Transfer learning for spatio-temporal transferability of real-time crash prediction models.

Real-time crash prediction is a heavily studied area given their potential applications in proactive traffic safety management in which a plethora of statistical and machine learning (ML) models have been developed to predict traffic crashes in real-time. However, one of the fundamental issues relating to the application of these models is spatio-temporal transferability. The present paper attempts to address this gap of knowledge by combining Generative Adversarial Network (GAN) and transfer learning to examine the transferability of real-time crash prediction models under an extremely imbalanced data setting. Initially, a baseline model was developed using Deep Neural Network (DNN) with crash and microscopic traffic data collected from M1 Motorway in the UK in 2017. The dataset utilised in the baseline model is naturally imbalanced with 257 crash cases and 16,359,163 non-crash cases. To overcome data imbalance issue, Wasserstein GAN (WGAN) was utilised to generate synthetic crash data. Non-crash data were randomly undersampled due to computational limitations. The calibrated model was then applied to predict traffic crashes for five other datasets obtained from M1 (2018), M4 (2017 & 2018 separately) and M6 Motorway (2017 & 2018 separately) by using transfer learning. Model transferability was compared with standalone models and direct transfer from the baseline model. The study revealed that direct transfer is not feasible. However, models become transferable temporally, spatially, and spatio-temporally if transfer learning is applied. The predictability of the transferred models outperformed existing studies by achieving high Area Under Curve (AUC) values ranging between 0.69 and 0.95. The best transferred model can predict nearly 95% crashes with only a 5% false alarm rate by tuning thresholds. Furthermore, the performances of transferred models are on par with or better than the standalone model. The findings of this study proves that transfer learning can improve model transferability under extremely imbalanced settings which helps traffic engineers in developing highly transferable models in future.

Examining the relationship between impaired driving and past crash involvement in Europe: Insights from the ESRA study.

Driving under the influence of alcohol, drugs and fatigue are all important factors of crash causation. Exploring the link between driver attitudes and crash involvement provides understanding on these important issues. To that end, questionnaire answers of car drivers disclosing their attitudes on the impacts of driving under the influence of alcohol, drugs and fatigue, and their relationship with past crash involvement as car drivers were analysed. A two-step approach is adopted: Principal Component Analysis (PCA) was employed to consolidate relative questions in numeric factor quantities. Afterwards, binary logistic regression was implemented on the calculated component scores to determine the impact of perspectives of road users for each factor on past crash involvement of car drivers. Data from the international ESRA2015 survey were utilized. PCA indicated that it is possible to meaningfully merge 29 ESRA2015 questions relevant to driving under the influence of alcohol, drugs and fatigue into 8 informative components accounting for an adequate percentage of variance. Binary logistic analysis indicated that components involving overall personal and communal acceptance of impaired driving, overall and past year personal behaviour towards impaired driving and frequency of typical journey checks by traffic police were all quantities positively correlated with past crash involvement.

Investigation of accidents involving powered two wheelers and bicycles - A European in-depth study.

INTRODUCTION: The number of road fatalities have been falling throughout the European Union (EU) over the past 20 years and most Member States have achieved an overall reduction. Research has mainly focused on protecting car occupants, with car occupant fatalities reducing significantly. However, recently there has been a plateauing in fatalities amongst 'Vulnerable Road Users' (VRUs), and in 2016 accidents involving VRUs accounted for nearly half of all EU road deaths. METHOD: The SaferWheels study collected in-depth data on 500 accidents involving Powered Two-Wheelers (PTWs) and bicycles across six European countries. A standard in-depth accident investigation methodology was used by each team. The Driver Reliability and Error Analysis Method (DREAM) was used to systematically classify accident causation factors. RESULTS: The most common causal factors related to errors in observation by the PTW/bicycle rider or the driver of the other vehicle, typically called 'looked but failed to see' accidents. Common scenarios involved the other vehicle turning or crossing in front of the PTW/bicycle. A quarter of serious or fatal injuries to PTW riders occurred in accidents where the rider lost control with no other vehicle involvement. CONCLUSIONS: Highly detailed data have been collected for 500 accidents involving PTWs or bicycles in the EU. These data can be further analyzed by researchers on a case-study basis to gain detailed insights on such accidents. Preliminary analysis suggests that 'looked but failed to see' remains a common cause, and in many cases the actions of the other vehicle were the critical factor, though PTW rider speed or inexperience played a role in some cases. Practical Applications: The collected data can be analyzed to better understand the characteristics and causes of accidents involving PTWs and bicycles in the EU. The results can be used to develop policies aimed at reducing road deaths and injuries to VRUs.

