Comparison of automated vehicle struck-from-behind crash rates with national rates using naturalistic data

https://doi.org/10.1016/j.aap.2021.106056Get rights and content

Highlights

  • AVs are struck from behind 4.8 times more per mile than conventional vehicles.

  • Most of the crash rate difference was found in urban driving.

  • Different definitions of “urban” across data sets limit the significance.

Abstract

Automated vehicle developers in California are required to submit records of crashes and distances traveled in autonomous mode for all vehicles in their fleets. Several studies have investigated this database to compare automated vehicle crash rates with national rates. Although automated vehicles are struck from behind in 73 % of their autonomous mode crashes, this is the first study to compare automated vehicle struck-from-behind crash rates to national rates using equivalent crash definitions. Rear-end collisions have substantial public health and economic impacts, representing a third of all collisions and $3.9 B in annual economic costs. In this study, automated vehicles in autonomous mode were found to be struck from behind at 4.8 times the rate of human-driven vehicles in a naturalistic driving study. When controlling for driving environment, the rates for AVs were 5.0 times higher for urban driving and not significant for business/industrial driving, although these results are for different manufacturers, complicating the results. Automated vehicles were more likely to be struck when stopped than when moving compared to human-driven vehicles, suggesting that automated vehicles’ decisions about where and when to stop or remain stopped at intersections are more plausible contributing factors than unexpected rates of deceleration.

Introduction

Vehicles with automated driving features, defined here as combined lateral and longitudinal control or SAE Levels 3–5 (SAE International, 2018), have been tested on public roads in the United States nearly continuously since 2010 (Beiker, 2014). In September 2014, the California Department of Motor Vehicles began requiring companies wishing to test AVs on public roads to obtain permits (California Department of Motor Vehicles, 2018). All permit holders were required to report all crashes using form OL 316 within 10 days of the incident, regardless of severity or whether the vehicle was under human or computer control. Permit holders were also required report all test vehicles, miles driven in autonomous mode, and disengagements of the autonomous driving system by January 1st for the prior period of December to November. The combined public data sets of vehicle crashes and mileages for all autonomous vehicle testing in California proved a valuable resource for researchers, and several studies have shown that automated vehicle (AV) crash rates are lower than general public crash rates when controlling for crash severity (Blanco et al., 2016; Teoh and Kidd, 2017). Studies have also reported that a majority of AV crashes involve the automated vehicle being struck from behind (Leilabadi and Schmidt, 2019). As a proportion of total crashes, these are far higher than those reported by the driving public (Favarò et al., 2017).

Struck-from-behind crashes are a significant economic and public health concern. Rear end collisions account for 32.3 % of all crashes in the United States (National Highway Traffic Safety Administration, 2020a, p. 29). Economic losses from whiplash injuries were estimated as $2.7 B nationally in 2002 dollars ($3.9 B in 2020) (“Federal Motor Vehicle Safety Standards; Head Restraints,” 2010).

Previous studies have investigated the AV crash rates generally, but none have compared AV and human-driven struck-from-behind crash rates. The purposes of this study are to determine whether AVs are struck from behind at higher rates per distance traveled than conventional vehicles, and to investigate potential causes.

Section snippets

Literature review

Several studies have analyzed automated vehicle crash records, differing in the methods and metrics. Table 1 shows an overview of automated vehicle crash studies and comparisons in the literature. Some studies did not calculate crash rates but instead performed exploratory analysis (Das et al., 2020), text mined crash narratives (Alambeigi et al., 2020; Boggs et al., 2020b), or modeled crash severity (Wang and Li, 2019). Other studies calculated automated vehicle crash rates but did not compare

Materials and methods

Relevant data is publicly accessible and available for download (Goodall, 2020).

Results and discussion

Automated vehicle crash rates per million vehicle-miles traveled were compared with crash rates of human-driven vehicles from the SHRP 2 NDS dataset. Table 3 shows an overview of the crash rate comparisons.

Conclusions

Automation of the driving will profoundly affect transportation both in terms of mobility and safety. Understanding the performance of automated driving systems, especially early deployments, is of profound importance. While previous studies have found AV crashes to occur at lower rates than those of human-driven vehicles, this is the first study to investigate struck-from-behind crashes which represent 73 % of autonomous-mode crashes. While automated vehicles are struck from behind at higher

Funding

This work was sponsored by the Virginia Department of Transportation, United States.

CRediT authorship contribution statement

Noah J. Goodall: Conceptualization, Methodology, Software, Formal analysis, Investigation, Data curation, Writing - original draft, Writing - review & editing, Visualization.

