88 m, or 111 m from the stop line. (At 72 km/h these correspond to EET values of 1.6–5.6 s.) Crossing probabilities were 0.91, 0.41, 0.27, 0.01 and 0.0 respectively. This study was conducted with real vehicles, yet its setup was artificial, so in many ways it is closer to a designed experiment than to a naturalistic study. We note that using lab/simulator studies in traffic safety research has advantages as well as limitations. Primary advantages of driving simulator studies are: (i) the ability to obtain clear identification of crossing decisions, precise representation of the vehicle state at yellow onset, as well as many other measures of interest; and (ii) the ability to design the distributions of various attributes. Key limitations of simulator studies are: (i) driver behavior is not necessarily exactly the same as in reality; (ii) sample size is usually modest (typical samples are of up to 100 drivers and up to 1000 yellow crossing events); and (iii) since conditions are created to explore drivers’ decisions, outcome frequencies (e.g. RLR crossings) cannot be identified. 2.2. Naturalistic studies Köll et al. (2004) studied the impact of flashing green on yellow behavior by collecting naturalistic data at 10 intersections over 5000 cycles. Yellow durations were 3 s at all intersections, except for one intersection with 4 s yellow. Peak volumes ranged from 400 to 2500 veh/h. Speed limits ranged from 40 to 70 km/h. Video records covering 85 m upstream of the signal were analyzed by AutoScope to identify vehicles and determine their speeds. Additional software was used to trace the vehicles. In total, the researchers identified 1621 crossings during yellow and 133 during

Please cite this article in press as: Bar-Gera, H., et al., Quantifying the yellow signal driver behavior based on naturalistic data from digital enforcement cameras. Accid. Anal. Prev. (2015), http://dx.doi.org/10.1016/j.aap.2015.03.040

G Model AAP 3769 No. of Pages 11

H. Bar-Gera et al. / Accident Analysis and Prevention xxx (2015) xxx–xxx

red. In Austria (with flashing green) there were 271, 204, and 79 crossings during the first, second, and third second of the yellow, and 23 red crossings. In Munich, Germany (without flashing green) there were 186, 201, and 210 crossings2 during the first, second, and third second of the yellow, and 92 red crossings. In about half of the cycles (2661) a vehicle was observed at yellow onset. Within these relevant cycles, the probabilities for yellow crossing were 0.46 in Austria and 0.97 in Munich. Red crossing rates (per 100 relevant cycles) were 1.90 in Austria and 14.94 in Germany. The range from 0.8 to 0.2 crossing probability was 2.2 s in Austria and 1.2 s in Munich. Gates et al. (2007) studied six intersections, with yellow durations ranging from 3.5 to 5.0 s, all-red durations ranging from 1.0 to 3.0 s, and speed limits ranging from 40 to 80 km/h (25 to 50 mph). Video cameras were used to determine the main attributes of 1001 relevant vehicles (either last-to-go or first-tostop). In each site, 4 to 14 h of video were collected, during which 90–250 relevant vehicles were observed. These data were used to analyze response times, deceleration rates, and crossing probabilities. Logistic regression analysis showed that EET is the best single predictor of crossing probabilities, with 77% correct predictions, compared to 82% correct predictions for a full model that included: EET, yellow duration, action of adjacent vehicle, vehicle type, presence of side-street traffic, and cycle length. Papaioannou (2007) studied a T intersection in Greece, with traffic volumes ranging from 50 to 500 veh/h. During 126 h of video, 6480 cycles were monitored, 18,271 vehicles were observed, of which the arrival of 2452 occurred while the green phase changed to yellow. Their average approach speed was 51.9 km/h. The research results (Papaioannou, 2007, Fig. 4) indicate behavior uncertainty for drivers positioned 20–70 m from the intersection at yellow onset. Hurwitz et al. (2011) examined driver behavior at 10 approaches of 5 intersections with yellow durations of 3.5 or 4 s, all-red times of 2 or 3 s, speed limits ranging from 72 to 88 km/h (45 to 55 mph), and average daily traffic ranging from 4000 to 9000 vehicles. Data was collected over 48 to 72 h. Pneumatic detectors 122 m (400 ft) upstream of the stop line were used to measure speeds accurately (Hurwitz, 2014). Video cameras were used to determine stop/cross decisions as well as vehicle positions at yellow onset on a 15 m (50 ft) grid. This setup was used to obtain “a sample size of approximately 1900 vehicles that experienced an incursion with the change interval while approaching one of the signalized intersections.” The number of observations per approach ranged from 40 to 200. The authors estimated for each intersection the crossing probabilities as a function of the distance to the intersection at yellow onset. These examples of naturalistic studies demonstrate their potential, as well as the practical challenges of data collection. One critical challenge is the complexity of setting up video equipment, and analyzing its output. This challenge limits sample sizes, and practical values are relatively modest, up to 1000–2000 yellow crossings. Another challenge is the precision of speed measurements. In the papers mentioned here estimates of speed accuracy from video records were not provided. One study used pneumatic detectors in addition to video in order to obtain more accurate speed estimates (Hurwitz, 2014). This approach has a limitation as well, because the pneumatic detectors are placed at a specific fixed location,