Prediction of rear-end conflict frequency using multiple-location traffic parameters.

Traffic conflicts are heavily correlated with traffic collisions and may provide insightful information on the failure mechanism and factors that contribute more towards a collision. Although proactive traffic management systems have been supported heavily in the research community, and autonomous vehicles (AVs) are soon to become a reality, analyses are concentrated on very specific environments using aggregated data. This study aims at investigating -for the first time- rear-end conflict frequency in an urban network level using vehicle-to-vehicle interactions and at correlating frequency with the corresponding network traffic state. The Time-To-Collision (TTC) and Deceleration Rate to Avoid Crash (DRAC) metrics are utilized to estimate conflict frequency on the current network situation, as well as on scenarios including AV characteristics. Three critical conflict points are defined, according to TTC and DRAC thresholds. After extracting conflicts, data are fitted into Zero-inflated and also traditional Negative Binomial models, as well as quasi-Poisson models, while controlling for endogeneity, in order to investigate contributory factors of conflict frequency. Results demonstrate that conflict counts are significantly higher in congested traffic and that high variations in speed increase conflicts. Nevertheless, a comparison with simulated AV traffic and the use of more surrogate safety indicators could provide more insight into the relationship between traffic state and traffic conflicts in the near future.

Review and ranking of crash risk factors related to the road infrastructure.

The objective of this paper is the review and comparative assessment of infrastructure related crash risk factors, with the explicit purpose of ranking them based on how detrimental they are towards road safety (i.e. crash risk, frequency and severity). This analysis was carried out within the SafetyCube project, which aimed to identify and quantify the effects of risk factors and measures related to behaviour, infrastructure or vehicles, and integrate the results in an innovative road safety Decision Support System (DSS). The evaluation was conducted by examining studies from the existing literature. These were selected and analysed using a specifically designed common methodology. Infrastructure risk factors were structured in a hierarchical taxonomy of 10 areas with several risk factors in each area (59 specific risk factors in total), examples include: alignment features (e.g. horizontal-vertical alignment deficiencies), cross-section characteristics (e.g. superelevation, lanes, median and shoulder deficiencies), road surface deficiencies, workzones, junction deficiencies (interchange and at-grade) etc. Consultation with infrastructure stakeholders (international organisations, road authorities, etc.) took place in dedicated workshops to identify user needs for the DSS, as well as "hot topics" of particular importance. The following analysis methodology was applied to each infrastructure risk factor: (i) A search for relevant international literature, (ii) Selection of studies on the basis of rigorous criteria, (iii) Analysis of studies in terms of design, methods and limitations, (iv) Synthesis of findings - and meta-analysis, when feasible. In total 243 recent and high quality studies were selected and analysed. Synthesis of results was made through 39 'Synopses' (including 4 original meta-analyses) on individual risk factors or groups of risk factors. This allowed the ranking of infrastructure risk factors into three groups: risky (11 risk factors), probably risky (18 risk factors), and unclear (7 risk factors).

Impact of real-time traffic characteristics on crash occurrence: Preliminary results of the case of rare events.

