This poster session presents papers from ANB20 (Safety Data, Analysis, and Evaluation) that explore new models and methods for safety analysis and evaluation studies. The papers include a broad range of safety topics from secondary crashes to distracted driving to automated vehicles. Wear good shoes. There is a lot here!
Application of Extreme Value Theory for Before–After Road Safety Analysis
Lai Zheng, Harbin Institute of TechnologyShow Abstract
Tarek Sayed, University of British Columbia
Because of well-recognized quality and quantity problems associated with the historical crash data, traffic conflict techniques have been increasingly used in the before-after safety analysis in recent years. This study proposes to use extreme value theory (EVT) approach to conduct the traffic conflict-based before-after analysis. The capability of providing confident estimation of extreme events by the EVT approach drives the before-after analysis to shift from normal traffic conflicts to more serious conflicts, which are relatively rare but have more in common with actual crashes. The approach is applied to evaluate the safety effects of converting channelized right-turn lanes to smart channels, based on traffic conflicts defined by time to collision (TTC) collected from three treatment intersections and one control intersection in the city of Penticton, British Columbia. Odds ratios and treatment effects are calculated from extreme-serious conflicts, the frequencies of which are estimated from the Generalized Pareto distributions of traffic conflicts with TTC≤0.5s. The results show approximately 34% reduction in total extreme-serious conflicts (i.e., combining merging conflicts and rear-end conflicts), indicating overall a remarkable safety improvement following the smart channel treatment. This finding is consistent with the analysis result based on traffic conflicts with TTC≤3.0s. It is also found that the reduction in extreme-serious merging conflicts is small and insignificant. This is caused by the fact the TTC values of merging conflicts become smaller after the treatment, and it is possibly because drivers get more aggressive with the better view of approaching cross-street traffic provided by the smart channel.
Comparison of Empirical Bayes and Propensity Score Methods for Road Safety Evaluation: A Simulation Study
Haojie Li, Southeast UniversityShow Abstract
Daniel Graham, Imperial College London
Hongliang Ding, Southeast University
The evaluation of the effects of road safety measures on road accidents has gained continuous attention among researchers in recent years. Besides the common used empirical Bayes (EB) approach, the propensity score (PS) methods have been widely employed in road safety evaluation studies. However, the conditions under which these methods can provide valid estimates of treatment effects are not well understood. We conduct a simulation-based comparison study to provide insight into the performance of the EB and PS methods in settings with and without violation of the key assumptions of the EB and PS methods. The models investigated include the EB, inverse probability weighting (IPW), and the doubly robust (DR) methods with different model specifications and data conditions. The results suggest that most of the methods can provide unbiased estimates of the treatment effect when the models are correctly specified, although the bias of the effect estimates increases slightly for all IPW models and most DR models with a small data sample, indicating that the propensity score methods are “data hungry”. The DR method is less affected by the omission of covariates and consistently provides unbiased estimates even in the scenarios with incorrect model specification, indicating its superiority to other two methods.
Application of Random Effects Negative Binomial Models with Clustered Dataset for Vehicle Crash Frequency Analysis
Haitao Gong, Jackson State UniversityShow Abstract
Shontria Dent, Jackson State University
Feng Wang, Jackson State University
Bin Zhou, Central Connecticut State University
For the past few years, vehicle crash frequency analysis has been one of the study areas of great interests in highway safety research. One of the major challenges is how to deal with the unobserved heterogeneity of crash data. While statistical models of crash frequency analysis based upon single probability distributions are constantly improving, several researchers discovered that multiple distribution models might better describe crash frequency data and capture more unobserved heterogeneity. Based upon the hypothesis that total crash counts occurring at an intersection may be affected by different unique sets of contributing factors, this research proposes a two-step approach to study the crash contributing factors at intersections in Mississippi Coast which is one of the most frequent crash areas in the State of Mississippi. In this study, the crash data are first clustered into subpopulations with the application of a hierarchical clustering method, and then a Random Effects Negative Binomial model is applied to each component at the intersection level. A model with no data clustering is also estimated to serve as the comparison benchmark. The analysis results show that this two-step approach can reveal more information about crash contributing factors and have increased predictive power and goodness of fit.
Analysis of Accident Injury-Severity Outcomes: The Zero-Inflated Hierarchical Ordered Probit Model with Correlated Disturbances
Grigorios Fountas, Edinburgh Napier UniversityShow Abstract
Panagiotis Anastasopoulos, University at Buffalo, The State University of New York
In accident injury-severity analysis, an inherent limitation of the traditional ordered probit approach arises from the a priori consideration of a homogeneous source for the accidents that result in a no-injury (or zero-injury) outcome. Conceptually, no-injury accidents may be subject to the effect of two underlying injury-severity states, which are more likely to be observed in accident datasets with excessive amounts of no-injury accident observations. To account for this possibility along with the possibility of heterogeneity stemming from the fixed nature of the ordered probability thresholds, a zero-inflated hierarchical ordered probit approach with correlated disturbances is employed, for the first time – to the authors’ knowledge – in accident research. The latter consists of a binary probit and an ordered probit component that are simultaneously modeled in order to identify the influential factors for each underlying injury-severity state. At the same time, the model formulation accounts for possible correlation between the disturbance terms of the two model components, and allows for the ordered thresholds to vary as a function of threshold-specific explanatory variables. Using injury-severity data from single-vehicle accidents that occurred in the State of Washington, from 2011 to 2013, the implementation potential of the proposed approach is demonstrated. The comparative assessment between the zero-inflated hierarchical ordered probit approach with correlated disturbances and its lower-order counterparts highlights the potential of the proposed approach to account for the effect of underlying states on injury-severity outcome probabilities and to explain more with the same amount of information.
Net-Social and Net-Private Benefits of Some Existing Vehicle Crash Avoidance Technologies
Corey D. Harper, Booz Allen HamiltonShow Abstract
Abdullah Khan, Carnegie Mellon University
Chris Hendrickson, Carnegie Mellon University
Constantine Samaras, Carnegie Mellon University
Most light-duty vehicle crashes occur due to human error. Many of these crashes could be avoided or made less severe with the aid of crash avoidance technologies. These technologies can assist the driver in maintaining control of the vehicle when a possibly dangerous situation arises by issuing alerts to the driver and in a few cases, responding to the situation itself. This paper estimates the social and private benefits and costs associated with three crash avoidance technologies, blind-spot monitoring, lane departure warning, and forward-collision warning, for all light duty passenger vehicles in the U.S. for the year 2015. The three technologies could collectively prevent up to 1.6 million crashes each year including 7,200 fatal crashes. In this paper, the authors estimate the net-social benefits to the overall society from avoiding the cost of the crashes while also estimating the private share of those benefits that are directly affecting the crash victims. For the first generation warning systems, net-social benefits and net-private benefits are positive. Moreover, the newer generation of improved warning systems and active braking should make net- benefits even more advantageous.
