Applying interpretable machine learning to classify tree and utility pole related crash injury types

In spite of enormous improvements in vehicle safety, roadway design, and operations, there is still an excessive amount of traffic crashes resulting in injuries and major productivity losses. Despite the many studies on factors of crash frequency and injury severity, there is still further research to be conducted. Tree and utility pole/other pole related (TUOP) crashes present approximately 12 to 15% of all roadway departure (RwD) fatal crashes in the U.S. The count of TUOP crashes comprise nearly 22% of all fatal crashes in Louisiana. From 2010 to 2016, there were 55,857 TUOP crashes reported in Louisiana. Individually examining each of these crash reports is not a realistic option to investigate crash factors. Therefore, this study employed text mining and interpretable machine learning (IML) techniques to analyze all TUOP crashes (with available crash narratives) that occurred in Louisiana from 2010 to 2016. This study has two major goals: 1) to develop a framework for applying machine learning models to classify injury levels from unstructured textual content, and 2) to apply an IML framework that provides probability measures of keywords and their association with the injury classification. The present study employed three machine learning algorithms in the classification of injury levels based on the crash narrative data. Of the used modeling techniques, the eXtreme gradient boosting (XGBoost) model shows better performance, with accuracy ranging from 0.70 to 24% for the training data and from 0.30% to 16% for the test data.     

An analysis of motorcyclists' injury severities in work-zone crashes with unobserved heterogeneity

Work-zone related crashes in Florida are increasing steadily in recent years. With such growing concern, work-zone is recognized in Florida Strategic Highway Safety Plan. The severity and complexity of motorcycle crashes in work-zones is critically important and worth investigating. However, the resulting effect of work-zone on motorcyclists' injuries in work-zone crashes is not fully understood. The purpose of study is to identify the contributing factors of motorcyclists' injury severity sustained in the work-zone crashes in Florida. Recognizing the relatively higher risk of motorcyclists in work-zones with respective to non-work-zones, this study further uncovers the contributing factors for single- and multi-vehicle motorcycle crashes in Florida work-zones. This study investigated motorcyclists' injury severity applying random parameter multinomial logit with possible heterogeneity in means and variances of the random parameters for single-motorcycle and multi-vehicle motorcycle crashes. This study utilizes the Crash Analysis Reporting (CAR) system in Florida over a period of five years from 2012 to 2016 (inclusive). The model result indicates a complex relationship between dark condition, old-aged motorcyclist (50–65), requirement and absence of endorsement, partial ejection, straight roadway segment, shoulder width (up to 1.22 m (4 ft), and 2.74–3.66 m (9–12 ft), urban interstate, activity area, and lane closure and work on shoulder-median work-zone types. The effect of work-zone on single-motorcycle crashes tends to have much more in resulting injury severities relative to multi-vehicle motorcycle crashes. It is more important to investigate the injury severity by single- and multi-vehicle crashes involving motorcycles in work-zones. These risk factors identified in the study are expected to provide more insights for the countermeasures specific to engineering (roadway design) and policy (motorcycle training), which can be considered to improve motorcycle safety in Florida.     

Road characteristics and environment factors associated with motorcycle fatal crashes in Malaysia

This study aims to determine risk factors contributing to traffic crashes in 9,176 fatal cases involving motorcycle in Malaysia between 2010 and 2012. For this purpose, both multinomial and mixed models of motorcycle fatal crash outcome based on the number of vehicle involved are estimated. The corresponding model predicts the probability of three fatal crash outcomes: motorcycle single-vehicle fatal crash, motorcycle fatal crash involving another vehicle and motorcycle fatal crash involving two or more vehicles. Several road characteristic and environmental factors are considered including type of road in the hierarchy, location, road geometry, posted speed limit, road marking type, lighting, time of day and weather conditions during the fatal crash. The estimation results suggest that curve road sections, no road marking, smooth, rut and corrugation of road surface and wee hours, i.e. between 00.00 am to 6 am, increase the probability of motorcycle single-vehicle fatal crashes. As for the motorcycle fatal crashes involving multiple vehicles, factors such as expressway, primary and secondary roads, speed limit more than 70 km/h, roads with non-permissible marking, i.e. double lane line and daylight condition are found to cause an increase the probability of their occurrence. The estimation results also suggest that time of day (between 7 pm to 12 pm) has an increasing impact on the probability of motorcycle single-vehicle fatal crashes and motorcycle fatal crashes involving two or more vehicles. Whilst the multinomial logit model was found as more parsimonious, the mixed logit model is likely to capture the unobserved heterogeneity in fatal motorcycle crashes based on the number of vehicles involved due to the underreporting data with two random effect parameters including 70 km/h speed limit and double lane line road marking.     

