Does road environment aesthetics influence risky driving behavior of autonomous vehicles? An evaluation on road readiness using explainable machine learning and random parameters multinomial logit with heterogeneity.

Abstract

Aesthetics has always been an advanced requirement in road environment design, because it can provide a pleasant driving experience and guide better driving behavior for human drivers. However, it remains unknown whether aesthetics-based road environment design also has an impact on autonomous vehicles (AVs), resulting in that current evaluation models on road readiness for AVs (RRAV) do not consider road environment aesthetics. Therefore, this study aims to explore the relationship between road environment aesthetics and risky driving behavior of AVs (RDBAV) and propose an RRAV evaluation model from the new perspective of road environment aesthetics. Using real autonomous driving data, 1,491 longitudinal RDBAV events and 225 lateral RDBAV events are acquired together with corresponding road environment images. A novel quantitative model of road environment aesthetics is developed and 38 relevant feature variables are extracted from four aspects, including Naturalness, Vividness, Variety, and Unity. Then, an explainable machine learning that combines XGBoost (eXtreme Gradient Boosting) with SHAP (SHapley Additive exPlanation) is employed to establish an evaluation model of RRAV, by treating the occurrence of RDBAV as the dependent variable and feature variables of road environment aesthetics as independent variables. The results show that this XGBoost-based RRAV evaluation model performs better than other commonly-used methods, with accuracies of 96.9% and 91.8% for longitudinal and lateral RDBAV prediction, respectively. Due to the advantages of SHAP, the influence degrees of aesthetic features of road environments on RDBAV are calculated and explained based on global and individual feature contributions. In addition, a random parameters multinomial logit model with heterogeneity in means and variances reveals that the indicator of left visual curve length in the "middle scene" and the indicator of dominant color have significant heterogeneity for the analyses of longitudinal RDBAV. The findings of this study might contribute to the accurate evaluation of RRAV from the new viewpoint of aesthetics, the development of human-like visual perception systems of AVs, and the optimization of aesthetics-based road environment design.