Considerable efforts have been made from researchers and policy makers in order to explain road crash occurrence and improve road safety performance of highways. However, there are cases when crashes are so few that they could be considered as rare events. In such cases, the binary dependent variable is characterized by dozens to thousands of times fewer events (crashes) than non-events (non-crashes). This paper attempts to add to the current knowledge by investigating crash likelihood by utilizing real-time traffic data and by proposing a framework driven by appropriate statistical models (Bias Correction and Firth method) in order to overcome the problems that arise when the number of crashes is very low. Under this approach instead of using traditional logistic regression methods, crashes are considered as rare events In order to demonstrate this approach, traffic data were collected from three random loop detectors in the Attica Tollway ("Attiki Odos") located in Greater Athens Area in Greece for the 2008-2011 period. The traffic dataset consists of hourly aggregated traffic data such as flow, occupancy, mean time speed and percentage of trucks in traffic. This study demonstrates the application and findings of our approach and revealed a negative relationship between crash occurrence and speed in crash locations. The method and findings of the study attempt to provide insights on the mechanism of crash occurrence and also to overcome data considerations for the first time in safety evaluation of motorways.

How many crashes are caused by driver interaction with passengers? A meta-analysis approach.

INTRODUCTION: Conversation and other interactions with passengers while driving induce a level of distraction to the person driving. METHOD: This paper conducts a qualitative literature review on the effect of passenger interaction on road safety and then extends it by using meta-analysis techniques. RESULTS: The literature review indicates that the distraction due to passengers is a very frequent risk factor, with detrimental effects to various driving behavior and safety measures (e.g., slower reaction times to events, increased severity of injuries in crashes), associated with non-negligible proportions of crashes. Particular issues concern the effect of passenger age (children, teenagers) on which the literature is inconclusive. Existing studies vary considerably in terms of study methods and outcome measures. Nevertheless, a meta-analysis could be carried out regarding the proportion of crashes caused by this distraction factor. The selection of studies for the meta-analysis was based on a rigorous method including specific study selection criteria. The findings of the random-effects meta-analyses that were carried out showed that driver interaction with passengers causes a non-negligible proportion of road crashes, namely 3.55% of crashes regardless of the age of the passengers and 3.85% when child and teen passengers are excluded. Both meta-estimates were statistically significant, revealing the need for further research, especially considering the role of passenger age. PRACTICAL APPLICATIONS: Stakeholders could make good estimates on future crash numbers and causes and take action in order to counter the effects of passenger interaction.

Meta-analysis of the effect of road work zones on crash occurrence.

There is strong evidence that work zones pose increased risk of crashes and injuries. The two most common risk factors associated with increased crash frequencies are work zone duration and length. However, relevant research on the topic is relatively limited. For that reason, this paper presents formal meta-analyses of studies that have estimated the relationship between the number of crashes and work zone duration and length, in order to provide overall estimates of those effects on crash frequencies. All studies presented in this paper are crash prediction models with similar specifications. According to the meta-analyses and after correcting for publication bias when it was considered appropriate, the summary estimates of regression coefficients were found to be 0.1703 for duration and 0.862 for length. These effects were significant for length but not for duration. However, the overall estimate of duration was significant before correcting for publication bias. Separate meta-analyses on the studies examining both duration and length was also carried out in order to have rough estimates of the combined effects. The estimate of duration was found to be 0.953, while for length was 0.847. Similar to previous meta-analyses the effect of duration after correcting for publication bias is not significant, while the effect of length was significant at a 95% level. Meta-regression findings indicate that the main factors influencing the overall estimates of the beta coefficients are study year and region for duration and study year and model specification for length.

Incorporating real-time traffic and weather data to explore road accident likelihood and severity in urban arterials.

