This study reviews the Artificial Intelligence and Machine Learning approaches developed thus far for driver profile and driving pattern recognition, representing a set of macroscopic and microscopic behaviors respectively, to enhance the understanding of human factors in road safety, and therefore reduce the number of crashes. It provides a definition of the two scientific fields in terms of safety, and identifies the most efficient approaches used regarding methodology, data collection and driving metrics. Results show that K-means and Neural Networks are the most commonly used methodologies for driver profile identification, and Dynamic Time Warping for driving pattern detection. Most studies discovered driver profiles related to aggressiveness, considering mainly speed and acceleration as driving metrics. Based on the gaps and challenges identified, this paper provides a new framework for combining microscopic and macroscopic driving behavior analysis, instead of examining them separately as is the state-of-theart. Such combined results can potentially improve the development of traffic risk models, which could be exploited in applications that monitor drivers in real-time and provide feedback. These models will represent human behavior more accurately, which can eventually lead to the recognition of 'optimal' human driving patterns that Automated Vehicles (AV) could 'mimic' to become safer.

@en

Driver behavior analytics is an important concept that plays a significant role in the understanding of road crashes. This paper investigates the optimal number of driver profiles to understand the most important characteristics that differentiate drivers and extract useful insights on the value of using different clustering approaches in profile recognition. To this end, two Machine Learning clustering algorithms, the K-Means and OPTICS algorithms, are applied on driving data from a large naturalistic experiment using almost 18 K trips recorded from 130 drivers. The results revealed 3 profiles, the less risky drivers, the modest drivers and the more aggressive drivers. Clustering was based on 3 important driving behavior characteristics, namely the number of speeding, headway and harsh events per 100 km. The less risky drivers profile was revealed by both algorithms, whereas drivers of higher aggressiveness are distinguished by K-Means based on the driving feature that dominates the rest. The OPTICS algorithm showed that many drivers, especially the aggressive ones, present unique behavior that cannot be grouped together with other drivers. The interpretability of driver profiles resulting from the application of these unsupervised learning techniques is worsened as the number of clusters increases. The association between driver profiles and individual characteristics leads to the conclusion that aggressiveness is mainly driven by personality traits and less by specific characteristics such as gender, age or past accident history. The results of this study can be potentially used to develop profile-specific applications that provide feedback to drivers and reduce their crash risk.@en

Driving pattern recognition has been applied for the purposes of driving styles identification and harsh driving events detection. However, the evolution of driving behavior around and especially before such events has not been investigated at a microscopic level. The objective of this research is to reveal existing driving patterns around harsh events at the driving ‘pulse’ level i.e. a few seconds before and after the event. For that purpose, a time-series clustering approach is applied on speed and heart rate metrics of individual drivers using data collected from a large naturalistic driving study. Results show that there are distinct speed patterns before harsh braking, harsh acceleration, and harsh cornering events. A deceleration is identified shortly before most harsh acceleration and cornering events, which possibly indicates reckless behavior, i.e. drivers not dedicating enough time to smoothly brake before cornering, or of a brief ‘decision-making’ moment before the harsh manoeuvre. On the contrary, speed seems to be steady before harsh braking events. Regarding heart rate, the analysis revealed certain patterns only after raw data were cleansed and filtered. These patterns may show increasing, decreasing or variable heart rate trends, which may correspond to different stress patterns of drivers around harsh events. Finally, we introduce the concept of driving pattern consistency, which can reveal the share of individual drivers that follow the same harsh event pattern. It is indicated that more than half of the drivers are not consistent, suggesting that driving patterns around harsh events may be more context-related than driver personality-related.@en

