Benefit-Cost Approach for Using Continuous Friction Measurements to Choose a Pavement Surface Treatment
Ross McCarthy, Virginia Polytechnic Institute and State University (Virginia Tech)Show Abstract
Gerardo Flintsch, Virginia Polytechnic Institute and State University (Virginia Tech)
Edgar de Leon Izeppi, Virginia Tech Transportation Institute
Motor vehicle crashes continue to be a leading cause of death in the United States, which makes them the primary driving force behind the development of highway safety programs that aim to reduce related fatal and serious injuries. A key component of these safety programs is that they are driven by safety data, which must be collected and analyzed in order to identify and choose sites and countermeasures that reduce crash risk in a cost-effective manner. Skid resistance is an example of a parameter that can be routinely monitored with testing equipment and used as a measure of safety. The relationship between crash risk and skid resistance is well grounded in decades of literature and serves as the foundation for establishing investigatory levels of skid resistance. Investigatory levels are used in pavement friction management programs found in countries such as Australia, New Zealand, and the United Kingdom to identify low-friction sites with a high potential for skidding crashes. Low-friction sites are investigated to determine the need for treatment. For the investigation, friction measurements are included in regression models, such as safety performance functions (SPFs), and used along with the Empirical Bayes (EB) method in a benefit-cost analysis to predict the potential benefits of improving friction with surface treatment at low-friction sites. In this paper, it is shown that a pilot application of this methodology on a small network can predict crash reductions and assist in choosing friction improvement treatment.
Modeling the Severity of Rural Run-Off-Road Crashes with Latent Class Analysis: Accounting for Differences in Driver Behavior
Shiyi Liu (firstname.lastname@example.org), University of VirginiaShow Abstract
Tianyang Han, University of Tokyo
Michael Fontaine, Virginia Transportation Research Council
Run-off-road (ROR) crashes represent the majority of fatalities on rural roadways. Identifying factors contributing to ROR crashes and their influence on injury severity has been a significant focus of past studies. Typically, past studies of ROR crashes have assumed that crashes are the outcomes of a variety of factors related to roadway design, traffic operations, pavement conditions, and environmental characteristics, while driver behavior factors have often been neglected. In this study, driver behavior is examined using 7 crash report fields pertaining to the driver: improper actions, driving speed, defects, distraction, safety equipment deployment, and alcohol/drug usage. Considering both conventional factors used in previous studies and driver behavior factors, the crash severity of ROR crashes are modeled. After comparing various modeling techniques, a combination of Multinomial Logit (MNL) and Latent Class Analysis (LCA) are employed in this study. The crash data from rural areas of Virginia from 2019 are utilized for model development. The MNL model estimations show the significance of the driver behavior impact on ROR severity. Results also show that the combined LCA and MNL models can effectively discover the underlying patterns behind crash data that are not apparent from the MNL model developed using the entire dataset. The LC-MNL model identified two distinct classes of events, each of which had distinct factors that influenced crash severity. The findings from these classes and their implications for countermeasure selection are discussed.
Investigating the Impact of Road Cross-Section Elements on Crash Occurrence in Urban Areas
Muhammad Khattak, Universiteit GentShow Abstract
Hans De Backer, Universiteit Gent
Pieter De Winne, Universiteit Gent
Tom Brijs, Universiteit Hasselt
Ali Pirdavani (email@example.com), Hasselt University
Appropriate roadway cross-section design is critical due to its impact on safety, capacity, and function of the facility. While it is generally straightforward to assess this impact on capacity and function, it is not always easy for safety evaluation. Literature shows contradictory observations concerning the complex relationship between roadway cross-section elements and crashes, particularly in urban areas. Another important issue is the presence of on-street parking and their safety implications in urban areas. In the current study, safety performance functions were developed to investigate the impact of roadway cross-section elements and on-street parking on crash occurrence using negative binomial distribution framework. A database consisting of six-year crash records, traffic data, and road geometry of urban roads of Antwerp, Belgium was created for modeling. This paper reports how cross-section elements, on-street parking, and exposure contribute to crash occurrence in urban areas and discusses whether the results could be used to improve safety performance of road segments. The results indicated that the effects of number of lanes, segment length, and traffic volume on crash occurrence were significant while that for lane width was not. Parking variable (parking arrangement) was significantly related to “injury”, and “injury & fatal” crashes. Roads with higher number of lanes experience more crashes than roads with fewer lanes. Roads with parking were more prone to injury & fatal crashes than no parking settings. To conclude, these findings showed that road cross-section elements and parking settings play an important role in crash occurrence on road segments in urban areas.
A Comparative Approach of Crash Frequency Modelling in Two Lane Rural Roads
Moataz Bellah Ben Khedher, Korea UniversityShow Abstract
Dukgeun Yun, Korea Institute of Civil Engineering and Building Technology (KICT)
Road accidents are now one of the leading causes of death in the world. Investigating the underlying factors that contribute to increased risk of these accidents is an essential procedure to take effective countermeasures. In this study, we take a particular interest in two lane rural roads in South Korea. Six count data regression models were developed and evaluated for the goodness of fit. Traditionally, the evaluation is performed using information criterion such as Akaike Information Criterion. In this research, assessment of different models performances was carried using additional methods that include machine learning techniques, i.e. data splitting, and graphical tools, i.e. rootgrams . Based on the results of every evaluation technique, negative binomial hurdle model clearly outperformed all other regression models. Therefore, three variables were identified to have a significant impact on crash occurrence in two lane rural roads. These features are safety barrier, shoulder width and Annual Average Daily Traffic.
Modeling Vehicle Collision Instincts Over Road Midblock Using Deep Learning
Shubham Patil, Sardar Vallabhbhai National Institute of TechnologyShow Abstract
Narayana Raju, Sardar Vallabhbhai National Institute of Technology
Shriniwas Arkatkar (firstname.lastname@example.org), Sardar Vallabhbhai National Institute of Technology
Said Easa, Ryerson University
The study aimed to understand the vehicle safety of heterogeneous (mixed) and homogenous traffic flow over road midblock. In addressing the limitations of existing safety frameworks, the paper proposes a new safety framework that includes the collision instincts caused by the surrounding vehicles using the conventional time-to-collision (TTC) measure. An automated trajectory data tool is developed using advanced image processing concepts to generate trajectory data over the study sections. In the proposed framework, the lateral movement of vehicles is accurately modeled using deep learning. Further, the proposed framework is tested using the developed trajectory datasets. The results show that, in mixed traffic, the collision points occur over the entire study section. In the case of homogeneous traffic, the collision instincts are clustered toward the median lanes. With advanced technologies, trajectory data can be implemented in real-time within the proposed safety framework. The application of the proposed methodology can flag critical areas over the road network for better treatment to improve road safety.
