We need more pedestrian crashes!

By Maryam Hasanpour, Ph.D. Candidate 

Maryam Hasanpour is a Ph.D. candidate in Transportation Engineering at Toronto Metropolitan University (formerly Ryerson University). Her research interest is traffic safety, focusing on conflict-based analysis using new and innovative machine learning methods. In addition to being a CARSP and ITE Canada member, she is part of the ITE Canada National Technical Program Committee and Events Coordinator of the CARSP Young Professionals' Committee.

Crashes are rare, so gathering enough data for pedestrian safety analysis is challenging. To overcome this issue, surrogate measures of safety have been proposed, including microsimulation and video analytics. Microsimulation is appropriate for evaluating the safety and operational measures for scenarios that may be contemplated but do not exist, such as different scenarios that might be implemented in the optimum development of Leading Pedestrian Intervals (LPIs). An LPI allows pedestrians to enter and establish their presence in the crosswalk at a signalized intersection 3-7 seconds before vehicles are given a green indication.  It turns out that this small change can make a significant positive impact on safety, as documented in past studies. But can LPI always make streets safer for pedestrians without causing significant vehicular traffic delay or congestion? In the first part of our study (1), we addressed this fundamental question using microsimulation. We found out that we can optimize LPI to achieve the maximum safety benefit for pedestrians at a satisfactory operational Level of Service for vehicles. Then, we discussed that using microsimulation may not always be the best approach. In order to examine the relationship between pedestrian conflicts and crashes more precisely, we can use video-driven traffic conflict. In past studies, however, many researchers focused solely on frequency indicator, possibly because of the complexity associated with jointly estimating both frequency and severity indicators. In our on-going research study, we aim to fill this gap by creating a new way to integrate conflict indicators by taking into account both how often they happen and how severe they might be. This may lead to a significant leap toward understanding how to improve pedestrian safety.

Researchers who work on pedestrian safety analysis often lament: "We need more pedestrian crashes!" Wait, what? Don't worry, dear readers – there is a twist to this seemingly puzzling statement. In this article, we journey into the world of pedestrian safety to explain!  

Crashes are (fortunately) rare, so gathering enough data for pedestrian safety studies can take a long time. As Elvik (2) mentioned, "If a sample is very small and/or has a very low mean number of accidents, it is just impossible to fit an accident model to it". But fear not, as researchers have responded with innovative solutions and introduced an approach that flips the script on pedestrian safety. Surrogate measures of safety, derived from microsimulation and video analytics, provide alternative ways to measure pedestrian safety. Let's embark on a journey to see how. 

Imagine if we could analyze vehicle-pedestrian interactions on a computer before they happen in real life. That is the essence of microsimulation which can be used to assess the frequency and severity of "near misses" or "near collisions", also known as "traffic conflicts". With this high-tech trick, we can figure out how to make crosswalks safer for people without causing traffic congestion or delay. In addition, microsimulation is most appropriate for evaluating the safety and operational measures for scenarios that may be contemplated, but do not exist such as the effects of varying site characteristics like the length of the leading pedestrian interval. Leading Pedestrian Intervals (LPI) is one of such countermeasures that has emerged in recent years to achieve safety equitability for pedestrians at signalized intersections. LPIs give pedestrians a little head start at crossings; pedestrians get 3-7 seconds before vehicles are given the green light, which makes them more visible to drivers and reduces their exposure time to left or right-turning vehicles (3). It turns out that this small change can make a significant impact on safety based on completed research evaluation studies. 

For example, Fayish and Gross (4) quantified the safety effects of LPI for ten signalized intersections in State College, Pennsylvania. A reduction of 58.7% was estimated for vehicle-pedestrian crashes. Although traffic and pedestrian volumes varied significantly from site to site and over the course of the day, all of the intersections had the same LPI interval of 3 seconds at all times. In addition to the safety evaluation, an economic analysis compared the mean comprehensive cost of vehicle-pedestrian crashes with the cost of implementing the LPI and determined that LPI was highly cost-effective. Another crash-based study was conducted by Goughnour et al. (5), who used data from 56 treated intersections in Chicago, 42 treated sites in New York City, and 7 treated sites in Charlotte, North Carolina to evaluate the safety effect of LPI installations. The crash modification factor (CMF) for vehicle-pedestrian crashes for all cities combined was 0.87 (a reduction of 13%). This study did not use a homogenous dataset such as New York City entirely prohibited right-turn-on-red (RTOR) at treated sites, while Chicago allowed this movement in most cases. Prohibiting or allowing RTOR has a significant impact on pedestrian exposure and conflicts, as noted below, thereby it should be accounted for in pedestrian safety analysis.