Incorporating Route Safety in the Pathfinding Problem Using Big Data
Nima Hoseinzadeh, University of Tennessee, KnoxvilleShow Abstract
Ramin Arvin, University of Tennessee, Knoxville
Asad Khattak, University of Tennessee, Knoxville
Lee Han, University of Tennessee, Knoxville
With the emergence of the internet of things, pathfinding problems have recently received a significant amount of attention. Various commercial applications provide automated routing by considering travel time, travel distance, fuel consumption, complexity of the road, etc. Unfortunately, many of these prospective applications do not consider route safety. Because connected vehicles (CV) generate enriched “Big Data”, researchers have opportunities to develop new transportation methods. The goal of this study is to address safety aspects in pathfinding problems by developing a methodological framework that simultaneously considers safety and mobility. To reach this goal, the concept of “driving volatility” is utilized as a surrogate safety performance measure. The proposed framework uses CV big data and real-time traffic data to obtain calculate safety indices and travel times. Measured safety indices include 5-year crash history, route speed and acceleration volatility, and driver volatility. Travel time and safety shape a cost function called “route impedance”. The algorithm has the flexibility for the user to predefine the weight for safety consideration. It also uses driver volatility to automatically increase weights of safety considerations for volatile drivers. In order to illustrate the algorithm, an origin-destination pair in Ann Arbor Michigan is selected and more than 42 million CV observations from around 2,800 CVs from the Safety Pilot Model Deployment were analyzed. Finally, this paper shows suggested routes for multiple scenarios to demonstrate the outcome of the study. The results revealed that the algorithm might suggest different routes when considering safety indices and not just travel time.
Modeling Highly Unbalanced Crash Injury Severity Data by Ensemble Methods and Global Sensitivity Analysis
Liming Jiang, University of Massachusetts, LowellShow Abstract
Yuanchang Xie, University of Massachusetts, Lowell
Tianzhu Ren, University of Massachusetts, Lowell
Due to its significance, Crash Injury Severity (CIS) has been extensively studied and numerous methods have been developed for investigating the relationship between crash outcome and explanatory variables. CIS data is often characterized by highly unbalanced injury distributions, with most crashes in the Property-Damage-Only (PDO) category and the severe injury category making up only a fraction of the total observations. Existing methods tend to favor crash outcome categories with the most observations. This often leads to a high modeling accuracy for PDO crashes but poor prediction accuracies for other injury categories. This research introduces three ensemble methods to model unbalanced CIS data: random forest, AdaBoost, and Gradient Boost. A more reasonable performance metric, F1 score, is used for model selection. It is found that AdaBoost and Gradient Boost clearly outperform the remaining methods and generate more balanced prediction accuracies. Additionally, a global sensitivity analysis method is adopted to determine the individual and joint impacts of various CIS impact factors on crash injury outcome. Grade percentage, driver restraint, accident type, road characteristics, and truck percentage are found to be the most influential factors. Finally, a simulation-based approach is adapted to further study how the impact of a particular factor (e.g., horizontal curve) may vary with respect to different value ranges.
Investigating the Characteristics of Connected and Autonomous Vehicle Involved Crashes
Chengcheng Xu, Southeast UniversityShow Abstract
ZIJIAN DING, Southeast University
Chen Wang, Southeast University
This study aimed to investigate the characteristics and patterns of the connected and autonomous vehicle (CAV) involved crashes. The crash data were collected from the reports of CAV involved crash submitted to the California Department of Motor Vehicles between 2015 and 2018. The descriptive statistics analysis was employed to investigate the characteristics of CAV involved crashes in terms of crash location, weather conditions, driving mode and vehicle movement before crash occurrence, vehicle speed, collision type, crash severity and damage locations of involved vehicles. The bootstrap based binary logistic regressions were then developed to investigate the factor contributing to the collision type and severity of CAV involved crashes. The results suggested that the CAV driving mode, collision location, roadside parking, rear-end collision, and one-way road are the main factors contributing to the severity level of CAV involved crashes. The CAV driving mode, CAV stopped or not, CAV turning or not, normal vehicle turning or not, and normal vehicle overtaking or not are the factors affecting the collision type of CAV involved crashes.
A Comprehensive Review of Secondary Crash Studies
Armana Huq, Florida International UniversityShow Abstract
Priyanka Alluri, Florida International University
Secondary crashes (SCs) could have resulted from primary incidents for their complex interaction between roadways, vehicles, traffic and environmental conditions. However, several researchers are still in doubt whether the principal cause of SCs are from primary incident or from recurring congestion. Compared to primary crashes, there have been very few studies that focused on SCs. This review paper focuses the existing literatures on SCs occurred on freeways and identify possible influential risk factors associated with these crashes. The current practices to identify SCs are first discussed in detail. Static, dynamic, and spatial analysis tools are discussed particularly. The models to predict the probability of secondary crash occurrences are presented next. Finally, a thorough investigation has been done to identify influential risk factors associated with SCs. The lessons learned from this comprehensive literature are eventually presented a number of research gaps with recommendations for manifesting potential mechanisms in analyzing SCs.
Hit and Run Crashes: An Application of Correlated Random Parameter Probit Model Using Real-Time Crash Data
cassandra Villa, California State Polytechnic University, PomonaShow Abstract
Gurdiljot Gill, California State Polytechnic University, Pomona
Wen Cheng, California State Polytechnic University, Pomona
Xudong Jia, California State Polytechnic University, Pomona
Jiao Zhou, 1978
The issue of unobserved heterogeneity in crash data has been highlighted by many recent traffic safety studies. The safety literature has demonstrated the capability of the full random parameters approach to address the issue of unobserved heterogeneity. However, such approach has been mostly restricted to the investigation of general crash frequency models. The current study provides the application of this approach to a concerning crash behavior of Hit and run (HR) by extending the conventional random parameter model to allow the correlation between parameters. This study also focuses on utilizing the real-time traffic data to predict the HR crash risk. Additionally, three other models are developed, representing the current safety literature, to compare the performance of the proposed correlated random parameter model. The results from the posterior model estimates demonstrated the evidence of parameters varying with observations. The model fit results illustrated the worst performance for the traditional probit model while the random parameters model was relatively superior. However, the model with correlated random parameters exhibited the best performance, potentially due to its advantage to replicate the realistic scenario where the explanatory variables may act as confounding factors due to their interactions. The results for model performance based on predictive accuracy were monitored by using ROC (receiver operating characteristic) curves. The results corroborated the model fitness trends and revealed that the accommodation of correlations for random parameters improved the model prediction performance, especially at threshold levels generally adopted by safety practitioners. Keywords: correlated random parameters, hit and run, real-time, probit
Spatial Local Effect Analysis of Traffic Accident Size Using Geographically Weighted Structural Equation Modeling
Taekyoung Kim, Korea Road Traffic AuthorityShow Abstract
Jaewoong Yun, Yonsei University
Jin-Hyuk Chung, Chung Ang University
Considering spatial factors in the analysis of data is an approach that can better reflect the real world. In fact, there have been studies on spatial analysis using spatial metric model and structural equation models, but each approach has limitations; the effect of specific independent variables on dependent variables does not reflect differences in regions. For example, at two points with perfectly identical point properties, certain events can occur at different levels. This difference is defined as a spatial local effect in this study, and previous research has grappled with such effects. In this study, we aimed to develop a complementary model, and propose a new approach combining a spatial metric model and a structural equation model. Through this model, we can move away from interpreting only the global effect, which is the effect of certain factors on the total system. In other words, it is possible to identify differences in the influence of certain factors on specific area and to develop customized actions for that areas. In this context, we confirmed its applicability by applying it to traffic accident data of Korea. In particular, the effect of spatial factors on the size of the traffic accidents was analyzed. Through this, this study will identify what factors should be controlled to reduce size of traffic accidents at specific points and help to establish appropriate measures for each point.
Wrong-Way Driving Crashes: A Random Parameters Ordered Probit Analysis of Injury Severity
Mohammad Jalayer, Rowan UniversityShow Abstract
Ramin Shabanpour, University of Illinois, Chicago
Mahdi Pour Rouholamin, DKS Associates, Inc.