A Proactive Safety Approach to Assess Overtaking Behavior and Crash Risk of Drivers under Time Pressure Situations
Nishant Pawar, Indian Institute of Technology, BombayShow Abstract
Nagendra Velaga (email@example.com), Indian Institute of Technology, Bombay
The aim of the current study is to assess the driving behavior, overtaking and crash probabilities of drivers during a car-following situation. Three different time pressure conditions (i.e., No Time Pressure (NTP), Low Time Pressure (LTP), and High Time Pressure (HTP)) were considered to analyze driving behavior during car-following and overtaking as well as crash probabilities. Minimum Time-to-Line Crossing (TLC) and Coefficient of Variation in Speed (CVS) were considered to examine driving behavior while following the lead vehicle. Further, minimum TLC and CVS were considered as explanatory variables to explore their influence on overtaking and crash probabilities. Minimum TLC was modelled using parametric survival analysis. CVS, overtaking and crash probabilities were modelled using Generalized Linear Mixed Models. The results showed that minimum Time-to-Line crossing reduced by 36.7% and 63.8% in LTP and HTP driving conditions, respectively. The Coefficient of Variation in Speed increased by 3.437% in HTP (no significant effect in LTP). The drivers using a car for work purpose and non-professional drivers showed aggressive driving behavior with low minimum Time-to-Line Crossing and Coefficient of Variation in Speed. The increase in overtaking probability (with time pressure) exposed drivers to greater collision risks which increased the likelihood of crashes. In general, male drivers showed more risky driving decisions than female drivers under time pressure conditions. However, it was observed that female drivers were more prone to crashes than male drivers. Overall, the results suggest that drivers take more risk to complete the driving task under time pressure conditions.
Secondary Crash Identification Using Crowdsourced Waze User Reports
Zhihua Zhang, University of Tennessee, KnoxvilleShow Abstract
Yuandong liu, Oak Ridge National Laboratory
Lee Han, University of Tennessee
Brad Freeze, Tennessee Department of Transportation
Secondary crashes are considered to be crashes that occur as a result of the noncurrent congestion originating from primary crashes, which always has a greater impact on safety and traffic than a single crash. A better understanding of secondary crashes would benefit traffic incident management, and this requires accurate identification of secondary crashes. In this study, we explored using crowdsourced Waze user reports to identify secondary crashes. A network-based clustering algorithm was proposed to extract the primary crash cluster, including all user reports originating from the primary crash, and any crash that occurred within the cluster would be the secondary crash. This method worked as a filter to select accurate primary-secondary relationships, thus identifying the exact secondary crashes. Then, we performed a case study for crashes occurring from June to December 2019 on a 30-mile stretch of I-40 in Knoxville. A static threshold method (crash duration and 10 miles), was used to pre-select the potential primary-secondary crash pairs. We pre-selected 75 out of 708 crashes as potential secondary crashes. Based on the pre-selected primary-secondary crash pairs, 17 secondary crashes were obtained with our method. We compared the results of our method with one of the commonly used methods, the speed contour plot method. Though our method captured fewer secondary crashes, it did identify several secondary crashes that could not be observed with the speed contour plot method. The results showed the applicability of our method and the potential of crowdsourced Waze user reports.
TWO-LANE HIGHWAY CRASH SEVERITIES: CORRELATED RANDOM PARAMETERS MODELING VERSUS INCORPORATING INTERACTION EFFECTS
Ahmed Farid, University of WyomingShow Abstract
Anas Alrijjal, University of Wyoming
Khaled Ksaibati (firstname.lastname@example.org), University of Wyoming
Two-lane highways represent the majority of highways in the U.S. and their safety is of a critical concern. Even though road safety researchers intensively evaluated two-lane highway safety, past studies were challenged by a methodological hindrance, namely that of correlated random parameters modeling methods. Random parameters models capture unobserved heterogeneity effects of crash contributing factors while correlated random parameters models offer the additional benefit of taking into account correlations among variables inducing such unobserved heterogeneity effects. However, correlated random parameters models do not permit excluding statistically insignificant variables describing cross-correlations among specific regressors. Therefore, in this research, both the correlated random parameters ordinal probit model and the uncorrelated random parameters ordinal probit model with interaction effects were compared, in terms of fit, when assessing the injury severity risks of two-lane highway crashes in Wyoming. With that, both models captured the combined effects of parameters. The results of this research indicated that the latter model exhibited a better fit. Furthermore, speeding, head-on collisions, sideswipe opposite-direction collisions, intersecting direction collisions, motorcycle involvement, impaired driving, distracted driving, the interaction effect of speeding with motorcycle involvement, that of head-on collisions with impaired driving and that of head-on collisions with commercial vehicle involvement all raised the likelihood of severe injuries. On the other hand, leaving the scene of the crash, proper seatbelt use, wet road surfaces and the interaction effect of impaired driving with motorcycle involvement were attributed to reduced severe injury risk. Mitigation measures were recommended based on this research’s findings.
Crash Injury Severity Prediction and Analysis Based on Comparison of Four Machine Learning Methods
Yanyan Chen, Beijing University of TechnologyShow Abstract
Yuntong Zhou, Beijing University of Technology
Jianming Ma, Texas Department of Transportation
Traffic crashes remain a major concern and challenge in countries worldwide. Numerous research studies have attempted to identify contribution factors of traffic crashes and reduce traffic crash injury severity. This paper compares four machine learning (ML) models to predict traffic crash injury severity and investigates factors affecting injury severity of crashes. The analysis uses 2015 crash data from police records in Beijing, China, including mobile location data and points of interest (POIs) data from Google Maps. The model results suggest that the Extremely Randomized Trees (ET), and the Random Forest (RF) perform similarly better compared to the Gradient Boosting Decision Tree (GBDT) and the Adaboost. Meanwhile, the highest overall detection accuracy was found to be 78.07%, while the recall was 42.84% using the ET model when all the features were included in the model. Specifically, minor injuries are the most difficult to distinguish and predict. Moreover, the density of roadway networks, resident population, employment, as well as the density of supermarkets were found to be factors that have a significant impact on traffic crash injury severity. As a result, based on the findings of this study, several countermeasures are recommended.