INTRODUCTION: The effective treatment of road accidents and thus the enhancement of road safety is a major concern to societies due to the losses in human lives and the economic and social costs. The investigation of road accident likelihood and severity by utilizing real-time traffic and weather data has recently received significant attention by researchers. However, collected data mainly stem from freeways and expressways. Consequently, the aim of the present paper is to add to the current knowledge by investigating accident likelihood and severity by exploiting real-time traffic and weather data collected from urban arterials in Athens, Greece. METHOD: Random Forests (RF) are firstly applied for preliminary analysis purposes. More specifically, it is aimed to rank candidate variables according to their relevant importance and provide a first insight on the potential significant variables. Then, Bayesian logistic regression as well finite mixture and mixed effects logit models are applied to further explore factors associated with accident likelihood and severity respectively. RESULTS: Regarding accident likelihood, the Bayesian logistic regression showed that variations in traffic significantly influence accident occurrence. On the other hand, accident severity analysis revealed a generally mixed influence of traffic variations on accident severity, although international literature states that traffic variations increase severity. Lastly, weather parameters did not find to have a direct influence on accident likelihood or severity. CONCLUSIONS: The study added to the current knowledge by incorporating real-time traffic and weather data from urban arterials to investigate accident occurrence and accident severity mechanisms. PRACTICAL APPLICATION: The identification of risk factors can lead to the development of effective traffic management strategies to reduce accident occurrence and severity of injuries in urban arterials.

Investigation on occupant injury severity in rear-end crashes involving trucks as the front vehicle in Beijing area, China.

PURPOSE: Rear-end crashes attribute to a large portion of total crashes in China, which lead to many casualties and property damage, especially when involving commercial vehicles. This paper aims to investigate the critical factors for occupant injury severity in the specific rear-end crash type involving trucks as the front vehicle (FV). METHODS: This paper investigated crashes occurred from 2011 to 2013 in Beijing area, China and selected 100 qualified cases i.e., rear-end crashes involving trucks as the FV. The crash data were supplemented with interviews from police officers and vehicle inspection. A binary logistic regression model was used to build the relationship between occupant injury severity and corresponding affecting factors. Moreover, a multinomial logistic model was used to predict the likelihood of fatal or severe injury or no injury in a rear-end crash. RESULTS: The results provided insights on the characteristics of driver, vehicle and environment, and the corresponding influences on the likelihood of a rear-end crash. The binary logistic model showed that drivers' age, weight difference between vehicles, visibility condition and lane number of road significantly increased the likelihood for severe injury of rear-end crash. The multinomial logistic model and the average direct pseudo-elasticity of variables showed that night time, weekdays, drivers from other provinces and passenger vehicles as rear vehicles significantly increased the likelihood of rear drivers being fatal. CONCLUSION: All the abovementioned significant factors should be improved, such as the conditions of lighting and the layout of lanes on roads. Two of the most common driver factors are drivers' age and drivers' original residence. Young drivers and outsiders have a higher injury severity. Therefore it is imperative to enhance the safety education and management on the young drivers who steer heavy duty truck from other cities to Beijing on weekdays.

Investigation of powered 2-wheeler accident involvement in urban arterials by considering real-time traffic and weather data.

OBJECTIVE: Understanding the various factors that affect accident risk is of particular concern to decision makers and researchers. The incorporation of real-time traffic and weather data constitutes a fruitful approach when analyzing accident risk. However, the vast majority of relevant research has no specific focus on vulnerable road users such as powered 2-wheelers (PTWs). Moreover, studies using data from urban roads and arterials are scarce. This study aims to add to the current knowledge by considering real-time traffic and weather data from 2 major urban arterials in the city of Athens, Greece, in order to estimate the effect of traffic, weather, and other characteristics on PTW accident involvement. METHODS: Because of the high number of candidate variables, a random forest model was applied to reveal the most important variables. Then, the potentially significant variables were used as input to a Bayesian logistic regression model in order to reveal the magnitude of their effect on PTW accident involvement. RESULTS: The results of the analysis suggest that PTWs are more likely to be involved in multivehicle accidents than in single-vehicle accidents. It was also indicated that increased traffic flow and variations in speed have a significant influence on PTW accident involvement. On the other hand, weather characteristics were found to have no effect. CONCLUSIONS: The findings of this study can contribute to the understanding of accident mechanisms of PTWs and reduce PTW accident risk in urban arterials.