Modeling two-way stop-controlled intersection crashes with zero-inflated models on Louisiana rural two-lane highways

Intersection safety continues to be a crucial issue throughout the United States. In 2016, 27% of the 37,461 traffic fatalities on U.S. roadways occurred at or near intersections. Nearly 70% of intersection-related fatalities occurred at unsignalized intersections. At such intersections, vehicles stopping or slowing to turn create speed differentials between vehicles traveling in the same direction. This is particularly problematic on two-lane highways. Research was performed to analyze safety performance for intersections on rural, two-lane roadways, with stop control on the minor roadway. Roadway, traffic, and crash data were collected from 4148 stop-controlled intersections of all 64 Parishes (counties) statewide in Louisiana, for the period of 2013 to 2017. Four count approaches, Poisson, Negative Binomial (NB), Zero-inflated Poisson (ZIP) and Zero-inflated Negative Binomial (ZINB) were used to model the number of intersection crashes for different severity levels. The results indicate that ZIP models provide a better fit than all other models. In addition to traffic volume, larger curve radii of major and minor roads and wider lane widths of major roads led to significantly smaller crash occurrences. However, higher speed limits of major roads led to significantly greater crash occurrences. Four-leg stop-controlled intersections have 35% greater total crashes, 49% greater fatal and injury crashes, and 25% greater property damage only (PDO) crashes, relative to three-leg intersections.     

Severity analysis of tree and utility pole crashes: Applying fast and frugal heuristics

A roadway departure (RwD) crash is defined as a crash that occurs after a vehicle crosses an edge line or a center line, or otherwise leaves the designated travel path. RwD crashes account for approximately 50% of all traffic fatalities in the U.S. Additionally, crashes related to roadside fixed objects such as trees, utility poles, or other poles (TUOP) make up 12–15% of all fatal RwD crashes in the U.S. Data spanning over seven years (2010–2016) shows that TUOP crashes account for approximately 22% of all fatal crashes in Louisiana, which is significantly higher than the national statistic. This study aims to determine the effect of crash, geometric, environmental, and vehicle characteristics on TUOP crashes by applying the fast and frugal tree (FFT) heuristics algorithm to Louisiana crash data. FFT identifies five major cues or variable threshold attributes that contribute significantly to predicting TUOP crashes. These cues include posted speed limit, primary contributing factor, highway type, weather, and locality type. The balanced accuracy is around 56% for both training and test data. The current model shows higher accuracies compared to machine learning models (e.g., support vector machine, CART). The present findings emphasize the importance of a comprehensive understanding of factors that influence TUOP crashes. The insights from this study can help data-driven decision making at both planning and operation levels.     

A novel deep ensemble based approach to detect crashes using sequential traffic data

Crash is one of the leading causes of death in the United States. Real time detection of crashes plays a pivotal role in increasing safety of highways. In this study, a deep ensemble modelling approach is proposed in which we first employed three powerful deep learning techniques, Long Short-Term Memory (LSTM), Gated Recurrent Units (GRUs), and Deep Neural Network (DNN), then these three models are combined using eight ensemble techniques to detect crashes in real time. Loop detectors' traffic condition, crash information, and weather condition are the main sources of data used in this study. In addition, since the dataset of this study includes 241 crash and 6038 non-crash cases, Synthetic Minority Over-sampling Technique (SMOTE) is used to overcome problem of imbalanced data before training the models. The results show that while deep learning models are performing well in detecting crashes, ensemble of these three deep learning models using Multilayer Perceptron (MLP) and Random Forest Classifier (RFC) can improve detection performance. Interestingly, MLP and RFC ensemble models achieved the highest detection rate and lowest false alarm rate values, respectively, among all the studied models. Comparing the models regarding area under curve (AUC) of ROC curve also shows that the best five models are MLP ensemble, RFC ensemble, DNN, GRU, and LSTM, respectively, with AUC of 97.2%, 95.2%, 93.7%, 91.8%, and 91.3%.     