Nima Golshani, Georgia Institute of Technology (Georgia Tech)
Huaguo Zhou, Auburn University
In the context of traffic safety, whenever a motorized road user moves against the proper flow of vehicle movement on physically divided highways or access ramps, this is referred to as wrong-way driving (WWD). WWD is notorious for its severity rather than frequency. Based on data from the NHTSA, an average of 355 deaths occur in the U.S. each year due to WWD. This total translates to 1.34 fatalities per fatal WWD crashes, whereas the same rate for other crash types is 1.10. Therefore, WWD crashes, and specifically their severity, must be meticulously analyzed using the appropriate tools to develop sound and effective countermeasures. The objectives of this study were to use a random-parameters ordered probit model to determine the features that best describe WWD crashes and to evaluate the severity of injuries in WWD crashes. This approach takes into account unobserved effects that may be associated with roadway, environmental, vehicle, crash, and driver characteristics. To that end and given the rareness of WWD events, 15 years of crash data from the states of Alabama and Illinois were obtained and compiled. Based on this data, a series of contributing factors including responsible driver characteristics, temporal variables, vehicle characteristics, and crash variables are determined, and their impacts on the severity of injuries are explored. An elasticity analysis was also performed to accurately quantify the effect of significant variables on injury severity outcomes. According to the obtained results, factors such as driver age, driver condition, roadway surface conditions, and lighting conditions significantly contribute to the injury severity of WWD crashes.
Identifying Contributing Factors to Crash Severity: Analysis of Gender Differences
Hesamoddin Razi Ardakani, Sharif University of TechnologyShow Abstract
Amin Ariannezhad, University of Arizona
Mohammad Kermanshah, Sharif University of Technology
This study aims to investigate the factors affecting the severity of urban crashes among male and female drives. Traffic crashes occurred in Tehran, Iran during the year 2009 were analysed in this research. The observations were divided into two groups based on driver’s gender. Two crash severity models were estimated for each group using binary logit model in which dependent variable included property damage only crashes and injury/fatal crashes. Pseudo-elasticity values were calculated to better compare male and female drivers and understand the exact effect of each variable. The results indicated that factors such as increased driver’s age, non-educated drivers, weekend, nighttime and intersections increase the crash severity in both male and female drivers. Among human factors which affect the crash severity, violating the traffic rules and speeding increased the crash severity in male drivers, while lack of driving experience and vehicle mechanical defect increased the females’ crash severity. Furthermore, crashes involving male drivers were more severe in clear and foggy weather conditions, while female drivers were observed to have more severe crashes in rainy weather. Significant differences in contributing factors to severity of crashes in male and female drivers prove that separate policies for these two groups of drivers should be adopted. The findings suggest that, educational programs could help encourage drivers to obey driving rules and improve roads safety.
Determining Optimal Segment Lengths for Traffic Safety Analysis Based on Spectral Analysis
Xi Zhao, Clemson UniversityShow Abstract
Yichuan Peng, Tongji University
Xinzhi Zhong, Tongji University
The Highway Safety Manual (HSM) presents a variety of methods for quantitative network segmentation. Existing approaches to determine segment lengths for safety analysis require engineering judgement and are subject to a lack of standard metrics for assessing segmentation performance. This paper presents a novel methodology that determines optimal segment lengths and innovates network segmentation methods for reliable safety analysis. The methodology is based on spectral analysis of crash density in the spatial frequency domain (SFD) in which low frequency components represent trends while high frequency components represent details and randomness. By proposing the one-dimensional spatial frequency domain analysis (SFDA), this paper discovered the characteristic of power spectral concentration within the low frequency band. Based on this finding, this paper further proposes the power spectral segment length (PSSL) for determine optimal segment lengths and the power spectral percentage (PSP) for assessing the segmentation performance. The methodology extended the knowledge of network segmentation and aggregation of crash data from a non-traditional perspective. It leads to the low-pass filtering method that outperforms the sliding window method, and an improved wavelet-based method that identifies high-risk segments properly.
Sensitivity Analysis of Bayesian Semiparametric Spatial Crash Frequency Models
Wen Cheng, California State Polytechnic University, PomonaShow Abstract
Gurdiljot Gill, California State Polytechnic University, Pomona
Jiao Zhou, California State Polytechnic University, Pomona
Tom Vo, Southern California Association of Governments
Frank Wen, Southern California Association of Governments
This study focused on the sensitivity analysis of Bayesian semiparametric spatial models which combine the strengths of spatially structured random effects and the Dirichlet mixture to account for the unobserved heterogeneity of crash counts. The three-year bicycle crash data from the city of Irvine in California aggregated at the transportation planning level of Traffic Analysis Zones (TAZ) were utilized for model development. Various evaluation criteria were employed to compare the performance of models with varying spatial weight matrices and precision parameters (alpha). The results demonstrate that there exists strong correlation among the posterior number of clusters (K), alpha, the fraction of variation explained by the spatial random effect, and different evaluation criteria. Even though the increased upper bound value of alpha does not necessarily lead to the enhanced model performance, the models with the full flexibility to choose the desirable amount of clustering tend to perform better than those with limited flexibility due to smaller allowable mass components. Compared with the precision parameter, no obvious trend is illustrated for the different evaluation criteria along the varying spatial weight matrices. However, the existence of significant performance variation among the models suggests the need to explore various spatial neighboring structures for the potential better modeling results.
Identification of Secondary Crash Risk Factors Using Penalized Logistic Regression Model
Angela Kitali, Florida International UniversityShow Abstract
Priyanka Alluri, Florida International University
Thobias Sando, University of North Florida
Wensong Wu, Florida International University
Secondary Crashes (SCs) have increasingly been recognized as a major problem leading to reduced capacity and additional traffic delays. However, the limited knowledge on the nature and characteristics of SCs has largely impeded their mitigation strategies. There are two main issues with analyzing SCs. First, relevant variables are unknown, while at the same time, most of the variables considered in the models are highly correlated. Second, only a small proportion of incidents results in SCs, making it an imbalanced classification problem. This study aims to develop a reliable SC risk prediction model using the Least Absolute Shrinkage and Selection Operator penalized logistic regression model with Synthetic Minority Over-sampling TEchnique-Nominal Continuous. The proposed model is considered to improve the predictive accuracy of the SC risk model since it accounts for the asymmetric nature of SCs, performs variable selection, and removes correlated variables. The study data were collected on a 35-mile I-95 section for three years in Jacksonville, Florida. SCs were identified based on real-time speed data. The results indicated that real-time traffic variables and primary incident characteristics significantly affect the likelihood of SCs. The most influential variables included mean of detector occupancy, coefficient of variation of equivalent hourly volume, mean of speed, primary incident type, percent of lanes closed, incident occurrence time, shoulder blocked, number of responding agencies, incident impact duration, incident clearance duration, and roadway alignment. The study results can be used by agencies to develop SC mitigation strategies, and hence improve the operational and safety performance of freeways.