Recent research in transport safety focuses on the processing of large amounts of available data by means of intelligent systems, in order to decrease the number of accidents for transportation users. Several Machine Learning (ML) and Artificial Intelligence (AI) applications have been developed to address safety problems and improve efficiency of transportation systems. However exchange of knowledge between transport modes has been limited. This paper reviews the ML and AI methods used in different transport modes (road, rail, maritime and aviation) to address safety problems, in order to identify good practices and experiences that can be transferable between transport modes. The methods examined include statistical and econometric methods, algorithmic approaches, classification and clustering methods, artificial neural networks (ANN) as well as optimization and dimension reduction techniques. Our research reveals the increasing interest of transportation researchers and practitioners in AI applications for crash prediction, incident/failure detection, pattern identification, driver/operator or route assistance, as well as optimization problems. The most popular and efficient methods used in all transport modes are ANN, SVM, Hidden Markov Models and Bayesian models. The type of the analytical technique is mainly driven by the purpose of the safety analysis performed. Finally, a wider variety of AI and ML methodologies is observed in road transport mode, which also appears to concentrate a higher, and constantly increasing, number of studies compared to the other modes.@en

Road safety research is largely focused on prediction and prevention of technical, human or organisational failures that may result in critical conflicts or crashes. Indicators of traffic risk aim to capture the passage to unsafe states. However, research in other industries has shown that it is meaningful to analyse safety along the whole spectrum of behaviours. Knowing the causes and patterns of “successful” interactions, rather than failures, could give new insights on the complexity of the system and the adaptability and resilience of its users in handling the inherent risks. The concept is known as Safety-II and has been extensively explored in the aviation, healthcare and process engineering domains. In this paper, we explore a new Safety-II paradigm for road safety research. We briefly review Safety-II applications in other sectors. We then present a Safety-II model for road safety, by means of an inverse version of Hyden's “safety pyramid”. Furthermore, we discuss a number of key road safety goals, theories, analysis methods and data sources and map them into a tentative taxonomy of Safety-I and Safety-II applications. It is concluded that there can be opportunities and benefits from adopting this new mindset, in order to complement existing approaches.

@en

This paper attempts to shed light on the temporal evolution of driving safety efficiency with the aim to acquire insights useful for both driving behavior and road safety improvement. Data exploited herein are collected from a sophisticated platform that uses smartphone device sensors during a naturalistic driving experiment, at which the driving behavior from a sample of two hundred (200) drivers during 7-months is continuously recorded in real time. The main driving behavior analytics taken into consideration for the driving assessment include distance travelled, acceleration, braking, speed and smartphone usage. The analysis is performed using statistical, optimization and machine learning techniques. The driver's safety efficiency index is estimated both in total and in several consecutive time windows to allow for the investigation of safety efficiency evolution in time. Initial data analysis results to the most critical components of microscopic driving behaviour evolution, which are used as inputs in the k-means algorithm to perform the clustering analysis. The main driving characteristics of each cluster are identified and lead to the conclusion that there are three main driving groups of the a) moderate drivers, b) unstable drivers and c) cautious drivers.@en

This paper deals with the problem of improving the existing optimization techniques for Data Envelopment Analysis (DEA). The algorithm proposed herein is a combination of the "quickhull algorithm" and a DEA algorithm written in Python programming language. To the best of the authors' knowledge no prior effort has been made to date to propose a methodology for reducing the running time of a DEA problem that incorporates multiple inputs and outputs. The algorithmic implementation is applied on the existing problem of driving efficiency evaluation by exploiting a driving data sample of 10,088 trips collected from smartphone devices. Results indicate that the proposed algorithm is performing relatively well for Big Data compared to other existing DEA algorithmic methodologies that yield the same optimal solution such as Standard DEA and RBE DEA methodologies. The results obtained are calculated for the test sets of 100, 500, 1000, 5,000 and 10,088 Decision-Making-Units (DMUs) and compared in terms of running time of each of the algorithms applied. The results of per trip analysis can be exploited in order to classify trips into different efficiency categories (such as efficient, less efficient, non-efficient) and present their main characteristics.@en

