2021

Jackson Goetz, Western Kentucky University
Kirolos Haleem, Western Kentucky University

This study investigates the safety impact of distracted driving (“texting while driving”) for different roadway configurations (i.e., “intersections, segments, freeways, and roundabouts”, “urban, suburban, and rural sections”, and “straight and curved road cross-sections”) and various lighting conditions (nighttime and daytime) using a driving simulator setting. The novelty of this study is investigating distracted driving at roundabouts. The simulator study took place at Western Kentucky University in Bowling Green, KY. Participants in this study included two main age groups, young adults (between 19 and 25 years) and middle-age groups (between 26 and 64 years). Overall, the standard deviation of speed and lane position of all participants was greater while texting and driving (i.e., when distracted). Both age groups drove significantly different from each other in terms of “lane position keeping” while texting and driving. “Driver speed” was a significant factor impacting “texting while driving” along rural straight and curved road sections. Texting at nighttime on the freeway for middle-age drivers had about 87% higher speed variance than texting at daytime. Texting at nighttime caused the highest speed variance than any daytime operation (either texting or not texting) for both age groups. At roundabouts, middle-age adults had greater speed and lane variances compared to their younger counterparts. Furthermore, there existed a relatively high speed and lane variances for both age groups, raising concerns about texting and driving at roundabouts. Useful recommendations include displaying message signs on the road that alert drivers on the dangerous effect of texting and driving, especially at nighttime.

Subin Park, Hanyang University, Ansan
Cheol Oh, Hanyang University
Shinhye Joo, Korea Transportation Safety Authority
Sungmin Hong, Korea Transportation Safety Authority

The evaluation of traffic safety levels is a fundamental aspect of supporting the establishment of safety policies and technical countermeasures by local governments. Existing safety indices that use actual crash data have limitations for the achievement of more active safety enhancement because long-term data collection is required to obtain sufficient samples. A promising alternative is to use indirect safety measures, which can be further upgraded in the era of big data, in the evaluation of traffic safety. This study proposes a novel safety indicator based on in-vehicle hazardous driving event data that is obtained from on-board devices, called digital tachographs (DTG), in Korea. The DTG-based indicatorex for evaluating traffic safety (DIETS), which is a probabilistic measure for quantifying the safety levels of local governments, was developed based on binary logistic regression (BLR) analyses. Hazardous driving events identified from DTG data were analyzed to derive independent variables in BLR modeling. DIETS is expected to facilitate the effective decision-making of local governments for the development and implementation of safety policies.

Sijun Shen (sshen@g.clemson.edu), Nationwide Children's Hospital
Simon Lin, Research Institute at Nationwide Childrens Hospital
Motao Zhu, Research Institute at Nationwide Childrens Hospital

Background: The prevalence of mobile sensing platforms allows researchers to evaluate individual driver safety using vehicle telemetric data. However, no study has assessed the feasibility of using aggregated vehicle telemetric data to predict crash spots. Objectives: The objective of this study is to determine if the aggregated fine geo-resolution vehicle telemetric data can be used to predict crash likelihood for roadway segments. Methods: The telemetric data from the GEOTAB company were used. The GEOTAB company recorded the frequency of harsh acceleration, harsh braking, harsh cornering, and the average magnitude of those harsh events for every 150×150 meter2 roadway segments within Columbus, Ohio between January and April 2018. Crash history were obtained from the 2018-2019 Ohio Policy Accident Report. Regularized logistic regression (RLR) with lasso penalty and boosted decision tree (BDT) algorithms were used to develop the predicting models. Results: Aggregated vehicle telemetric data provided effective predictions for crash spots (Area under curve [AUC] ≥ 0.73). Models’ predictive performance can be further improved if both vehicle telemetric variables and crash history were included in the models (AUC ≥ 0.77). The BDT models had superior predictive performance than the RLR models, due to its capability of incorporating complex relationships (e.g., non-linearity and all-way interactions) between predicting and predicted variables. Conclusion: Our study demonstrates the utility of geo-resolution vehicle telemetric data to predict crash spots. Aggregated vehicle telemetric data provide valuable information for crash likelihood monitoring and thereby, enable implementation of timely safety interventions by police and city planner.