Determinants of Crash Type and Severity Using Generalized Structure Equation Modeling
Kyung(Kate) Hyun, University of Texas, ArlingtonShow Abstract
Suman Mitra, University of California, Irvine
Kyungsoo Jeong, Massachusetts Institute of Technology (MIT)
Yeow Chern Tok, University of California, Irvine
Crash type is an informative indicator to infer driving behaviors and conditions that cause a crash. In particular, rear-end and sideswipe crashes are typically caused by improper vehicle interaction such as sudden lane-changing or speed control while hit-object crashes are likely the result of single driver’s mistake. This study developed vehicle grouping measures to represent the vehicle interaction considering that the vehicles could affect each other when travelling as a group. Then, the effects of vehicle interaction on crash type and severity were investigated using Generalized Structure Equation Modeling (GSEM). The proposed GSEM captures the complex relationships among the various crash factors such as traffic condition, driver characteristics, environmental conditions, and vehicle interaction to the crash attributes including type and severity. Vehicle interaction and resulting driving behaviors are observed from microscopic traffic data. This study collected over 3 million individual vehicle data and matched to 1,360 crash reports. Results showed that the vehicle grouping measures have significant impacts on crash types. The proportions of vehicles forming a homogenous or heterogeneous group positively affect rear-end and sideswipe while speed difference in the heterogeneous group had a positive effect on hit-object crashes. In addition, truck involvement is identified as a significant influential factor for sideswipe crashes while human factors such as age and gender play important roles in all type of crashes. Crash severity was negatively affected by total flow, and rear-end were more likely to result severe crashes than hit-object crashes. Keywords: vehicle group, interaction, crash type, severity, GSEM
Before–After Analysis of Safety Effects of Variable Speed Limit System Using Full Bayesian Models
Ziyuan Pu, University of WashingtonShow Abstract
Xiaoyu Guo, Texas A&M Transportation Institute
Zhibin Li, Southeast University
Ying Jiang, University of Washington
Yinhai Wang, University of Washington
Chao Zhang, Tsinghua University
Variable speed limits (VSL) have been increasingly used to improve traffic safety on freeway mainlines. The primary objective of this study was to evaluate the safety impacts of the VSL system implemented on Interstate 5 in Seattle, United States since 2010. A Full Bayesian (FB) before-after analysis was conducted based on 9,787 crashes that occurred in a 72-month study period. The analysis was conducted for all crashes, crash severity levels, crash types and crash causes. The FB before-after results implied that the total crash count was reduced by 32.3% with a standard deviation of 3.58% after the implementation of VSL system on the target freeway. The decrease in number of no injury crashes is greater than the decrease in crashes with severe injury and possible injury. The effect with respect to reducing head-on, face and leading-end crashes was with the most beneficial among all crash types, while the effect on rear-end crash was the least. The study also compared the traffic speed features in the before and after periods in order to fully evaluate the impacts of the VSL system on traffic operations. The result indicated that, the difference in speed was apparently reduced with the VSL system deployed The results of this study are particularly valuable for policy making and cost-benefit evaluation associated with VSL system implementations.
Modeling the Effects of Lake-Effect, Snow-Related Weather Conditions on Daily Traffic Crashes: A Time Series Count Data Approach
Bandhan Dutta Ayon, Western Michigan UniversityShow Abstract
Benjamin Ofori-Amoah, Western Michigan University
Lei Meng, Western Michigan University
Jun-Seok Oh, Western Michigan University
Kathleen Baker, Western Michigan University
Winter weather in many parts of North America is characterized by heavy snowfall that affects traffic safety. Lake Effect Snow (LES) in the Great Lakes region exacerbates the problem by increasing snowfall totals and severity of winter weather locally. Past studies investigating the effects of winter weather on traffic crashes have mainly focused on site-specific weather conditions and overlooked mesoscale meteorological phenomena. Therefore, the primary objective of this paper is to develop a crash count model establishing the relationship between LES and winter traffic crashes. Daily crash data, traffic exposure data and meteorological data from State of Michigan are modelled to examine the impact of meteorological characteristics behind LES formation on the observed counts. Additionally, this paper introduces a relatively new class of time series models known as Negative Binomial Integer-valued Generalized Autoregressive Conditional Heteroscedastic (NB-INGARCH) model. NB-INGARCH offers an alternative to the integer-valued time series models and accounts for the overdispersion, non-negativity, and time interdependencies. The performance of the NB-INGARCH model is compared with Poisson INGARCH model using the Probability Integral Transformation (PIT) histogram, marginal calibration plot and scoring rules. The resultant models were quite similar in terms of coefficient estimates and goodness of fit. The results suggest that several predictor variables for LES formation are significantly related to crash data. However, NBINGARCH model exhibits better predictive performance than Poisson INGARCH by addressing overdispersion and unobserved heterogeneity issues.
Incorporating Spatial Effects into Temporal Dynamic of Traffic Fatality Risks: A Case Study on Lower States of the United States, 1975–2015
Hanchu Zhou, Central South UniversityShow Abstract
Fangrong Chang, Central South University
Pengpeng Xu, University of Hong Kong
Mohamed Abdel-Aty, University of Central Florida
Helai Huang, Central South University
Road traffic fatality rate has long served as a regular indicator to evaluate and compare road safety performances for different administrative divisions. This article introduced a novel method known as spatial Markov chains model to incorporate the spatial effects into the temporal dynamic of the fatality rates. Comparing with the traditional Markov chains model, the proposed spatial Markov chains model can quantify the influence of neighboring sites explicitly in the transition process. A case study using a long time span dataset from 1975 to 2015 in the 48 lower states of the United Sates was conducted to illustrate the proposed model. The fatality rates were measured as the number of traffic fatalities per 100 million vehicle miles or per 10,000 residents. Our results show that the probability of transition for one state between different levels of traffic fatality risks depends largely on the context of its surrounding neighbors. Another important finding is that relative to the estimates of traditional Markov chains model, states surrounded by neighborhoods with relatively low fatality rates takes a longer time to transform to a higher level of fatality risk in the spatial Markov chains model, whereas those with high risk neighborhoods takes less time to deteriorate. These findings confirm that it is imperative to incorporate spatial effects when modeling the temporal dynamic of safety indicators to assess and monitor the safety trends of the areas of interests.
Real-Time Crash Risk Prediction Using Long Short-Term Memory Recurrent Neural Network
Jinghui Yuan, University of Central FloridaShow Abstract
Mohamed Abdel-Aty, University of Central Florida
Yaobang Gong, University of Central Florida
Qing Cai, University of Central Florida
With the help of widely deployed traffic detectors along arterials and intersections, real-time traffic data are collected and updated in a very short time period, which enables us to conduct real-time analysis at signalized intersections. Among them, real-time crash risk prediction is one of the most promising and challenging research topics. This study attempts to predict real-time crash risk by considering time series dependency with the employment of Long Short-Term Memory Recurrent Neural Network (LSTM-RNN) algorithm. Also, the Synthetic Minority Over-Sampling Technique (SMOTE) was utilized in this study to generate a balanced training dataset for algorithm training. In comparison, a conditional logistic model was developed based on matched case control design. It is worth pointing out that both models were evaluated based on the real-world unbalanced test dataset rather than artificially balanced dataset. The comparison results indicate that the LSTM-RNN with SMOTE outperforms the conditional logistic model. The methods and findings of this study attempt to verify the feasibility of real-time crash risk prediction by using LSTM-RNN with over-sampled dataset (SMOTE).
