Real-Time Estimation of Drivers' Behaviour

The increasing market penetration of in-vehicle sensor technology, e.g., cameras and lidars, is likely to provide more and more insights on drivers' behaviours. Real time knowledge of drivers' behaviour can be useful to a wide range of applications, such as driving assistance in traffic situations identified as critical, detection of tiredness levels or identification of drivers for insurance purposes. In situations where drivers only interact with other vehicles, for instance on highways, the drivers' behaviours can be modelled through microscopic car-following models. Those models are known to be able to capture the physical features of traffic flow, while accurately describing specific drivers' behaviour via behavioural parameters. A lot of research to date has focused on fitting the parameters of these models to real trajectory datasets using off-line techniques, with the objective to reproduce drivers variability in simulation. However very little has been published on online parameter estimation of car-following models. This paper investigates how in-vehicle sensors could help estimate car-following behavioural parameters. The focus is on the Intelligent Driver Model. A discussion of the sensitivity of the estimation is provided in light of previous work, and it is argued that, depending on the traffic state, not all the parameters need to be calibrated. An extended Kalman filter with physical inequality constraints on the behavioural parameters is then formulated for the class of time continuous car-following models. The filter exhibits systematic convergence for real trajectories of leaders and synthetic-simulated-trajectories of followers, for randomly picked parameters within the physical bounds, and for fixed and dynamic observation time steps ranging from 0.1 s up to 1 s. This demonstrates that online parameter estimation of behavioural parameters has great potential.