When and Where does the Next Crash Occur? A Discretized Duration Based Modeling Approach
Rajesh Paleti Ravi Venkata Durga (email@example.com), University of Texas, AustinShow Abstract
Asif Mahmud, Pennsylvania State University
Vikash Gayah, Pennsylvania State University
Abdul Pinjari, Indian Institute of Science
This paper develops real-time crash prediction modeling system that seeks to predict the time until the next crash is expected along a roadway facility. The proposed model incorporates both the parsimonious parametric probability structure of hazard duration models and the flexibility of discrete choice models for incorporating time-varying covariates and unobserved heterogeneity. Another useful feature of the proposed model is that it does not require case-control or sampling of time-intervals that are ubiquitous in existing real-time crash prediction models. The model was applied to estimate inter-crash durations along I-405 in California. The statistical fit measures on 80% estimation sample as well as validation metrics and policy analysis on 20% test sample demonstrate the practical applicability of the proposed modeling system.
Exploration of Various Spatio-temporal Interactions in Crash Frequency Models
Mankirat Singh, California State Polytechnic University, PomonaShow Abstract
Wen Cheng, California State Polytechnic University, Pomona
Yihua Li, Central South University of Forestry and Technology
Dean Samuelson, California Department of Transportation (CALTRANS)
Edward Clay, California State Polytechnic University, Pomona
Extensive research efforts have been put forth to improve the prediction of crash frequencies by employing the spatio-temporal models. Despite the large number of studies exploring various spatial and temporal effects, there is still a lack of conclusive findings related with the performance of spatio-temporal interactions. The current study bridges the gap by performing a comprehensive comparison of different spatio-temporal interactions with distinct temporal treatments in crash frequency models. Fifteen spatio-temporal models were developed which can be clustered from different perspectives: (1) three groups of models based on temporal treatments containing linear time trend, autoregressive-1 (AR1), and random walk-1 (RW1); (2) models with and without spatio-temporal interaction terms; (3) four different types of interactions including both structured and unstructured spatial or temporal random effects. To estimate the model parameters, the present study employed a fast Bayesian inference approach, or, Integrated Nested Laplace Approximation (INLA). The predictive accuracy of alternative models was assessed by employing various evaluation criteria which include deviance information criterion (DIC), log pseudo marginal likelihoods (LPML), and Probability Integral Transform (PIT). The results illustrated that the models with spatiotemporal interaction perform better than the models without spatiotemporal interactions. The dynamic temporal effects, RW1 and AR1, were found to perform almost the same, with both being superior to the non-dynamic parametric linear trend. With respect to the average performance among all interactions, the interaction of both unstructured spatial and temporal effects was found to outperform others.
Reliability-based Assessment of Potential Risk for Lane Changing Maneuvers
Yang-Jun Joo, Seoul National UniversityShow Abstract
Ho-Chul Park, Myongji University
Seung-Young Kho, Seoul National University
Dong-Kyu Kim (firstname.lastname@example.org), Seoul National University
Despite the urgent need for continuous risk assessments during autonomous driving, achieving reliable assessment results is still challenging due to the unpredictable behaviors of adjacent human drivers and the resultant complexity. Such complexity especially increases during lane changes, because several vehicles need to interact with other vehicles. This paper proposes a new framework to analyze lane changing risk on freeways considering the forecastability in adjacent vehicles. Virtual lane change scenarios are constructed based on historical maneuvers in adjacent vehicles, and the risk of potential lane change is evaluated through the safety evaluation result of the scenario. Adjacent vehicles’ future maneuvers are predicted using a multivariate Bayesian structural time series (MBSTS) model, and the forecastability is estimated as the standard error of the predicted values. The failure probability of those lane changing scenarios is obtained through the first-order reliability method (FORM) assuming that failure occurred when any time to collisions (TTC) value for adjacent vehicles was less than a threshold during the lane change. This study compares six scenarios with different levels of uncertainty to show the effect of uncertainty in the level of risk. The proposed framework differentiates itself from existing methods by estimating higher risk in more significant uncertainties in adjacent vehicles. It is expected that the outcome of this study will be valuable in developing reliable lane change strategies in autonomous driving.
Spatio-Temporal Accident Prediction: Effects of Negative Sampling on Understanding Network-Level Accident Occurrence
Jeremiah Roland, University of Tennessee, ChattanoogaShow Abstract
Peter Way, University of Tennessee, Chattanooga
Mina Sartipi, University of Tennessee, Chattanooga
Osama Osman, University of Tennessee, Chattanooga
In projects centered around rare event case data, the challenge of data comprehension is greatly increased due to insufficient data for deriving insight and analysis. This is particularly the case with traffic accident occurrence, where positive events (accidents) are rare with, in most cases, no data set existing for negative events (non accidents). One method to increase available data is negative sampling. In this work, four negative sampling techniques are presented with varying ratios of negative to positive data. These types of techniques are based on spatial, temporal, and a mixture of the two types of data, with the data ratios acting as class balancing tools. The best performing model found was with a negative sampling technique that shifted temporal information and had an even 50/50 data split, with an F-1 score of 93.68. These results are promising for ITS applications to inform of potential accident locations in an entire area for proactive measures to be put in place.
