Simulating Naturalistic Driving Environment for Autonomous Vehicles

Abstract

Autonomous Vehicle (AV) technology has the potential to revolutionize the future transportation landscape. Simulation plays an indispensable role in the development and testing of AVs, due to its unmatched advantages of controllability, repeatability, and cost-effectiveness. To build the simulation system, one key problem is how to model background agents behavior, in order to construct the Naturalistic Driving Environment (NDE). To ensure the trustworthiness of simulation results, the NDE must meet the novel and demanding criterion, i.e., statistical realism, in terms of simulation fidelity. The simulated NDE needs to be statistically representative of the real-world traffic environment, particularly for those long-tail safety-critical events, which are critical for AV safety. Unfortunately, the real-world NDE is spatiotemporally complex, highly interactive, and with only rare occurrences of safety-critical events. Therefore, how to build a high-fidelity simulator is a long-standing problem. This dissertation aims to provide systematic methods for simulating high-fidelity NDE for AV development and testing, leveraging large-scale Naturalistic Driving Data (NDD). We first identify the statistical realism requirement on NDE, a new requirement for microscopic traffic simulators brought by AV applications. To achieve this, a data-driven method is proposed to characterize human car-following and lane-changing behavior distributions. The proposed method is validated using real-world data in the simulation of multi-lane highway driving environments. In contrast to highway driving environments, urban environments usually involve more complex interactions between multiple agents. Therefore, we further develop NeualNDE, a Deep Learning (DL)-based NDE modeling framework. The results validate that NeuralNDE can achieve accurate normal driving statistics (e.g., vehicle speed/distance/yielding behavior distributions, etc.), and more importantly, safety-critical driving statistics (e.g., crash rate/type/severity and near-miss statistics, etc.), as demonstrated in the simulation of real-world urban driving environments. One important application of high-fidelity NDE is AV safety performance evaluation. We discover that sparse but adversarial adjustments to the NDE, resulting in the Naturalistic and Adversarial Driving Environment (NADE), can significantly reduce the required test miles without loss of evaluation unbiasedness. The results show that, compared with directly evaluating AV in NDE, the proposed NADE environment can accelerate the evaluation process by multiple orders of magnitude. Consequently, we provide a complete pipeline for accurate and efficient simulation-based AV testing. In summary, this dissertation presents methodologies for building high-fidelity NDE and uses AV testing as an example to demonstrate the importance of NDE simulation. These proposed methods pave the way for enhancing AV safety performance, which is beneficial for all stakeholders, including AV developers, customers, and regulators, and contributes to the long-term development of AV technology.