Inferring Traffic Signal Phases From Turning Movement Counters Using Hidden Markov Models

This work poses the problem of estimating traffic signal phases from a sequence of maneuvers. We model the problem as an inference problem on a discrete-time hidden Markov model (HMM) in which maneuvers are observations and signal phases are hidden states. The model is calibrated from maneuver observations using either the classical Baum-Welch algorithm or a Bayesian learning algorithm. The trained model is then used to infer the traffic signal phases on the data set via the Viterbi algorithm. When training with the Bayesian learning algorithm, we set the prior distribution as a Dirichlet distribution. We identify the best parameters of the prior distribution for both fixed-time and sensor-actuated signals using numerical simulations and employ them in the field experiments. It is shown that when the model is trained by the Bayesian learning method with appropriate prior parameters from the Dirichlet distribution, the inferred phases are more accurate in both numerical and field experiments. Because the best set of prior parameters for a fixed-time intersection is different from those for sensor-actuated signals, a classification strategy to distinguish between these two types of signals is proposed. The supporting source code and data are available for download at https://github.com/reisiga2/TrafficSignalPhaseEstimation.