Safety Analysis of Human Car-following Models

Abstract

The emerging Intelligent Transportation System (ITS) involves a variety of collision avoidance systems to achieve the goal of zero fatality on the road. A thorough safety analysis is mandatory as a system failure can result in serious injuries and even loss of lives, and simulation is extensively used for this purpose. The simulation requires the consideration of human drivers in the system as the future traffic will be consisting of autonomous vehicles and human-driven vehicles. Hence, car-following models are increasingly used in the research on ITS to capture human driving behavior. Meanwhile, the quality of system safety is evaluated using safety indicators that can detect potential collisions between vehicles. For reliable safety validation, the car following models should be realistic. In particular, precisely capturing the possibility of collision in human driving is paramount. An accurate simulation of a car-following model can be viewed in two aspects:one is to incorporate random human factors in the car-following model to improve the model behavior, and the other is to use formal modeling languages to avoid undefined simulation behavior. Although each of these aspects is considered in civil and computer engineering separately, there has been not many attempts to marry these two aspects. On the other hand, another problem is related to the safety indicator calculation algorithms because these include several assumptions that can mislead the safety evaluation. In this thesis, we aim to address these two shortcomings. Despite the long history of car-following models, it is still questionable that whether these models can reproduce the risk of collision realistically. Therefore, we conduct a critical analysis of two widely used car-following models called Intelligent Driver Model (IDM) and Optimum Velocity Model (OVM) to determine the usefulness of these models for safety validation. An extensive experimental dataset obtained from human drivers is used to compare with the IDM and OVM simulation. We perform the analysis in three steps. The first step analysis shows that these models are reasonably accurate for simulating vehicle dynamics. However, the second step analysis reveals that IDM and OVM significantly underestimate the risk of collision (i.e., extremely safer than the reality). The last step analysis shows the simulation of a vehicle platoon of 10 vehicles. Through these analyses, we clearly address the weaknesses of IDM and OVM, and motivate the need for improvements. We develop a novel compositional car-following model called Modal car-following model (MCFM) based on Hybrid Input Output Automata (HIOA). It integrates an existing car-following model, such as IDM and OVM, with a human factor model which captures three distinct human factors. This work contributes to bridging the gap between civil and computer engineering. HIOA allows modeling of the physical system combined with discrete mode switches, which is ideal for describing piecewise continuous phenomena. Thus, HIOA models offer a succinct framework for the specification of car-following behavior. The human factors considered in our model are the human perception error, human reaction delay, and temporal anticipation. An extensive benchmarking shows that MCFM can significantly improve IDM and OVM simulation. More precisely, the root-mean-squared error of the following vehicle position in the simulation is reduced by up to 48.8% for IDM and 7.04% for OVM. Furthermore, MCFM precisely captures the risk of collision in human driving. The other aspect that this thesis considers is related to the safety indicator calculation algorithm. The safety validation of ITS not only relies on the accurate modeling and simulation of the system but also an accurate safety indicator calculation algorithm that can objectively evaluate the system safety. Furthermore, the safety indicators are widely used in collision avoidance systems as safety constraints. While there are many existing algorithms available, they make unrealistic assumptions in the calculation that can lead to failure in detecting a potential collision.This is critical as it may lead to an actual collision. Therefore, in this thesis, we clearly demonstrate this issue and solve it by formulating an optimization problem for calculating the safety distance. The assumptions in our approach are more realistic than existing algorithms. Through simulations, we reveal that our safety distance calculation detects potential collisions more accurately and produces the optimal safety distance. This thesis concludes with the future work towards the application of our work with a vision for simulation of the future traffic consisting of autonomous and human driven vehicles.