3

and therefore the speed of the vehicle is known when crossing the detectors, rather than at yellow onset. 2.3. The need for a large sample size Sample sizes of previous studies were sufficient to obtain many meaningful results, such as measures of “width” for the yellow behavior uncertainty. Yet there are important questions that require larger sample sizes by orders of magnitude. For example: 1. What functional form is the most appropriate one for predicting

crossing probability? Should it be four-parameters logistic or standard (exponential) logistic? Is there symmetry between the transition from nearly certain crossing and the transition to nearly certain stopping? 2. How do estimated parameters change when the data is divided into categories such as weekdays vs. weekends, daytime vs. nighttime, peak vs. off-peak, by weather conditions, variations over time, etc. To the best of our knowledge, to date there have not been any studies that evaluated 100,000 yellow crossings or more, which is needed in order to address questions of the type mentioned above. In other contexts, loop detectors have been used to collect extensive traffic data, with much larger sample sizes (e.g. PEMS, 2004). In this research we demonstrate how loop detector data can be used also for quantifying driver behavior at yellow signals. 3. Methodology The actual entrance time (AET) of vehicle i (time elapsed from yellow onset until entrance) is denoted by ti. Given a chosen time resolution, dt (0.01 s in our actual data), the frequencies of entrances with AET in the interval [t,t + dt) are defined by: yðtÞ ¼ jfi : t T red (9) Ci ¼ TRUE ti  T red Given a chosen time resolution, dt (0.01 s in our data), the frequency of arrivals with EET in the interval [t, t + dt) can be defined by: ye ðtÞ ¼ jfi : t  te;i < t þ dtgj

(10)

Similarly, the frequencies of arrivals which result in crossing with EET in the interval (t, t + dt) can be defined by: yec ðtÞ ¼ jfi : C i ¼ TRUE and t  te;i < t þ dtgj

(11)

6. Conceptual framework for comparing entrance time distribution with conditional crossing probability In Section 1 two primary alternatives for quantifying yellow signal driver behavior were mentioned: one by the conditional probability of stopping as a function of EET, and the other by the relative frequency of entrances as a function of the time after the yellow onset. These characterizations are related but not equivalent. We show here that under certain naïve assumptions the two are in fact equivalent, enabling the use of both approaches interchangeably. Otherwise, appropriate conversions may be needed for relaxing these assumptions. 6.1. The framework The definition of actual entrance time (AET), denoted by ti for vehicle i, was presented in Section 3, as well as entrance frequency, y(t), and its expected value mðtÞ. The primary variables describing a state of vehicle i at yellow onset are its speed, si, and its distance from the stop line, di (as discussed in Section 2). The ratio between them, te,i = di / si, is viewed as the extrapolated entrance time (EET).

