Learning traffic signal phase and timing information from low-sampling rate taxi GPS trajectories

Traffic signal phase and timing (TSPaT) information is valuable for various applications, such as velocity advisory systems, navigation systems, collision warning systems, and so forth. However, the acquisition of the TSPaT information in the city-scale is very challenging. In this paper, we propose a framework to learn the TSPaT information from low-sampling rate taxi GPS trajectories. Specifically, our framework could learn: the phasing scheme, i.e., the number of phases and the assignment of traffic movements to phases; timing plans, including the cycle length and green lengths of phases within a cycle, for each given fixed-time signalized intersection. In our framework, the cycle length is the first important parameters to be learned. We formalize the cycle length estimation problem as a general approximate greatest common divisor (AGCD) problem, and propose the most frequent AGCD (MFAGCD) algorithm to solve the problem. The MFAGCD algorithm is robust to noises and outliers, and could estimate the cycle length with a high accuracy using a small number of green-start times extracted from taxi GPS trajectories. Based the correlation between phases, we propose an all-direction joint determination method to jointly estimate green lengths using green-start times and cross-over times from all phases. The effectiveness of our framework is experimentally evaluated on three selected fixed-time signalized intersections in Shanghai, China.