连续数据环境下的道路交通事故风险预测模型

针对现有研究多基于病例对照的欠采样方法,即每起事故从连续交通流数据中按一定比例抽取对照的非事故数据构建模型,而该类模型在连续数据环境中的预测精度存在缺陷的状况,对城市交通连续观测并动态调控的技术环境（简称连续数据环境）开展道路交通事故风险预测模型构建研究。首先提出基于全样本交通流数据,结合“调整事故分类阈值”的方法解决事故风险预测研究中的非平衡数据分类问题;而后采用上海市城市快速路2014年5,6月的线圈检测交通流数据及历史事故数据开展实证研究,以受试者工作特征曲线下面积为评价指标,对比基于全样本和抽样样本构建的常用事故风险预测模型（逻辑回归、随机森林）的整体预测能力;以灵敏度和特异度的几何均数为评价指标,对比3种分类阈值计算方式（约登指数法、事故占比法和交叉点法）对事故/非事故综合预测精度的影响。结果表明：在连续数据环境下,采用全样本数据建模能使模型整体预测能力提高13.06%;基于约登指数法进行分类阈值计算可使模型的事故/非事故综合预测精度最佳。     

Non-fatal spine injuries resulting from motorcycle crashes

This study aims to determine spinal injury patterns and identify crash factors commonly associated with serious spinal injury as a result of motorcycle crashes. Data was retrospectively collected from motorcyclists sustaining spinal injuries from road crashes treated at Kuala Lumpur Hospital, Malaysia, over the 5-year period from 2005 to 2009. Each patient's injuries were analyzed by reviewing his or her medical records for radiographic imaging and computed tomography scans.A total of 151 patients were included in this study, of which, males accounted for over 87%. The first lower lumbar (L1) was the most commonly injured vertebral level, followed by the adjacent thoracic vertebra (T12). Fracture to the vertebral body without dislocation was found to be the most frequently observed spinal injury pattern. Injury severities for a majority of patients (65%) were measured at Maximum Abbreviated Injury Scale (MAIS) of 2. Serious spinal injury was associated with thorax or upper-extremity injury.Prevalence of lumbar spinal injury in the study reflects a predominantly low-speed crash among the motorcyclist in the region. Motorcyclists are at greater odd to sustain severe spinal injury when directly striking an object compare to striking the ground during the crash event.     

Factors affecting severity of motorcycle accidents on Thailand's arterial roads: Multiple correspondence analysis and ordered logistics regression approaches

Thailand was classified as a middle-income country and ranked second highest in terms of road traffic fatality rate in the world in 2015. By 2018, this ranking went up to ninth in world which may be because of various earnest safety policies implementation, supporting road safety research and establishing a road safety directing center. However, crash fatality rate has considerably remained high until recent year, indicating a clear need for further related research. Considering severity of the crashes, the majority of fatal crashes involved the motorcycle road user. Therefore, motorcycle crashes are important issues and should be considered to mitigate fatality due to immoderate proportion of motorcycle road user and motorcyclist fatality. This study aims to identify factors that influence the severity of motorcycle accidents on Thailand's arterial roads by employing ordered logistic regression and multiple correspondence analysis. The results demonstrated that although both analyses were relatively different, they provided similar results. Age, road lanes, and helmet wearing were significant factors that influenced the severity of motorcycle accidents. The results could serve as reference for planning strategies or organizing campaigns to reduce and prevent death owing to road traffic accidents, which may enhance the overall image of road traffic safety in Thailand.     

