Article Text
PDF
Abstract
Objectives To examine the patterns and associated factors of road traffic injuries (RTIs) involving autonomous vehicles (AVs) and to discuss the public health implications and challenges of autonomous driving.
Methods Data were extracted from the reports of traffic crashes involving AVs. All the reports were submitted to the California Department of Motor Vehicles by manufacturers with permission to operate AV test on public roads. Descriptive analysis and χ2 analysis or Fisher’s exact test was conducted to describe the injury patterns and to examine the influencing factors of injury outcomes, respectively. Binary logistic regression using the Wald test was employed to calculate the OR, adjusted OR (AOR) and 95% CIs. A two-tailed probability (p<0.05) was adopted to indicate statistical significance.
Results 133 reports documented 24 individuals injured in 19 crashes involving AVs, with the overestimated incidence rate of 18.05 per 100 crashes. 70.83% of the injured were AV occupants, replacing vulnerable road users as the leading victims. Head and neck were the most commonly injured locations. Driving in poor lighting was at greater risk of RTIs (AOR 6.37, 95% CI 1.47 to 27.54). Collisions with vulnerable road users or incidents happening during commute periods led to a greater number of victims (p<0.05). Autonomous mode cannot perform better than conventional mode in road traffic safety to date (p=0.468).
Conclusions Poor lighting improvement and the regulation of commute-period traffic and vulnerable road users should be strengthened for AV-related road safety. So far AVs have not demonstrated the potential to dramatically reduce RTIs. Cautious optimism about AVs is more advisable, and multifaceted efforts, including legislation, smarter roads, and knowledge dissemination campaigns, are fairly required to accelerate the development and acceptance.
- engineering
- public health
- social marketing
- motor vehicle occupant
Statistics from Altmetric.com
Introduction
Global fatalities resulting from crashes climbed to 1.35 million in 2016 with an increase of 8% since 2012. Road traffic injury (RTI) is the leading cause of death for children and young adults aged 5–29 years.1 Among crashes of manned vehicles, driver errors are estimated to account for 94%, of which recognition errors comprise approximately 41%, decision errors 33% and performance errors 11%.2 Despite plenty of regulatory and educational interventions having been carried out, human drivers’ mistakes remained the predominant cause of RTIs. Target 3.6 of Sustainable Development Goals to halve road traffic deaths by 2020 may not be achieved if the same reduction pace continues, and even the mortalities in certain regions are rising.3–5 Emerging technologies, especially artificial intelligence, were expected to eliminate driver-related road traffic hazards. Autonomous vehicles (AVs), also known as ‘robotic’, ‘driverless’ or ‘self-driving’ vehicles, evolved from the preceding advanced driving assistant system and had improved target distinguishing arithmetic, route planning and active control. It is anticipated that AVs will be a revolutionary technological innovation with the potential to lower the permit threshold of driving, reduce travel costs, mitigate environmental pollution and improve road safety.6–8 A survey showed that the public were likely to be receptive to AVs when provided with information relating to public health benefits.9 Autonomous taxis may be introduced more rapidly to realise the anticipant implications because they were deemed to be cost-effective.7 Some researchers proposed that the public health sector’s expertise should be harnessed to assist in the adoption of AVs with research and evidence-based advocacy.10
However, there are substantial challenges for vehicular automation, which are not only technical but also ethical and public health-related issues.11 12 Six levels (L0–L5) of vehicular automation were specified based on partial or complete dynamic driving task by the Society of Automotive Engineers.13 L0 refers to no automation and relies on continuous driver operation. The driver completes pivotal operations in L1 (driver assistance) and L2 (partial automation). L3, also known as conditional automation, allows the operator to intervene only when the automation system disengages. Higher level automation, L4 and L5 included, can independently accomplish entire operations without human participation, and the latter is suitable for every driving scenario. Redefining driver roles from operators to supervisors signifies that AV occupants delegate authority to the system instead of being replaced, which inevitably leads to adaptive issues like over-reliance on automation, loss of alertness and impaired driving performance when regaining control and even skill degradation over time.14 15 Emerging characteristics of crashes involving AVs in the evolution of vehicular automation may also arise. To date, AVs in levels 1 and 2 have become public commodities and have raised extensive public concerns on account of AV-involved crashes having caused six deaths.16
To facilitate the development of vehicular automation, the California Department of Motor Vehicles (CA DMV) allowed manufacturers to test AVs on public roads with permission according to the regulations approved in January 2014. Subsequently, test drivers were required in September 2014. As of 28 January 2019, there were 62 test permit holders which were obliged to submit a crash report involving AVs within 10 days of the occurrence. The reports provide public availability of raw data on AV-involved crashes and RTIs. Some articles have explored the crash characteristics, correlative factors and AV–driver interaction based on those open-access datasets.17–20
The number of disengagements which were counted as the shift of vehicle control from automated driving to manual driving had a significantly high correlation with the autonomous miles travelled.19 It should be noted that AV disengagements played a safety-critical role and were triggered by various causes, including human factors, system failure, and external conditions.20 Driver reaction time of regaining the steering wheel was a major concern, of which remarkable distinction was observed in diverse companies, disengagement types, roadway conditions, and autonomous miles travelled.19 Compared with conventional crashes, crash frequency (eg, the miles per crash) involving AVs were much higher, and rear end and collisions at an intersection were the most common.17 18 Other studies were conducted from the perspectives of sociology, ergonomics, and legislation using models, investigations, and reviews, fewer relating to traffic injuries on basis of field data. To discover the potential defects in road safety regarding AVs, injury patterns and associated factors of AV-involved crashes during field test were examined and discussed in this paper.
