Estimating Traffic Signal States by Exploiting Sparse Low-Frequency Floating Car Data

In this thesis, a novel methodology for the traffic signal state prediction based on real-world Floating Car Data (FCD) has been developed. In order to understand the characteristics of a real-world, commercially available FCD source, a one year dataset obtained from the ADAC (General German Automobile Club) is analyzed at two major intersections in Munich (Germany). Motivated by sparse and likewise low-frequently sampled trajectory volumes, this work developed a comprehensive methodology that covers all aspects of the data processing to infer hourly estimates of the signal timing parameters, such as the cycle length and the green intervals typically used on workdays or weekends. These parameters allow the prediction of future traffic signal states with respect to different constraints. The thesis describes a novel hypothesis-based cycle length estimation procedure that estimates the typical cycle length for an hour by analyzing vehicle stop bar crossing events. These events are projected into the time scale of potential cycle length candidates by the help of a modulo operation. Only the correct cycle length causes a concentration of crossing events within the cycle time, which is exploited by the hypothesis test. The concept has been analyzed for a fixed-time controlled intersection by means of a Monte-Carlo-Simulation. The thesis describes moreover the development of a green interval estimation methodology that is capable to deduce the green interval by exploiting sparse crossing events as well as stopped vehicles. The concept relies on an iterative classification procedure that infers the green interval by analyzing the number of counted crossing events per cycle second. A rule-based and iterative self-verification process checks the consistency of the green interval, which makes the method robust when working with small samples sizes. In addition, the thesis closed the gap of a missing green interval verification that checks overlapping green intervals of conflicting movements, caused by traffic actuation. As the developed concept requires ideally green interval estimates on all conflicting movements, the thesis developed also a novel cluster-based approach that tries to identify hours with very similar signal timing parameters. This time clustering allows the signal timing estimation even on movements with small trajectory samples sizes. Finally, the complete concept was calibrated and validated at four real-world intersections in Munich (Germany).