Utilizing Transfer Learning for Temporal and Spatial Transferability of Real-Time Crash Prediction Models
Cheuk Ki Man (email@example.com), Loughborough UniversityShow Abstract
Mohammed Quddus, Loughborough University
Athanasios Theofilatos, Loughborough University
Real-time crash prediction is a heavily studied area given its potential in proactive traffic safety management. A plethora of statistical and machine learning models have been developed to predict crashes in real-time. However, limited studies have been conducted to assess and improve the transferability of models. This paper attempts to address this gap by combining Generative Adversarial Network (GAN) and transfer learning, in order to examine the transferability of real-time crash prediction model under an extremely imbalanced data setting. Initially, a baseline model was developed using Deep Neural Network with crash and macroscopic traffic data collected from M1 Motorway in United Kingdom in 2017. The dataset utilized in the baseline model is naturally imbalanced with 257 crash cases and 16,359,163 non-crash cases. To overcome data imbalance issue, GAN was utilized to generate synthetic crash data. Non-crash data were randomly undersampled due to computational limitations. Weights obtained from the trained model were then fitted to five other datasets obtained from M1 (2018), M4 (2017 & 2018 separately) and M6 Motorway (2017 & 2018 separately) using transfer learning. Transferability results were compared with direct transfer through testing from the baseline model. This study revealed that direct transfer is not feasible. However, models become transferable temporally, spatially and spatio-temporally under transfer learning. The predictability of the transferred models is satisfactory with high AUCs ranging between 0.76 to 0.87 relative to existing studies. The best transferred model can predict nearly 80% crashes with 20% false alarm rate by tuning thresholds.
Exploring Factors Affecting Injury Severity of Crashes in Freeway Tunnel Groups
Amjad Pervez (firstname.lastname@example.org), Central South UniversityShow Abstract
Jaeyoung Lee, Central South University
In mountainous freeways, some tunnels are located adjacent to each other resulting in a tunnel group where the safety conditions are more challenging compared to the single tunnels. However, limited research efforts have been made to investigate traffic safety in the tunnel groups. This study aims to investigate factors influencing the injury severity of crashes in the freeway tunnel groups. The analysis is based on five years of police-reported data (2012-2016) collected from six tunnel groups in Hunan Province, China. A mixed logit model was developed as an alternative to the conventional logit model to account for the unobserved heterogeneity. Results indicate that the daytime, weekdays, entrance zone, downgrades, speeding, fatigue driving, and rollover collisions are positively while winter, curves, and sideswipe are negatively associated with severe crashes and have signs consistent with engineering intuition. More importantly, due to the complex driving environment of the tunnel groups the summer, access zone, connecting zones and drivers with less driving experience tends to increase the likelihood of severe crashes. Multiple countermeasures are recommended to improve the safety in tunnel groups, including provision of variable message signs to provide information to the drivers regarding the distance to the tunnel and speeding limitations, periodic maintenance of the illumination according to the light guidelines in the different zones of the tunnel groups, implementation of the automatic section speed control for speeding, and public awareness about the complex driving environment of the tunnel groups.
Application of Poisson-Tweedie Regression Approach for Modelling Crossing Conflicts at Un-Signalized T-intersections under Mixed Traffic Conditions
Jaydip Goyani, Sardar Vallabhbhai National Institute of TechnologyShow Abstract
Ninad Gore, Sardar Vallabhbhai National Institute of Technology
Shriniwas Arkatkar, Sardar Vallabhbhai National Institute of Technology
Post Encroachment Time (PET) is the most acceptable surrogate safety measure (SSM) for analyzing crossing conflicts, where accident data is of inferior quality. The present study assesses, and models crossing conflicts using Poisson-Tweedie regression approach. Five un-signalized T-intersections located in different regions of India with varying geometry and traffic control characteristics were selected. The Poisson-Tweedie distribution offers a unified structure to model zero-inflated, over-dispersed, under dispersed, count data, as well as multiple response variables. The flexibility of the distribution lies in the domain of p, which includes positive real number values. Initially, conflicts were bifurcated into critical and non-critical conflicts using PET values. Results revealed that location of the intersection, intersection geometry, time of the day, and volume of the offending stream as the most significant variables that influence the number of conflicts at the intersection. Crossing conflicts were observed to increase with an increase in the volume of the offending stream. For similar traffic flow conditions, a lesser number of conflicts were observed at an intersection in a rural area compared to an urban area. Similarly, an average reduction of 25-31% in the number of conflicts was observed for intersections with Central Island compared to intersections without Central Island. Further, out of the total conflicts, lesser critical conflicts were observed during off-peak hours compared to non-critical conflicts. The developed conflict prediction model can assist city planners and traffic engineers in facilitating appropriate measures to enhance traffic safety at un-signalized T-intersections.
County-Level Crash Data Exploration Using Principal Component Analysis Based Agglomerative Hierarchical Clustering
Hari-Krishnan Melempat-Kalapurayil, University of Texas, San AntonioShow Abstract
Amit Kumar (email@example.com), University of Texas, San Antonio
This study aims to conduct a county-level data exploration to identify the county groups or clusters in relation to socio-demographics, road infrastructure, traffic and crash features using clustering algorithm from the stream of unsupervised machine learning. Hierarchical clustering algorithm was adopted in this study to group the 254 counties in Texas, USA into identical groups with reference to the socio-demographics, road, and crash features. Such analysis can assist decision-makers in delivering efficient and effective resources allocation and policy analysis for priority regions. Three separate analyses are performed for crashes related to motor vehicles, pedestrians, and pedal cyclist. The study also presents an analysis framework that attempts to address the issue of low sample size high dimension to capture maximum variation in the data by utilizing the concepts from principal component analysis (PCA). Primary results from the clustering analysis shows two distinct groups of counties which shows high variation with respect to the crash features. The cluster tree results from this study indicate small variations in cluster members for motor vehicle crashes and pedestrian crashes whereas large variations for pedal cyclist crashes.