However, as mentioned, it is not always feasible to work with crash data. Therefore, several studies evaluated the safety impact of LPI installations based on traffic conflicts as crash surrogates. These included Hubbard et al. (6), who evaluated the safety impacts of LPI installations using traffic conflict data derived from the recorded video at suburban intersections. Their results suggest that LPI could not improve pedestrian safety in the suburban environment without prohibiting RTOR. They observed that pedestrian crossings conflicting with right-turn vehicles during the walk interval increased after implementing LPI. Moreover, they found, perhaps logically, that the right-turn volume is a key indicator of the safety impact of an LPI installation, so much so that restricting RTOR with LPI implementation was proposed to improve the safety efficiency of LPI. In another study based on video-derived traffic conflict data, Guo et al. (7) evaluated the safety effect of LPI installations and their results indicated a reduction of between 18.1% and 20.9% in severe vehicle-pedestrian conflicts. 

But here's the catch – can we always make streets safer for pedestrians at crosswalks without causing traffic delay or congestion for vehicles? It is like finding the perfect recipe for a delicious and balanced meal. Think of LPI as a treatment that may work best in certain situations, like crowded pedestrian intersections or specific times of the day. So, in essence, the question is: Where, when, and under what circumstances can LPIs work most effectively? In our recent study (1), we addressed this fundamental question by reviewing the relevant literature and then presenting the research from the application of microsimulation to fifteen Toronto intersections where LPIs of 5 seconds have been implemented. Different hypothetical scenarios were defined based on the indications from previous studies to examine the potential influence of factors such as turning volumes, crossing width, length of the LPI interval, pedestrian volumes, and whether or not RTOR is allowed. The microsimulation involved using a recently released module for accommodating LPI phasing in the PTV Vistro software. PTV Vistro is a traffic analysis software that provides operational analysis of the network. In our study, the vehicular delay incurred due to an LPI installation was measured by creating a layout in PTV Vistro containing all relevant parameters, including vehicles, pedestrians, and signal timing. Using this software, the Level of Service (LOS), vehicle delay, and maximum queue length can be calculated for each scenario. LOS indicates the quality of traffic operations at an intersection, ranging from LOS A to LOS F, with LOS A indicating little or no traffic delays and LOS F indicating poor operation with very long delays. Fortunately, the PTV Vistro implementation of LPIs requires no additional signal phases to model exact LPI operations for the traffic controller, thereby facilitating LPI application in the software. In Vistro, LPI can be mimicked by utilizing the “Delayed Vehicle Green” parameter (8).

Our study found that using LPIs can make crossing the street safer for pedestrians while still keeping traffic flowing at a satisfactory level of service of C or D. But it's not a one-size-fits-all solution. The effects of LPI depend on various factors, ranging from individual behaviours such as perception and reaction time to how many cars are turning, or how long people have to walk. Statistical models were then developed to estimate the effects of LPI installations on vehicle-pedestrian conflicts after controlling for pedestrian and turning vehicle volumes. These models indicated about 50% reductions in severe conflicts. Considering the most recent crash-based CMF estimate of 0.87 by Goughnour et al., 2021(5) these numbers would imply that a 10% reduction in pedestrian conflicts would be associated with a 2% reduction in pedestrian crashes. Although there are no available crash prediction models for vehicle-pedestrian conflicts, there can be some assurance that these effects are reasonably consistent with indications from crash prediction models currently available for vehicle-vehicle conflicts (9,10). The practical application benefits of the research lie in its contribution to evaluating the economic and operational impacts of LPI installations. Specifically, economic impacts can be optimized by identifying practical application circumstances to achieve the maximum safety benefit for pedestrians at a satisfactory operational LOS for vehicles. 

So what's next? 

The first part of this research study helped us figure out the best ways to use LPIs and ensure they help both pedestrians and drivers. But wait, there is more to discover. This study is just the beginning. The researchers want to examine the relationship between pedestrian conflicts and crashes more precisely. The first problem in drawing a relationship between conflict and crashes is that microsimulation may not be precise, especially when pedestrians are involved; This is because these simulations require careful adjustment, calibration, and validation to be reliable. Even when we do a good job checking them, there is still a possibility that the simulated pedestrians and vehicles behaviours in the computer do not exactly match how they behave in the real world. So, to get a more precise picture, it is better to look at real video footage of near-misses instead of relying solely on computer simulations. 