Real-Time Crash Risk Analysis for Signalized Intersections
Jinghui Yuan, University of Central FloridaShow Abstract
Mohamed Abdel-Aty, University of Central Florida
This study attempts to investigate the relationship between crash occurrence at signalized intersections and real-time traffic, signal timing, and weather characteristics based on 23 signalized intersections in Central Florida. The intersection and intersection-related crashes were collected and then divided into two types, i.e., within intersection crashes and intersection entrance crashes. Bayesian conditional logistic models were developed for these two kinds of crashes, respectively. For the within intersection models, the model results showed that the through volume from “A” approach (the traveling approach of at-fault vehicle), the left turn volume from “B” approach (near-side crossing approach), and the overall average flow ratio (OAFR) from “D” approach (far-side crossing approach), were found to have significant positive effects on the odds of crash occurrence. Moreover, the increased adaptability for the left turn signal timing of “B” approach and more priority for “A” approach could significantly decrease the odds of crash occurrence. For the intersection entrance models, average speed was found to have significant negative effect on the odds of crash occurrence. The longer average green time and longer average waiting time for the left turn phase, higher green ratio for the through phase, and higher adaptability for the through phase can significantly improve the safety performance of the intersection entrance area. These results are important in real-time safety applications at signalized intersections in the context of proactive traffic management and adaptive signal control.
Improving Intersection Safety with RCUT: Louisiana Experience
Xiaoduan Sun, University of Louisiana, LafayetteShow Abstract
Ming Sun, University of Louisiana, Lafayette
M. Ashifur Rahman, University of Louisiana, Lafayette
Kenneth McManis, University of Louisiana, Lafayette
Donghui Shan, CCCC First Highway Consultants Co., Ltd
Destiny Armstrong, University of Louisiana, Lafayette
The safety of intersections on major corridors is always a concern because of the high-risk vehicle maneuvers intertwined with a high operating speed. It is especially a problem for intersections with a two-way stop-sign control where vehicles on a low-speed minor roadway must cross multiple lanes and the median before merging into the path of high-speed traffic. To improve the intersection safety, Restricted Crossing U-turns (RCUT) has been constructed in Louisiana since 2011. Because of the short history of the RCUT implementation, there are limited studies available that address the RCUT safety effectiveness. This paper evaluates the safety benefit of six RCUTs in Louisiana including five intersections in urban and suburban areas. Unlike the previous studies, this investigation covers both the RCUT intersection only and RCUT system (consisting of the intersection, two U-turns and segments in between). The crash analysis shows a 100% reduction in fatalities, 41.5% in injuries and 22.3% in property damage only crash for the RCUT intersection only, and less impressive reductions for the RCUT system. The review of the original crash reports greatly benefits the investigation on why the crashes increased at few locations, thus, provides the valuable information on how to correct these crash problems through the detailed design and traffic control. The safety improvement plus the high ratio of benefit to cost strongly demonstrate that the RCUT is an effective and economically justified countermeasure on high-speed roadways in both rural and urban areas.
Effect of Vehicular Defects on Crash Severity: A Bayesian Data Mining Approach
Subasish Das, Texas A&M Transportation InstituteShow Abstract
Anandi Dutta, Texas A&M University
Srinivas Geedipally, Texas A&M Transportation Institute
Chaolun Ma, Texas A&M University
Zachary Elgart, Texas A&M Transportation Institute
Vehicle defects have an adverse effect upon overall roadway safety. Although vehicles with safety and emission related issues are more prone to crash occurrences, the sensitivity of crashes to vehicle defects is minimal. The National Motor Vehicle Crash Causation Survey (NMVCCS), conducted from 2005 to 2007, showed that an estimated 44,000 crashes occurred due to vehicular defects—about 2 percent of the NMVCCS crashes. Louisiana is one of the states that has a vehicular safety inspection in place. However, the recent traffic crash statistics showed a higher percentage of vehicle defect related crash fatalities in Louisiana (around 3 percent of all traffic fatalities). This fact called for an in-depth analysis of the vehicle defect related crashes in Louisiana. This study used seven years (2010-2016) of traffic crash data from Louisiana to investigate the association between crash severity and vehicle defect types. A Bayesian data mining approach is applied to identify the key associations. The findings showed that vehicle age is associated with severe injury crashes. Worn tires and defective brakes are the over-represented vehicle defect categories. The Empirical Bayes Geometric Mean (EBGM) scoring method, which is used to determine the relationship between vehicle defects and crash severity types, produced several top rules that require further attention. The findings of this study can be used by different stakeholders to enhance roadway safety and reduce vehicular defect associated crashes.
Analysis of Factors Affecting the Frequency of Crashes on Interstate Freeways by Vehicle Type and Severity Incorporating Weather Prediction Models
Cristopher Aguilar, Northern Arizona UniversityShow Abstract
Brendan Russo, Northern Arizona University
Amin Mohebbi, Northern Arizona University
Simin Akbariyeh, California Polytechnic State University
Since the introduction of the interstate system in 1956, motorists have relied heavily on these roadways for both personal and commercial travel. However, the interstate system experiences a large number of traffic crashes which cause property damage, injuries, fatalities, and non-recurring delay. Understanding what causes these crashes at a system wide level is of vital importance for all users. This study utilized seven years of crash data from the State of Arizona, examining factors affecting the frequency of crashes with a focus on different vehicle types and simulated precipitation data. Vehicle type categories included passenger vehicles, freight vehicles, motorcycles, and buses/recreational vehicles/trailers. The study utilized statewide crash data along Arizona interstates including I-8, I-10, I-17, I-19 and I- 40, along with roadway geometric data and traffic data. Additionally the Weather Research and Forecasting (WRF) model was used to simulate precipitation data to analyze precipitation effects on crash frequency and provide an example of how this validated data can be used in traffic safety and operational management. Random parameters negative binomial models were developed for different vehicle types and crash severities, and the results show that several roadway and traffic variables, as well as precipitation, are associated with crash frequency and the results vary among different vehicle types and crash severities. Ultimately, the findings provide important insights into factors affecting interstate freeway crash frequency for different vehicle types, and may be useful in planning countermeasures in efforts to improve safety on these freeways.
Enhancing Real-Time Crash Risk Prediction Performance Considering Spatial and Temporal Correlations in Support Vector Machine
Tong Liu, Southeast UniversityShow Abstract
Zhibin Li, Southeast University
Ziyuan Pu, University of Washington
Chengcheng Xu, Southeast University
Meng Li, Southeast University
Unobserved heterogeneity in crash data could affect the predicting accuracy of crash risks. Such effects can be considered within the spatial and temporal correlation to improve the model prediction performance. This study aims at proposing an enhanced support vector machine (SVM) model that involves the spatial and temporal weight features in the model structure to address the spatial and temporal proximity in the real-time crash risk predictions. A total of 254 crash data on the Interstate 80 were obtained. Traffic flow data 5 min before the occurrence of each crash were extracted to be the case database. Non-crash traffic flow data were randomly extracted from the collision free periods to be the control database. The Receiver Operating Characteristics (ROC) curves were drawn to evaluate and compare the prediction performance of different models. The results showed that by incorporating the spatial and temporal correlations in the SVM, the model fitness was improved: the predicting accuracy was increased from 79.8% to 86.5% as compared to the basic SVM model. Two weight matrixes of spatial and temporal correlation in the SVM were tested, and the models with the 0-1 first order weight feature had the highest predicting accuracy. We also tested the modeling accuracy for different ratios of training and testing sample sizes. Findings of this study suggest that the proposed SVM model with the spatial and temporal correlation can effectively improve the predicting accuracy of real-time crash risks based on the traffic variables from loop detector stations.