Evaluating the correlation between risky riding behaviour and self-reported crashes in India: Minimizing unobserved heterogeneity

Riding Behaviour is found to be the main cause of Powered Two-Wheeler (PTW) crashes in more than90% of the crash events. The high percentage of PTW crashes resulting in fatalities has sought a serious need for research to examine risky riding Behaviour. A widely used instrument for measuring the self-reported riding Behaviour of PTW riders is the Motorcycle Rider Behaviour Questionnaire (MRBQ). In this study, exploratory factor analysis of the MRBQ revealed a four-factor solution viz., traffic errors, control errors, speed violations, and stunts. Despite the popularity of MRBQ, it is capable of covering only a small fraction of the large number of elements that affect the riding Behaviour. Many other elements remain overlooked in the analysis, resulting in unobserved heterogeneity. Therefore, the present study uses a random parameter negative binomial (RPNB) model to minimize the effect of unobserved heterogeneity. It was inferred from the RPNB model that variables like gender, control error, and speed violation have a randomly distributed regression coefficient. Further, it is found that traffic errors are the most significant predictor of crash risk. Additionally, results depict that male riders are positively associated with crashes, and they are more likely to involve in crashes as compared to female riders. The finding of this paper will be valuable for policymakers and decision-makers to improve the rider training program, licensing system, and design road safety campaigns.     

Sampling bias and weight factors for in-depth motorcycle crash data in Thailand

Motorcycle crashes are documented in Thailand's national records but are underreported and lacking detail. In-depth motorcycle crash data, collected by Thailand Accident Research Center (TARC), contains a smaller number of motorcycle crashes but more detail. However, to draw conclusions at a national level, representativeness of the TARC in-depth data is currently unknown, and the correction of sampling biases may be required. In this study, the Capture-recapture method was used to examine the underreporting in the national crash data (from the government insurance company). It was found that 69% of fatal and 70% of non-fatal injuries were underreported, respectively. The in-depth crash data was found to be biased. The weighting methods post-stratification and iterative proportional fitting were applied to compensate for the bias and are shown to improve the representativeness of the in-depth motorcycle crash data. Weighted in-depth crash data appears to be suitable to draw conclusions on motorcyclist safety in Thailand.     

Inclusion of speed and weather measures in safety performance functions for rural roadways

The research on relationships among vehicle operating speed, roadway design elements, weather, and traffic volume on crash outcomes will greatly benefit the road safety profession in general. If these relationships are well understood and characterized, existing techniques and countermeasures for reducing crash frequencies and crash severities could potentially improve, and the opportunity for new methodologies addressing and anticipating crash occurrence would naturally ensue. This study examines the prevailing operating speeds on a large scale and determines how traffic speeds and different speed measures interact with roadway characteristics and weather condition to influence the likelihood of crashes. This study used three datasets from Washington and Ohio: 1) Highway Safety Information System (HSIS), 2) the National Performance Management Research Dataset (NPMRDS), and 3) National Oceanic and Atmospheric Administration (NOAA) weather data. State-based conflated databases were developed using the linear conflation of HSIS and NPMRDS. The results show that certain speed measures were found to be beneficial in quantifying safety risk. Annual-level crash prediction models show that increased variability in hourly operating speed within a day and an increase in monthly operating speeds within a year are both associated with a higher number of crashes. Safety practitioners can benefit from the current study in addressing the issue of speed and weather in crash outcomes.     

Extremely serious crashes on urban roadway networks: Patterns and trends

Extremely serious traffic crashes, defined as having a death toll of two and greater than two, have become a serious safety concern on urban roadways in Louisiana. This study examined the different contributing factors of these crashes to determine significant trends and patterns. We collected traffic crash data from Louisiana during the period of 2013 to 2017 and found that a total of 72 extremely serious crashes (around 2% of all traffic fatalities) occurred on Louisiana urban roadway networks. As crash data contain an enormous list of contributing factors, there was an issue of ‘more features than data points’ in solving the research problem. Most of these variables are categorial in nature. We selected a dimension reduction tool called Taxicab Correspondence Analysis (TCA) to investigate the complex interaction between multiple factors under a two-dimensional map. Findings of the study reveal several key clusters of attributes that show patterns of association between different crash attributes. The conclusions of this study are exploratory, and the results can help in better visualizing the association between key attributes of crashes. The findings have potentials in designing suitable countermeasures to reduce extremely serious crashes.     