There are two main ways to extrapolate crashes using traffic conflict data. One way is to find patterns and connections between near-misses and crashes. The other way is to see whether near-misses actually lead to crashes or serve as alternatives to crashes. The first method aims to create a statistical link between the number of conflicts and actual crashes, while the second approach views a near-crash as a situation that occurs when a crash is narrowly avoided. Consequently, both crashes and near-crashes can be represented as alternative results within a binary choice model (11). While each method has its strengths and weaknesses, the first approach (e.g., finding a link between near-misses and crashes) for estimating statistical models that relate conflicts to crashes can be more logically appealing. One great advantage of these models is that conflicts can capture many different factors that contribute to crashes, thereby avoiding model-fitting issues that we normally face when accounting for them while working directly with crash data, such as model-fitting issues rooted in excessive zeros and imbalanced crash observations.

To understand the relationship between conflicts and crashes, we need some specific safety measures. Two of the most important and logical measures in conflict-based studies are conflict frequency and conflict severity. Now, here is the interesting part. To date, most researchers have focused on just one safety indicator, possibly because of the complexity associated with jointly estimating both conflict frequency and severity indicators. In our on-going research, we aim to contribute to filling this gap by creating a new way to combine near-miss data, taking into account both how often they happen and how severe they might be. To do this, we are using a machine learning model that learns from data. In the process, we are "learning", interestingly enough, from methods used in the aviation safety field, where crashes are even rarer (12,13).

Stay tuned for the results!

References

1. Hasanpour M, Persaud B. Using microsimulation to investigate the optimal deployment of leading pedestrian intervals at signalized intersections. TSR. 2022 Dec 27;3:000022. 
2. Elvik R. Assessing causality in multivariate accident models. Accident Analysis & Prevention. 2011 Jan;43(1):253–64. 
3. Saneinejad S, Lo J. Leading Pedestrian Interval: Assessment and Implementation Guidelines. Transportation Research Record. 2015 Jan;2519(1):85–94. 
4. Fayish AC, Gross F. Safety Effectiveness of Leading Pedestrian Intervals Evaluated by a Before–After Study with Comparison Groups. Transportation Research Record. 2010 Jan;2198(1):15–22. 
5. Goughnour E, Carter D, Lyon C, Persaud B, Lan B, Chun P, et al. Evaluation of Protected Left-Turn Phasing and Leading Pedestrian Intervals Effects on Pedestrian Safety. Transportation Research Record. 2021 Nov;2675(11):1219–28. 
6. Hubbard SML, Bullock DM, Thai JH. Trial Implementation of a Leading Pedestrian Interval: Lessons Learned. Institute of Transportation Engineers (ITE) Journalurnal. 2008 Oct;78(10):pp 32, 37-41. 
7. Guo Y, Sayed T, Zheng L. A hierarchical bayesian peak over threshold approach for conflict-based before-after safety evaluation of leading pedestrian intervals. Accident Analysis & Prevention. 2020 Nov;147:105772. 
8. PTV Vistro embraces pedestrian safety goals with the new Leading Pedestrian Interval feature. [cited 2021 May 17]; Available from: https://www.ptvgroup.com/en/solutions/products/ptv-vistro/knowledge-base/feature-learning-center/leading-pedestrian-interval/
9. Saleem T, Persaud B, Shalaby A, Ariza A. Can Microsimulation be used to Estimate Intersection Safety?: Case Studies Using VISSIM and Paramics. Transportation Research Record. 2014 Jan;2432(1):142–8. 
10. Peesapati LN, Hunter MP, Rodgers MO. Can post encroachment time substitute intersection characteristics in crash prediction models? Journal of Safety Research. 2018 Sep;66:205–11. 
11. Jovanis PP, Aguero-Valverde J, Wu KF, Shankar V. Analysis of Naturalistic Driving Event Data: Omitted-Variable Bias and Multilevel Modeling Approaches. Transportation Research Record. 2011 Jan;2236(1):49–57. 
12. Basora L, Bry P, Olive X, Freeman F. Aircraft Fleet Health Monitoring with Anomaly Detection Techniques. Aerospace. 2021 Apr 7;8(4):103. 
13. Olive X, Basora L. Detection and identification of significant events in historical aircraft trajectory data. Transportation Research Part C: Emerging Technologies. 2020 Oct;119:102737.