This research aims to correlate drivers' characteristics with their safety performance. In order to achieve this objective, two different data sets were used deriving from 12 drivers who participated on an on-road driving experiment while being assessed by a safety behaviour expert. Drivers participating in the experiment also responded to the questionnaire administered, which mainly included self-evaluation questions along with other that concerned the condition of their vehicle, demographics, driving experience etc. The analysis of drivers' behavior was carried out using the statistical methods of factor analysis and linear regression analysis. Three groups that characterize drivers' perception of careless, aggressive and cautious driving behavior were derived from factor analysis. Moreover, linear regression analysis revealed that driving experience, headways, self-reported driving skill and defensive driving positively affect the overall on-road driving performance score. More precisely, it was ascertained that driving experience leads to statistically significant increase in overall on-road driving performance score. Other factors such as defensive driving and keeping safe distances from other vehicles. The findings of this research could be exploited for providing feedback on drivers regarding their driving behavior in order to further reduce their driving risk.@en

The aim of this paper is the development of driver speed models based on detailed driving data collected from smartphone sensors. More specifically, this research investigates to which extent various driving behaviour parameters (harsh acceleration and deceleration events, driving distance, percentage of driving time per different road types, etc.) interact with each other and how these might potentially serve as driving speed predictors. Real time driving behaviour data collection was carried out using a smartphone application. Data were collected from 100 drivers and a total of 18,853 trips between July and December 2016. Six different linear regression models are developed to predict average driving speed under different driving conditions. One model for each road type (urban, rural and highways), one model for the risky hours period and one for the rest of the day, and a general model. The results indicated that the distance covered by the driver, the number of the harsh events occurred and the average deceleration are the strongest predictors of the average speed of a driver.@en

The aim of this paper was to provide a methodological framework for estimating the amount of driving data that should be collected for each driver in order to acquire a clear picture regarding their driving behavior. We examined whether there is a specific discrete time point for each driver, in the form of total driving duration and/or the number of trips, beyond which the characteristics of driving behavior are stabilized over time. Various mathematical and statistical methods were employed to process the data collected and determine the time point at which behavior converges. Detailed data collected from smartphone sensors are used to test the proposed methodology. The driving metrics used in the analysis are the number of harsh acceleration and braking events, the duration of mobile usage while driving and the percentage of time driving over the speed limits. Convergence was tested in terms of both the magnitude and volatility of each metric for different trips and analysis is performed for several trip durations. Results indicated that there is no specific time point or number of trips after which driving behavior stabilizes for all drivers and/or all metrics examined. The driving behavior stabilization is mostly affected by the duration of the trips examined and the aggressiveness of the driver.@en

Introduction: Technological advancements during recent decades have led to the development of a wide array of tools and methods in order to record driving behavior and measure various aspects of driving performance. The aim of the present study is to present and comparatively assess the various driver recording tools that researchers have at their disposal. Method: In order to achieve this aim, a multitude of published studies from the international literature have been examined based on the driver recording methodologies that have been implemented. An examination of more traditional survey methods (questionnaires, police reports, and direct observer methods) is initially conducted, followed by investigating issues pertinent to the use of driving simulators. Afterwards, an extensive section is provided for naturalistic driving data tools, including the utilization of on-board diagnostics (OBD) and in-vehicle data recorders (IVDRs). Lastly, in-depth incident analysis and the exploitation of smartphone data are discussed. Results: A critical synthesis of the results is conducted, providing the advantages and disadvantages of utilizing each tool and including additional knowledge regarding ease of experimental implementation, data handling issues, impacts on subsequent analyses, as well as the respective cost parameters. Conclusions: New technologies provide undeniably powerful tools that allow for seamless data handling, storage, and analysis, such as smartphones and in-vehicle data recorders. However, this sometimes comes at considerable costs (which may or may not pay off at a later stage), while legacy driver recording methods still have their own niches to fill in research. Practical Applications: The present research supports researchers when designing driver behavior monitoring studies. The present work enables better scheduling and pacing of research activities, but can also provide insights for the distribution of research funds.@en

