Driver behaviour prediction

Team: 2 people;
Start: September 2014;
Duration: 36 months;
Technologies: C++, ROS, Python, MATLAB, IBEO LIDAR, Mobileye EPM3/4, ARS350 Radar, Bayes Net Toolbox, Bayesian Networks with Gaussian Mixture Models, RT3000 GPS;

Curiosity: The free and open-source robot operating system (also known as ROS) actually works quite nicely for real-world autonomous vehicles, too!

One of our doctoral research projects is the intersection of prediction and strategy. Its aim is to increase safety and enable more convenient driving in the future with advanced driver assistance systems (ADAS) or autonomous cars. To achieve this, the (semi-) autonomous systems need to know in advance what other drivers on the road might do in the near future (within approximately the next 10 seconds). But every prediction of the future inherently includes some uncertainty as to the outcome. So the next step is for the uncertain predictions to be considered with regards to the level of uncertainty and potential reactions to it so that the best reaction can be chosen by the (semi-) autonomous vehicle. Two potential use cases are considered in more detail in this research. The first is a highway scenario in which a longitudinal controller enhanced with the predictions will react a lot earlier to potential cut-in vehicles than for example current adaptive cruise control systems (ACC). This enables greater safety and comfort due to earlier braking manoeuvres, which therefore need to be less abrupt. Another use case is zip-merge situations in slow moving traffic. The prediction estimates which gaps a given vehicle might use to change lane and the autonomous system uses that prediction to let other drivers merge cooperatively.