Attitudes of Greek Drivers with Focus on Mobile Phone Use While Driving.

OBJECTIVE: This article investigates the attitudes and behavior of Greek drivers with specific focus on mobile phone use while driving. METHODS: The research is based on the data of the pan-European SARTRE 4 survey, which was conducted on a representative sample of Greek drivers in 2011. Analysis of the drivers' behavior was carried out by the statistical methods of factor and cluster analysis. RESULTS: According to the results of factor analysis, Greek drivers' responses in the selected questions were summarized into 4 factors, describing road behavior and accident involvement probability as well as their views on issues concerning other drivers' road behaviors, fatigued driving, enforcement of road safety, and mobile phone use while driving. The results of cluster analysis indicated 5 different groups of Greek drivers--the moderate, the optimistic, the conservative, the risky, and the reasonably cautious--and the characteristics of each group where identified. CONCLUSIONS: These results may be useful for the appropriate design of targeted road safety campaigns and other countermeasures.

A review of powered-two-wheeler behaviour and safety.

Powered-two-wheelers (PTWs) constitute a very vulnerable type of road users. The notable increase in their share in traffic and the high risk of severe accident occurrence raise the need for further research. However, current research on PTW safety is not as extensive as for other road users (passenger cars, etc.). Consequently, the objective of this research is to provide a critical review of research on Power-Two-Wheeler behaviour and safety with regard to data collection, methods of analysis and contributory factors, and discuss the needs for further research. Both macroscopic analyses (accident frequency, accident rates and severity) and microscopic analyses (PTW rider behaviour, interaction with other motorised traffic) are examined and discussed in this paper. The research gaps and the needs for future research are identified, discussed and put in a broad framework. When the interactions between behaviour, accident frequency/rates and severity are co-considered and co-investigated with the various contributory factors (riders, other users, road and traffic environment, vehicles), the accident and injury causes as well as the related solutions are better identified.

A review of the effect of traffic and weather characteristics on road safety.

Taking into consideration the increasing availability of real-time traffic data and stimulated by the importance of proactive safety management, this paper attempts to provide a review of the effect of traffic and weather characteristics on road safety, identify the gaps and discuss the needs for further research. Despite the existence of generally mixed evidence on the effect of traffic parameters, a few patterns can be observed. For instance, traffic flow seems to have a non-linear relationship with accident rates, even though some studies suggest linear relationship with accidents. On the other hand, increased speed limits have found to have a straightforward positive relationship with accident occurrence. Regarding weather effects, the effect of precipitation is quite consistent and leads generally to increased accident frequency but does not seem to have a consistent effect on severity. The impact of other weather parameters on safety, such as visibility, wind speed and temperature is not found straightforward so far. The increasing use of real-time data not only makes easier to identify the safety impact of traffic and weather characteristics, but most importantly makes possible the identification of their combined effect. The more systematic use of these real-time data may address several of the research gaps identified in this research.

Motorcycle riding under the influence of alcohol: results from the SARTRE-4 survey.

Riding a motorcycle under the influence of alcohol is a dangerous activity, especially considering the high vulnerability of motorcyclists. The present research investigates the factors that affect the declared frequency of drink-riding among motorcyclists in Europe and explores regional differences. Data were collected from the SARTRE-4 (Social Attitudes to Road Traffic Risk in Europe) survey, which was conducted in 19 countries. A total sample of 4483 motorcyclists was interviewed by using a face-to-face questionnaire. The data were analyzed by means of multilevel ordered logit models. The results revealed significant regional differences (between Northern, Eastern and Southern European countries) in drink-riding frequencies in Europe. In general, declared drinking and riding were positively associated with gender (males), increased exposure, underestimation of risk, friends' behaviour, past accidents and alcohol ticket experience. On the other hand, it was negatively associated with underestimation of the amount of alcohol allowed before driving, and support for more severe penalties.