Fig. 8. Day to night ratios: contribution to effective green time vs. the entrance frequency reduction duration (29 intersections).

Please cite this article in press as: Bar-Gera, H., et al., Quantifying the yellow signal driver behavior based on naturalistic data from digital enforcement cameras. Accid. Anal. Prev. (2015), http://dx.doi.org/10.1016/j.aap.2015.03.040

G Model AAP 3769 No. of Pages 11

H. Bar-Gera et al. / Accident Analysis and Prevention xxx (2015) xxx–xxx

9

Given ye ðtÞ ¼ k we have yec ðtÞ  Binomðk; PðtÞÞ, hence Eðyec ðtÞjye ðtÞ ¼ kÞ ¼ k  PðtÞ: A general fact in probability theory is that EðXÞ ¼ EðEðXjYÞÞ. Therefore, the connection between the crossing probability function and the expected frequencies is given by:

intersection4 , and the second is the effect of gaps between extrapolated entrance time and actual stop line crossing time on differences between the frequencies of crossings. We claim that the latter effect is in a sense “residual”. We will illustrate this via a specific (hypothetical) example where gaps between AET and EET exist, but they do not cause any differences between expected frequencies of crossings, i.e. mðtÞ ¼ mec ðtÞ, over a certain range of values of t. To present the example we need the following additional notations:

(16) vt;k ¼ j i : t  ti  t þ dt; t þ k  dt  te;i  t þ ðk þ 1Þ  dt j

mec ðtÞ ¼ Eðyec ðtÞÞ ¼ E½Eðyec ðtÞjye ðtÞÞ ¼ Eðye ðtÞ  PðtÞÞ ¼ PðtÞme ðtÞ

ct;k ¼ P t  ti  t þ dtjt þ k  dt  te;i  t þ ðk þ 1Þ  dt

We assume that both ye(t) and yec(t) are random variables, with expected values of me ðtÞ and mec ðtÞ respectively. The conditional crossing probability function, given extrapolated entrance time t, is: PðtÞ ¼ ProbðC ¼ TRUEjt  te;i < t þ dtÞ

(12)

(13)

We want to examine the connection between mðtÞ, me ðtÞ, and mec ðtÞ. In the considered range (t < Tred) all vehicles counted in y(t) crossed the intersection. Now consider the following assumptions: (i) P(0) = 1. In other words, a vehicle that reached the stop line

right at yellow onset will cross. m(t) = mec(t). In other words, the expected frequency of stopline-crossings when the actual entrance time is equal to t is the same as the expected frequency of stop-line-crossings when the extrapolated entrance time is equal to t. See additional discussion in Subsection 6.2. (iii) me(t) = l for all t. That is, the expected frequency of arrival times is constant for all t. This is the most problematic assumption of the three, as will be discussed in Subsection 6.2. (ii)

Based on these assumptions we define the relative expected frequency of stop-line-crossing times at time t as: Pr ðtÞ ¼ mðtÞ=mð0Þ



The key assumptions leading to mðtÞ ¼ mec ðtÞ are that ct;k is symmetric in k, and that mec ðtÞ is a linear function of t. It is natural for mec ðtÞ to be a decreasing function of t. To simplify the exposition we consider a specific (arbitrary) illustrative example where ct;k ¼ 0:05  0:0025jkj, regardless of t, for 0:5  t  2:5 and jkj  20; and mec ðtÞ ¼ mec ð0Þ  ð1  t=3Þ. Gaps between EET and AET in this case are as high as 20  dt, i.e. 0.2 s if dt = 0.01 s as is the case in our data, thus corresponding to 10% speed changes. We now show that in this case mec(t) = m(t) for 0.5  t  2.5.

mðtÞ ¼

20 X

mec ðt þ k  dtÞ  ct;k

k¼20

(14)