Modeling lane-change-related crashes with lane-specific real-time traffic and weather data

This study investigates the relationship between lane-change-related crashes and lane-specific, real-time traffic factors. It is anticipated that the real-time traffic data for the two lanes—the vehicle's lane (subject lane) and the lane to which that a vehicle intends to change (target lane)—are more closely related to lane-change-related crashes, as opposed to congregated traffic data for all lanes. Lane-change-related crash data were obtained from a 62-mile long freeway in Southeast Wisconsin in 2012 and 2013. One-minute traffic data from the 5- to 10-minute interval prior to the crashes were extracted from an immediately upstream detector station and two immediately downstream stations from the crash location. Weather information was collected from a major historical weather database. A matched case-control logistic regression was used for analysis. Results show that the following factors significantly affect the probability of a lane-change-related crash: average flow into the target lane at the first downstream station, the flow ratio at the second downstream station, and snow conditions. Additionally, the average speed in the target lane at the first downstream station contributes to the occurrence of lane-change crashes during snowy conditions. According to the model, the probability of a lane-change-related crash under real-time traffic conditions can aid in flagging potential crash-prone conditions. The identified contributing factors can help traffic operators select traffic control and management countermeasures to proactively mitigate lane-change-related crashes.     

Crash prediction model for two-lane rural highways in the Ashanti region of Ghana

Crash Prediction Models (CPMs) have been used elsewhere as a useful tool by road Engineers and Planners. There is however no study on the prediction of road traffic crashes on rural highways in Ghana. The main objective of the study was to develop a prediction model for road traffic crashes occurring on the rural sections of the highways in the Ashanti Region of Ghana. The model was developed for all injury crashes occurring on selected rural highways in the Region over the three (3) year period 2005–2007. Data was collected from 76 rural highway sections and each section varied between 0.8 km and 6.7 km. Data collected for each section comprised injury crash data, traffic flow and speed data, and roadway characteristics and road geometry data. The Generalised Linear Model (GLM) with Negative Binomial (NB) error structure was used to estimate the model parameters. Two types of models, the ‘core’ model which included key exposure variables only and the ‘full’ model which included a wider range of variables were developed. The results show that traffic flow, highway segment length, junction density, terrain type and presence of a village settlement within road segments were found to be statistically significant explanatory variables (p < 0.05) for crash involvement. Adding one junction to a 1 km section of road segment was found to increase injury crashes by 32.0% and sections which had a village settlement within them were found to increase injury crashes by 60.3% compared with segments with no settlements. The model explained 61.2% of the systematic variation in the data. Road and Traffic Engineers and Planners can apply the crash prediction model as a tool in safety improvement works and in the design of safer roads. It is recommended that to improve safety, highways should be designed to by-pass village settlements and that the number of junctions on a highway should be limited to carefully designed ones.     

Development and validation of side-impact crash and sled testing finite-element models

Side-impact collisions are the second leading cause of death and injury in the traffic accidents after frontal crashes. Side-impact airbags, side door bars and other protection techniques have been developed to provide occupant protection. To confirm the effectiveness of protection equipment installed in vehicles, studying the degree of impact is fundamental to understand the effect of automobile collisions on the human body. Therefore, the dynamic response of the human body to traffic accidents should be analyzed to reduce the level of occupant injuries. Generally, the experimental method is complex and expensive. Recently, numerical crash simulations have provided a valuable tool for automotive engineers. This work presents full-scale and sled side-impact test finite-element (FE) models - based on the Federal Motor Vehicle Safety Standard No. 214 - that simulate a side-impact accident. The crash simulations utilized the LS-DYNA finite-element code. The human body's dynamic response to crashes is discussed herein. Additionally, occupant injuries were measured. To verify the accuracy of the proposed crash test and sled test FE models, simulation results are compared with those obtained from experimental tests. The comparison results indicate that the proposed crash test and sled test FE models have considerable potential for assessing a vehicle's crash safety performance and assisting future development of safety technologies.     

Modeling injury severity of crashes involving trucks: Capturing and exploring risk factors associated with land use and demographic in addition to crash,driver, and on-network characteristics

