Investigating the long- and short-term driving characteristics and incorporating them into car-following models

Highlights

• The LSTD model is proposed to describe long-and short-term driving characteristics.

• The long-term characteristics are classified into three types via cluster analysis.

• The short-term characteristics are identified and divided into two types.

• The six types of LSTD characteristics are incorporated into IDM and Gipps’ model.

• Significant improvement is demonstrated by incorporating LSTD into CF models.

Abstract

This study provides a new method for better incorporating human factors in modeling car-following behavior. As the primary decision maker and vehicle operator, human driver is the vital component of the driving process. During the driving process, an external stimulus may trigger short-term psychological changes, and these changes are considered as the endogenous cause of many abnormal driving behaviors, which often lead to unsafe traffic disturbances and even crashes. In this paper, we investigate the intrinsic long-term driving characteristics and its short-term changes after driver experiences an external stimulus. A long- and short-term driving (LSTD) model is proposed to incorporate such changes into car-following driving behavior modelling. The long-term driving characteristics are extracted through a cluster analysis, and the changes after an external stimulus are identified and measured as the indicator of the short-term driving characteristics. NGSIM data are used to demonstrate the existence of LSTD characteristics, and the soundness of the LSTD model. Two classical car-following models (i.e. the intelligent driver model, Gipps’ model) are integrated with the LSTD model, and the integrated models show a promising performance as the errors decrease by 36.7% and 35.7%, respectively.

Introduction

The car-following (CF) modeling aims to describe the longitudinal interactions of vehicles on the road, and is the core component of microscopic traffic flow modeling and simulation. Over the past decades, numerous CF models have been developed. Most of these models attempted to understand the traffic characteristics by approximating vehicles’ physical dynamics in the traffic flow, and can be categorized as engineering CF models (Saifuzzaman and Zheng, 2014). In these models, all the movements are the continuous application of a single control law, whereas no emotional fluctuation or psychological changes of drivers are reflected. Just like a ‘driving machine’, drivers always drive according to the unified initial parameters.

The observable driving behavior parameters in the real world, such as speed, acceleration, trajectory, are the joint impact results of drivers’ physiological and psychological characteristics, vehicle dynamic characteristics and external traffic environment (van Lint and Calvert, 2018). However, as the primary decision maker and vehicle operator in the driving process, the human drivers were rarely studied in depth from the psychological perspective and incorporated in the CF models (Boer, 1999, Saifuzzaman and Zheng, 2014).

Recently, some studies highlighted the significance of considering human factors (HF) in CF modeling to characterize realistic CF behavior in complex driving situations. Saifuzzaman and Zheng (2014) reviewed recent developments and research needs of CF models from the perspectives of both engineering and HF. They pointed out that most studies attempt to incorporate these HF into CF models by calibrating one or two indirect parameters such as modified reaction time and estimation errors (Treiber et al., 2006a, Treiber et al., 2006b, Lindorfer et al., 2017), while other studies focus on taking into account individual phenomenon and driving errors, such as distraction (Bevrani et al., 2012, Przybyla et al., 2012, Saifuzzaman et al., 2015, van Lint and Calvert, 2018) and fatigue driving (Yang and Peng, 2010), and make efforts to enable CF models to describe these specific behaviors. There also exist some studies (Taylor et al., 2015, Kazemi and Abdollahzade, 2016, Zheng et al., 2019) investigated the driver heterogeneity, some studies (Treiber and Helbing, 2003, Treiber et al., 2006a, Treiber et al., 2006b) introduced dynamic parameters (e.g., desired speed, desired time headway, etc.) changes, and some others studies (Jiang et al., 2014, Huang et al., 2018a, Huang et al., 2018b) quantified that might change the desired time gap. However, the psychological characterization (such as short-term changes in driving behavior caused by psychological fluctuations) was largely neglected. The temporary changes, such as road-rage driving, aggressive driving, are considered as the intrinsic cause of many abnormal movements, traffic turbulences and even accidents (AAA, 2009, AAA, 2016). Some puzzling traffic flow phenomena, such as capacity drop, stop-and-go oscillations, and traffic hysteresis are also caused by fluctuating human driving behaviors (Saifuzzaman and Zheng, 2014, Sun et al., 2018, Sun et al., 2014a, Sun et al., 2014b, van Lint and Calvert, 2018). Therefore, it is essential to investigate short-term changes in driving behavior in CF modeling.

For a realistic and reliable CF model with HF, the observations/information about driver is the best research resources. Unfortunately, starting from the early days of traffic flow modelling, it has been a constant struggle for researchers to obtain such a complete set of data. In our understanding, this is also a main obstacle that prevents researchers to explicitly incorporate human factors into car following models. In the field of CF behavior, the most commonly used method is to investigate the driving style through driving behavior expressed by the vehicle. Generally, the driving styles can be categorized as timid, neutral and aggressive (Zheng et al., 2013, Bergasa et al., 2017, Martinez and Cao, 2018), and are regarded as long-term driving characteristics in this paper, as the 3 classes shown in Fig. 1(a). For each driver, the long-term driving characteristic is the intrinsic driving style under normal traffic conditions, and it is constant and unchanged. However, many studies (Ossen and Hoogendoorn, 2011, Taylor et al., 2015, Kazemi and Abdollahzade, 2016, Zheng et al., 2019) have indicated that the CF process has a time-varying nature. During the driving process, a driver could be influenced by many environmental factors, which are called stimuli, including traffic congestion, weather, other vehicles’ movement, etc. In this study, we named the time-varying nature as short-term driving characteristic, and it is a temporary change in driving behavior after external stimulus and psychological fluctuation. After the short-term changes, the driver will return to their long-term driving characteristics. The whole process is shown in Fig. 1(a). The “long-term” has fixed attribute while the “short-term” has temporary attribute, and the two jointly constitute a driver's driving attribute. But even in response to the same stimuli, different drivers can display various short-term characteristics, as shown in Fig. 1(b). Some drivers may temporarily change their driving styles, and then resume their initial and normal driving patterns (i.e. long-term driving characteristics) after different periods of short-term driving.

Given the significance of HF in CF modeling, in this paper, we firstly study drivers’ characteristics of intrinsic driving and temporarily changing. We then define the long- and short-term driving (LSTD) characteristics and incorporate them into CF models to accurately depict the human driving behavior and the subtle changes caused by external stimuli and psychological fluctuations. The main contributions are:

• Modeling concepts: through the analysis of driving behavior, we find that drivers have different LSTD characteristics which are classified into three long-term and two short-term, and totally six types of LSTD characteristics. From the field of psychology to human driving behavior, the LSTD model is proposed;

• Different findings: under the same stimulus, drivers with the same long-term characteristics also show differences in short-term changes, which is explicitly analyzed;

• Application performance improvement: the LSTD model is integrated with the two classical CF models (i.e. IDM, Gipps’ model), and demonstrates significant performance improvement as the errors decrease by 36.7% and 35.7%, respectively.

The remaining of this paper is structured as follows: Section 2 reviews the studies on the HF in driving behavior and CF models; Section 3 investigates the LSTD characteristics and proposes the LSTD model; Section 4 describes the data used in this paper; Section 5 integrates LSTD with IDM and Gipps’ model, respectively; Section 6 interprets and discusses the results; and Section 7 summarizes the main conclusions and provides suggestions for future research.