¼ mec ð0Þ

20  X

k¼20

6.2. Plausibility of assumptions The analysis in Subsection 6.1 is based on three quantitative assumptions: (i) P(0) = 1; (ii) mðtÞ ¼ mec ðtÞ; and (iii) me ðtÞ ¼ l for all t, and two qualitative assumptions: (iv) all relevant vehicles are free to choose to cross if they wish to do so; and (v) all vehicles that cross the stop line will enter the intersection and cross it. In most countries the transition from green to yellow is instantaneous. Under such circumstances the first assumption is almost trivial, as drivers who entered the intersection at yellow onset cannot stop. In Israel (the case here), at all non-urban intersections the green ends with 3 s of flashing green, followed by 3 s of yellow. It is therefore feasible for drivers to make the decision to stop even before yellow onset. As discussed in Section 3.1, we can examine whether entrance frequencies are stable during the beginning of the yellow phase. If they are stable, it can be interpreted as a supporting evidence for the conjecture that during the same interval crossing probabilities are stable, and possibly close to one. Assumption (ii) involves two issues. The first is whether a vehicle that crossed the stop line will necessarily enter the

1

t þ k  dt 3

  ð0:05  0:0025jkjÞ

  t ¼ mec ð0Þ 0:05  1  3  20  X t þ k  dt t  k  dt þ1 1  ð0:05  0:0025jkjÞ þ 3 3 k¼1

(15)

The first transition is by assumption (ii), the second transition is by assumption (i), and the third transition is by assumption (iii). Note that Pr(t) can be calculated from the data while P(t) cannot. The quality of Pr(t) as an approximation of the crossing probability function, P(t), is therefore dictated by the degree to which each of the three assumptions holds.

(17)

Note that: vt;k jðyec ðt þ k  dtÞ ¼ nÞ  Binomðn; ct;k Þ;     ¼ E E vt;k jyec ðt E vt;k jðyec ðt þ k  dtÞ ¼ n ¼ n  ct;k ;Eðvt;k Þ   þk  dtÞÞÞ ¼ E yec ðt þ k  dtÞ  ct;k ¼ mec ðt þ k  dtÞ  ct;k ; and P  P1 mðtÞ ¼ EðyðtÞÞ ¼ E 1 v m ðt þ k  d tÞ  c ¼ ec t;k k¼1 t;k k¼1

If the above three assumptions are satisfied, then: Pr ðtÞ ¼ mðtÞ=mð0Þ ¼ mec ðtÞ=mec ð0Þ ¼ mec ðtÞ=me ð0Þ ¼ mec ðtÞ=me ðtÞ ¼ PðtÞ



"

#    X  20 t t ¼ mec ð0Þ 0:05  1  2 1 þ  ð0:05  0:0025jkjÞ 3 3 k¼1 #   " 20 X t 2  ð0:05  0:0025jkjÞ  0:05 þ ¼ mec ð0Þ  1  3 k¼1   t ¼ mec ð0Þ  1  ¼ mec ðtÞ 3 Differences vanish in this case due to the combination of linearity of mec ðtÞ, the symmetry of the gap distribution (ct;k ) with respect to k, and its independence with respect to the EET value. In general, a similar mechanism may reduce part of the influence of gaps on differences in frequencies of crossings, leaving only “residual” differences. The question is when frequency differences can be

4

The digital camera system in Israel takes two pictures (1 s apart) in every event that a vehicle crossed the stop line 1 s or more into the red phase. Out of these documented events, 45% lead to citations. One of the reasons for not issuing a citation is if the pictures show that the vehicle stopped after the stop line, and did not enter the intersection. This is an indication that crossing the stop line while not entering the intersection is not a negligible phenomenon.

