The risk of drivers engaging in distracting activies is increasing as in-vehicle technology and carried-in devices become increasingly common and complicated. Consequently, distraction and inattention contribute to crash risk and are likely to have an increasing influence on driving safety. Analysis of police-reported crash data from 2008 showed that distractions contributed to an estimated 5,870 fatalities and 515,000 injuries. This paper assesses the extent to which vision-based algorithms can detect different types of driver distraction under different driving conditions.

Data were collected on the National Advanced Driving Simulator from 32 volunteer drivers between the ages of 25 and 50. Participants drove through representative situations on three types of roadways (urban, freeway, and rural) twice:​ once with and once without distraction tasks. The order of the drives was counterbalanced. The three distraction tasks included a reaching task, a visual-manual task and a cognitive task which were repeated eight times throughout the drive.

Four different vision-based algorithms were evaluated. All of them performed significantly better than chance (random) performance . There was little difference between the approaches for the visual-manual bug task which required the most eyes-offroad time. The algorithm that estimated level of distraction by combining percent of glances to the road and long glances away from the road performed best for the arrows task, and was also the only algorithm that detected cognitive impairment. Differences across road types were also observed. Trade-offs exist between ensuring distraction detection and avoiding false alarms that complicate determining the most promising algorithm for detecting distraction. The differences in the algorithms’ abilities across evaluation criteria, road type, and distraction task type demonstrate critical trade-offs in capabilities that need to be considered. Depending on how feedback is presented to drivers, high false alarm rates may undermine drivers’ acceptance of the system. The study shows the importance of designing and testing algorithms with a variety of challenges to assess performance across a range of representative road and task types.