Modeling Longitudinal Driving Behavior of Vehicles Under Disordered Traffic Conditions

Abstract

Over the last few decades, the problem of increased motorization has led to high levels of traffic congestion, traffic hazards, and environmental pollution. It is necessary to understand and model the driving behaviors of vehicles in the traffic stream to overcome these issues and achieve a safe and sustainable transportation system. In Western countries, traffic conditions are observed to be ordered, i.e., vehicles are predominantly cars that follow lane discipline. In contrast, the traffic stream in Southeast Asian countries (e.g., India and China) is disordered in nature, composed of a wide mix of vehicles maneuvering without lane-discipline, giving rise to different regimes in driving behavior. Longitudinal driving behavior in disordered traffic is characterized mainly by two regimes: the following regime, where the vehicle follows a single leader, and the non-following regime, where the leaders change frequently. The existing studies used the driving behavior models developed for ordered traffic conditions; however, due to the lack of microscopic traffic datasets and the exhausting data collection procedures, this topic of research has not been explored well under disordered traffic conditions. Therefore, the present study aims to address these shortcomings in modeling longitudinal driving behavior in disordered traffic by calibrating the existing physics based driving behavior models and integrating them with the data-driven models. For this purpose, the vehicle trajectory data are extracted from videos captured using a swarm of Unmanned Aerial Vehicles (UAVs) on an urban arterial road in Chennai city. The length of the study section chosen is comparatively higher than that of studies. However, the physics-based models may be constrained in their assumptions and have fixed variable space. In order to incorporate the inferences from data and flexibility in modeling, the data driven models such as Artificial Neural Network (ANN) and Random Forest (RF) models, are trained to predict the behavior of vehicles in