Please cite this article in press as: Bar-Gera, H., et al., Quantifying the yellow signal driver behavior based on naturalistic data from digital enforcement cameras. Accid. Anal. Prev. (2015), http://dx.doi.org/10.1016/j.aap.2015.03.040

G Model AAP 3769 No. of Pages 11

10

H. Bar-Gera et al. / Accident Analysis and Prevention xxx (2015) xxx–xxx

neglected, or alternatively whether some conversion can be used to obtain mec ðtÞ from m(t) and vice versa. These issues are beyond the scope of the present paper. The main concern is assumption (iii), of uniform extrapolated entrance times. Various factors may lead to deviations from this assumption, including: signal coordination, actuated or semiactuated traffic signal control, cycle failures, etc. Quantitative analysis of each of these effects, and development of methodologies to address them, remain a subject for future studies. While we cannot rule out the possibility of substantial deviations from the uniformity assumption, we can examine its plausibility by considering the stability of the frequencies of stopline-crossings during the beginning of the yellow phase. If the uniformity assumption is valid, we should expect stability of initial frequencies. Indeed such stability is not a proof that the assumption of uniformity holds, but it is a positive indicator that can be determined practically. Specific options to examine this stability were discussed in Subsection 3.1. Only intersections for which the stability assumption was found reasonable were included in the analyses. 7. Conclusions and future research This paper shows how loop detector data can be used to naturalistically quantify driver behavior following the onset of the yellow traffic signal phase. The dataset used was obtained from 37 digital enforcement cameras at non-urban signalized intersections in Israel, over a period of nearly two years. The data contain more than 200 million vehicle entrances, of which 2.3% (5 million vehicles) entered the intersection during the yellow phase. Piecewise linear regression was used to check whether the distribution of the entrance times is stable at the beginning of the yellow phase. The results show that in most intersections under study the frequencies of entrance time after the yellow onset are relatively stable during the beginning of the yellow phase. For these intersections logistic regression fits the data well, and the fit of exponential logistic regression is better than the fit of four parameters logistic regression. The duration of frequency reduction from 90% to 10% varies considerably across intersections, and ranges from 1.9 s to 2.9 s. Yet in most intersections the night-time behavior is rather similar to the day-time behavior. The methodology presented here opens many opportunities for future research. It can be used to examine the influence on driver behavior of changes in enforcement strategies. Examples for such studies include transferring cameras between active and nonactive installations, or evaluating the influence of new installations (assuming that the loops are installed before the camera). Furthermore, it can be used to examine causes of heterogeneity, both within and across locations, which may include geometric and traffic characteristics, as well as signal timing schemes. Another possible opportunity for future research is to examine the connection between crash data and driver behavior as quantified by the proposed method. It should be mentioned that the specific results presented here are not necessarily transferrable, due to the fact that at the intersections examined in this study the green phase ends with 3 s flashing green. Yet the underlying behavior could be quite similar, and therefore the methodology is likely to be applicable in other countries too. Comparisons of results from different places using this methodology can potentially lead to interesting insights. Possible improvements in the data collection configuration include: (1) identification of entrance frequencies towards the end of the green, and during the flashing green; and (2) record signal phase duration and cycle times as potential explanatory variables.

Another possibility for collecting distributions of vehicle entrance times after the yellow or red onset for off-line analysis may be by utilizing existing loop detectors connected to controllers at actuated and semi-actuated traffic signals. This might be a challenging direction, because existing controllers may or may not have the abilities to record individual vehicle entrance times, and because the location of the loop detectors may not necessarily be ideal for the purpose of identifying intersection entrance times. In addition, if such configuration is used to identify red-lightrunning, it is also necessary to verify by video or otherwise that the vehicles indeed crossed the intersection and did not stop after the detector. Precise entrance time records may be useful not only for scientific purposes, but also in order to address several other needs: 1. Site-specific determination of the necessary all-red duration; 2. Off-line signal timing performance evaluation; and 3. Accident investigations.