This paper aims to provide a methodological framework for the comparative evaluation of driving safety efficiency based on Data Envelopment Analysis (DEA). The analysis considers each driver as a Decision Making Unit (DMU) and aims to provide a relative safety efficiency measure to compare different drivers based on their driving performance. The last is defined based on a set of driving analytics (e.g. distance travelled, speed, accelerations, braking, cornering and smartphone usage) collected using an innovative data collection scheme, which is based on the continuous recording of driving behavior analytics in real time, using smartphone device sensors. Safety efficiency is examined in terms of speed limit violation, driving distraction, aggressiveness and safety on urban, rural and highway road and in an overall model. DEA models are identifying the most efficient drivers that lie on the efficiency frontier and act as peers for the rest of the non-efficient drivers. The proposed methodological framework is tested on data from fifty-six (56) drivers during a 7-months period. Findings help distinguish the most efficient drivers from those that are less efficient. Moreover, the efficient level of inputs and outputs to switch from non-efficiency to the efficiency frontier is identified. Results also provide a potential for classification of the driving sample based on drivers’ comparative safety efficiency. The main characteristics of the most and less efficient drivers are consequently analyzed and presented. Most common inefficient driving practices are identified (aggressive, risky driving, etc.) and driving behavior is comparatively evaluated and analyzed.@en

This paper aims to investigate which parameters affect users’ willingness to pay for alternative usage-based motor insurance pricing schemes such as Pay-as-you-drive (PAYD) and Pay-as-how-you-drive (PHYD). For that reason, a dedicated questionnaire was designed and administered to 100 participants including both revealed and stated preference questions and proposed scenarios regarding current and alternative insurance schemes. In order to account for unobserved heterogeneity, a mixed logit model was applied to analyze vehicle insurance choice. Candidate variables include the effect of driving characteristics, drivers’ demographics and the price of vehicle insurance premiums. Two distinct mixed logit models were developed; one mixed logit model to investigate the factors influencing the choice of present insurance policy over PAYD and one for present insurance policy over PHYD. Results indicated that women and smartphone owners are more likely to choose a new insurance schemes. Kilometers and cost reduction were also found to affect similarly the choice for both Usage-Based-Motor Insurance (UBI.) Moreover, the higher the speed reduction imposed to the user, the lower the probability of the UBI scheme to choose it. It was also found that people over 40 years old are less likely to choose PHYD insurance.@en

The objective of this paper is to provide a review of the most popular and often implemented methodologies related to Usage-based motor insurance (UBI). UBI schemes, such as Pay-as-you-drive (PAYD) and Pay-how-you-drive (PHYD), are a new innovative concept that has recently started to be commercialized around the world. The main idea is that instead of a fixed price, drivers have to pay a premium based on their travel and driving behaviour. Despite the fact that it has been implemented only for a few years, it appears to be a very promising practice with a significant potential impact on traffic safety as well as on traffic congestion mitigation and pollution emissions reduction. To this end, the existing literature on UBI schemes is reviewed and research gaps are identified Findings show that there is a multiplicity and diversity of several research studies accumulated in modern literature examining the correlation between PAYD (based on driver's travel behaviour and exposure) and PHYD (based on driving behaviour) schemes and crash risk in order to determine crash risk. Moreover, there is evidence that UBI implementation would eliminate the cross-subsidies phenomenon, which implies less insurance costs for less risky and exposed drivers. It would also provide a strong motivation for drivers to improve their driving behaviour, differentiate their travel behaviour and reduce their degree of exposure by receiving feedback and monitoring their driving preferences and performance, which would result in crash risk reduction both totally and individually. The paper finally discussed the current and emerging challenges on this research field.

