i-DREAMS: D3.2 Toolbox of recommended data collection tools and monitoring methods and a conceptual definition of the Safety Tolerance Zone

Abstract

This deliverable aims to present the practical conceptualisation of the Safety Tolerance Zone (STZ) in order for the project to transition from a theoretical framework for operational design into the practical implementation of the STZ estimation in the subsequent Work Packages (WPs) of the project. In order for this transition to be outlined, the proposed measurements and technologies for driver monitoring and evaluation need to be contrasted with the sensing capabilities of the technologies available within the project and an appropriate modelling framework must be defined for the STZ. In order to assure the real-time estimation of the STZ levels and promptly/swiftly trigger adequate interventions, deviations from normal driving must also be identified. Accordingly, a detailed description of driver monitoring measurements which help to determine the STZ levels as well as identify the abnormal driving, is provided within the deliverable. Where applicable, recommendations on measurements along with the corresponding thresholds for detection of events per mode are provided. More specifically, risk factors (e.g. actual speed, harsh acceleration and braking, or aggressiveness) associated with the STZ as well as indicators of abnormal driving (e.g. ECG, hands on the wheel, fatigue, sleepiness) are initially specified. To obtain available thresholds in order to convey the idea of creating a starting point for defining the STZ levels and abnormal driving, a literature review was conducted. The review demonstrated that thresholds are mostly employed detecting high speeds, short time headways and harsh acceleration or braking events in cars. However, limited information on thresholds was found for trucks, buses and rails. Additionally, considerations on how to exploit the available technologies (i.e. CardioID, OSeven, Mobileye) in the experimental setup for all transport models are highlighted. The final section of the deliverable deals with the mathematical formulation of the STZ in an appropriate modelling framework. Following a thorough literature review of models dealing with driver behavior and collision risk modelling in real-time, the most prominent approaches were found to be Dynamic Bayesian Networks or DBNs (a probabilistic graphical time-series model) and Long Short-Term Memory networks or LSTMs (a deep neural network formulation). In order to allow for more flexibility, and keeping in mind that within the project, post-trip driver evaluations are also to be designed, two approaches, namely Structural Equation Models (SEMs) and Discrete Choice Models (DCMs) were also proposed that provide “static” predictions, in contrast with DBNs and LSTMs which work dynamically (i.e. in real-time). For each of the aforementioned methods or techniques, a brief description of their underpinning procedure is presented, followed by their application for the identification of the STZ levels along with abnormal driving. The most significant practical considerations concerning the modelling of the STZ include the experimentation of the classification algorithm once data become available, the flexibility of the risk indicators with their respective thresholds as well as the problem of data labelling and the specification of driving scenarios, in which STZ levels are most distinctive. Finally, the project subsequent steps comprise the coding of the models, in an appropriate programming framework, and an extensive experimental testing and tuning of the models using data from driving simulator and on-road trials, in order to guarantee the effective and correct real-time identification of the STZ levels as well as the proper triggering of interventions for road safety enhancement.