With the advance of L2+ automated driving, more and more drivers pay less attention to the road while driving, which poses a great risk to the road safety, and so does drowsiness. In order to combat this problem, I developed this prototype for monitoring the driver's attention, issuing an audio alarm when necessary and saving the video periodically.
A single web camera
Mediapipe (deep learning tool developed by Google, Apache-2.0 license), predicts a face mesh made up of a list of 468 points. With the known indices of eyebrow, eye, iris, noise and face edge, one can easily connect them and use relevant method to monitor them.
30FPS and higher
- Iris visibility
- Eye blink frequency
- Eye Aspect Ration (EAR, similar to ECG)
- Face orientation
- Yawn
- EAR pulse pattern analysis
- Head movement track and analysis
- voice recorder
- automatic parameter tuning
Saved data
- video clips (periodically saved in case of crash)
- Audio
- Analysis results
- EAR
- MAR
- Face orientation
- Warnings
There are many hyperparameters needing fine tuning to work for every person. One way to make this system self-adpating is to assume that the driver is sober at the beginning of driving, and use these images to tune the relevant parameters.
Due to light condition,or wearing a sunglass, eyes may not be detected properly, which likely fail the drowsiness detection. One of the remedy is to track head micro-movement.
This implementation of driver attention monitor system was deeply inspired by Asadullah Dal for Iris visbility calculation and Nicolai Nielsen for face pose estimation.