@en

This study analyzed road, traffic, and human factors of pedestrian crossing behavior through the development of integrated choice and latent variables models. The analysis used recent research as a starting point, in which a two-stage approach was successfully tested, including a separate estimation of human factors and choice models. Data from a dedicated field survey were used: pedestrian field observations of road crossing behavior in different road and traffic scenarios were combined with a questionnaire on pedestrian attitudes, perceptions, motivations, and declared behaviors. The integrated choice and latent variables models were developed for four road types: major urban arterials, main roads, secondary roads, and residential roads. Results suggest that the effect of traffic conditions on pedestrian crossing choices was more important on main and secondary urban roads, whereas on major urban arterials and on residential roads it was nonsignificant. In regard to the effects of human factors, a risk latent variable was found to enhance the explanatory power of most of the models. This variable was estimated on the basis of different indicators in each case, reflecting a clear risk-taking tendency on major and main roads and an optimization tendency on minor roads. Overall, it is indicated that the integration of human factors in pedestrian crossing models provides meaningful and insightful results, and they may be advantageous compared with the two-stage approach.@en

The objective of this paper is to provide a critical review of the most popular and often implemented methodologies related to Usage-based motor insurance (UBI). UBI schemes, like Pay-as-you-drive (PAUD) and Pay-how-you-drive (PHUD), are a new innovative concept that has recently started to be commercialized around the world. The main idea is that instead of a fixed price, drivers have to pay a premium based on their driving behaviour and degree of exposure. Despite the fact that it has been implemented only for a few years, it is proven to be a very promising practice with a significant potential impact on traffic safety. This is achieved by the financial incentive given to drivers in order to improve their driving behaviour such as reducing the number of harsh braking and acceleration events taking place or reducing their degree of exposure such as their annual mileage, the time of the day travelling etc. and therefore reduces traffic risk. It can also be beneficial towards other social objectives such as traffic congestion and pollution emissions reduction. To this end, the existing literature on UBI schemes is critically reviewed and research gaps are identified. Findings show that there is a multiplicity and diversity of several research studies accumulated in modern literature examining the correlation between PAUD (based on driver's exposure) and PHUD (based on driving behaviour) schemes and traffic risk in order to determine accident risk. Moreover, it seems that UBI implementation would eliminate the cross-subsidies phenomenon, which implies less insurance costs for goods and less exposed drivers. Moreover, it would also provide a strong motivation for drivers to improve their driving behaviour and reduce their degree of exposure by receiving feedback and monitoring their driving performance and exposure which would result in traffic risk reduction both totally and individually.@en

This study compares the performance of statistical and Bayesian combination models with classical single time series models for short-term traffic forecasting. Combinations are based on fractionally integrated autoregressive time series models of travel speed with exogenous variables that consider speed's spatio-temporal evolution, and volume and weather conditions. Several statistical hypotheses on the effectiveness of combinations compared to the single models are also tested. Results show that, in the specific application, linear regression combination techniques may provide more accurate forecasts than Bayesian combination models. Moreover, combining models with different degrees of spatio-temporal complexity and exogeneities is most likely to be the best choice in terms of accuracy. Moreover, the risk of combining forecasts is lower than the risk of choosing a single model with increased spatio-temporal complexity.@en

We examine the effects of incident occurrence on freeway traffic. Although the true influence of a freeway incident may not be directly observed, it may be identified using the maximum spatial extent of the disturbance induced to upstream traffic. Spatial and temporal extent is susceptible to various traffic, weather, geometry and incident specific factors. The above framework is implemented using a Multiple Indicators-Multiple Causes (MIMIC) latent variable model. Results using data from Athens, Greece indicate that the MIMIC model is able to accurately determine the influence of an incident on upstream traffic with 72% probability of producing mean square errors less than 0.05. Speed, lane volume, alignment, rainfall intensity, clearance time and whether the incident is a secondary or a primary one are among the most influential factors for assessing the anticipatory effect of incidents to traffic.@en