Shan Jiang (sj576@scarletmail.rutgers.edu), Rutgers University
Mohsen Jafari, Rutgers University
Mohammad Jalayer, Rowan University

ABSTRACT The use of systematic techniques with historical crash data and qualitative measures has long been a common practice to identify the problematic road features and develop countermeasures to mitigate the crash risk in crash-prone locations. This paper proposes a novel approach, Safe Route Mapping (SRM) model that integrates crash-based estimates with conflict risks computed from driver-based data to score the safety of roadways. An advanced Safety Performance Function (SPF) estimates the number of crashes, and a driver-based model calculates dynamic conflict risk measures from driver and traffic data. In real-life implementations of the proposed methodology, the driver-based data and traffic data can be collected from vehicles or infrastructure-based data sources, including smartphones. We demonstrated the methodology using real historical crash data and simulated driver-based data obtained from VISSIM and SSAM. We show safety risk heat maps for the example roadway and illustrate how these maps change with driver types and traffic volumes. The proposed methodology fills the existing gaps in the use of near real-time dynamic data to designate safe corridors, dispatch law enforcement, and plan safety projects. Drivers can also use the road heat maps for situational awareness and trip planning. Keywords: Safe Route Mapping, Crash prediction, Real-time risk scoring, Safety Performance Function, Dynamic risk heat map.

Yanli Ma (mayanli@hit.edu.cn), Harbin Institute of Technology
Jieyu Zhu, Harbin Institute of Technology
Ke Chen, Heilongjiang Provincial Highway Survey and Design Institute

To reduce traffic accidents at signal intersections, it is significant to investigate the conflict identification between left-turning vehicles and straight vehicles in the opposite direction. The trajectory data of vehicles can be used to identify real-time conflicts in intersections. To perform such identification, accurate vehicle localisation should be obtained to clearly recognise the conflicts between left-turning vehicles and straight vehicles in the opposite direction at the signal control intersection. On the basis data collection of coordinate position, velocity, acceleration and yaw Angle of vehicles, Kalman filter algorithm was used to estimate the vehicle trajectory to obtain the vehicle kinematics information via the on-board system. The traffic conflict areas of the left-turning vehicles and straight vehicles in the opposite direction were determined through vehicle trajectory extrapolation, and the left-turn collision at the signal intersection was identified using the post-encroachment time algorithm and vehicle movement information. In addition, Anderson–Darling and modified Kolmogorov–Smirnov tests were performed to verify the goodness of fit of the data. Results show that the vehicle speed and localisation errors of the proposed method decreased by 66.67% and 83.33% compared with the results before filtering, respectively. Moreover, the results of the conflict recognition method based on trajectory reconstruction is consistent for both goodness of fit tests. This study can provide driving decision for drivers of left-turning vehicles and provide technical support for the research and development of left-turn anti-collision systems.

Hyeonseo Kim, Hanyang University, Ansan
Kyeongju Kwon, Hanyang University, Ansan
Nuri Park, Hanyang University, Ansan
Juneyoung Park, Hanyang University, Ansan
Mohamed Abdel-Aty, University of Central Florida

The main objective of this study is to evaluate the safety effects caused by altering drivers of the lengths of deceleration and acceleration lanes at rest areas on expressways in Korea. Although general conclusions can be found through crash-based safety analysis, to examine more specific optimal conditions considering various traffic conditions, this study proposes a novel framework to explore and evaluate crash-based and simulation-based safety performances. For this purpose, the safety performance function (SPF) and crash modification factor (CMF) were developed to reflect real-world safety impacts. To consider nonlinear trends of the parameters, nonlinearizing link functions were introduced into the analysis. Two types of simulation analyses were conducted to 1) find the combination of surrogate safety measures (SSMs) that best fit with the crash-based safety performance results and 2) determine the optimal lengths of deceleration lane and acceleration lanes for different traffic conditions. The results showed that the best length of deceleration lane of a rest area is between 240 m and 260 m, depending on the traffic conditions. The results also indicated that the optimal length of acceleration lane of a rest area is between 385 m and 400 m, depending on the traffic parameters. The findings of this study could be used to determine the safety solutions with a micro-traffic simulator.

Margaret Hunter, Purdue University
Enrique Saldivar-Carranza, Purdue University
Jairaj Desai, Purdue University
Jijo Mathew, Purdue University
Howell Li, Purdue University
Darcy Bullock, Purdue University

Typical safety improvements at signalized intersections are identified and prioritized using crash data over 3-5 years. Enhanced probe data that provides date, time, heading, and location of hard-braking events has recently become available to agencies. In a typical month, over six million hard-braking events are logged in the state of Indiana. This study compared rear-end crash data over a period of 4.5 years at 8 signalized intersections with weekday hard-braking data from July 2019. Using Spearman’s rank-order correlation, results indicated a strong correlation between hard-braking events and rear-end crashes occurring more than 400 ft upstream of an intersection.  The paper concludes that hard-braking events occurring at a far distance from the stop bar may be a useful tool to screen potential locations of rear-end crashes and follow up with mitigation measures quicker than the 3-5 year cycle used by agencies that rely on crash data.  Now that hard-braking data is commercially available in the United States, these techniques scale quite easily to state and national levels for near immediate implementation.

