Identification of Cognitive Load in Naturalistic Driving

Over the last decade, naturalistic driving studies have provided significant insight into issues pertaining to the roles of distraction, inattention, and drowsiness in crash risk. New techniques have been developed for coding naturalistic data, mining it, and analyzing it. The application of these techniques has begun to deepen our understanding of how different types of secondary tasks may vary in their effects on crash risk. However, one methodological gap that still remains in the field concerns techniques for identifying periods of cognitive load within streams of naturalistic driving data. Cognitive load is a critical element of current concerns about driver distraction, particularly with regards to intentional tasks of listening and conversing while driving, but also as related to spontaneously occurring processes like daydreaming and becoming lost in thought, which may also take place while driving. The research reported here was undertaken to develop a methodology for identifying epochs of cognitive activity occurring while driving by using indicators of cognitive load that are based on eye behavior. It furthermore set out to test the ability of these metrics, when used together as part of a model or algorithm, to discriminate pre-identified epochs of cognitive task load from “just driving” baselines drawn from naturalistic driving data. To accomplish these goals, the project mined an existing naturalistic driving database to extract the following types of driving epochs: those containing conversation, which was operationally defined as involving cognitive load; those containing visual-manual interaction; and those containing just driving. Data from these epochs were augmented with new measures related to three hypothesized cognitive load indicators. The measures which were explored as behavioral markers of cognitive load included the following and were derived from the acquired naturalistic data: (a) very long glances to the forward road, coupled with (b) increased concentration of glances to the forward road, (c) reduced breadth of active scanning, and (d) changes in blink rate relative to other types of task loads as well as just driving baselines. Results of mixed-model analyses confirmed that cognitive epochs corresponded with the hypothesized patterns on two of the four indicators. However, the analysis revealed a surprising finding on all metrics: the “just driving” baseline epochs behaved as if they, too, contained significant amounts of cognitive load, perhaps in the form of daydreaming or mind-wandering. This finding complicated the modeling effort, which, although completed, yielded inconclusive results. The logistic regression approach showed much promise as a technique, but the predictor variables in the final model were difficult to interpret, and again seemed consistent with the notion that the just driving baselines, which were intended to serve as comparison epochs, were instead similarly composed of cognitive load periods. These latter epochs’ cognitive loads were simply of a different type—namely, daydreaming, thinking, or mind-wandering activity. This finding, while unexpected, has several important implications, both for future efforts toward the development of algorithms used to identify types of task load from naturalistic data, and for the use of baselines in epidemiological risk estimation.