Identifying Underreported Work Zone Crash Using Machine Learning Techniques − A Comparative Study
Md Abu Sayed (firstname.lastname@example.org), University of Wisconsin, MilwaukeeShow Abstract
Xiao Qin, University of Wisconsin, Milwaukee
Rohit J Kate, University of Wisconsin, Milwaukee
D M Anisuzzaman, University of Wisconsin, Milwaukee
Zeyun Yu, University of Wisconsin, Milwaukee
The manifold data format of traffic crash report and the scared use of advance machine learning technologies in crash classifications are the reasons for many work zone(WZ) crashes remain unreported in crash statistics. This study aimed to automate the process of finding these underreported work zone crashes from police reported crash narratives. Seven state-of-the art of machine learning techniques; (1) multinomial naive bayes(MNB), (2) logistic regression(LGR), (3) Support Vector Machine(SVM), (4) Random Forest(RF), (5) K-nearest neighbor(K-NN), (6) Gated recurrent Unit(GRU) and (7) NoisyOR, were applied to recover underreported WZ crashes from reported non work zone(NWZ) crash narratives. As an experimental study, three-years crash narratives were collected from Wisconsin DOT, of which 70% were noises (false positive or false negative) and most of the words in narratives were irreverent to work zone. Our investigation using a test sample of top-scorer narratives of seven models showed that GRU and NoisyOR outperformed the rest of the classifiers in detecting underreported WZ crashes. Moreover, GRU and NoisyOR were successful finding WZ related words to use them for classifications, whereas the second-best performers; LGR and SVM, were mostly affected by irreverent words. Further investigation using a large test sample size for GRU and NoisyOR revealed that GRU detected a few more underreported WZ crashes than NoisyOR. However, the detection rate changes a bit abruptly for GRU whereas it changes uniformly for NoisyOR. Both models can be used to automatically detect underreported WZ crashes from crash narratives without human intervention.
A Full Bayesian Space-time Random Effect Approach for Hexagon-based Crash Frequency Modeling
Haipeng Cui (email@example.com), National University of SingaporeShow Abstract
Kun Xie, Old Dominion University
Although spatial and temporal correlations of crash observations have been well addressed in recent studies, the interactions between them are rarely studied in the safety literature. To accommodate the complex spatiotemporal data structures, Markov Chain Monte Carlo (MCMC) method is generally used, resulting in high computational expense for model estimation. This paper proposes a Bayesian spatiotemporal interaction (BSTI) approach for crash frequency modeling with an integrated nested Laplace approximation (INLA) method to greatly expedite the estimation process. Manhattan is selected as the study area. Hexagons are used as the basic geographic units to capture crash, transportation, land use, and demo-economic data. A series of Bayesian spatiotemporal models are developed and compared. Results show that the BSTI model with type II interaction outperforms the other models by showing the ability to capture both spatial and temporal correlations as well as spatiotemporal interactions. It is also found that the unobserved heterogeneity is mostly attributed to the spatial effects instead of temporal effects. The findings show the necessity of addressing the spatiotemporal interactions in crash frequency modeling.
Crash Prediction for Advanced Driver Assistance Systems: Development and Comparative Analysis of Advanced Deep Learning Techniques
Osama Osman (firstname.lastname@example.org), University of Tennessee, ChattanoogaShow Abstract
Mustafa Hajij, Santa Clara University
Motor vehicle crashes have claimed the lives of 38,800 lives and caused 4.4 million injuries in 2019 alone. Studies have shown that 94% of these crashes are because of driver errors. Such a huge contribution of driver errors to crashes points out that efforts to improving safety should be directed towards both vehicles and drivers through advanced driver assistance systems (ADAS) and vehicular technologies. This study investigates the potential realtime data collected through vehicular technologies on driver behavior offer to predict crashes as a first line of defense to avoid them. Three deep learning models were developed including multilayer perceptron neural networks (MLP-NN), long-short-term memory networks (LSTMN), and convolutional neural networks (CNN) using vehicle kinematics time series data extracted from the Second Strategic Highway Research Program Naturalistic Driving Study (SHRP 2 NDS) dataset. The study builds on the hypothesis that crashes are preceded by turbulences that take place over time (turbulence horizon). If these turbulences are detected in a timely manner they can help predict and avoid crashes. Several values were tested for the turbulence horizon and the prediction horizon (how long before the crash impact it can be predicted) to identify the optimal values. The results showed that the CNN model can predict all crashes with a 100% accuracy and zero false alarms 3 seconds before the crash impact time, when a 6-second turbulence horizon is used. This outstanding performance presents the developed model as a promising tool for implementation in ADAS.
Multiple Correspondence Analysis of Wrong-Way Driving Fatal Crashes on Freeways
Yukun Song (email@example.com), Auburn UniversityShow Abstract
Huaguo Zhou, Auburn University
Qing Chang, Auburn University
Mohammad Jalayer, Rowan University
The objective of this study is to identify clusters of contributing factors associated with the occurrence of wrong-way driving (WWD) fatal crashes on freeways using the multiple correspondence analysis (MCA) method based on the Burt matrix with an adjustment of inertias. A total of 14 years (2004–2017) of WWD fatal crash data was extracted from the National Highway Traffic Safety Administration (NHTSA) Fatality Analysis Reporting System (FARS) data set. A standard procedure was developed to extract the WWD crash information (including a total of 3,817 crashes) on freeways from the FARS. Each crash contains various characteristics of crashes, vehicles, and persons, e.g., crash time, crash location, vehicle type, driver age, etc. The MCA analysis used a total of 18 key variables with 57 defined categories. The results of this study indicate that four clusters of factors—(1) younger drivers, driving under the influence (DUI), weekends, nighttime, summer, urban areas, clear weather conditions, dry road conditions, and street lighting; (2) older drivers, no DUI, and daylight; (3) dark no light, winter, and rural; and (4) ice/snow surface, rain/snow/sleet/hail/fog, and wet streets—when combined might contribute to the occurrence of some WWD fatal crashes.
Modeling Injury Severity of Unconventional Vehicle Occupants: A Hybrid of Latent Segment and Random Parameters Logit Model
Bijoy Saha, University of British Columbia, OkanaganShow Abstract
Mahmudur Fatmi, University of British Columbia
Md. Mizanur Rahman, Bangladesh University of Engineering and Technology
Unconventional vehicles such as human-pulled and engine-operated three-wheeler vehicles are popular travel modes in developing countries. These slower moving vehicles have limited safety features, posing significant injury risks to their occupants. This study investigates injury severity of unconventional vehicle occupants (UVOs). A hybrid latent segmentation-based random parameters logit (LSRPL) model is developed utilizing 5-year police reported collision records from Dhaka, Bangladesh. LSRPL model captures multi-dimensional heterogeneity by allocating victims into discrete latent segments (i.e. inter-segment heterogeneity) and allowing a continuous distribution of parameters within the segments (i.e. inter-segment heterogeneity). The model is estimated for two segments using victim and crash attributes: segment one is a lower-risk segment, and segment two is a higher-risk segment. The model results suggest that victim and driver profile, crash attributes, environmental factors, road network attributes, and transportation infrastructure and land use attributes influence injury severity of UVOs. For example, human-pulled three-wheeler vehicle and engine-operated three-wheeler paratransit passengers, head-on and right-angle collisions, and crashes at 3-way and 4-way intersections have a higher likelihood to result in severe injury. The model confirms the existence of significant inter-segment heterogeneity. For example, mid-block crashes are more likely to result in severe injury in higher-risk segment, and show a lower likelihood for severe injury in lower-risk segment. The model further confirms intra-segment heterogeneity for higher mixed land use. For example, in the case of mid-block crashes, higher mixed land use shows significantly lower mean for high-risk segment, revealing a lower likelihood of severe injury in higher mixed land use areas.
