Understanding and Predicting Human Driving Behaviors via Machine Learning Models (Multiple Year Project)

Principal Investigator: Yi-Ching Lee, PhD, Children's Hospital of Philadelphia

Below is an executive summary of this line of research. Please note that each summary describes results and interpretation that may not be final. Final interpretation of results will be in the peer-reviewed literature.

Among teen drivers, speed management is a significant factor that contributes to crashes. To better understand how technology can help reduce crash risk, this study used driving simulators to analyze and learn a human behavior – speed management – using a technique called machine learning. Such techniques are used to create predictive driving models that may be incorporated into smart-car or in-vehicle monitoring systems. The long-term goal of this research was to fully understand young drivers’ risky behaviors in order to advance safety awareness and management of driving risks.



The long-term objective of this study is to utilize state-of-the-art experimental and analytical techniques to create accurate models of teenage drivers’ behavior in order to inform the development and testing of new technology and training methodologies to improve teen driving and reduce risk. The broad objective is to examine the potential for the personalized feedback to improve driving behavior and reduce dangerous behavior, specifically in the context of speed management of teen drivers.


This indirect model can be used to predict accurate long-term driving behaviors. The figure shows  four predictions, in four different situations. For each prediction, two trajectories are shown: The blue dots represent the trajectory that the driver actually took, while the red dots show the predicted trajectory by our model.

Drawing on phase one results, the objectives of phase two were to determine: (1) which parameters of modeling techniques were most suitable for producing predictions of unsafe and dangerous speed management behaviors, and (2) the effect of feedback on behavioral changes in speed management in subsequent simulator drives.

Researchers upgraded the machine learning modeling techniques, including testing and comparing machine learning approaches for accuracy and scalability; improving model accuracy; reducing prediction errors; and redesigning the feedback interface. In parallel, researchers developed and tested driving scenarios in the driving simulator to best align with the collection of training and testing scenarios for the machine learning modeling work. The driving contexts were programmed in a specific way and presented in a specific order so that the machine learning models could capture participants’ driving styles in various driving contexts. This ensured that the models had sufficient training data with which to output predictions on how each driver would perform in new driving scenarios.

The third component of the project involved collecting data from young drivers. Drivers in the feedback condition received model predictions as feedback half-way through a study visit. Drivers in the control condition did not receive any feedback. This allowed researchers to examine the effect of receiving feedback on subsequent driving behaviors in similar or different driving contexts.

Preliminary results show that the indirect model outperforms all other modeling approaches. This work paves the way for future in-vehicle machine learning systems that can provide drivers with real-time monitoring and feedback to help reduce crash risk.



In phase one of the study, researchers determined the feasibility of using computational approaches to automatically model speed management. The project utilized a subset of an existing dataset from the development of a Simulated Driving Assessment. This Assessment included driving scenarios and situations that reflected the most common crash configurations among young drivers as identified in the results of the National Motor Vehicle Crash Causation Survey (NMVCCS). The dataset contained behavioral and eye-tracking data from four simulator drives, a driver education instructor’s subjective ratings of driving behaviors, images from three video cameras, and self-report measures from the teen participants.

Data from 17 novice teen drivers (licensed for less than 90 days) were analyzed using three modeling techniques. The modeling techniques predicted the behavior of the driver at the same instant, as represented by a given sample, and the future behavior of the driver. Additionally, a certified driver education instructor reviewed video recordings of the participant’s simulator drives and evaluated each participant’s crash likelihood and driving skill levels in comparison to all drivers, as well as teenage drivers.

Results suggest that modeling was effective at predicting vehicle control behaviors up to one second into the future. These predictions were consistent with the skill proficiency and crash likelihood ratings given by a professional driving education instructor.

Co-Principal Investigator

Santiago Ontanon, Drexel University

Project Team Members

Chelsea Ward, MS, Children’s Hospital of Philadelphia (Y3);Flaura Winston, MD, PhD, Children’s Hospital of Philadelphia (Y1,Y2);Catherine McDonald, PhD, RN, The University of Pennsylvania (Y1);Avelino Gonzalez, PhD, University of Central Florida (Y1,Y2);Dana Bonfiglio, BS, Children’s Hospital of Philadelphia (Y1).


Molly Tiedeken, Drexel University (Y1,Y2, Y3);Sam Snodgrass, Drexel University (Y1,Y2,Y3);Leif Malm, Drexel University (Y3);Celia Cheng, Drexel University (Y2).

IAB Mentors

Doug Longhitano, American Honda Motor Co., Inc. (Y2,Y3);Melissa Miles, State Farm Mutual Automobile Insurance Company (Y3);Tim Hsu, FCA US LLC (Y2);Noelle Lavoie, Parallel Consulting (Y1);Steve Roberson, State Farm Mutual Automobile Insurance Company (Y1);Dan Glaser, General Motors Holdings LLC (Y1);Clayne Woodbury, Realtime Technologies Inc. (Y1);Sara Seifert, Minnesota HealthSolutions (Y1).

Publication References