Visual sampling of in -vehicle displays while driving: Empirical findings and a queueing network cognitive model.

Abstract

Safety concerns increase as the number of installed in-vehicle information systems rapidly grows. The increased cognitive and visual demands associated with such systems result in frequent and prolonged glances away from the road, which make crashes more likely to occur. The goal of this work was to model how the visual demands of driving and of in-vehicle map-reading tasks affect the allocation of visual attention between the road and an in-vehicle display. In a driving-simulator, the momentary visual demands of driving on curves of several radii were quantified with the visual occlusion method. Experimental findings revealed that perceived visual demand (expressed as the proportion of time spent looking at the road) increased linearly with the reciprocal of curve radius (Demand = 0.39 + 33/Radius). In a follow-up experiment, the duration of glances to an in-vehicle display while driving on constant radius curves was measured. As the visual demand of driving increased from straight roads to sharp curves (194 m radius), subjects made more glances to the display (2.6 to 3.5) but glances were shorter (1.8 to 1.2 s). Total glance time at the display remained unchanged, indicating that drivers adjusted the timing of glances to the increased load of driving. In a third experiment, similar findings were made for a longer in-vehicle map-reading task. Additionally, the demand imposed by continuous map rotation resulted in longer glances at the display (2.5 to 2.0 s). A computational cognitive model, QN-MHP (the Queueing Network-Model Human Processor) was used to model some of these experimental findings while generating realistic steering actions in a driving simulator. The model predicted well the timing of task performance (e.g., task time predictions were within 1 s of actual times) for a dual-task condition of map-reading while driving. The approach for modeling task-switching is novel among existing human performance models in relying on local, rather than central, processes and is dynamic, rather than predetermined. The model, though preliminary, has the prospect of serving as both a human factors tool for the design of in-vehicle information systems and as a tool for better understanding human performance under such conditions.