A note from Flaura Koplin Winston, MD, PhD, CIRP@CHOP scientific director: Today we are pleased to welcome a guest blog post from Santiago Ontañón, PhD, a research scientist and assistant professor at the College of Computing and Informatics at Drexel University. Dr. Ontañón led a study, “Learning to Predict Driver Behavior from Observation,” and recently presented on the findings at the Association for the Advancement of Artificial Intelligence (AAAI) Spring Symposia on Learning from Observation of Humans at Stanford University.
I’ve been working with a group of researchers from CHOP and George Mason University to develop machine learning techniques to improve driving behaviors to help reduce crash risk for young adults. These techniques provide algorithms to help predict and manage certain teen driving behaviors, but previous ones fared poorly in predicting these behaviors. Although accurate for predicting the behaviors of a driver for less than one second into the future, they quickly degrade and are basically indistinguishable from random predictions.
We developed a novel machine learning technique called indirect prediction, which showed the ability to produce much more accurate predictions over longer spans of time:
- From a machine learning perspective, indirect prediction circumvents the main challenge that standard supervised machine learning techniques face when asked to predict sequential decision-making tasks, such as driving a vehicle.
- From a practical point of view, we showed that this approach accurately predicted the behavior that specific drivers would exhibit in situations, such as approaching a STOP sign (being able to predict, for example, that a certain driver would make a full stop, or that another driver would not fully stop, but drive through the STOP sign at about 5mph).
For our latest study we collected data from 32 subjects, with each driving five times in the CHOP driving simulator: one practice drive and four experimental drives that represented various traffic situations and allowed for dynamic interactions with each participant’s vehicle. We trained a machine learning system by showing it data from the first experimental drive. Then we evaluated it by comparing the predictions made by the machine learning system on how drivers behave in each experimental drive against the actual behaviors.
An example demonstration of the machine learning tool we developed for this work can be seen in this video.
Moreover, although predictions are reasonably accurate, an open question is whether it’s possible to achieve predictions of similar accuracy with on-road data collected from real cars, in real situations. Developing systems that can predict human driving behavior sufficiently in advance would have numerous applications in future semi-autonomous driving vehicles, giving them the ability to anticipate dangerous situations and provide appropriate assistance in a timely manner.
**Like what you’ve read? Subscribe to Research in Action to have the latest in child injury prevention delivered to your inbox.**