Highly automated vehicle safety driver monitoring support through Bayesian MCMC Logistic Regression

As of 2019, in the US alone, there were more than 1,400 highly automated vehicles (HAVs) in testing by more than 80 companies spread across 36 states. While in their testing stages, HAVs include the presence of a safety driver. These specially trained drivers must take back control if the vehicle is not responding correctly to the ever-changing roadway environment. Research suggests that monitoring the driving task can lead to an increase in fatigue, followed by a decrease in vigilance, and a decrease in situation awareness.

As part of California DMV AV data policies, Waymo publicly released disengagement request and mileage data on its HAV testing in 2018. From these data, which were represented in mileage, a temporal metric was created which revealed that safety drivers were monitoring for 150-250 hours without a disengagement request in Waymo HAVs. While there are still many unknowns, these results suggest that safety drivers testing HAVs may be susceptible to decreased monitoring performance. To mitigate these effects, the benefits of training and performance support systems are discussed.

From the first analysis, as well as a literature review, a gap was identified which could be filled by providing in-vehicle support for safety driver monitoring. It was found that performance support systems, which aid in the monitoring task and decide warning timing, can help with the shortcomings of human monitoring. To quantify this, Bayesian Logistic Regression Models were created using vehicle trajectory data from the SHRP2 NDS data set to predict forward collision, with hopes of implementation into future performance support systems that aid safety driver vigilance. The results indicate that the best model can perform with an accuracy of 88% and a sensitivity of 93%. The data set was limited to predicting with the use of 29 crash events. Future models may be able to utilize the posterior probability distributions from this model to produce more accurate results. Furthermore, next generation models will also be able to utilize more data, and therefore more predictor variables, to anticipate a forward collision.

From an extensive literature review, there is reason to believe that Likelihood Alarm Systems (LAS) may pose as a strong performance support system for safety drivers, as they have been shown to help drivers sustain situation awareness through automation transparency without creating fatigue or alarm annoyance, even when used with imperfect collision prediction systems. The imperfect model created in this paper had less false alarms when implemented into an LAS, creating a better monitoring environment for the safety driver. Future work should be done to create performance support systems and collision prediction models which aid in the monitoring task and create HAV transparency in order to optimize safety driver monitoring performance.