Spotlighting distraction in artificial intelligence driver assistance systems

Abstract

As artificial intelligence driver monitoring systems gain momentum in intelligent mobility, it is critical to analyse how distraction is defined and induced. This systematic review was specifically focused on studies conducted in driving simulators. A Boolean query was iteratively developed to retrieve articles from Scopus that fulfil the following criteria: (1) being an empirical study, (2) addressing driver distraction, (3) using a driving simulator, (4) aiming at developing an artificial intelligence monitoring system. After screening, 34 articles remained and were analysed according to four general themes: definition of distraction, characteristics of the scenarios used in the driving simulator, sampling of participants, and procedures. Results showed that the most common definitions of distraction consider it as a shift in the driver’s attention towards a secondary task, which implicates in a degradation of the execution of the primary task (i.e., driving the vehicle), and, consequently, a reduction in driving safety. Most articles described the scenarios used in the simulator in greater detail and, in some cases, variations in traffic density, visibility, and environmental conditions were observed. Furthermore, scripted critical events in the scenario (e.g., car in front of the participant breaking) were also used. Recruitment and samples varied greatly between studies, with the smallest population consisting of two and the largest of 97 participants. Despite the sample size, participants still needed to meet eligibility criteria such as having a driver’s license, possessing minimum driving experience, health prerequisites, being part of a specific group, age, and gender. Procedures and tasks were not always described in detail. However, several studies described an initial moment where participants could familiarize themselves with the simulator without taking measurements, while fewer reported that participants were allowed to familiarize themselves with the tasks. Session length varied from eight to 90 minutes. Regarding the operationalization of distraction in experiments, some studies required drivers to perform a single type of distraction-inducing task (mental calculations, use of In-Vehicle Information System (IVIS), cell phone operation, and manual tasks) with varying difficulty levels. Still, most studies relied on a combination of different tasks, such as cell phone use, physical tasks (e.g., drinking, moving objects, and applying makeup), and IVIS use. Results showed studies favour the description of the digital systems over the experiment design and procedures and a preference for locating the studies at the individual level of analysis, precluding a broader understanding of human behaviour as socially constructed and signified. We argue that articulation with higher levels of analysis would bring relevant explanations for actual road behaviour and personal and social factors should be considered when developing driver monitoring systems aimed at reducing distraction. Our results may assist future studies within the same scope, guiding the definition of effective experimental designs to test artificial intelligence driving monitoring systems, while contributing to a more holistic understanding of driver’s behaviour.