Real-time travel time prediction using particle filtering with a non-explicit state-transition model

Abstract

The research presented in this paper develops a particle filter approach for the real-time short to medium-term travel time prediction using real-time and historical data. Given the challenges in defining the particle filter time update process, the proposed algorithm selects particles from a historical database and propagates particles using historical data sequences as opposed to using a state-transition model. A partial resampling strategy is then developed to address the degeneracy problem by replacing invalid or low weighted particles with historical data that provide similar data sequences to real-time traffic measurements. As a result, each particle generates a predicted travel time with a corresponding weight that represents the level of confidence in the prediction. Consequently, the prediction can produce a distribution of travel times by aggregating all weighted particles. A 95-mile freeway stretch from Richmond to Virginia Beach along I-64 and I-264 is used to test the proposed algorithm. Both the absolute and relative prediction errors using the leave-one-out cross validation concept demonstrate that the proposed method produces the least deviation from ground truth travel times, compared to instantaneous travel times, two Kalman filter algorithms and a K nearest neighbor (k-NN) method. Moreover, the maximum prediction error for the proposed method is the least of all the algorithms and maintains a stable performance for all test days. The confidence boundaries of the predicted travel times demonstrate that the proposed approach provides good accuracy in predicting travel time reliability. Lastly, the fast computation time and online processing ensure the method can be used in real-time applications.