Examining Multi-Layer, Perceptron-Based Machine Learning Method to Predict Imbalanced Sample of Traffic Crash
Chongyi Li, Tongji UniversityShow Abstract
Jinzhao zhou, South China University of Technology
Yichuan Peng, Tongji University
Jian Lu, Tongji University
This paper combined a data processing method with imbalanced sample distribution and a machine learning method based on multi-layer function approximator was employed to deal with the prediction of crash severity, especially when the sample size of the crashes is small. Severe injury and caused to death crashes are needed to be dedicated to avoid. However, few study focused on improving the prediction accuracy of the few but more devastating severe injury crashes. The purpose of this research is to improve the prediction accuracy of each level of severity of crashes. It can effectively reduce the severity of crashes and mitigate the harm caused by traffic crashes by combining the prediction results to take effective countermeasures. This research first analyzed the distribution of the severity of traffic crash injuries in California State in 2010. Seventeen important influencing factors were selected through spearman's correlation analysis. After that, the data was equalized and the multi-layer neuron network was applied to predict the severity of the crashes. Finally, the prediction results were compared with Support Vector Machine. It was shown from modeling results that the utilized sample distribution balancing processing method and multi-layer function approximators based machine learning method can be more efficient in predicting the severity of crash injuries.
An Empirical Analysis on Temporal Stability of Factors in Work Zone Crashes in Florida: A Random Parameters Heterogeneity-in-Means Approach
Mouyid Islam, USF Center for Urban Transportation ResearchShow Abstract
Chanyoung Lee, Center for Urban Transportation Research at USF
Work-zone crashes in Florida have increased recently, particularly from 2012 to 2017. This study investigates factors leading to work-zone crashes in Florida in two distinct economic time periods in Florida—the recession-induced period (2012–2014) and the post-recession period (2015–2017). The main focus of this study is to estimate two separate time period models focusing on injury severity of work-zone crashes with mixed logit model incorporated with a heterogeneity-in-means approach. The study examines the temporal stability of contributing factors in work-zone crashes considering two time periods with a log likelihood test. Marginal effects of individual parameter estimates on work-zone crash severity were assessed to study the temporal stability of the effect of individual parameters on the likelihood of work-zone crash severity. The variables extracted from Florida’s Crash Analysis Reporting System (CARS) encompass a wide variety of factors related to crash, vehicle, roadway geometry, traffic volume, driver demographics, spatial and temporal characteristics affecting the injury severity of work-zone crashes. The model results indicate significant temporal instability resulting from a possible complex interaction with macroeconomic conditions over the years from larger-scope and higher-budgeted work-zone projects in Florida with evolving driving behavior, traffic volume, and crash reporting practice in traditional state crash data. Mixed logit models on injury severity with a heterogeneity-in-means approach on work-zone crashes open a promising frontier of future research. This novel effort recognizes the possibility of uncovering complex interactions from underlying extensive and multiple data sources that otherwise expose the limitations of traditional crash databases and their management.
Safety Performance of Displaced Left Turn Intersections Case Studies in San Marcos, Texas
Yi Qi, Texas Southern UniversityShow Abstract
Qiao Sun, Texas Southern University
Qun Zhao, Texas Southern University
Tao Tao, Texas Southern University
Wenrui Qu, Qilu University of Technology
Intersections with the displaced left turn (DLT) design are innovative intersections that are designed to increase the mobility of vehicles by relocating the left turn lane (lanes) to the far-left side of the road upstream of the main signalized intersection. Since DLT is a relative new design and very limited crash data are available, previous studies have focused mainly on analysis of the design’s operational performance rather than its safety performance. To fill this gap, in this study we investigated the safety performance of two DLT intersections located in San Marcos, Texas. Crash data from 2011 to April 2018 were extracted from the TxDOT Crash Record Information System (CRIS). These crash data were analyzed using two different approaches, i.e., 1) statistical analysis and 2) collision diagram based analysis. The results of this study indicated that the DLT design has reduced conflicts related to left turns significantly. Also, some safety problems associated with traffic signage, geometric design, and access management of the DLT design also were identified. As a result of these analyses, recommendations were provided for safe implementation of the DLT design in the future.
Crash Severity Effects of Adaptive Signal Control Technology: Insights from Pennsylvania and Virginia
Zulqarnain H. Khattak, University of VirginiaShow Abstract
Michael Fontaine, Virginia Transportation Research Council
Jiaqi Ma, University of Cincinnati
Brian Smith, University of Virginia
Adaptive signal control technology (ASCT) is an intelligent transportation systems (ITS) technology that optimizes signal timings in real time to improve corridor flow. While several past studies have examined the impact of ASCT on crash frequency, little is known about its effect on injury severity outcomes. This paper used ordered probit models to estimate the injury severity outcomes resulting from ASCT deployment using 8 years of crash data from 42 intersections in Pennsylvania and 11 years of crash data from 49 intersections in Virginia. A unique aspect of this data was the availability of before and after deployment characteristics for two different ASCT technologies. The estimation results revealed that both ASCT systems were associated with a reduced propensity for injury crashes. The best fit model also revealed a similar trend towards reductions in severe crashes. This model performed well on validation data with low forecast error of 0.301 and was also observed to be spatially transferable. These results encourage the consideration of ASCT deployments at intersections with high crash severities and have practical implications for aiding agencies in making future deployment decisions about ASCT.
Predicting the Frequency of Secondary Crashes Caused by One Primary Crash Using Zero-Inflated Ordered Probit Regression
Chengcheng Xu, Southeast UniversityShow Abstract
Shuoyan Xu, Southeast University
Chen Wang, Southeast University
Jing Li, Southeast University
This paper aimed to investigate the effects of real-time traffic flow conditions on the frequency of secondary crashes caused by one primary crash on freeways. The zero inflated ordered probit (ZIOP) regression model was developed to link the probability of multiple secondary crashes after the occurrence of one primary crash with real-time traffic flow, geometric, weather and primary crash characteristics. The ZIOP regression model analyzed the probability of secondary crash frequency after one primary crash by separating it into two states. One is a secondary-crash-free state that determines whether the occurrence of a crash will lead to one or more secondary crashes, and the other is a secondary-crash-prone state that determines the secondary crash frequency caused by one primary crash. The average speed, average traffic volume, and the difference between the numbers of on-ramp and off-ramp are the significant variables in the secondary-crash-free state. In the secondary-crash-prone state, the significant variables affecting the probability of multiple secondary crashes include average detector occupancy, rainy weather, primary crash severity, and hit-and-run primary crash. The ROC curves were used to test predictive performance of the ZIOP model. The test results suggested that the ZIOP model provide reasonably good predictive accuracy of multiple secondary crashes caused by one primary crash.
Impact of Built Environment on the Severity of Vehicle Crashes Caused by Distracted Driving
Zhenhua Chen, Ohio State UniversityShow Abstract
Youngbin Lym, Ohio State University
This study evaluates the influences of built environment on the severity of vehicle crashes with focuses on a comparative analysis between the crashes caused by distracted driving and non-distracted driving. Using a comprehensive dataset with 1.4 million crash records in Ohio for the period 2013-2017 as an example, the relationships between built environments and the severity of vehicle crashes caused by distracted driving were examined using the generalized order logit regression method. The outcomes of severity analysis confirm that distracted driving related crashes tend to be more severe than non-distracted driving related crashes. In particular, the crashes by distracted driving were found to be much more severe if the accident occurs at work zones or on interstate highways. On the other hand, roundabout was confirmed effective in reducing crash severities in general with a more significant effect on mitigating severity for DD distracted driving related crashes.
Prediction and Factor Identification for Crash Severity: Comparison of Discrete Choice and Tree-Based Models
Xinyi Wang, Georgia Institute of Technology (Georgia Tech)Show Abstract
Sung Hoo Kim, Georgia Institute of Technology (Georgia Tech)
In the traffic safety area, crash severity is a widely studied topic, using various types of models. The aims of this study are twofold: (1) to identify factors contributing to crash severity, including road-environment factors, human factors, and vehicle factors, and (2) to compare the prediction performance and the interpretation ability of discrete choice and tree-based models. Specifically, we compare the multinomial logit (MNL) model and the random forest (RF) model. This study employs 2017 Maryland crash data, which are publicly available from the Department of Maryland State Police. The estimated models identify contributing variables such as collision type, occupant age, and speed limit. For the given dataset, RF outperforms MNL based on multiple measures (precision, recall, and F1-score). Two models indicate some variables that significantly affect crash severity such as collision type and vehicle body type. Based on sensitivity analyses, in general, MNL is more sensitive to the change of variables than RF. In addition, RF can automatically capture the nonlinear effects of continuous variables, reduce the influence of collinearity relationships existing among explanatory variables, and automatically consider variable interactions.
Applications of Measurement Error Correction Approaches in Statistical Road Safety Modeling
Anusha Musunuru, Kittelson & Associates, Inc. (KAI)Show Abstract
Richard Porter, VHB
Road safety modelers frequently use average annual daily traffic (AADT) as a measure of exposure in regression models of expected crash frequency for road segments and intersections. Recorded AADT values at most locations are estimated by state and local transportation agencies with significant uncertainty, often by extrapolating short-term traffic counts over time and space. This uncertainty in the traffic volume estimates, often termed in a modeling context as measurement error in right-hand-side variables, can have serious effects on model estimation, including: 1) biased regression coefficient estimates, and 2) increases in dispersion. The structure and magnitude of measurement error in AADT estimates are not clearly understood by researchers or practitioners, leading to difficulties in explicitly accounting for this error in statistical road safety models, and ultimately in finding solutions for its correction. This study explores the impacts of measurement error in traffic volume estimates on statistical road safety models by employing measurement error correction approaches, including Regression Calibration and Simulation Extrapolation. The concept is demonstrated using crash, traffic, and roadway data from rural, two-lane horizontal curves in the State of Washington. The overall results show that the regression coefficient estimates with a positive coefficient were larger and a negative coefficient were smaller (i.e., more negative) when the measurement error correction methods were applied to the regression models of expected crash frequency. Future directions in applications of measurement error correction approaches to road safety research are provided.
A Taxonomy of Naturalistic Driving Errors and Violations and Its Variations Across Different Land Use Contexts: A Path Analysis Approach
Asad Khattak, University of Tennessee, KnoxvilleShow Abstract
Behram Wali, University of Tennessee, Knoxville
Numan Ahmad, University of Tennessee
Driver errors and violations are highly relevant to the safe systems approach as human error tends to dominate crash occurrence, contributing to a good 80% to 90% of crashes. To understand errors and the contexts in which errors and violations occur, this study harnessed unique data from the Naturalistic Driving Study (NDS)-SHRP2. A systematic taxonomy is first developed to classify driver errors and violations based on their presence during the perception-reaction process and to analyze their contribution in safety critical events. The NDS data provides a unique opportunity to observe pre-crash behaviors of drivers in diverse spatio-temporal contexts. Given safety critical events such as crashes and near-crashes, recognition errors were predominant in almost all types of locations. A rigorous multinomial logit and ordered probit based path analysis technique is applied to conceptualize the direct relationships between key built-environment factors and crash propensity, as well as the indirect relationships between built environment factors and crash propensity through the mediating errors and violations. The empirical framework allows us to explore certain land-use and roadway environments associated with different types of errors along with their direct and indirect effects on crash propensity. Detailed results are discussed in the paper, along with implications.
Multivariate Copula Modeling of Intersection Crash Consequence Metrics: A Joint Estimation of Injury Severity, Crash Type, Vehicle Damage, and Driver Error
Kai Wang, University of ConnecticutShow Abstract
Tanmoy Bhowmik, University of Central Florida
Shamsunnahar Yasmin, University of Central Florida
Shanshan Zhao, University of Connecticut
Naveen Eluru, University of Central Florida
Eric Jackson, Connecticut Transportation Safety Research Center
This study employs a copula-based multivariate ordered probit model to simultaneously estimate the four common intersection crash consequence metrics – driver error, crash type, vehicle damage and injury severity – by accounting for potential correlations due to common observed and unobserved factors. To this end, a comprehensive literature review of relevant studies was conducted; four different cupula model specifications including Frank, Clayton, Joe and Gumbel were estimated to identify the dominant factors contributing to each crash consequence indicator; and specific countermeasures were recommended for each of the contributing factors to improve the intersection safety. The model goodness-of-fit illustrates that the Joe copula model with the parameterized copula parameters outperforms the other models, which verifies that the injury severity, crash type, vehicle damage and driver error are significantly correlated due to common observed and unobserved factors and, accounting for their correlations, can lead to more accurate model estimation results. The parameterization of the copula function indicates that their correlation varies among different crashes, including crashes that occurred at stop-controlled intersections and four-leg intersections and crashes which involved drivers younger than 25. The model coefficient estimates indicate that the driver’s age and gender, driving under the influence of drugs and alcohol, intersection geometry and control types, adverse weather and light conditions and the vehicle type are the most critical factors contributing to severe crash outcomes. It is anticipated that this study can shed light on identifying intersection safety issues, and help develop effective countermeasures to improve intersection safety.
Contributing Factors for Focus Crash Types and Facility Types
Taha Saleem, UNC Highway Safety Research CenterShow Abstract
Richard Porter, VHB
Raghavan Srinivasan, University of North Carolina, Chapel Hill
Daniel Carter, UNC Highway Safety Research Center
Scott Himes, VHB
Thanh Le, VHB
This paper describes efforts to identify focus crash types, focus facility types, and associated crash contributing factors to inform applications of systemic safety improvements. Systemic safety improvements—when selected and targeted appropriately—provide a tremendous opportunity to proactively reduce crashes and their resulting harm. The main objectives of this study are to (1) select reliable and applicable data resources, statistical methodologies, analysis procedures, and tools, (2) conduct data analysis to identify and validate focus crash types and facility types and their associated contributing factors, and (3) identify potential low-cost safety strategies that may effectively be used as systemic safety improvements. The study used intersection data from Washington and Ohio and non-intersection data from California and Ohio. The data were enhanced with information from the National Oceanic and Atmospheric Administration and US Census Bureau. Random forest algorithm was adopted to analyze the data and identify the contributing factors to the focus crash types and facility types. The roadway factors uncovered by the analysis as influencing the frequencies of the different crash types were generally consistent with what was expected based on previous research and existing practice. Findings related to the socioeconomic and weather-related factors showed promise, but there is not yet a significant amount of theory to support or refute the socioeconomic- and weather-related results of this effort. A six-step countermeasure selection process is also identified to use the contributing factor findings to assist safety practitioners in making informed choices regarding countermeasures to address the focus crash types.
A Multimodal Approach for Monitoring Driving Behavior and Emotions
Arash Tavakoli, University of VirginiaShow Abstract
Arsalan Heydarian, University of Virginia
Vahid Balali, California State University, Long Beach
Studies have indicated that emotions can significantly be influenced by environmental factors; these factors can also significantly influence driver’s emotional state and, accordingly, driving behavior. Furthermore, as the demand for autonomous vehicles is expected to significantly increase within the next decade, a proper understanding of the driver/passenger(s)’ emotions, behavior, and preferences will be needed in order to create an acceptable level of trust with humans. This paper proposes a novel semi-automated approach for understanding the effect of environmental factors on driver’s emotions and behavioral changes through a naturalistic driving study. This setup includes a frontal road and facial camera, smart watch for tracking physiological measurements, and a Controller Area Network (CAN) serial data logger. The results suggest that the driver’s emotion is highly affected by the type of road, presence of a passenger, and weather condition, which potentially can change the driving behaviors. For instance, by defining emotions metrics as valence and engagement, there exist significant differences between human emotion in different weather conditions and road types. Participant’s engagement was higher in rainy and clear weather compared to cloudy weather. Moreover, his engagement was higher in city streets and highways compared to one lane roads and two lane highways. In addition, presence of a passenger increases the amount of engagement of the driver.
Identification and Prediction of Severity-Based Crash Hot Spots for Occupants of Different Age Groups in Various Time Intervals of a Day
Somayeh Mafi, Florida A&M UniversityShow Abstract
Yassir Abdelrazig, Florida A&M University
The identification and prediction of crash hotspots is an essential task in the highway safety management, particularly when highway officials have a limited budget for roadway mitigations. Implementing suitable methods for crash hotspot identification and prediction can result in the efficient employment of federal, state and local government resources for enhancing transportation safety. This paper aims to conduct GIS-based hotspot analysis to identify the crash-prone locations for various occupant age groups during different time intervals of a day and predict the location of these hotspots using statistical and machine learning models. For this purpose, first, the crash-prone locations for different occupant age groups and various time intervals of a day (twelve combinations) were identified by using severity-weighted crash hotspots analyses on a case study in Tampa Bay region (Florida, District 7). Since the number of crash hotspots in each dataset was so limited compared to non-hotspots, undersampling was used in order to adjust the class distribution of each dataset before implementing the classifiers. Then, binary logit models (BLM) were implemented to predict crash hotspots and investigate the influence of a range of parameters on the probability of creating a crash hotspot. In the end, the prediction performance of BLMs was compared with the C4.5 machine learning models. Results showed that C4.5 machine learning models outperformed BLMs in accurately predicting crash hotspots. Moreover, the models displayed substantial differences in crash hotspot determinants and their coefficients across the occupants’ age groups and time intervals of a day.
Alternative Model Structures for Multivariate Crash Frequency Analysis: Comparing Simulation-Based Multivariate Model with Copula-Based Multivariate Model
Tanmoy Bhowmik, University of Central FloridaShow Abstract
Moshiur Rahman, University of Central Florida
Shamsunnahar Yasmin, University of Central Florida
Naveen Eluru, University of Central Florida
In safety literature, there are two ways to incorporate the potential correlation between multiple crash frequency variables: (1) simulation-based approach and (2) analytical closed form approach. The current research effort proposed a comparison between simulation-based multivariate model and copula-based closed form approach to analyze zonal level crash counts for different crash types. The empirical analysis is based on traffic analysis zone (TAZ) level crash count data for both motorized and non-motorized crashes from Central Florida for the year 2016. A comprehensive set of exogenous variables including roadway, built environment, land-use, traffic, socio-demographic and spatial spillover characteristics are considered for the analysis. The resulting data fit and prediction performance offered by the copula-based approach clearly highlights the copula-based approach’s superiority over the simulation-based multivariate model. The applicability of the model for hot zone identification is illustrated by generating plots identifying hot and cold zones by crash type in the Central Florida region.
Analyzing Automated Vehicle Crashes in California: Application of a Bayesian Binary Logit Model
Alexandra Boggs, University of Tennessee, KnoxvilleShow Abstract
Asad Khattak, University of Tennessee, Knoxville
Behram Wali, University of Tennessee, Knoxville
Automated vehicles (AVs) represent an opportunity to reduce the number of crashes by eliminating driver error as safety studies reveal human error contributes in 94% of crashes. However, existing literature lacks an understanding of the contributing factors of AV crashes. To provide insights on these crashes, this study created a unique database from California Department of Motor Vehicles (DMV) 66 manufacturer-reported Traffic Collision Reports (OL 316). The gathered information includes text mining of narratives in the reports and answers to close-ended crash questions. Results indicate that AV technology was faulty once of the 66 crashes (1.52%); the most frequent AV crash type is rear-ended (58%; N=38)—but in all cases, except one manually driven AV, the AV was struck by a conventional vehicle. This noteworthy outcome motivated us to analyze rear-end collisions by estimating assorted Bayesian models rigorously. The results indicate that most AV collisions occurred in the fully automated mode (65.2%), and the odds of AVs being struck were higher compared to vehicle takeover before impact and conventionally driven vehicles. Furthermore, the odds of an AV being rear-ended were substantially higher at an intersection than any other location, owing to the complexity of movements and conflicts at intersections. Given a crash, AV-involved rear-end crashes were more likely on one-way streets and when AVs were in motion. Within the constraints of the available data, the results highlight risk factors, given AV-involved crashes on public roadways. This study helps us understand the interactions of AVs and human-driven conventional vehicles in complex urban environments.
Determination of the Driver At-Fault Using Possibility Theory-Based Classification
Shabnam Nazmi, North Carolina A&T State UniversityShow Abstract
Saina Ramyar, North Carolina A&T State University
Abdollah Homaifar, North Carolina A&T State University
With the advent of driving assistant systems and the emerging capabilities for analyzing large amounts of data, various driving-related problems are revisited in the past decade. Determining the driver at-fault is one aspect that has traditionally been handled based on expert evaluations and state laws. However, integrating expert knowledge with the information available in measurements from driving experiments make it possible to exploit both sources of information simultaneously. In this study, a possibility theory-based classifier, namely possibility rule-based classifier using function approximation, is employed to capture the uncertainty in expert knowledge due to incompleteness. In this approach, a model is inferred from the 100-Car naturalistic driving dataset that demonstrates the uncertainty which is inherent in making decisions based on expert evaluations. In this experiment, the objective is to predict a degree to which each driver is at-fault in rear-end collision events. It is shown that the proposed approach can efficiently utilize the expert information and provide a graded fault evaluation for each driver engaged in the accident. This graded evaluation can either be used for further interpretations by an expert or utilized to determine the most plausible prediction.
Investigating the Effect of Driver, Vehicle, and Road-Related Factors on Location-Specific Crashes Using Naturalistic Driving Data
Grace Ashley, Louisiana State UniversityShow Abstract
Osama Osman, Virginia Polytechnic Institute and State University
Sherif Ishak, Old Dominion University
Julius Codjoe, Louisiana Department of Transportation and Development
According to NHTSA, traffic accidents cost the country billions of U.S. dollars each year. Intersection accidents alone account for 23% of the 32,675 motor crash deaths in 2014. With the advent of the largest naturalistic driving dataset in the US collected by the SHRP2 NDS project, this study performs a crash-only analysis to identify driver, vehicle and roadway-related factors that affect the driving risk at different location types using a machine learning tool. The second objective is to analyze the most important factors obtained from the machine learning analysis to identify how it affects crash risk. The results showed that the order of importance of variables was driver behavior, locality, lane occupied, alignment and through travel lanes. Also, drivers who violated traffic signals were 4 times more likely to be involved crash than drivers who did not. Those who violated stop signs were 2 times more likely to be involved in crashes than those who did not. Drivers performing visual-manual tasks at uncontrolled intersections were 2.7 times more likely to be involved in crashes than those who did not engage in these tasks. At non-intersections, drivers who performed visual-manual tasks were 3.4 times more likely to be involved in crashes than drivers who did not. These findings add to the evidence that the institution of safety awareness programs geared towards intersection safety is imperative.