Panos Prevedouros, University of Hawaii
Braxton Chong, University of Hawai'i, Manoa
Luana Pereira,, University of Hawai'i, Manoa

Partly due to the pre-Covid-19 booming economy and the increasing numbers of distracted 37 motorists and other road users, safety risk (and insurance premiums) increased substantially 38 making liability reduction via monitoring, evaluation and coaching necessary for professional 39 drivers. The CEE Department at the UH, in collaboration with taxi and tour operators in Hawaii 40 have deployed state-of-the-art in-vehicle video monitoring equipment; the UH team assists in the 41 creation of a naturalistic driving database based on accelerometer-triggered "harsh events" that 42 record video clips starting ten seconds before the harsh event and ending ten seconds after the 43 event. Near misses are manually inspected and then coded recorded harsh events with an 44 observable traffic safety risk which could have caused property damage or more serious 45 outcomes. A number of variables are recorded for each near miss event including description of 46 the event, the vehicles and people involved, environmental and roadway factors, and 47 accelerometer/onboard data such as g-force and speed profile. The goal of this research was to 48 analyze the speed profiles by type of incident and level of speed: up to 20 mph, 21 to 35 mph, 49 and over 35 mph and present observable trends and differences in the speed profiles. Nearly 300 50 speed profiles were subjected to a preliminary analysis herein. The accumulation of more cases 51 is necessary for statistically significant inferences.

Pei Li (peili@knights.ucf.edu), University of Central Florida
Mohamed Abdel-Aty, University of Central Florida

Real-time crash likelihood prediction plays a crucial role in the proactive traffic safety management system. Most of the existing studies obtained traffic data from fixed devices, such as loop detectors, Bluetooth detectors, cameras, etc. However, these devices are not flexible enough to deploy at a large-scale and collect city-wide traffic data. With the help of mobile sensing technologies, vehicle-based data (e.g., GPS trajectory, connected vehicles data) are becoming more popular. Nevertheless, only a few studies investigated the application of this novel data for crash likelihood prediction with limited vehicle types. In addition, crash likelihood prediction using deep learning methods, especially Recurrent Neural Networks (RNNs), has received much attention in recent years. However, temporal attention, a powerful mechanism for learning time-series data, was ignored by all of the studies related to crash likelihood prediction. This paper utilized data fusion techniques to integrate two real-world trajectory datasets with a variety of vehicles. The traffic conditions of urban arterials were described with various speed-related features (e.g. average speed, standard deviation of speed, etc.). To predict the crash likelihood for arterials, this paper designed a deep learning architecture (TA-LSTM-CNN) containing a Long Short-term Memory (LSTM) with temporal attention and a Convolutional Neural Network (CNN). Experimental results indicated that the proposed method could achieve outstanding performance (e.g. high sensitivity and low false alarm rate) for the real-time crash likelihood prediction with the help of trajectory data fusion. Further, model comparison results suggested that the proposed model outperformed other state-of-the-art models in terms of various metrics.

Parisa Hosseini, Rowan University
Mohammad Jalayer, Rowan University
Subasish Das, Texas A&M University
Huaguo Zhou, Auburn University

Wrong-way driving (WWD) has been a long-lasting issue for transportation agencies and law enforcement, since it causes pivotal threats to road users. Notwithstanding being rare, crashes occurring due to WWD are more severe than other types of crashes. It is time consuming to identify true WWD crashes from large crash database. It often involves a large man-hours to review hardcopy of crash reports. Otherwise, it may cause overestimation or underestimation of WWD crash frequencies. To fill this gap, the present study aims at identifying WWD crashes from other motor vehicle crashes in police reports. By applying text mining techniques, useful information can be extracted from the crash report narratives. In order to distinguish real WWD crashes from other motor vehicle crashes, machine learning methods were implemented to develop classification algorithms. In this study, four classification algorithms, including Naïve Bayes (NB), Random Forest (RF), Support Vector Machine (SVM), and Neural Network (NN) were implemented to categorize crash reports as WWD and non-WWD crashes. Hardcopies of crash reports were used to evaluate the performance of each classification algorithm. Results indicated that RF outperformed in identifying true WWD crashes in comparison with other algorithms with the highest accuracy of 98%.