Extracting Rules from AV involved Crashes by applying Decision Tree and Association Rule Methods
Md Tanvir Ashraf, West Virginia UniversityShow Abstract
Kakan Dey (firstname.lastname@example.org), West Virginia University
Sabya Mishra, University of Memphis
Autonomous Vehicles (AVs) can dramatically reduce the number of traffic crashes and associated fatalities by eliminating the avoidable human-errors related crash contributing factors. Many companies have been conducting pilot tests on public roads in several states in the US and other countries to fast-track AV mass deployment. AV pilot operations on California public roads caused 251 AV involved crashes (as of February 2020). These AV involved crashes provide a unique opportunity to investigate AV crash risks in the mixed traffic environment. This study collected the AV crash reports from the California DMV and applied the Decision Tree (DT), and Association Rule methods to extract the pre-crash rules of AV involved crashes. Extracted rules revealed that the most frequent AV involved crash type was rear-ended crash and predominantly occurred at intersections when AVs were stopped and engaged in the autonomous mode. AV and Non-AV manufacturers, and transportation agencies can use the findings of this study to minimize AV related crashes. AV companies could install a distinct signal/display to inform the operational status of the AV (i.e., autonomous or non-autonomous mode) to human drivers around the AVs. Moreover, the Automatic Emergency Braking (AEB) system in non-AV could avoid a significant number of rear-end crashes as often rear-end crashes occurred due to the failure of non-AV’s timely slow down behind AVs. Transportation agencies can consider separating the AVs from the non-AVs by assigning “AV only lanes” to eliminate the excessive rear-end crashes due to the mistakes of human-drivers in non-AVs
Missing Data Problem Treatment in Crash Data: A Genetic Algorithm-based Clustered Weighting Method
Sina Asgharpour (email@example.com), Sharif University of TechnologyShow Abstract
Mohammadjavad Javadinasr, University of Illinois, Chicago
Zeinab Bayati, Sharif University of Technology
Amir Samimi, Sharif University of Technology
Addressing the missing data problem, as one of the most commonplace challenges in the crash data studies, is of vital importance. The detrimental effects induced by a missing-containing dataset on the subsequent analyses seem inevitable. In this sense, missing data treatment techniques endeavor to compensate for the bias engendered by missing-ness. Considering that the missing-ness potentially causes variables’ distribution to undergo drastic deviations from their unbiased state, in this study, we proposed a new weighting method in an optimization model framework to minimize the overall deviation from the unbiased state of the dataset, with respect to a predefined set of essential variables (i.e., the control variables ). In this regard, we devised a heuristic algorithm based on Genetic Algorithm, as well as a boosting process, called clustering, to circumvent the calculation burden. Moreover, we used Iran’s nation-wide yearly crash data to appraise the performance of the proposed algorithm. According to the results, the overall deviation, in case of having 10 control variables, decreases from 10% (associated with the Complete Deletion) to approximately 0.9% (i.e., 92% improvement). Besides, the statistical significance tests in the before/after stages demonstrate the high performance of the method in reducing the deviation. Another finding reveals the inverse relationship between the performance of the algorithm and the number of the control variables with the reduction of the performance index from 92 to roughly -40, subsequent to a decrease in the number of the control variables from 10 to 1.
Influence of Incident Spatiotemporal Estimation Method in Secondary Crash Identification
Angela Kitali, Florida International UniversityShow Abstract
Priyanka Alluri, Florida International University
Thobias Sando, University of North Florida
Accurate estimation of the primary incident spatiotemporal impact area is essential and imperative for successfully mitigating secondary crashes. This study presents a data-driven approach to automatically determine the spatiotemporal impact areas of primary incidents and hence detect secondary crashes that occurred within the affected area. The proposed approach considered how the queue caused by the primary incident grows and dissipates along each roadway segment upstream of the incident. The effectiveness of the proposed approach was compared with the approach that assumed that the impact of the incident along all the impacted segments is the same, referred to as the base approach . A majority of secondary crashes occurred under congested traffic conditions. Incidents with a major impact on traffic were the primary contributors to secondary crashes. The comparison of the secondary crashes detected by the improved and the base approach indicated that the base approach identified 54% more secondary crashes than the improved approach . These additional crashes were found to occur mostly under less traffic congestion such as during off-peak hours, on weekends, and when the primary incident had only a minor impact on traffic. Although the improved approach identified fewer secondary crashes, it is more precise because it considers segment-based traffic conditions. The proposed method more accurately identifies secondary crashes since it better reflects the changes in traffic characteristics caused by the primary incident.
Hybrid Artificial Intelligence Models for Work Zone Crash Frequency Analysis at Bridge Locations
Seyedmirsajad Mokhtarimousavi (firstname.lastname@example.org), Florida International UniversityShow Abstract
Jason Anderson, Portland State University
Mohammed Hadi, Florida International University
Atorod Azizinamini, Florida International University
Crash characteristics differ from location to location, as well as over time, along with varying features of participants at fault, environmental and geometrical conditions. Although the impact of work zone presence on crash frequency has been investigated in previous studies, the risk factors associated with work zone crash frequency at bridge locations are not fully understood. With this in mind, this study utilizes a Negative Binomial (NB) regression and a Support Vector Regression (SVR) models trained by Artificial Bee Colony (ABC) optimization algorithm for modeling work zone crash frequency. Incorporating three years of crash records from 2015-2017, road inventory data, bridge geometric and location specification, into a data-driven analysis, work zone crash frequency were investigated on a number of 60 bridge locations in Miami-Dade County. A sensitivity analysis was also conducted considering the black-box characteristic of the SVR and compared to the effects of variables indented through the NB modeling framework. The prediction performance of the developed models was evaluated by three commonly-used criteria including the coefficient of determination (R 2 ), the Mean Absolute Deviation (MAD), the Mean Square Error (MSE), and the Root Mean Square Error (RMSE). The results demonstrated that the proposed SVR models predict work zone crash data more effectively and accurately than traditional NB models. In addition, bridge median type, law enforcement, horizontal curve, bridge surface width indicators were among the most important factors that affects the number of work zone related crashes on bridges.
