Abstract-This paper considers the traffic flow estimation problem for the purposes of on-line traffic prediction, mode detection and ramp-metering control. The solution to the estimation problem is given within the Bayesian recursive framework. A particle filter (PF) is developed based on a freeway traffic model with aggregated states and an observation model with aggregated variables. The freeway is considered as a network of components, each component representing a different section of the traffic network. The freeway traffic is modelled as a stochastic hybrid system, i. [12].In the present paper we develop a particle filter for freeway traffic flow estimation. Particle filters (PFs) are appropriate for traffic state estimation because they can cope with large, and highly nonlinear models as well as non-Gaussian signals. The structure of the particle filter corresponds closely to compositional modelling, and it allows a parallelization for different sections of the road, thus allowing a reduction in the computational time.In [21] a solution to highway traffic estimation is proposed using a sequential Monte Carlo algorithm, based on first-order traffic models (only dynamics of the traffic density is modelled), distinguishing between the free-flow mode and the congestion mode. The traffic mode is estimated via a Monte Carlo technique, the so called mixture Kalman filter [5] which is applied for recursively estimating the traffic density. In contrast to [21], the traffic in the present paper is described by a second-order model, and we develop a filter that estimates both traffic density and speed.The particle filter presented in Section IV uses aggregated traffic and observation models. The state model is a recently developed compositional model [4] for freeway traffic. Sending and receiving functions describe respectively the downstream and upstream propagation of perturbations. The freeway is considered as a network of components (Fig. 1), each component representing a different section of the traffic network. Several sections form a link. Sensors are available only at some boundaries between sections. Measurements are averaged over regular or irregular time intervals before being transmitted to the estimation agent. Within the estimation agent an update of the conditional distribution of the traffic density and speed estimate is performed using Monte Carlo simulation (for all sections concurrently and possibly via parallel processing). Whenever a new measurement is received a Bayesian update of this conditional distribution is evaluated via the PF. The PF can easily be extended to models where for the very important sections a detailed model is used, while for other sections a more aggregate, coarse model is employed.
43rd IEEE Conference on Decision and Control