Qinghong Chen (chenqinghong@csu.edu.cn), Central South University
Ruifeng Gu, Central South University
Helai Huang, Central South University
Jaeyoung Lee, Central South Universty
Xiaoqi Zhai, Central South University
Ye Li, Central South University

This study aims to investigate contributing factors to potential collision risks during lane-changing processes from the perspective of vehicle groups and explore the unobserved heterogeneity of individual lane-changing maneuvers. Vehicular trajectory data, extracted from the Federal Highway Administration’s Next Generation Simulation dataset, are utilized and 579 lane-changing vehicle groups are examined. Stopping distance indexes are developed to evaluate the potential collision risks of lane-changing vehicle groups. Three binary logit models and three mixed binary logit models are established based on different perception reaction time. Model estimation results show that the mixed binary logit models outperform their counterparts regardless of the perception reaction time type. Several variables significantly affect the risk status of lane-changing vehicle groups, including the mean values of clearance distance and speed differences between the leading vehicle in the current lane and subject vehicle, standard deviations of clearance distance and speed differences between these two vehicles, as well as standard deviations of the speed difference between subject vehicle and the following vehicle in the target lane. Interestingly, the influences of the last three variables differ considerably across the observations. Since one of the explanations is individual heterogeneity, personalized designs for advanced driver assistance system would be an effective measure to reduce the risk.

Qinghong Chen (chenqinghong@csu.edu.cn), Central South University
Helai Huang, Central South University
Xiaoqi Zhai, Central South University
Ye Li, Central South University

The lane-changing maneuver has critical effects on roadway safety. This study evaluates the risky factors associated with the safety status of lane-changing vehicle groups and investigates the unobserved heterogeneity. Specifically, the instability of different lane-changing vehicle patterns is also investigated. A naturalistic vehicle trajectory dataset HighD was employed and 4,842 lane-changing vehicle groups were obtained. These vehicle groups were divided into 16 patterns according to the vehicle type, and four patterns with relatively large sample sizes were selected. A lane-changing risk index was developed to evaluate the risk level of vehicle groups. Random parameter ordered logit models were established based on the four selected patterns. Likelihood ratio tests were conducted to examine the stability of model estimates across different patterns, which indicated statistically significant instability among most patterns. The model results show that different patterns have different significant variables and the effects of these variables are found to vary across patterns. Generally, the risk level of vehicle groups are highly associated with (1) the longitudinal velocities and accelerations of vehicles; (2) the lateral velocity and acceleration of lane-changing vehicle; and (3) the gap distances between vehicles. And the effects of gap distances are found to vary across vehicle groups. This study suggests that advanced driver assistance system should develop more targeted strategies and provide different services according to vehicle patterns.

Subasish Das, Texas A&M University

The formulation of generation is based on the criteria that people within a common birth year range will experience the same major events in their lifetime. In crash data analysis, age related factors are commonly based on some age ranges or even sometimes in a simpler way by designating people or drivers as young, mid-age, or elderly. It is mostly due to the suitability of the interventional design. It is anticipated that seven generations are still playing role as driver, non-motorist, or occupant: the Greatest generation, the Silent generation, Baby boomers, Millennials, Generation Z, and Post Z. This study used nine years (2010-2018) of Fatality Analysis Reporting System (FARS) data to provide some insights about traffic fatalities and different generations. This study provides some intuitive exploratory analysis with data visualization tools. As Baby Boomers, Generation X, and Millennials are the top three generations in terms total number of fatalities, odds ratios have been developed to determine the specific traits that are associated with each generation. The findings of this unique analysis will provide more contexts to the generation-based studies.

Nicolette Formosa (n.formosa@lboro.ac.uk), Loughborough University
Mohammed Quddus, Loughborough University
Stephen Ison, De Montfort University

