| dc.description.abstract | Driver behaviour models are widely used in the traffic engineering literature and practice. They are used for understanding drivers’ manoeuvring decisions in traffic streams. They also form the building blocks of microscopic traffic simulation tools, which are employed for traffic flow analysis and capacity estimation necessary for the design and operation of traffic facilities and evaluation of operational strategies. Most driver behaviour models in the literature assume homogeneous and orderly traffic conditions, characterised by homogeneity (i.e., only passenger cars comprise the traffic streams) and orderliness (i.e., vehicles move only in the longitudinal direction, except when changing lanes). Models developed with such assumptions cannot be applied to analyse heterogeneous, disorderly (HD) traffic conditions. This is because HD traffic streams, unlike homogeneous traffic streams, comprise a wide variety of vehicle classes with considerably different physical and operational characteristics. Moreover, driving in HD traffic streams is characterised by weaker lane discipline due to a greater extent of lateral movements than that in homogeneous traffic streams.
This dissertation aims to formulate and apply driver behaviour models for HD traffic streams on uninterrupted traffic facilities while considering the following aspects – (1) the multi-vehicle anticipation (MVA) behaviour, where drivers’ manoeuvring decisions are influenced by multiple vehicles around them, as opposed to a single lead vehicle ahead, (2) the treatment of driver behaviour as a combination of different manoeuvring decisions, such as the decision of whether to accelerate, decelerate, or remain in same speed (represented by a discrete variable) and the decision of the extent of acceleration or deceleration (represented by continuous variables) – as opposed to a single, continuous variable representing all these facets of driver behaviour, (3) the incorporation of stochasticity due to the errors drivers make in perceiving the traffic environment, and (4) the consideration of drivers’ intentions (which are typically latent to the analyst) and two-dimensional movements of vehicles simultaneously while also incorporating MVA behaviour. Specifically, the following driver behaviour models are formulated and applied to understand driver behaviour in empirical trajectory datasets from Chennai (HD traffic) and California (homogeneous traffic):
1. The first model presented in this dissertation is an MVA-based discrete-continuous choice modelling framework to model vehicles’ longitudinal movements in HD traffic streams. In this model, driver behaviour at a given time instance is treated as a combination of (a) the driver’s choice of whether to accelerate, decelerate, or maintain a constant speed – represented by a discrete variable – and (b) the extent of acceleration or deceleration – represented by continuous variables. The discrete and continuous variables representing driver behaviour are modelled using a simultaneous econometric framework. The proposed model is used to examine driver behaviour in the HD traffic dataset from Chennai. The empirical analysis reveals the significance of the MVA effect on driver behaviour. Specifically, drivers consider the relative speeds and space gaps with respect to multiple vehicles within an influence zone around their vehicle. In addition, it is found that the influence of the traffic environment on drivers’ discrete choices (whether to accelerate, decelerate, or maintain a constant speed) is not the same as that on their choices of how much to accelerate or decelerate.
2. The second model is an extension of the above model to recognise the panel data nature of vehicle trajectory datasets typically used for estimating the parameters of driver behaviour models. This model recognises the role of vehicle- and driver-specific unobserved factors (latent to the analyst), such as aggressiveness that influence driving behaviour, and such influence persists across all observations of a vehicle. Doing so helps in reducing the confounding effects of unobserved factors when the proposed model is applied to different datasets to compare driving behaviour in different traffic streams. The panel data model is used to understand and compare longitudinal driving behaviour between the HD traffic dataset of Chennai and the homogeneous traffic dataset of California. The empirical analysis reveals the presence of MVA effect on driving behaviour in the homogeneous traffic setting, too. However, drivers in the HD traffic stream are influenced by more vehicles in their vicinity than those in the homogeneous traffic stream.
3. In the third model formulation, a mixed multinomial logit-based framework is developed to recognise stochasticity in driver behaviour models due to drivers’ errors in perceiving the traffic environment. For this model, an econometric analysis is undertaken to evaluate two different ways of specifying errors in variables in discrete choice models – additive errors and multiplicative errors. It is shown that the multiplicative specification of errors has a better behavioural basis and allows better identification of parameters representing variability due to drivers’ perception errors. An application of this model to the HD traffic dataset reveals different levels of variability due to errors in the perception of different traffic environment variables. It is found that drivers may pay greater attention to (which results in lower variability in) perceiving space gaps and relative speeds with respect to vehicles directly ahead of them than those not directly ahead.
4. The fourth and final model formulation is a two-dimensional, MVA, and multi-stimuli-based latent class framework to analyse motorcyclists’ two-dimensional movements in HD traffic streams. This formulation conjectures that drivers manage their cognitive load by dividing their driving decisions into two steps – (a) higher-level, strategic intentions (of whether to accelerate, decelerate, or maintain a constant speed and whether to steer to the left of, right of, or keep straight along the longitudinal direction), which are not fully observable from vehicle trajectories (hence latent to the analyst), and (b) lower-level, tactical decisions that can be observed in vehicle trajectories, such as the specific angle of movement and the specific extent of acceleration or deceleration executed. When applied to the HD traffic dataset of Chennai, the proposed model suggests that drivers’ higher-level intentions are more strongly influenced by the microscopic traffic environment variables than their lower-level decisions, perhaps because drivers invest a greater extent of cognitive resources for making higher-level intentions than that for lower-level decisions.
Finally, a traffic simulator is developed to simulate traffic streams using the models developed in this dissertation. The simulation experiments using this simulator demonstrate that all the microscopic driver behaviour models developed in this dissertation reflect the typically observed macroscopic properties of vehicular traffic steams. | en_US |
