Learning Traffic Light Parameters with Floating Car Data

The knowledge of traffic light parameters, such as cycle plan or future signal phase and timing information (SPaT) of traffic lights is the base for a vast number of use scenarios. A few examples are traffic signal adaptive routing, green light optimal speed control, red light duration advisory or efficient start-stop control. The basis for all these functionalities is the knowledge on the correct traffic light cycle time, i.e. the periodicity of the traffic light's signaling sequence. With a correct cycle time given, green start and end times can be derived from periodically reoccurring movement patterns. In this paper, we propose a method to reconstruct a traffic light's cycle plan through the interpretation of the recorded information on a vehicle's movement pattern (trajectory) in the intersection area. The recorded trajectories are temporarily sparse and and the cycle plan changes frequently. Therefore, we propose a model that focuses on the performance on very limited available trajectory data and yet is robust with regard to estimation errors. We show that our approach is able to detect the correct cycle time with already 30 trajectories at an accuracy of 99%.