Employs a spatio-temporal graph to infer the interaction between adjacent road segments as well as the temporal dynamics of the time series data.

Able to predict traffic speed of road networks different from the network used to train, regardless of the network topology.

Outperforms the image-based approaches using the CapsNet and CNN, requiring the smaller, constant number of parameters to train.

Addresses the problems of 1) information of the dynamic environment dissipating over time; 2) uncertainty of cost in optimization, where the location of the operational area is different from the location of sensor deployment.

Reducing the planning time and distributing the computational burden by online optimization realize real-time strategy and outperform the "optimal" solution.

Cramer Rao Bound, Dubins path, and gradient descent for convex optimization are used.

Addresses the problem of measurement ambiguity in particle filtering, which increases the covariance after measurement update stage.

Key observation in this study is that the posterior distribution is contributed by the prior distribution as well as the measurement. The proposed method skips the ambiguous measurement update and later employs the skipped measurement.

The proposed method outperforms the standard particle filter (PF), auxiliary particle filter (APF), mixture particle filter (MPF), and receding horizon Kalman filter (RHKF) in an application to terrain-referenced navigation.