Connected and Autonomous Vehicles (CAVs) that rely on Artificial Intelligence to autonomously navigate are expected to deliver a step change in safety. Existing Collision Avoidance Systems (CASs) assess and quantify the threat level surrounding the ego-vehicle. However, they are not able to plan the best response to a fully unexpected dangerous situation while driving. Therefore, it is important that the algorithm has the ability to cope with uncertainties since not all situations are ‘car-following’. Previous research has not taken this uncertainty into account, so it is desirable to develop robust systems which are not restricted by the predefined movement patterns of the vehicle. In fact, the readily available CAS estimates the threat level based only on one factor: Time-To-Collision (TTC). This approach is limited since it cannot handle all scenarios and ignores all uncertainties. To overcome these limitations, this paper uses deep learning to develop multiple CASs to identify the optimal factors to estimate the threat level under uncertainty. Comparative analyses were undertaken by incorporating new varying input factors to each system (e.g. surrogate safety measures, vehicle parameters, macroscopic traffic data and hybrid of factors). Experiments based on real-world data highlight which factors are important and validate that adding more factors increases sensitivity of the systems. Results also show that systems considering uncertainty, lower the false alarm rate and extends the systems’ application for a wider spectrum of traffic scenarios. This is paramount for CASs as uncertainties are inherent in any real-world deployment of CAVs in a mixed traffic stream.

Xiaoyan Xu, Tongji University
Xuesong Wang (wangxs@tongji.edu.cn), Tongji University
Xiangbin Wu, Intel Labs China
Yu Lin, Shanghai SH Intelligent Automotive Technology Co., Ltd.
Xinyi Hou, Traffic Police Headquarters of Shanghai Public Security Bureau

Conflict severity is the outcome of complex interactions between roadway and environmental characteristics, and vehicle motion. Understanding how and to what extent a vehicle is influenced by roadway and surrounding road users during a conflict can help to analyze the causal mechanisms of collisions, thus providing insights into roadway safety improvement countermeasures. This study utilized the NGSIM Peachtree Street vehicle trajectory dataset to achieve the objective of investigating causal factors of conflicts at intersections by exploring roadway-to-vehicle and vehicle-to-vehicle interactions. In order to remove the outliers and white noise existing in the raw data, vehicle trajectories were reconstructed by discrete wavelet transform and Kalman filtering. The generalized time-to-collision was adopted to detect and measure the severity of conflicts, and 423 conflict events were finally extracted. Path analysis models were then established to explore in exactly which ways the roadway-to-vehicle and vehicle-to-vehicle interactions were related to conflict severity. Various roadway and environmental characteristics such as traffic flow’s average speed, percentage of trucks and intersection skew angle were included in the models. The results indicate the roadway and environmental characteristics have both direct and indirect effects on conflict severity; while for the indirect effects, the conflict vehicle’s kinematics such as the average and standard deviation of speed play an intermediate role in linking roadway factors and conflict outcome. The framework of this study can be applied to assessing roadway readiness for both human-driven and automated vehicles.

Qiangqiang Shangguan, Tongji University
Ting Fu (tingfu@tongji.edu.cn), Tongji University
Rui Jiang, Shandong Hi-Speed Group
Shou'en Fang, Tongji University

Traditional surrogate measures of safety (SMoS) do not fully consider the crash mechanism or fail to consider the crash probability and consequences at the same time. In addition, driving risks are constantly changing with driver's personal characteristics and environmental factors. However, few studies have considered the impact of driver’s behavior characteristics and environmental factors on driving risks. To address the above research gaps, 16,905 car-following events were extracted from Shanghai Naturalistic Driving Database. A new SMoS, named rear-end crash risk index (RCRI), was proposed to identify driving risks. Performance of RCRI and traditional SMoS were compared. Using this measure, a risk comparative analysis was conducted to study the impact of factors from different facets in terms of weather, temporal variables and traffic conditions. Then, a mixed-effects linear regression model was applied to investigate the relationship between various influencing factors and driving risks. Results show that RCRI has better performance over other traditional SMoS because it can describe dynamic changes in risks and can be applied to any car-following scenarios. The comparative analysis indicates that high traffic density, workdays and morning peaks lead to higher risks. Moreover, results from the mixed-effects linear regression model suggest that driver’s behavior characteristics, traffic density, day-of-week (workday v.s. holiday) and time-of-day (peak hour v.s. off-peak hour) had significant effects on driving risks. The study provides a SMoS that can better identify driving risks in a more reliable way. Results can be applied to real-time risk prediction and traffic management to improve driving safety.

