In the US all vehicular crashes that produce property damage over certain dollar-equivalents (each state varies somewhat), personal injury or death are reported and data are stored in large data sets. Some states include a "causal factor" or "contributing variable" that includes "asleep at the wheel", "driver asleep", etc. There is no uniform reporting of this that I know of, however, some of the states, such as California, Illinois, Michigan , New York or Pennsylvania have large enough populations and many years of data collection to produce significant volumes for analysis. Each state would need to be approached to determine if such variables are routinely collected at crash sites and to determine if data are available for research purposes.

I have used such data in the past and it is the only way to ensure comprehensive, longitudinal data for such jurisdictions. I doubt that insurance companies would share such data from claims, however that might be worth exploring.

I'm not sure how practical this is, but the SHRP2 Naturalistic Driving Study (~3000 vehicles instrumented with sensors/cameras over a ~12 month period) includes crash and near-crash events that might also be coded to fatigue level, but the data include video of the drivers' faces as well as a lot of other drivnig/vehicle data in real-time. Further information is online at: http://insight.shrp2nds.us/

Thank you so much for your suggestion. I need a driver facial expression data set that may contains images or videos.

Will that not bias your data for very new vehicles?

The distribution of vehicle model years is shown at:

https://insight.shrp2nds.us/data/category/vehicles#/graph/7

The vehicles are as new as 2013 (when the study was being run) with about 600 vehicles (approximately 20%?) being from year 2002 or earlier. They also tried to include a wide range of driver age as well.

Are you trying to generalize to a "normal" driving population?

There is a new public dataset that is based on realistic drowsiness for 3 levels of awareness. In consists of 60 participants.

It is called UTA-RLDD

I provide the links here:

Dataset:

https://sites.google.com/view/utarldd/home

Paper:

https://arxiv.org/abs/1904.07312

Code:

https://github.com/rezaghoddoosian/Early-Drowsiness-Detection

Demo:

https://youtu.be/3psnER2oVUA

I hope it helps!

http://cv.cs.nthu.edu.tw/php/callforpaper/datasets/DDD/

This is one good dataset which can be used for non commercial purpose.

It is important to use samples and “drowsiness “ interventions that are able to produce external validity. Way too many laboratory studies have been blindly applied to road safety policies and programs based on fine laboratory work that has little external validity; such good research then fails to impact road safety. Road safety research, if nothing else, must be based on an intentional effort to produce real answers to serious problems.