Perhaps the combined motivations may justify the necessary investment in controller technology. Acknowldgement This research was partially funded by the Israeli National Road Safety Authority. Appendix A. We claim that t 0.5 is a reasonable approximation for the yellow contribution to effective green time. This claim can be demonstrated in two ways: intuitive justification and formal derivation. Both types of arguments rely on the consideration of the asymptotic logistic behavior fitted to the data as representative of the flow during the last seconds of the green phase, i.e. the count per one second of green is K 0 ¼ K=dt. This is based on the assumption of stability during the beginning of the yellow phase. The intuitive argument is based only on the symmetry property of the logistic functional form around the inflection point t 0.5, i.e. the fact that mðt 0:5 þ DÞ þ mðt 0:5  DÞ ¼ K. The overall total number of crossings after the green, S, can be divided into two components: S1 – the number of crossings from t ¼ 0 until t ¼ 2t 0:5 , and S2 – the number of crossings from t ¼ 2t 0:5 þ dt onwards. Due to the exponential nature of the logistic regression, the tail in S2 is usually relatively small, and can be neglected. Therefore: X

2t 0:5=dt

S1 ¼

mðk  dtÞ ¼ K  t 0:5 =dt ¼ K 0  t 0:5

(18)

k¼0

and the resulting approximation for effective green contribution is geff ¼

S S1 ¼ t 0:5 K0 K0

(19)

To make a formal argument we switch to continuous time, and let yc ð½t1 ; t2 Þ be the count of crossings during a general interval [t1, t2]. The expected value is mc ð½t1 ; t2 Þ ¼ Eðyc ð½t1 ; t2 ÞÞ, and its density with respect to t is md ðtÞ ¼ K 0 =½1 þ expðBðt  MÞÞ. These definitions are consistent with the estimated discrete model because tþdt Z

mc ð½t; t þ dtÞ ¼

tþdt Z

md ðtÞ  dt ¼ t

K0 dt 1 þ expðBðt  MÞÞ

t

K0 K dt  ¼ ¼ mðtÞ 1 þ expðBðt  MÞÞ 1 þ expðBðt  MÞÞ

(20)

Please cite this article in press as: Bar-Gera, H., et al., Quantifying the yellow signal driver behavior based on naturalistic data from digital enforcement cameras. Accid. Anal. Prev. (2015), http://dx.doi.org/10.1016/j.aap.2015.03.040

G Model AAP 3769 No. of Pages 11

H. Bar-Gera et al. / Accident Analysis and Prevention xxx (2015) xxx–xxx

We want to find S ¼ mc ð½0; T Red Þ mc ð½0; 1Þ. For this purpose, consider the general integral Zt2

mc ð½t1 ; t2 Þ ¼

Zt2

md ðtÞ  dt ¼ t1

K0 dt 1 þ expðBð1  MÞÞ

(21)

t1

This integral can be computed by substituting h ¼ expðBðt  MÞÞ, as follows: dh ¼ B  expðBðt  MÞÞ  dt

(22)

dh ¼ dt Bh

(23) Zh2

mc ð½t1 ; t2 Þ ¼ h1

lnð1 þ hÞhh21

 Zh2  K 0 dh K0 1 1 K0  ¼  dh ¼  ½lnðhÞ h 1þh 1 þ hB  h B B h1

    K0 h2 1 þ h1 þ ln ¼  ln B 1 þ h2 h1

(24)

where h1 ¼ expðBðt1  MÞÞ and h2 ¼ expðBðt2  MÞÞ. Substituting t1 ¼ 0 and t2 ! 1, i.e. h1 ¼ expðBMÞ and h2 ! 1, we get:   h2 (25) ¼ lnð1Þ ¼ 0 lim In 1 þ h2 h2 !1

S ¼ mc ð½0; 1Þ ¼ ¼

    K0 1 þ h1 K0 1  ln ¼  ln þ1 h1 B h1 B

K0  lnðexpðBMÞ þ 1Þ B

(26)