Jenna Kirsch, Wayne State University
Stephen Remias (sremias@wayne.edu), Wayne State University
Steven Lavrenz, Wayne State University

Historically, crash frequency models have not directly incorporated reliability measures for roadway mobility and delay. However, recent developments in crowd-sourced probe vehicle data, which allow for the systematic measurement of travel speeds on arterial and highway segments, are poised to change this. In this study, probe vehicle data is used to investigate the impact of level of travel time reliability (LOTTR) on crashes. Crash and mobility data were collected from 2,945 interstate highway, state trunkline highway, and U.S. highway segments in the State of Michigan. Pavement characteristics, geometric data, and traffic volumes were also obtained. Two negative binomial models were estimated to determine the significance of LOTTR on crash frequency, one with and one without the added variable. The LOTTR variable was found to be statistically significant and the model with LOTTR had a significantly better fit to the data than the model without LOTTR. The results suggest that there is future value in including LOTTR and other reliability metrics in crash frequency models. The findings also suggest that roadway operators who focus on improving LOTTR could observe reduced crashes as well.

Nipjyoti Bharadwaj, NRC Research Associateship
Praveen Edara (EdaraP@missouri.edu), University of Missouri, Columbia
Carlos Sun, University of Missouri, Columbia, Ellis Library

Car crashes can occur in a variety of ways. A common type of collision involving two vehicles is when one vehicle rear-ends another vehicle. This study describes a methodology to understand the braking judgments in rear-end safety-critical events, i.e., crashes and near-crashes, using naturalistic driving study (NDS) data collected as part of USDOT's Second Strategic Highway Research Program (SHRP 2). A small subset of the rear-end events involving younger drivers (16-19 years old) were used to illustrate the proposed methodology. Kinematic measures such as the inverse time-to-collision (TTC) were used to compare braking performances for crashes and near-crashes. The deceleration of the follower vehicle was modeled using Linear Regression. The independent variables are inverse TTC, deceleration of the lead vehicle, and speed of the follower vehicle. The study also demonstrated that range vs. range rate plots are useful to identify the follower’s reaction (i.e., onset of braking) in response to the lead vehicle’s deceleration. The average TTC value for the crash and near-crash events at the onset of braking was 0.9 seconds and 1.92 seconds, respectively. All safety-critical events analyzed in this study experienced TTC values lower than 3 seconds. This study also developed representative plots of range versus range rates during braking that clearly delineate the boundaries between a crash and a near-crash. The kinematic differences between a crash and a near crash discerned in this study can be useful for designing collision avoidance and driver assistance systems.

Zulqarnain Khattak (khattakzh@ornl.gov), Oak Ridge National Laboratory
Michael Fontaine, Virginia Transportation Research Council
Asad J. Khattak, University of Tennessee

This study developed a deep learning approach based on 1D convolutional neural networks (CNN) for detection of safety critical events (SCEs) using large-scale naturalistic driving vehicle kinematics data. The data are unique in the sense that such accurate pre-crash data at high fidelity are not available in traditional crash repositories. This study contributes to the literature by providing a first attempt at predicting responses to SCEs by applying Artificial Intelligence techniques. Specifically, the study develops deep learning-based CNN architectures for identification of SCEs using driving volatility based kinematic thresholds. The key contribution lies in developing a CNN input layout that is acceptable to CNN schemes and represents the motion kinematics such as speed acceleration and volatility measures. Several 1D-CNN architectures were developed using layers numbers of convolutions, layer patterns, and kernels. Shallow and deep architectures were tested, revealing higher accuracy of shallow architectures in detecting SCEs. The optimal number of epochs were identified using an early stopping method while the CNN performance was improved by increasing the number of epochs. The ensemble CNN had the highest predictive accuracy of 95.6%, which was 2.5% higher than the optimal CNN using test data. The ensemble CNN also outperformed classical machine learning models and model performance reported in past studies on detection of SCEs. These results have implications for identification of safety hotspots and providing real-time alerts and warnings in connected and automated vehicle environment.

Hiba Nassereddine (nassereddin2@wisc.edu), University of Wisconsin, Madison
Kelvin Santiago-Chaparro, University of Wisconsin, Madison
David Noyce, University of Wisconsin, Madison

Traffic conflicts and surrogate safety measures have been used as an alternative to crash-based methods to study road safety. The extreme value theory offers a modeling framework for safety surrogates. Using video data, time-to-collision (TTC) values between right-turn-on-red (RTOR) vehicles and through vehicles were evaluated. Using trajectory data collected from radars, post-encroachment-time (PET) values were computed. In this study, six models were developed based on RTOR conflicts; two Univariate Generalized Extreme Value (UGEV), each using TTC and PET, respectively, and two Univariate Generalized Pareto (UGP) models, each using TTC and PET, respectively. Additionally, the two models Bivariate Generalized Extreme Value (BGEV) and Bivariate Generalized Pareto (BGP) which both jointly use TTC and PET were also applied. Using the resulting estimates, the number of crashes was estimated for each model. For the univariate models, the results show that the estimated crashes from UGEV models are closer to the observed number of crashes than those from UGP models. As for the bivariate models, the estimated crashes from BGP are closer to the observed crashes than those from BGEV models. The more accurate BGP crash estimates are attributed to the efficient use of data. UGP and BGEV models underestimated crash estimation while UGEV using PET and BGP has a similar performance estimating crashes. The results show that the UGEV model using TTC performed the best followed by the UGEV model using PET and the BGP model. This study presents a step forward in developing safety models based on several safety surrogates.

