Presenter: Dr. Feng Guo
The high-frequency, high-resolution telematics driving data provide valuable information on both long-term driver behavior as well as instantaneous driving conditions. This project uses telematics driving data to modeling driving risk with three primary objectives: 1) characterize the high-frequency kinematic signatures for safety critical events; 2) modeling driver level crash risk prediction based on kinematics features; and 3) instantaneous crash risk assessment.
The research team propose a state-of-the-art approach for characterizing the high-frequency kinematic signatures. We developed several features representing driver behavior and underlying driving risk. These features were applied to several large-scale ride-hailing data and naturalistic driving study data to predict driving risk.
The high-frequency kinematic data coupled with the rarity of crashes demand novel modeling approaches. We developed deep learning-based models and variational inference-based rare event modeling to predict crashes from normal driving as well as predict high risk drivers. A convolutional neural network and long short-term memory network is developed to predict crashes, near-crashes, and normal stopping behaviors. We developed a novel Variational Information for Extremal (VIE) framework for modeling rare events through deep learning models.
This project addressed key methodological challenges in predicting driving risk using high frequency telematics data. The findings of the project will benefit driving data processing at scale, driver safety management program, and real-time risk prediction.
For more information about the Project or Safe-D visit the link below: