Abstract
Accurate traffic state estimation is essential for active traffic management and intelligent transportation systems. However, fixed detectors suffer from limited spatial coverage, while floating car data (FCD) are sparse and randomly distributed in time and space. In addition, spatial heterogeneity in traffic flow further challenges traditional data assimilation methods. To solve these problems, a robust traffic state estimation framework is proposed that includes the heterogeneous data fusion and parameter adaptation. Differential Evolution (DE) is applied in offline calibration of key parameters of METANET for the first time. Next, in the framework of Extended Kalman Filtering (EKF), segment-level free-flow speed is modeled as a state and therefore included to account for spatial heterogeneity. Moreover, a dynamically variable observation Jacobian matrix is designed to fuse high-frequency fixed detector data with low-frequency sparse FCD. On the basis of both controlled simulations and the NGSIM I-80 dataset, they demonstrate that the proposed approach is capable of providing accurate traffic state reconstruction in the blind spots of the sensors and increasing estimation accuracy and robustness for meeting challenging real-world scenarios.