Search

The aim of this paper is to gain insights into the level of heterogeneity in car-following behavior in real traffic. We use a large sample of trajectory observations collected by means of a helicopter to identify differences between the car-following behaviors of: (1) passenger car drivers, (2) passenger car drivers and truck drivers and (3) passenger car drivers following a passenger car and passenger car drivers following a truck. We thereto calibrate eight car-following models making different assumptions about the way in which drivers follow their leader(s) on the same lane. We show that considerable behavioral differences exist between passenger car drivers. Different passenger car drivers do not only consider different stimuli (like speed difference(s) with the leading car(s) and distance headway(s) to leading car(s)) but also the extents to which these stimuli influence their behavior differ. Truck drivers turn furthermore out to adopt in general a more robust car-following behavior than passenger car drivers. Their speeds show, for example, less variation over time. We also find indications that the desired headways of passenger car drivers are lower when following a truck than when following a passenger car.

Parameter values obtained by calibrating a car-following model using trajectory observations can provide important new insights into longitudinal driving behavior that cannot be derived from double loop detector data. However, that holds only if these behavioral parameters can be reliably estimated. This important aspect of calibration has been generally neglected in scientific research. In this contribution, therefore, calibrations are done to analyze several aspects influencing the degree to which behavioral parameters can be reliably estimated. In these calibrations trajectory data created by the authors with known characteristics as well as empirical trajectories are used. It is shown that traffic conditions during observation determine which model parameters can be reliably estimated. That measurement errors do negatively influence the reliability of parameter estimates is also demonstrated. The same negative effect appears to occur when a car-following model not fully able to capture the dynamics of the observed driver is calibrated. Finally, it is found that the extent to which a specific parameter can be reliably estimated is dependent on the degree to which it affects model predictions and thereby the value of the calibration objective. From this contribution it is concluded that to answer a specific research question on longitudinal driving behavior, it is essential to collect trajectories during appropriate traffic conditions. The results furthermore stress that even when trajectories contain enough information to calibrate the behavioral parameters of interest, whether it is reliable enough to draw inferences on car-following behavior needs to be determined for every parameter value separately.

Interest in calibration of car-following models by using real-life microscopic trajectory data is increasing. However, more information is needed on the influence of methodological issues on calibration results as well as on the influence of practical issues related to the use of real-life data. In particular, the influence of measurement errors on parameter estimates has not yet been considered in detail. To gain insight into the influence of measurement errors on calibration results, synthetic data were created to which several types of measurement error are introduced. These data are input to a validated calibration procedure, after which it is studied how well the parameters used for creating the data can be identified from the erroneous data. The sensitivity of the objective function to small changes in the optimal parameters also is assessed. The calibrations are repeated by using different variables in the objective. The three main findings are that (a) measurement errors can yield a considerable bias in the estimation results, (b) parameters minimizing the objective function do not necessarily capture following dynamics best, and (c) measurement errors substantially reduce the sensitivity of the objective function and consequently reduce the reliability of estimation results. The extent to which these problems caused by measurement errors can be avoided by smoothing the data carefully before use is assessed and discussed.

A dedicated trajectory data collection method using a helicopter enabled a range of in-depth empirical studies of car-following behavior. These studies found a high degree of heterogeneity in car-following behavior; that is, drivers' driving styles turned out to be highly different because different modeling approaches were needed to model these behaviors satisfactorily. Therefore the impact of heterogeneity in car following on modeling traffic dynamics is examined to gain insight into the effect of incorporating different types and degrees of heterogeneity in car-following behavior on the dynamics of a simulated traffic flow. The microsimulation approach that was adopted focused on two case studies: the first case study focused on heterogeneity in parameter values by comparing stability results for heterogeneous platoons and homogeneous platoons, which could in fact be seen as a solid preparation for the second case study. The second study was a simulation of a fixed stretch of road on which, from a certain point on, a speed limit was imposed for the drivers. In this case study the link between the empirical results and the simulations was strengthened as several types of heterogeneity (e.g., different model specifications for different drivers and different parameter settings and combinations of them) were explored and compared with the empirically estimated parameters. Both analyses show clear differences between simulations with homogeneous drivers and simulations with heterogeneous drivers.

This paper examines the car-following behavior of individual drivers in real traffic on the basis of vehicle trajectory data extracted from high-resolution digital images collected at a high frequency from a helicopter. These data are used to cross-compare seven car-following models regarding their average performances as well as their specific performances for all individual drivers observed using a simulation approach. The prime objective of this cross-comparison is to study interdriver differences; both optimal parameter settings and model performances are compared between drivers. Average model performances reveal that the simplest models are generally not able to capture the dynamics of car-following behavior correctly, whereas individual estimates show that the performances of more elaborate models differ between drivers. The most important contribution of this paper is that analysis shows that interdriver differences cannot be caught by different parameter settings alone; driving styles of individual drivers appear to be inherently different in that various car-following models are needed to model them satisfactorily.

Using vehicle trajectories for a motorway collected from a helicopter, and a newly developed approach to parameter identification, this paper provides new empirical evidence of multi-anticipative car-following by estimating the driver-specific parameters of the different multi-anticipative car-following models. We investigate the nature of the multi-leader stimuli, providing insight into the number of vehicles ahead to which drivers react and the kind of stimuli drivers respond to. The paper also shows that there is large inter-driver variability in multi-leader driving behavior. The implications of our research findings for microscopic modeling are discussed as well.

Microscopic simulation models have become widely applied tools in traffic engineering. Nevertheless, parameter identification remains a difficult task. This is for one caused by the fact that parameters are generally not directly observable from common traffic data. The second difficulty stems from the fact that real driving behavior is variable in time and space, etc. This paper puts forward a new approach to identify changing parameters of delayed car-following models, i.e. models that include a reaction time. The approach is based on the unscented particle filter approach, which is generalized to enable estimation of reaction times. The estimation of this true delay is achieved without linearization. Besides the methodological contribution, we show empirical evidence for changing driving behavior by applying the approach to real-life microscopic traffic data