Declaration of Competing Interest

The author has no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

Special thanks to Jeremy Sudweeks for clarifying total mileage figures in the SHRP 2 Naturalistic Driving Study, Dr. Michael Fontaine for comments on an earlier draft of this paper, and Dr. Subasish Das for providing a database of automated vehicle crashes in California through March 2020.

References (36)

  • M. Blanco et al.

    Automated Vehicle Crash Rate Comparison Using Naturalistic Data

    (2016)
  • L.J. Blincoe et al.

    The Economic and Societal Impact of Motor Vehicle Crashes, 2010 (Revised) (No. DOT HS 812 013)

    (2015)
  • California Department of Motor Vehicles

    Title 13, Division 1, Chapter 1, Article 3.7 – Testing of Autonomous Vehicles

    (2018)
  • California Department of Motor Vehicles

    Autonomous Vehicle Collision Reports [WWW Document]

    (2020)
  • California Department of Motor Vehicles

    Disengagement Reports [WWW Document]

    (2020)
  • Center for Data Reduction and Analysis Support

    Release May 2020 SHRP2 Phase I Error Correction - Change Log [WWW Document]. SHRP 2 NDS

    (2020)
  • S. Das et al.

    Automated vehicle collisions in California: applying Bayesian latent class model

    IATSS Res.

    (2020)
  • M. Davis et al.

    National Telephone Survey of Reported and Unreported Motor Vehicle Crashes (No. DOT HS 812 183)

    (2015)
  • Cited by (20)

    • Advancing investigation of automated vehicle crashes using text analytics of crash narratives and Bayesian analysis

      2023, Accident Analysis and Prevention
      Citation Excerpt :

      From those studies, rear-end collision was found to be the most frequent type of collision, while injury severity of AV-involved crashes tended to be lower than that of human-driven vehicle crashes (Schoettle and Sivak 2015, Favarò et al. 2017, Ashraf et al. 2021, Kutela et al. 2022a, McCarthy 2022). Especially, it was shown that AV-involved crashes had a higher proportion of rear-end collisions than conventional vehicle crashes, while AVs had been struck from behind 4.8 times more frequently than human-driven vehicles (Petrović et al. 2020, Goodall 2021). Besides, some studies have attempted to identify influential factors concerning AV crash types, injury, or vehicle damage (Wang and Li 2019, Xu et al. 2019, Boggs et al. 2020b).

    • What can we learn from autonomous vehicle collision data on crash severity? A cost-sensitive CART approach

      2022, Accident Analysis and Prevention
      Citation Excerpt :

      Many researchers have focused on the analysis of AV safety using AV road test data. For instance, studies have addressed the comparison between AV and conventional human-driving vehicle (HV) crash rate (Goodall, 2021; Favarò et al., 2017; Schoettle and Sivak, 2015; Banerjee et al., 2018); AV disengagement (Wang and Li, 2019a; Boggs et al., 2020a; Dixit et al., 2016; Favarò et al., 2018); descriptive analysis (Favarò et al., 2017), statistical analysis (Das et al., 2020; Boggs et al., 2020b), crash sequence analysis (Song et al., 2021), and text analysis (Alambeigi et al., 2020; Boggs et al., 2020b) of AV collision report; AVs’ safety reliability (Li and Zhai, 2019; Kalra and Paddock, 2016). Nevertheless, limited research has been dedicated to AV crash severity analysis.

    • Mining patterns of autonomous vehicle crashes involving vulnerable road users to understand the associated factors

      2022, Accident Analysis and Prevention
      Citation Excerpt :

      Similar to pedestrian-AV studies presented earlier, the bicyclists-AV studies used either survey questionnaires or simulation data. Relatively few studies have utilized actual AV crash data collected from the California Department of Motor Vehicle to evaluate the safety implications of AVs (Alambeigi et al., 2020; Boggs et al., 2020; Das et al., 2020a; Favarò et al., 2017; Goodall, 2021). However, the literature on the safety implication of AV on VRU is still limited.

    • Driving examiners’ views on data-driven assessment of test candidates: An interview study

      2021, Transportation Research Part F: Traffic Psychology and Behaviour
      Citation Excerpt :

      At the same time, there is a growing realization that fully automated driving may not be achieved within the next three to five decades (Litman, 2021; Shladover, 2016; Tabone et al., 2021). While there have already been compelling demonstrations of automated driving without human intervention, even the most advanced prototypes to date need occasional human intervention or behave in unexpected manners (Boggs et al., 2020; Goodall, 2021), suggesting that for the coming decades, drivers still need to be trained and licensed. Although fully automated driving is not within immediate reach, cars are becoming increasingly computerized data-collection machines.

    View all citing articles on Scopus
    View full text