Simulating and Analyzing aggressive car-following behavior for testing autonomous vehicles

Abstract

In a mixed traffic flow, evaluating the operation safety of autonomous vehicles (AVs) is crucial under different aggressive car-following behavior of surrounding human driven-vehicles (HVs). To pursue this goal, this paper develops a machine-learning-based simulation method accompanied by a traditional car-following model. We integrate the unsupervised learning method, Approximate Bayesian computation, and Gipps car-following model to obtain different parameter distributions of the Gipps model. After utilizing the key parameter distribution, this paper employs a supervised learning method to predict the aggressive index of human-driven vehicles in each individual car-following event. Further, we verify the outputs of this simulation model with the NGSIM I-80 traffic dataset. After the validation, we develop an AV simulation study to analyze Avs’ performance in different aggressive HVs scenarios. The result indicates that the higher penetration rate of AV is essential for stabilizing AVs’ performance in terms of velocity and the probability of involving in a crash. Additionally, aggressive HV drivers significantly impact scenarios of low AV penetration rate regarding safety issues.