About

Road crashes remain one of the most prominent causes of unnatural death in Europe and worldwide. With the human driver being the most significant factor contributing to road crashes, the study of driving behavior is a primary research field within the research domain of road safety. Despite the extensive attention the domain already experiences, the research effort has been traditionally narrowed in scope by data scarcity. Additionally, the introduction of automated vehicles in the road environment, and advancements in intelligent traffic management systems architectures can unlock a great many new research directions and applications, further necessitating elaborate research on driving behavior and crash risk. Project i-TRAUM, funded by the Deutsche Forschungsgemeinschaft (DFG), aims to be the catalyst in this direction. Building on the status quo of driving behavior analytics, cutting-edge models are being developed to integrate the various exogenous factors affecting the driving task into one comprehensive predictive framework for crash risk. Special design consideration is input in rendering the framework suitable for near real-time applications, to enable integration of the insights acquired in commercial Advanced Driver Assistance Systems (ADAS).

The challenge

Human drivers have been traditionally categorized, based on their driving style, as conservative, normal, or aggressive. Recent research, however, recognizes that these categorizations are not inherent traits of individuals; drivers may transition from one behavioral state to another even within the same trip. Furthermore, the proneness to a crash cannot be the same for each driver: it is inevitable that personal driving competence and experience are aversion factors that kick in when needed to avoid a collision. This mechanism functions in different ways and intensities for each person. Up to now, the standard modeling practice has used predefined, static thresholds for what constitutes risky driving, with the imminence of danger reflected objectively by the successive lowering of these thresholds. To account for the individuality of the driving persona behind the wheel, soft thresholds must be adopted instead, which must be learned in real time, based on demonstrated driving behavior and past experience, providing adaptive, up-to-date risk cut-off values.

At the same time, the road environment itself involves a great degree of variability. The infrastructure, surrounding traffic, weather, and other related factors influence the potential for crashes. It is therefore the case that driving is a highly contextualized task, and, naturally, even for the same driving behavior, crash imminence varies with exogenous factors. All this should be considered within a unified modeling pipeline that operates in both online and offline modes, robustly fusing diverse data from various sources and sensors.

The i-TRAUM plan

The EU project i-DREAMS demonstrated the task complexity–coping capacity model for real-time crash risk estimation and highlighted the need for contextual consideration when modeling the driver–vehicle–infrastructure system. Leveraging this very extensive Naturalistic Driving Study (NDS) dataset, comprising high-resolution driving and biometric features of over three million kilometers-driven, project i-TRAUM forwards microscopic driving behavior research with clear contributions, aiming to develop an online context-aware crash risk estimation framework. This framework will be entirely modular, so that it remains functioning in the absence of data for some of its constituent factors. Crucially, it will consider real-time-occurring and historical information to produce the most accurate estimates. On the infrastructure side, this involves qualitative features related to road safety (alignment, road category, etc.), and dangerous locations that are responsible for past crashes. On the driver side, the framework will consider the past driving behavior, so that the individualized driving competence factor is also incorporated into the model, as well as an estimate of the current driver state and proneness to risky behavior. Lastly, the impact of the surrounding traffic will be considered, fusing information about the observed traffic state at the link level. By creating such a versatile framework, the influence of the individual behind the wheel and the driving environment is highlighted, resulting in soft thresholds as to what constitutes risky driving, and directly improving upon the status quo of static risk thresholds.