Challenges of Data-Driven Simulation of Diverse and Consistent Human Driving Behaviors

TL;DR

The paper addresses the challenge of creating statistically realistic driving environments for autonomous vehicles by proposing a data-driven framework that accommodates the complexity, diversity, and temporal consistency of human drivers. It critically compares theory-based IDM with driver-specific parameters to a data-driven policy (XGBoost) using the NGSIM dataset, finding that per-driver IDM offers improvements but remains suboptimal, while XGBoost struggles with one-step predictions and accumulative errors. The analyses reveal that human driving styles are multimodal and non-Gaussian and that driving behavior exhibits temporal consistency, implying the need for driver-aware stochastic models in simulations. These insights pave the way for more faithful, high-fidelity AV simulators capable of capturing real-world driving variability and stability over time.

Abstract

Building simulation environments for developing and testing autonomous vehicles necessitates that the simulators accurately model the statistical realism of the real-world environment, including the interaction with other vehicles driven by human drivers. To address this requirement, an accurate human behavior model is essential to incorporate the diversity and consistency of human driving behavior. We propose a mathematical framework for designing a data-driven simulation model that simulates human driving behavior more realistically than the currently used physics-based simulation models. Experiments conducted using the NGSIM dataset validate our hypothesis regarding the necessity of considering the complexity, diversity, and consistency of human driving behavior when aiming to develop realistic simulators.

Figures (5)

• Figure 1: We compare the velocity predictions of the two types of IDMs and a learned XGBoost predictor to the ground truth data in the simulated NGSIM environments. We find that replicating the human driving behavior is very difficult, even in this simplistic car-following setting.
• Figure 2: We compare the change in velocity predicted by XGBoost to the ground truth change in velocity. The plot indicates that human behavior may be inherently unpredictable, and capturing the behavior seems to be highly challenging.
• Figure 3: We plot the average acceleration and deceleration and the preferred minimum time headway, that is, the safety margin between two vehicles for each driver in each of the four NGSIM datasets. The plots show that human driving behavior is highly diverse. It can be fat-tailed and multi-modal and does not simply follow a Gaussian distribution.
• Figure 4: We plot the distribution of the IDM parameters for each of the four NGSIM datasets when each driver is represented by a separate set of parameters. The black vertical line can be interpreted as the human mean, whereas the red-shaded area depicts the expected noise in the IDM fitting process.
• Figure 5: We analyze the consistency of human drivers on the four datasets in the NGSIM data by fitting IDM to the longest trajectories in the dataset. Our results show that in all of the NGSIM datasets, the human driving style shows statistically significant consistency when we focus on the human drivers of which most data has been collected. The uncertainty plotted with error bars is one standard error of the mean.