Nearly 499,000 motor vehicle crashes involving trucks were reported across the United States in 2018, out of which 22% resulted in fatalities and injuries. Given the growing economy and demand for trucking in the future, it is crucial to identify the risk factors to understand where and why the likelihood of getting involved in a severe or moderate injury crash with a truck is higher. The focus of this research, therefore, is on developing a methodology, capturing and integrating data, exploring, and identifying risk factors associated with surrounding land use and demographic characteristics in addition to crash, driver, and on-network characteristics by modeling injury severity of crashes involving trucks. Crash data for Mecklenburg County in North Carolina from 2013 to 2017 was used to develop partial proportional odds model and identify risk factors influencing injury severity of crashes involving trucks. The findings indicate that dark lighting condition, inclement weather condition, the presence of double yellow or no-passing zone, road sections with speed limit >40 mph and curves, and driver fatigue, impairment, and inattention have a significant influence on injury severity of crashes involving trucks. These outcomes indicate the need for effective geometric design and improved visibility to reduce the injury severity of crashes involving trucks. The likelihood of a severe or moderate injury crash involving a truck is also high in areas with high employment, government, light commercial, and light industrial land uses. The findings can be used to identify potential risk areas, proactively plan and prioritize the allocation of resources to improve safety of transportation system users in these areas.     

Modeling the accuracy of traffic crash prediction models

Crash forecasting enables safety planners to take appropriate actions before casualty or loss occurs. Identifying and analyzing the attributes influencing forecasting accuracy is of great importance in road crash forecasting. This study aims to model the forecasting accuracy of 31 provinces using their macroeconomic variables and road traffic indicators. Iran's road crashes throughout 2011–2018 are calibrated and cross-validated using the Holt-Winters (HW) forecasting method. The sensitivity of crash forecast reliability is studied by a regression model. The results suggested that the root mean square error (RMSE) of crash prediction increased among the provinces with higher and more variant average monthly crashes. On the contrary, the accuracy of crash prediction improved in provinces with higher per capita GDP, and higher traffic exposure. A 1% increase in crash variability, average historical crash count, GDP per capita, and traffic exposure, respectively, resulted in a 0.65%, 0.52%, −0.38%, and −0.13% change in the RMSE of forecasting. The addition of traffic exposure and macroeconomic factors significantly enhanced the model fit and improved the adjusted R-squared by 14% compared to the reduced model that only used the historical average and variability of crash count as the independent variables. The findings of this research suggest planners and policymakers should consider the notable influence of macroeconomic factors and traffic indicators on the crash forecasting accuracy.     

Investigating fatal and injury crash patterns of teen drivers with unsupervised learning algorithms

Teenagers have been emphasized as a critical driver population class because of their overrepresentation in fatal and injury crashes. The conventional parametric approaches rest on few predefined assumptions, which might not always be valid considering the complicated nature of teen drivers' crash characteristics that are reflected by multidimensional crash datasets. Also, individual attributes may be more speculative when combined with other factors. This research employed joint correspondence analysis (JCA) and association rule mining (ARM) to investigate the fatal and injury crash patterns of at-fault teen drivers (aged 15 to 19 years) in Louisiana. The unsupervised learning algorithms can explore meaningful associations among crash categories without restricting the nature of variables. The analyses discover intriguing associations to understand the potential causes and effects of crashes. For example, alcohol impairment results in fatal crashes with passengers, daytimes severe collisions occur to unrestrained drivers who have exceeded the posted speed limits, and adverse weather conditions are associated with moderate injury crashes. The findings also reveal how the behavior patterns connected with teen driver crashes, such as distracted driving in the morning hours, alcohol intoxication or using cellphone in pickup trucks, and so on. The research results can lead to effectively targeted teen driver education programs to mitigate risky driving maneuvers. Also, prioritizing crash attributes of key interconnections can help to develop practical safety countermeasures. Strategy that covers multiple interventions could be more effective in curtailing teenagers' crash risk.     

基于多任务迁移学习的交通警情信息自动处理方法

从交通警情数据中自动获取信息对于快速处理交通事故和提高交通管理水平具有重要的意义。为此，提出了一种基于多任务迁移学习的交通警情信息自动处理方法，该方法上游采用文本预训练模型作为共享参数层，下游建立多任务并行学习方法，实现对交通警情中的关键信息、类型和语义自动处理。选取江苏省苏州市2年内共120 191条原始交通警情作为试验数据，通过自动处理方法构建了一套标准的交通警情信息数据库。试验结果表明：所建立的关键信息抽取方法可以更精准地提取警情数据中的时间、地址和车牌信息；交通警情分类模型性能优于现有的深度学习模型，分类准确率达93%；基于局部特征增强的警情语义分析方法重点识别了警情中事故的严重程度和救援需求，识别准确率达87%。研究结论显示交通警情自动化处理方法具有良好的可移植性和实用性。