Denis Monyo, University of North Florida
Henrick Haule, Florida International University
Angela Kitali, Florida International University
Thobias Sando, University of North Florida

Older drivers are prone to driving errors that can lead to crashes. The risk of older drivers making errors increases in locations with complex roadway features and higher traffic conflicts. Interchanges are freeway locations with more driving challenges than other basic segments. Because of the growing population of older drivers, it is vital to understand driving errors that can lead to crashes on interchanges. The knowledge can assist in developing countermeasures that will ensure safety for all road users when navigating through interchanges. The goal of this study was to determine driver, environmental, roadway, and traffic characteristics that influence older drivers' errors resulting in crashes along interchanges. The analysis was based on three years (2016-2018) of crash data from Florida. A two-step approach involving a latent class clustering analysis and the penalized logistic regression was used to investigate factors that influence driving errors made by older drivers on interchanges. This approach accounted for heterogeneity that exists in the crash data and enhanced the identification of contributing factors. The results indicated patterns that are not obvious without a two-step approach, including variables that were not significant in all crashes but specific clusters. These factors included driver gender and interchange type. Results also showed that all other factors, including distracted driving, lighting condition, area type, speed limit, time of day, and horizontal alignment, were significant in all crashes and few specific clusters.

Qikang Zheng (zhengqikang@seu.edu.cn), Southeast University
Pan Liu, Southeast University
Chengcheng Xu, Southeast University
Yao Xiao, Suzhou Planning & Design Research Institute Co., Ltd.

Numerous studies have been conducted to improve the prediction efficiency of crash risks. Nevertheless, the most important crash precursors were neglected. The primary purpose of this study is to identify optimal crash precursors for different segment types, as well as provide a threshold selection method for real-time crash risk models. The mainline was divided into basic sections, weaving areas, merging areas, and diverging areas. Bayesian logistic models were established for each type of segment, and significant factors were distinguished. A threshold selection method was proposed based on cost-benefit theory, and the threshold is determined as the value when the number of proactive safety interventions to prevent a crash is 5000. Models with one, two and three optimal variables were developed, and the prediction performance of the models was evaluated. Comparison results show that the minimum amount of parameters which can achieve the ideal prediction effectiveness is two.  In this situation, 25%, 50%, 20% and 20% of the crashes occurring at basic sections, weaving areas, merging areas and diverging areas can be accurately predicted respectively. Downstream average speed was recommended as the best crash precursor variable for all the segment types. Support Vector Machine (SVM) and Random Forest (RF) were utilized to confirm the conclusion. The results of this study can be applied to help reduce crash risk to a relatively economical level in practical applications.

Xiaoyu "Sky" Guo, Texas A&M University, College Station
Lingtao Wu (wulingtao@gmail.com), Texas A&M Transportation Institute
Xiaoqiang Kong, Texas A&M University, College Station
Yunlong Zhang, Texas A&M University, College Station

Distracted driving is one of the most significant factors contributing to crashes, and distraction‑affected crashes are increasing in recent years. Although researchers have developed safety performance functions (SPFs) for various crash types, SPFs for distraction-affected crashes are rarely reported in the literature. One possible reason is lack of critical distracted behavior information in the commonly used safety data (i.e., roadway inventory, traffic, and crash counts). Recently, drivers’ phone use while operating a vehicle (referred as phone use data) are recorded by mobile application companies and become available to safety researchers. The primary objective of this study is to examine if the phone use data can potentially benefit the development of SPFs for distraction-affected crashes. To fulfill the objective, the authors integrated phone use data with roadway inventory, traffic, and crash data in Texas. After that, the Random Forest (RF) algorithm is applied to examine if the frequency of phone use while driving is a significant factor for predicting the number of distraction-affected crashes. Further, this study developed two SPFs for distraction-affected crashes with and without the phone use information, and assessed model fitting and prediction performances of them. RF result reveals that the phone use information is the most important factor contributing to the number of distraction-affected crashes. Performance evaluation indicates that including phone use information in the SPFs consistently improves the model’s fitting and prediction abilities.

