Vehicular Traffic Modelling and Traffic Flow Estimation with Particle Filtering Methods
Friday 02 May 2008, 1430-1530
Lecture Theatre, Postgraduate Statistics Centre
Traffic flow on motorways is a nonlinear, many-particle phenomenon, with complex interactions between vehicles such as traffic jams, stop-and-go-waves. This talk Dr Mila Mihaylova from The Department of Communication Systems presents a stochastic macroscopic model of freeway traffic suitable for on-line estimation and control purposes.
The freeway is considered as a network of interconnected components. Stochastic equations describing the macroscopic traffic behaviour of each cell, as well as its interaction with neighbouring cells are derived. This model has been applied to estimation of traffic state in motorway networks.
We show one solution of the traffic estimation problem within Bayesian framework. A particle filter (PF) for traffic flow prediction has been developed and its performance compared with an Unscented Kalman filter (UKF). The ability of particle filtering to deal with highly nonlinear models and non-Gaussian signals makes them the suitable approach for the traffic flow estimation.
Additionally, the problem is characterised with sparse and missing data. Measurements (from video cameras or magnetic loops) are received only at boundaries between some segments and averaged within regular or irregular time intervals. This limits the measurement update in the PF and UKF to only these time instants when a new measurement arrives, with possibly many state updates in between consecutive measurement updates. The filters performance is validated and evaluated over synthetic and real traffic data.
In large traffic networks, the computational demands for traffic prediction are high and there is a need of decentralised or parallelised processing techniques. Finally, two parallelised particle filters for traffic flow estimation will be presented and open issues will be discussed.