Evolutionary Training Approaches for Machine Learning Models for Analyzing Classification Imbalanced Crash Datasets
Seyedmirsajad Mokhtarimousavi, Florida International UniversityShow Abstract
Mohammed Hadi, Florida International University
Eazaz Sadeghvaziri, Morgan State University
Atorod Azizinamini, Florida International University
Machine learning techniques have gained attention by safety researchers for use in predicting the attributes of crashes on transportation facilities. Crash events typically occur in rare instances which results in imbalanced datasets. Thus, the utilization of such techniques to predict crash outcomes require selecting the appropriate modeling techniques to cope with imbalanced dataset. Supervised algorithms consist of two major phases: training and testing. Learning from data to predict the outcome of interest on unseen data makes training procedure the most difficult challenge in the context of machine learning, in which finding the optimal set of parameters can highly affect the prediction performance of the model. In this study, three widely-used machine learning algorithms, Support Vector Machine (SVM), Artificial Neural Network (ANN), and K-Nearest Neighbour (KNN) are utilized for the classification of work zone crash severity. Three evolutionary optimization algorithms, including the Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Artificial Bee Colony (ABC) are used for training the developed models to enhance the performance of the models. The prediction performance enhancements are evaluated against the base models without using the optimization algorithms. The results indicated that while the best prediction performance was obtained when using the GA optimization combined with the SVM, utilizing PSO and GA to train and optimize the parameters of the ANN and KNN significantly improved their performances especially in dealing with imbalanced data.
California Autonomous Vehicle Crashes: Explanatory Data Analysis and Classification Tree
Seyedehsan Dadvar, CYFOR Technologies LLCShow Abstract
Mohamed Ahmed, University of Wyoming
Autonomous Vehicle (AV) is an evolving technology with many capabilities and limitations. The main safety attribute of AVs is eliminating human drivers from the driving process with the promise to decrease road crashes drastically. AV field tests are being conducted in several states in the US and in other parts of the world. California Department of Motor Vehicle (DMV) has mandated all AV crashes and disengagement incidents being publicly reported by permit holders since 2014. Several different studies had used the CA DMV data to investigate different aspects of AVs especially road safety attributes. In this study, 234 CA DMV AV-related crashes (2017-2020) were examined. Explanatory Data Analyses indicated that rear-end and side-swipe were the main collision types and based on geographic distribution of crashes, the majority of them happened in a relatively small area in San Francisco bay area usually surrounding the permit holder headquarters. The classification tree using Chi-square Automatic Interaction Detector (CHAID) method was developed for AV-related crashes based on driving mode and AV movement, company (permit holder), road surface, other vehicle movement, intersection / control type, and crash time was identified as a significant contributing factor. Results, limitations, and potential future work were discussed in the context of the AVs and roadway safety.
Vehicle Group Crash Risk Prediction based Active Traffic Management Strategies for Expressways
Ziliang He, Tongji UniversityShow Abstract
Wanjing Ma, Tongji University
Ling Wang (email@example.com), Tongji University
Hao Zhong, Tongji University
Chunhui Yu, Tongji University
The safety of expressways is important. This study developed variable speed limits (VSL) and ramp metering (RM) strategies based on the prediction of vehicle group crash risk to improve the safety of expressways. The goal of VSL was to minimize the crash risk of multiple vehicle groups in the next time period, and the updated speed limits were sent directly to the connected vehicles (CV) to adjust the speed of the vehicle group. As for the RM, the metering rate and opening time of ramp control were designed based on mainline occupancy, vehicle group crash risk, and predicted time of vehicle group arriving at the ramp. Considering the impact of RM on flow and the influence of the VSL on the capacity, the coordinated VSL and RM strategy (VSL-RM) was established . These strategies were tested in microsimulations. The crash risk index and the Surrogate Safety Assessment Model (SSAM) were utilized to evaluate the safety effect of these strategies. The results showed that the three strategies improved the safety of the expressway. Additionally, the higher the penetration rate of connected vehicles, the higher the safety benefits of VSL and VSL-RM. Moreover, the VSL-RM was superior to VSL and RM. Keywords : Connected Vehicle, Crash Risk, Variable Speed Limit, Ramp Metering, Coordination of Ramp Metering and Variable Speed Limit The safety of expressways is important. This study developed variable speed limits (VSL) and ramp metering (RM) strategies based on the prediction of vehicle group crash risk to improve the safety of expressways. The goal of VSL was to minimize the crash risk of multiple vehicle groups in the next time period, and the updated speed limits were sent directly to the connected vehicles (CV) to adjust the speed of the vehicle group. As for the RM, the metering rate and opening time of ramp control were designed based on mainline occupancy, vehicle group crash risk, and predicted time of vehicle group arriving at the ramp. Considering the impact of RM on flow and the influence of the VSL on the capacity, the coordinated VSL and RM strategy (VSL-RM) was established . These strategies were tested in microsimulations. The crash risk index and the Surrogate Safety Assessment Model (SSAM) were utilized to evaluate the safety effect of these strategies. The results showed that the three strategies improved the safety of the expressway. Additionally, the higher the penetration rate of connected vehicles, the higher the safety benefits of VSL and VSL-RM. Moreover, the VSL-RM was superior to VSL and RM. Keywords : Connected Vehicle, Crash Risk, Variable Speed Limit, Ramp Metering, Coordination of Ramp Metering and Variable Speed Limit
Exploring A Need to Model Two- and Multiple-vehicle Crashes Separately
Angela Kitali, Florida International UniversityShow Abstract
Emmanuel Kidando, Cleveland State University
Md Asif Raihan, Bangladesh University of Engineering and Technology
Boniphace Kutela, Texas A&M Transportation Institute
Priyanka Alluri, Florida International University
Thobias Sando, University of North Florida
Single-vehicle crashes have been shown to differ from two-plus vehicle crashes. Several studies have discussed the issues with modeling single- and two-plus vehicle crashes together. However, none of the empirical studies have attempted to study two-vehicle (2V), and multiple-vehicle (MV), i.e., three-plus, crash groups to understand their correlation and influencing factors. This study first investigates whether there is a need to develop separate safety performance functions for 2V and MV crashes, in addition to single-vehicle crashes. Then, the correlation and influencing factors of 2V and MV are evaluated. Three regression models – a correlated bivariate negative binomial regression (BNR) model, an uncorrelated bivariate negative binomial regression (NR) models, and a univariate negative binomial regression (UNR) model, are fitted and compared. The analysis is based on the 2011-2015 crash data that occurred on I-4 in Florida. Findings indicate that the BNR model significantly outperformed the NR and the UNR models. The model results suggest that disaggregating these crashes while allowing correlation between the groups for the latent effects in the model best describes the data. Traffic volume, posted speed limit, and median type were found significant in contributing to the occurrence of both 2V and MV crashes. Additional contributing factors included the presence of interchange influence area for 2V crashes and the presence of a vertical curve and the presence of a horizontal curve for MV crashes. Study findings could assist transportation officials implement specific safety countermeasures for road segments that are identified as hotspots for 2V and MV crashes.
