Bryan Reimer

The influence of partial automation systems on driver behavior may grow as their performance and design evolves. Using a large-scale naturalistic driving dataset, we examined the prevalence and frequency of task engagement with and without automation active, comparing behavior across vehicles equipped with different partial-automation driver systems.Video and vehicle CAN data were collected from over 12,800 miles of highway driving with and without the use of partial automation across three partial automation system types: (i) hands-on wheel collaborative steering with torque-based driver monitoring system, (ii) hands-on wheel with torque-based driver monitoring system without collaborative steering, and (iii) hands-free driving system with camera-based driver monitoring. Negative-binomial regression models used to estimate the prevalence, frequency, and the changes in prevalence over time with and without partial automation active. Drivers were found to spend over 30% of their time engaged in at least one in-vehicle task. Variation in behaviors associated with system design and vehicle layout included: (i) greater prevalence of manual electronic device (e.g., cellphone) use while using the hands-free and hands-on automation, (ii) greater prevalence and frequency of reading and watching the center screen in the vehicle with a single central screen-based interface, and (iii) differences in steering-wheel vehicle controls and voice interactions across vehicle designs.These findings suggest that the use of partial automation with its associated driver support designs may shape drivers’ task engagement. Understanding these and other driver behavior patterns can inform future design strategies for driver monitoring and support systems, as well as evidence-based safety policy.

In 2012, a major traffic safety organization tasked the MIT AgeLab with developing a data-driven system for rating the effectiveness of new technologies intended to improve safety. Such a system was envisioned as having the potential to educate and guide consumers towards more confident and strategic purchasing decisions, ideally encouraging adoption of technologies with demonstrated safety benefit. In addition, an evaluation of the status and extent of existing data was seen as a way of identifying research gaps in the state of knowledge about safety systems. The focus was on technologies as a class, not on a rating review of individual vehicle model implementations. As conceptualized, the system aimed to complement traditional NCAP style ratings as well as to provide consumers with transparent information on early stage and often improving safety technologies. Development of the rating system and identification of data was undertaken in consultation with a range of academic, industrial, consumer, NGO, and governmental experts as well as with representatives of many of the major automotive manufacturers and suppliers. A key observation that emerged was that data on objectively demonstrable real-world benefits were generally sparse and often lower than expectations based on theoretical considerations, simulation studies, or pre-production evaluations. A number of experts and industry representatives expressed some surprise at both the divergence between theoretical and observed benefits and the relative scarcity of data upon which to make objective assessments, while others were quite aware of these issues and the need for the development of objective data under real-world operating conditions. A number of factors that might be relevant to understanding why such differences between expected and observed benefits exist were identified. One outcome of this effort was the founding of the Advanced Vehicle Technology (AVT) consortium to collect and examine objective data under naturalistic driving conditions of how drivers interact with, engage or don't engage, various production safety and driver assistance systems. This ongoing effort is contributing to insights concerning actual benefits and reasons for benefit gaps. Drawing from our initial work, as well as newer sources of data, we argue that the evaluation and rating of safety and driver assistance technologies for informing the consumer and the public at large should consider both theoretical potential and existing demonstrated benefit of specific technologies. This position is increasingly relevant as the effectiveness of many newer technologies have the potential to actually improve over the lifecycle of a vehicle through software updates. The emphasis on ratings based on observed benefit for actual drivers under realworld conditions is proposed to be complementary, rather than competing with, ratings focused largely on controlled test track evaluations of engineered capability. In addition, a case is made for providing ratings that assesses benefit relative to overall crash, injury, and fatality rates-and in relation to the specific scenario / crash event type that a given technology is intended to address. This approach should aid consumers in considering the extent to which a specific technology is or is not relevant to their particular driving needs.

This document describes both subject training and the experimental administration of the auditory presentation-verbal response delayed digit recall task (n-back) used by the MIT AgeLab in a series of simulation and on-road driving studies. The full stimulus item set, training materials and instructions are provided to assist other researchers who are interested in using the task and methodology in other work.

With recent advances in wireless and computing technology, distracted driving has become commonplace on the roadway, extending far beyond the cell phone to other distractions affecting driver attention. This research builds upon a previous study that collected eye positions from a driver completing some routine tasks in an automobile. Traditional methods of eye movement analysis relied on manual identification of patterns from recorded scene video. However, these methods are not efficient when processing the large data sets acquired to discriminate significantly between a drivers' behavior under different road and distraction conditions. In this work, new procedures and tools for the analysis of recorded eye positions are proposed for the collected data. To overcome the time demands for manual classification methods, automated methods of analysis are developed as the major focus of this research. Signal analysis is considered as a method of identifying periods of eye movements that differ significantly from the "steady state". Methods for the analysis of eye movements in a static setting are developed and used for studying distracted driving and other tasks occurring in a dynamic setting. In a dynamic setting, fixations, smooth and saccadic eye movements are identified from the recorder eye positions and manually mapped to the actions being completed. This research focuses on developing a software tool for examining the link between the location and time a subject allocates their attention on a particular region of the visual field. Examples detail the feasibility of analyzing various types of recorded eye positions including laboratory experiments, simulation and real road driving experiments. Finally, Hidden Semi-Markov models are proposed to model how a driver interacts with in-vehicle devices.

Non-intrusive, real-time analysis of the dynamics of the eye region allows us to monitor humans’ visual attention allocation and estimate their mental state during the performance of real-world tasks, which can potentially benefit a wide range of human-computer interaction (HCI) applications. While commercial eye-tracking devices have been frequently employed, the difficulty of customizing these devices places unnecessary constraints on the exploration of more efficient, end-to-end models of eye dynamics. In this work, we propose CLERA, a unified model for Cognitive Load and Eye Region Analysis, which achieves precise keypoint detection and spatiotemporal tracking in a joint-learning framework. Our method demonstrates significant efficiency and outperforms prior work on tasks including cognitive load estimation, eye landmark detection, and blink estimation. We also introduce a large-scale dataset of 30 k human faces with joint pupil, eye-openness, and landmark annotation, which aims at supporting future HCI research on human factors and eye-related analysis.