By Taylor’s first order expansion lnðexpðBMÞ þ 1Þ lnðexpðBMÞÞ þ expðBMÞ ¼ BM þ expðBMÞ 1 K0 0   ðBM þ expðBMÞÞ K0 K B expðBMÞ M ¼Mþ B

geff ¼

(27)

mc ð½0; 1Þ

11

This derivation also provides a first order estimate for the precision of the approximation of geff by M ¼ t 0:5. For example, typical values in our results are M ¼ B ¼ 2. For these values the first order correction term is expðBMÞ=B 0:01s. References Devlin, A., Candappa, N., Corben, B., Logan, D., 2011. Designing safer roads to accommodate driver error. Curtin-Monash Accident Research Center. Project 09-006RSC. Erke, A., 2009. Red light for red-light cameras?: a meta-analysis of the effects of red-light cameras on crashes. Accid. Anal. Prev. 31 (5), 897–905. Gates, T.J., Noyce, D.A., Laracuente, L., Nordheim, E.V., 2007. Analysis of driver behavior in dilemma zones at signalized intersections. Transp. Res. Rec. J. Transp. Res. Board 2030 (1), 29–39. GATSO, 2014, GATSO MILLIA, http://www.gatso.com/upload/ 126904133751c0452706a09.pdf, (accessed 31.10.2014). Md. Mazharul Haque, Amanda D. Ohlhauser, Simon Washington, Linda Ng Boyle, 2013. Examination of distracted driving and yellow light running: analysis of simulator data. Transp. Res. Board 13–1710. HCM, 2010. The Highway Capacity Manual. Transportation Research Board, Washington. Hurwitz, D.S., 2014. Personal Communication . Hurwitz, D.S., Knodler, M. A, M.A., Nyquist, B., 2011. Evaluation of driver behavior in type II dilemma zones at high-speed signalized intersections. ASCE J. Transp. Eng. 137 (4), 277–286. IIHS, 2015. Insurance Institute for Highway Safety. Q&A: Red Light Cameras. http:// www.iihs.org/iihs/topics/t/red-light-running/qanda#red-light-running (accessed 04.03.2015). Köll, H., Bader, M., Axhausen, K.W., 2004. Driver behavior during flashing green before amber: a comparative study. Accid. Anal. Prev. 36 (2), 273–280. Miller, J.S., Khandelwal, R., Garber, J.N., 2006. Safety impacts of photo-red enforcement at suburban signalized intersections an empirical Bayes approach. Transp. Res. Rec. J. Transp. Res. Board 1853, 27–34. Papaioannou, P., 2007. Driver behavior, dilemma zone and safety effects at urban signalized intersections in Greece. Accid. Anal. Prev. 39 (1), 147–158. PEMS, 2004, Caltrans Performance Measurement System, http://pems.dot.ca.gov/ (accessed 31.10.2014). Rakha, H., El-Shawarby, I., Setti, J.R., 2007. Characterizing driver behavior on signalized intersection approaches at the onset of a yellow-phase trigger. IEEE Trans. Intell. Transp. Syst. 8 (4), 630–640. Retting, R.A., Ferguson, S.A., Farmer, C.M., 2008. Reducing red light running through longer yellow signal timing and red light camera enforcement: results of a field investigation. Accid. Anal. Prev. 40 (1), 327–333. Shin, K., Washington, S., 2007. The impact of red light cameras on safety in Arizona. Accid. Anal. Prev. 39 (1), 1212–1221.

(28)

Please cite this article in press as: Bar-Gera, H., et al., Quantifying the yellow signal driver behavior based on naturalistic data from digital enforcement cameras. Accid. Anal. Prev. (2015), http://dx.doi.org/10.1016/j.aap.2015.03.040

Quantifying the yellow signal driver behavior based on naturalistic data from digital enforcement cameras.

The yellow signal driver behavior, reflecting the dilemma zone behavior, is analyzed using naturalistic data from digital enforcement cameras. The key...
2MB Sizes 0 Downloads 4 Views