Relationship between motorcyclists' attitudes, behavior, and other attributes with declared accident involvement in Europe.

OBJECTIVE: The objective of this study is to investigate patterns of road safety attitudes and behaviors of motorcyclists in Europe on the basis of the results of the pan-European questionnaire-based survey SARTRE-4, carried out in late 2010 in 18 European countries and Israel. In addition, we attempt to explore the link between attitudes, behaviors, and other motorcyclist attributes with motorcyclist involvement in accidents in the past 3 years, in which someone, including the rider, was injured and received medical attention as stated in the motorcyclists' responses. METHODS: The various components of motorcyclist attitudes and behaviors such as reasons for driving a motorcycle, driving while impaired, perceived risk factors, and risk-taking behavior were determined by means of a principal component analysis (PCA) on 38 variables contained in the survey. A binary logistic regression model was then applied in order to link the attitudes and the stated behavior with the declared involvement in past accidents. RESULTS: The results revealed 8 components. Component 1 (driving while impaired and speeding accident factors), component 2 (motorcycle benefits), component 3 (perceived risk of maneuvers), component 4 (sensation seeking), component 5 (road, vehicle, and environmental risk factors), component 7 (no modal options), and component 8 (attitudes toward drinking and friends' drinking) are associated with stated preferences and attitudes, whereas component 6 (dangerous and angry behaviors) is associated with stated behavior. Moreover, it was found that motorcyclists who tend to have dangerous attitudes and behaviors as well as younger motorcyclists are more likely to have been involved in an accident. It was also showed that driving exposure is positively associated with increased probability of a past accident. CONCLUSIONS: The findings of the study provide some insight into the association between attitudes, behaviors, and declared past accident involvement. Furthermore, the analysis of such large databases with the inclusion of many different countries constitutes a step for further research in the field of motorcyclists' behaviors and safety. Supplemental materials are available for this article. Go to the publisher's online edition of Traffic Injury Prevention to view the supplemental file.

Factors affecting accident severity inside and outside urban areas in Greece.

OBJECTIVES: This research aims to identify and analyze the factors affecting accident severity through a macroscopic analysis, with a focus on the comparison between inside and outside urban areas. Disaggregate road accident data for Greece for the year 2008 were used. METHODS: Two models were developed, one for inside and one for outside urban areas. Because the dependent variable had 2 categories, killed/severely injured (KSI) and slightly injured (SI), the binary logistic regression analysis was selected. Furthermore, this research aims to estimate the probability of fatality/severe injury versus slight injury as well as to calculate the odds ratios (relative probabilities) for various road accident configurations. The Hosmer and Lemeshow statistic and other diagnostic tests were conducted in order to assess the goodness-of-fit of the model. RESULTS: From the application of the models, it appears that inside urban areas 3 types of collisions (sideswipe, rear-end, with fixed object/parked car), as well as involvement of motorcycles, bicycles, buses, 2 age groups (18-30 and older than 60 years old), time of accident, and location of the accident, seem to affect accident severity. Outside urban areas, 4 types of collisions (head-on, rear-end, side, sideswipe), weather conditions, time of accident, one age group (older than 60 years old), and involvement of motorcycles and buses were found to be significant. CONCLUSIONS: Factors affecting road accident severity only inside urban areas include young driver age, bicycles, intersections, and collision with fixed objects, whereas factors affecting severity only outside urban areas are weather conditions and head-on and side collisions, demonstrating the particular road users and traffic situations that should be focused on for road safety interventions for the 2 different types of networks (inside and outside urban areas). The methodology and the results of this research may provide a promising tool to prioritize programs and measures to improve road safety in Greece and worldwide.