Investigating the Typical Scenarios and Contributory Factors to Crash Severity of Autonomous Vehicle Involved Collisions Using Association Rule Analysis
Jing Zhang (firstname.lastname@example.org), Southeast UniversityShow Abstract
Chengcheng Xu, Southeast University
Autonomous vehicles (AVs) are considered to have the potential to bring considerable benefits to the transportation system, involving the improvement of traffic safety and traffic efficiency, as well as the reduction of congestion and emissions. However, the AVs are posing considerable uncertainty on road safety, especially when AVs and conventional vehicles are operating on the public road together. This study aimed to identify typical scenarios of collisions and contributory factors to human-injured collisions involved with an autonomous vehicle. The association rule analysis was used to identify common collision patterns using the autonomous vehicle involved collision reports from the California Department of Motor Vehicles (CDMV). Six typical scenarios have been identified based on collision types, including 1 broadside collision, 1 sideswipe collision and 4 rear-end collision scenarios. Moreover, down-slope, night time, multi vehicles, and high-density traffic were found to be the four main contributory factors to human-injured collisions involved with an autonomous vehicle. Based on the results of association rule analysis, potential implications for preventing AV involved crashes and reducing collision severity were identified.
Risk Analysis of Road Transport Accidents of Hazardous Materials by Machine Learning
Xiaoyan Shen, Chang'an UniversityShow Abstract
Shanshan Wei (email@example.com), Chang'an University
Yuqing Feng, Chang'an University
Fan Zhang, Chang'an University
The safe movement of hazardous materials is receiving increasing attention because various sudden and devastating hazardous material accidents have resulted in substantial injury to humans, damage to property and environmental pollution. The aim of this paper is to explore a suitable method for analyzing road transport accidents involving hazardous materials and to study the main risk factors for accidents of different severities (property damage only (PDO), injury (INJ) and fatality (FAT)). Initially, we assessed three classification algorithms, i.e., decision tree C5.0 (C5.0), support vector machine (SVM) and multilayer perceptron (MLP), using a hazardous material transportation accident dataset. The results reveal that the predictions of C5.0 algorithms are superior to those of SVM and MLP. Hence, C5.0 algorithm was applied to extract the probable risk factors and associations between these factors and 3 different severities of hazardous material transportation accidents. The results showed that direct accident form (DAF), indirect accident form (IAF), and road section (RS) all have significant effects on accidents involving only property damage.Direct accident form (DAF), indirect accident form (IAF), road type (RT), road segment (RS) and time (TIME) all have a substantial effect on injury accidents. Direct accident form (DAF), indirect accident form (IAF), hazardous material type (HMT) and road surface condition (SC) are important factors in the occurrence of fatal accidents. The above results provide a theoretical basis for discussing safety problems in hazardous materials transport activities and offer valuable suggestions for measures to reduce the severity of accidents.
Assessment of Crash Occurrence Using Historical Crash Data and A Random Effect Negative Binomial Model: A Case Study for a Rural State
Karla Diaz-Corro, University of Arkansas, FayettevilleShow Abstract
Leyla Coronel Moreno, University of Arkansas, Fayetteville
Suman Mitra, University of Arkansas, Fayetteville
Sarah Hernandez, University of Arkansas, Fayetteville
The objective of this work is to identify factors that influence crash occurrence within a Traffic Analysis Zone (TAZ) by accounting for serial and spatial correlation in longitudinal crash data. This is accomplished by applying a Random Effect Negative Binomial model (RENB). Unlike commonly used count models such as Poisson and Negative Binomial (NB), RENB accounts for heterogeneity and serial correlation in crash occurrence. A RENB was applied to 15 years (180 months) of crash data in Arkansas, a relatively rural state, with 1,817 TAZs. RENB estimated impacts were measured using the Incidence Rate Ratio (IRR). The significant causal factors found to contribute to increases in observed crashes include, in order of IRR-estimated magnitude: (i) average precipitation (a one unit increase in average precipitation results in a 134% increase in total monthly crashes for a TAZ), (ii) average wind speed (16%), (iii) urban designation (7%), (iv) traffic volume (2%), and (v) total roadway mileage (1% for each functional class). Snow depth and days of sunshine were found to decrease the number of accidents by 15% and 2%, respectively. Employment and total population had no impact on crash occurrence. Goodness-of-fit comparisons show that RENB provides the best fit among Poisson and NB formulations. Model diagnostics confirm the presence of over-dispersion and serial correlation indicating the necessity of RENB model estimation. The main contribution of this work is the identification of crash causal factors at the TAZ level for longitudinal data, which supports data-driven performance measurement requirements of recent federal legislation.
DISCLAIMER: All information shared in the TRB Annual Meeting Online Program is subject to change without notice. Changes, if necessary, will be updated in the Online Program and this page is the final authority on schedule information.