A Particle Filter Real-time Traffic State Estimation

Traffic state estimation and prediction is the foundation and the key technology for urban traffic flow guidance system, and it is also an important study domain of intelligent transportation systems. Traffic state estimation can prevent the city road networks from serious traffic jam, reduce vehicles' latency time and realize the efficiency distribute of traffic flow in road networks. A large number of research works for traffic state estimation were almost exclusively based on the methodology of Kalman filtering and its extensions for nonlinear systems. These algorithms and applications were reported for both short freeway sections and simple road networks, while the traffic scenario supposed were more and more complex. Research works proved that extended Kalman filter can estimate and predict the traffic state efficiently. Even now, as the traffic state estimation is a nonlinear and non-Gaussian issue, extended Kalman filtering method has obvious disadvantages.The disadvantages mainly come from the linearization of the model in extended Kalman filtering method and the Gaussian noise supposition. A powerful and scalable approach has recently been developed, known under the name as particle filters. It uses the Monte Carlo method to sample the posterior probability density function. To each particle a weight is assigned at current time. The weight and the value of all particles together approximate the conditional density function of the state. After the arrival of a new observation vector, the algorithm updates the weights. At last, the estimation result can be calculated by the weights and values of all particles. The main task of this paper is to develop a general approach for the real-time estimation of the complete traffic state based on particle filter and macroscopic traffic flow model.In this approach, the road stretch is divided into several segments one by one with the same length while the electronic sensors are usually placed between some segments and the measurements are less than states estimated. This paper uses the compact state-space method and treats the model parameters as the traffic states. In this paper, we also research the influences on the PF algorithm from the different estimation method. Simulation results and real data experiments prove that particle filtering can predict and track the state efficiently. It overcomes the disadvantages of EKF and UKF method. The complexity of this algorithm and the additional operation workload will not affect the applications for on-line road traffic management.