Vijay Govindarajan and Ruzena Bajcsy

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2017-226

December 15, 2017

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2017/EECS-2017-226.pdf

In order to design safe and effective human-in-the-loop systems, developing robust and useful models of human behavior is absolutely vital. However, this problem is highly difficult to address, given that humans often act unpredictably. We look at two approaches to model the human agent: one motivated by safety and capturing the likely behaviors across a wide range of human agents, and another motivated by personalization and adapting control to the specific human user.

In the first approach, we investigate the problem of determining prediction sets for human-driven vehicles using Hamilton-Jacobi reachability analysis and empirical observations from driving datasets. Given evaluation metrics of accuracy, precision, and risk, we optimize disturbance bounds to construct forward reachable sets with high precision that satisfy accuracy and risk constraints. To demonstrate the approach, we apply our framework to a lane changing scenario to provide set predictions that provide safety guarantees without being over-conservative. We show an example of this method that allows us to construct a reachable set with over 85% accuracy and under 25% risk.

In the second modeling approach, we seek to model the human as a collaborator and use the state of the driver to develop an adaptive assistance system. We focus on the problem of measuring the driver state under varying levels of cognitive workload using affective (i.e. emotion) sensing, including facial analysis and electroencephalography (EEG). This information is then used to help predict the brake reaction time of the driver, a key input in designing forward collision warning systems. We use online learning methods, as a way for the autonomous system to gradually learn from examples and improve predictions over time. We then demonstrate the results in a pilot study, which shows that while detecting the cognitive task is challenging, affective sensing can be used directly to reduce prediction error or speed up learning of brake reaction time for some individuals. We then end with some improvements that can be made to further strengthen the quality of the affective sensing-based prediction models in future work.

Advisors: S. Shankar Sastry and Ruzena Bajcsy

---

BibTeX citation:

@mastersthesis{Govindarajan:EECS-2017-226,
    Author= {Govindarajan, Vijay and Bajcsy, Ruzena},
    Title= {Human Modeling for Autonomous Vehicles: Reachability Analysis, Online Learning, and Driver Monitoring for Behavior Prediction},
    School= {EECS Department, University of California, Berkeley},
    Year= {2017},
    Month= {Dec},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2017/EECS-2017-226.html},
    Number= {UCB/EECS-2017-226},
    Abstract= {In order to design safe and effective human-in-the-loop systems, developing robust and useful models of human behavior is absolutely vital.  However, this problem is highly difficult to address, given that humans often act unpredictably.  We look at two approaches to model the human agent: one motivated by safety and capturing the likely behaviors across a wide range of human agents, and another motivated by personalization and adapting control to the specific human user.

In the first approach, we investigate the problem of determining prediction sets for human-driven vehicles using Hamilton-Jacobi reachability analysis and empirical observations from driving datasets.  Given evaluation metrics of accuracy, precision, and risk, we optimize disturbance bounds to construct forward reachable sets with high precision that satisfy accuracy and risk constraints.  To demonstrate the approach, we apply our framework to a lane changing scenario to provide set predictions that provide safety guarantees without being over-conservative.
We show an example of this method that allows us to construct a reachable set with over 85% accuracy and under 25% risk.

In the second modeling approach, we seek to model the human as a collaborator and use the state of the driver to develop an adaptive assistance system.  We focus on the problem of measuring the driver state under varying levels of cognitive workload using affective (i.e. emotion) sensing, including facial analysis and electroencephalography (EEG).  This information is then used to help predict the brake reaction time of the driver, a key input in designing forward collision warning systems.  We use online learning methods, as a way for the autonomous system to gradually learn from examples and improve predictions over time.  We then demonstrate the results in a pilot study, which shows that while detecting the cognitive task is challenging, affective sensing can be used directly to reduce prediction error or speed up learning of brake reaction time for some individuals.  We then end with some improvements that can be made to further strengthen the quality of the affective sensing-based prediction models in future work.},
}

EndNote citation:

%0 Thesis
%A Govindarajan, Vijay 
%A Bajcsy, Ruzena 
%T Human Modeling for Autonomous Vehicles: Reachability Analysis, Online Learning, and Driver Monitoring for Behavior Prediction
%I EECS Department, University of California, Berkeley
%D 2017
%8 December 15
%@ UCB/EECS-2017-226
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2017/EECS-2017-226.html
%F Govindarajan:EECS-2017-226