Rear-end collision risk estimation for autonomous vehicles under mixed traffic scenarios using real-world perception data

This is the authors' abstract. We don't add key points for this paper.

Using the Lyft and Waymo Autonomous Vehicle (AV) datasets, this study investigated AV’s relative rear-end collision risk for different car following (CF) scenarios. Traditional Extreme Value Theory (EVT) approach and Deep Learning (DL) method were combined to perform traffic conflict-based surrogate safety analysis, which can also be used for real-time crash risk estimation. Relative crash risks between different CF pairs show that HDV following AV scenarios had a two to three times higher risk of rear-end crash compared to the HDV following the HDV scenario. Surrogate safety analysis showed that AVs were more cautious and maintained higher time gaps and headways than HDVs. The AV→HDV scenarios indicate that AVs as a follower vehicle reduced the rear-end crash risk. These findings could be used to develop a warning system that can be installed in AVs to warn following HDVs when the predicted crash risk goes below a safe threshold.