Exploring the mechanism of crashes with automated vehicles using statistical modeling approaches.

Song Wang,Zhixia Li,Yanyong Guo

doi:10.1371/journal.pone.0214550

Abstract

Autonomous Vehicles (AV) technology is emerging. Field tests on public roads have been on going in several states in the US as well as in Europe and Asia. During the US public road tests, crashes with AV involved happened, which becomes a concern to the public. Most previous studies on AV safety relied heavily on assessing drivers’ performance and behaviors in a simulation environment and developing automated driving system performance in a closed field environment. However, contributing factors and the mechanism of AV-related crashes have not been comprehensively and quantitatively investigated due to the lack of field AV crash data. By harnessing California’s Report of Traffic Collision Involving an Autonomous Vehicle Database, which includes the AV crash data from 2014 to 2018, this paper investigates by far the most current and complete AV crash database in the US using statistical modeling approaches that involve both ordinal logistic regression and CART classification tree. The quantitative analysis based on ordinal logistic regression and CART models has successfully explored the mechanism of AV-related crash, via both perspectives of crash severity and collision types. Particularly, the CART model reveals and visualize the hierarchical structure of the AV crash mechanism with knowledge of how these traffic, roadway, and environmental contributing factors can lead to crashes of various serveries and collision types. Statistical analysis results indicate that crash severity significantly increases if the AV is responsible for the crash. The highway is identified as the location where severe injuries are likely to happen. AV collision types are affected by whether the vehicle is on automated driving mode, whether the crashes involve pedestrians/cyclists, as well as the roadway environment. The method used in this research provides a proven approach to statistically analyze and understand AV safety issues. And this benefit is potential be even enhanced with an increasing sample size of AV-related crashes records in the future. The comprehensive knowledge obtained ultimately facilitates assessing and improving safety performance of automated vehicles.

Highlights

Technological advancement has brought Autonomous Vehicles (AVs) into reality, with the fact that relationships between vehicles and drivers are likely to be reversed significantly in the twenty years [1]
This paper aims at quantitatively investigate into the significant and ruling factors that contribute to AV crashes with various severity levels and collision types
Among all 113 AV involved crashes, 76 crashes happened with the vehicle driving on Automated Driving (AD) mode. 37 of these crashes happened with the vehicle driving on conventional mode

Summary

Introduction

Technological advancement has brought Autonomous Vehicles (AVs) into reality, with the fact that relationships between vehicles and drivers are likely to be reversed significantly in the twenty years [1]. Twenty-nine states have enacted legislation to regulate AVs and approved AVs public tests [3,4]. Public tests of AV have already been underway in several states of the US such as California, Nevada, and Michigan, etc. The AV manufacturers that are testing AVs on public roads are either from traditional vehicle manufacturers (i.e., Toyota, Nissan, and General Motor), or technology companies (i.e., Google, Uber, and Baidu). These AV manufacturers have commonly adopted SAE’s six levels of autonomy. Most of the vehicles currently that are tested on public roads are either Level 3 (conditional automation) or Level 4 (high automation) AVs

Objectives

Methods

Results

Discussion

Conclusion