Driving a motorized vehicle on US roads can be dangerous to one's health. The US National Safety Council estimates that there were 38,800 traffic fatalities in the US in 2019. While these numbers have not grabbed the attention of the press or of our political leadership in the midst of the current Covid-19 pandemic, traffic fatalities remain a serious problem that takes a terrible annual toll in lives and cries out for innovative ways to reduce these losses. The work we describe here seeks to bring Artificial Intelligence to bear to reduce this annual loss of life, especially in young drivers.

Our concept recognizes the opportunities for identifying health issues in real time in a person while he/she drives a motor vehicle. Rather than intrusively instrumenting the driver to gather physiological data, our system analyzes his/her actual driving behavior on the road to identify his/her medical condition. We particularly target the identification of those drivers who may suffer from uncontrolled Attention Deficit/Hyperactivity Disorder (ADHD), a prevalent chronic medical condition that potentially increases the probability of a motor vehicle accident when uncontrolled.  Identification of this condition in real time could lead to notification of the proper authorities (i.e., police, EMTs, or parents if minors) prior to an accident and remove a serious risk from the roads.

Our work is the result of a five-year research program sponsored by the US National Science Foundation at four institutional partners: Drexel University, Children's Hospital of Philadelphia, George Mason University, and the University of Central Florida. Here we only describe the approach taken at the University of Central Florida, where we employed machine learning from observation (LfO) to build models of drivers. LfO is a type of machine learning that can be used to build a model of a person's behavior (actions) strictly through unobtrusive observation of said behavior. We call these models agents, and they are capable of prescribing a driving action to be taken in a just-in-time fashion, in reaction to the situation (i.e., the context) that a suite of  sensors perceive in real time. The actions prescribed by the agent are then compared against a driver's actual actions in real time to detect any serious discrepancies that would indicate a problematic medical condition. Our work uses LfO systems to generate two models of a driver's behavior: one under normal (i.e., medicated) conditions, and another one under abnormal (i.e., un-medicated) conditions.   

The agents were created from unobtrusive observation of human test subjects as they drove a simulated vehicle through a simulated road network. The use of a car simulator eliminated safety risks to the human test subjects, to our research staff, and to any other drivers and pedestrians who might have been present if we had gathered the data while driving an actual automobile on actual roads. This would have been particularly true when a test subject afflicted with ADHD was driving a car while un-medicated. 

When realized to its fullest potential, we envision this approach to work as follows: an at-risk driver who suffers from ADHD or other such medical condition is brought to a test center when in a normal state of health and asked to drive a simulated automobile through several different traffic situations (contexts). Her/his performance data are collected and an agent is created that reflects her/his normal driving style.  The model is thereafter placed on-board his/her car and it predicts what the driver would normally do under the traffic conditions being perceived by the agent through a suite of sensors. The conditions perceived would be the same as those perceived by the driver. Any discrepancies between the two that are consistent and/or severe could be flagged as indicative of an abnormal condition, and corrective action could be initiated by the computer by contacting the appropriate authorities before an accident occurs. 

We used two existing LfO systems to build our agents – Genetic Context Learning (GenCL) and Force-feedback Approach to Learning from Coaching and Observation with Natural and Experiential Training (Falconet). These systems had been previously developed in our laboratory as part of our prior research. Our results show that agents built with Falconet could be used to correctly identify a difference between driving normally (medicated) and driving abnormally (un-medicated) 71% to 80% of the time.  While these results fall short of the minimally acceptable accuracy for ultimate commercial application of our concept, they do provide great promise that such an approach could be made feasible with further research and development. More importantly, the contextually-centered nature of our work allowed us to discover situations (i.e., contexts) where the differences in behavior between normal and abnormal driving are more pronounced and thus easier to identify.  This could be further exploited in future research.

Last Modified: 10/05/2020
Modified by: Avelino J Gonzalez