Kang Wang, Tongji University
Ling Wang (wang_ling@tongji.edu.cn), Tongji University
Wanjing Ma, Tongji University
Hao Zhong, Tongji University
Hongge Zhu, China Highway Engineering Consultants Corporation

Previous studies have proven that the crash possibility and crash type on different segment types are different. However, there are not enough studies which have conducted microscopic crash mechanism analyses considering different types of segments. To comprehensively identify and analyze the heterogeneity in crash mechanism between four types of segments, i.e., merge, diverge, weaving, and basic segments, this study is proposed. Firstly, the segment type heterogeneity was analyzed from crash characteristics, significant variables, and variables importance aspects. Secondly, a method of variables selection was proposed to solve the “dimension disaster” in modeling. Thirdly, a nested logit model was built to quantitatively analyze the impact of crash contributing factors on the crash risk. The results showed that there exist statistically significant differences between four types of segments in crash characteristics, i.e., crash rate, number of vehicle(s) involved, crash type, and crash severity. Additionally, it was found that the most crash risk for merge and weaving segments is from the segments close to the target segment, but the most crash risk for diverge and basic segments is from vehicles traveling from upstream. Besides, it was found that the weather parameters had a similar effect on the crash risk between four types of segments, but it was different for geometry and traffic parameters, which indicated the heterogeneity of crash mechanisms for different segment types. Moreover, when the number of ramps upstream increases or when the distance between ramps and target segment decreases, the crash risk will increase.

Emmanuel Kidando, Cleveland State University
Angela Kitali, Florida International University
Boniphace Kutela, Texas A&M Transportation Institute
Alican Karaer, Florida State University
Mahyar Ghorbanzadeh, Florida A&M University-Florida State University College of Engineering
Mohammadreza Koloushani, Florida A&M University-Florida State University College of Engineering
Eren Ozguven, Florida A&M University-Florida State University College of Engineering

This study explored the use of real-time traffic events and signal timing data in understanding factors influencing the injury severity of vehicle occupants at intersections. The analysis was based on three years (2017-2019) of the crash and high-resolution traffic data. The best fit regression was first identified by comparing the conventional regression model with the hierarchical Bayesian logistic models. The hierarchical model with a heavy-tailed distribution was shown to best fit the dataset and was used in the variable assessment. The model results revealed that about 13.6% of the unobserved heterogeneity comes from site-specific variations, which underlines the need for the use of the hierarchical model. Among the real-time traffic events and signal based variables, approach delay, and platoon ratio were found to significantly influence the injury severity of vehicle occupants at 90% Bayesian credible interval. Additionally, manner of collision, occupant seat position, number of vehicles involved in a crash, gender, age, lighting condition, and day of the week were found to significantly impact the vehicle occupant injury. The study findings provide valuable insights to transportation agencies for developing countermeasures to proactively mitigate the crash severity risk.

Chaolun Ma, Texas A&M University System
Yongxin Peng, Texas A&M University
Lingtao Wu (wulingtao@gmail.com), Texas A&M Transportation Institute
Xiaoyu "Sky" Guo, Texas A&M University, College Station
Xiubin Bruce Wang, Texas A&M University System

Distraction occurs when a driver’s attention is diverted from driving to a secondary task. The number of distraction‑affected crashes are increasing in the recent years. Accurately predicting distraction-affected crashes is critical for roadway agencies to reduce distracted driving behaviors and distraction-affected crashes. Recently, emerging phone use data and machine learning techniques are becoming available to safety researchers, and can potentially improve prediction of distraction-affected crashes. Hence, this study first examined if phone use events provide important information for distraction-affected crashes. The authors developed two models with and without phone use events by a machine learning technique (i.e., XGBoost), and compared their performance with a conventional statistical model (i.e., logistic regression model). The measurement demonstrates the superiority of XGBoost over logistic regression with high dimensional dataset. Further, this study implemented SHAP (SHapley Additive exPlanation) to interpret the results and analyze the importance of individual features related to distraction-affected crashes, and tested its ability to improve the prediction accuracy. An XGBoost model is trained and its result achieves sensitivity of 91.59 %, specificity of 85.92% and accuracy of 88.72%, respectively. The results suggest that: (1) phone use information is an important factor associated the occurrences of distraction-affected crashes; (2) distraction-affected crashes are more likely to occur on roadway segments with higher exposure (i.e., length and traffic volume), unevenness of traffic flow condition, or with medium truck volume.