This paper proposes a new data-driven stochastic car-following model based on the principles ofpsycho-spacing or action-point modeling. It uses empirical or experimental trajectory data andmimics the main microscopic behavioral characteristics present in the data.In the action-point model, regions are defined in the relative speed - distance headway planein which the follower is likely to perform an action (increase or decrease acceleration) or not.These regions can be established empirically from vehicle trajectory data, yielding a joint cumulativeprobability distribution function of the action points. Furthermore, the conditionaldistribution of the actions (the size of the acceleration or deceleration given the current distanceheadway and relative speed or given the acceleration before the action) can be determined fromthese data as well.To assess the data correctly, a new filtering technique is proposed. The main hypothesisbehind this idea is that the speed profile is a continuous piecewise linear function: accelerationsare piecewise constant changing values at non-equidistant discrete time instants. The durationsof these constant acceleration periods are not fixed, but depend on the state of the followerin relation to its leader. The data analysis indeed illustrates that driving behavior shows nonequidistantconstant acceleration periods.The aforementioned distributions of the action points and the conditional accelerations formthe core of the presented data-driven stochastic model. The paper depicts the mathematicalformalization describing how these distributions can be used to simulate car-following behavior.Based on empirical data collected on a Dutch motorway, we illustrate the workings of theapproach and the simulation results.
展开▼