Methods
Data sources
Crash information was documented in prescribed report forms which were obtained from the open-access database on the official website of CA DMV.21 Simplified reports were employed before 2018, which contained five sections of key information: (1) manufacturer’s information, including manufacturer’s name and AV model; (2) status of vehicles and victims involved in the crash, including date, time of day, location of collision, road users involved, severity and areas of AV damage; (3) casualty, including the number of victims and their status; (4) description of crash details, including required driving mode options (autonomous or conventional) and non-normalised description, which was possibly related to traffic situation preceding collision, process details, injuries and postinjury treatment; and (5) certification to ensure the authenticity and accuracy of submitted information. A detailed version of the report came into effect in January 2018, whereas a few crashes were recorded in original forms. Compared with the simplified reports, more critical variables were attached to section 4 in the detailed format, in which weather, lighting, roadway surface, roadway conditions, movement preceding collision and type of collision were appended. Double entry and a joint decision with the third participation in case of divergence on the understanding of detail description were performed to ensure the accuracy of the data.
Variable definition and statistical analysis
Injury outcomes were directly from self-reports instead of the inclusion criteria of injury epidemiology due to not all crashes presenting post-injury treatment or absence from work. According to general office hours in America, commute periods were defined as 8:00~9:59, 12:00~12:59 and 17:00~18:59. The injured locations and post-injury treatment were extracted from the description of crash details. Eighteen states of vehicle movement preceding collision (eg, stopped, proceeding straight and merging) were merged into three categories based on the driving operation and moving trajectory at the time. Any substantial variation of moving direction, such as making right/left turn, changing lanes or merging was considered as direction changes. Except for rear end remaining the same, broadside and side swipe were classified as side collision, head-on and overturned as others. Variables of weather, lighting and roadway conditions were treated similarly, keeping the groups with maximum frequency constant. Driving at dusk, dawn or in dark streets with or without lights was regarded as pilot in poor lighting. Vulnerable road users referred to motorcyclists, bicyclists, scooterists, skaters and pedestrians in our study. The injury incidence rate was calculated as the ratio of the number of victims to that of crashes. Descriptive analysis and χ2 analysis or Fisher’s exact test via SPSS V.19.0 were conducted to describe the injury patterns and to examine the associations of specific factors with injury outcomes, respectively. Once the significant variables were identified in univariate analysis, a multivariate binary logistic regression, using entry method with criteria of p<0.05 for entry and p<0.10 for removal of independent variables, was performed to control the potential confounding factors. The Wald test was used to detect the remarkable factors of regression model with dependent variable as whether a crash causes injuries or not. OR, adjusted OR (AOR) and their 95% CIs were adopted to estimate the likelihood of road users suffering from traffic injuries. A two-tailed probability (p<0.05) was employed to indicate statistical significance.
Results
Patterns of AV-involved traffic injuries
A total of 133 reports were acquired from January 2017 to June 2019, including 93 in detailed version and 40 in the simplified version. Twenty-four victims in 19 crashes reported injuries on site or the next day, with the crude injury incidence of 18.05 per 100 crashes. As shown in table 1, the collisions during commute periods led to higher injury occurrence than that of non-commute periods, 31.58% (12/38 injured) for the former and 12.63% (12/95 injured) for the latter. The injury incidence rates ranged from 2.86% (1/35 injured) in the period of July–September to 29.03% (9/31 injured) in that of January to March. Most crashes happened on streets where the injury presence was 19.30% (22/114 injured), approximately 10 percentage points higher than that on expressways (2/19 injured). Of 126 crashes involving road users, 109 collided with other vehicles including AVs, totally resulting in 18 persons injured. The injury incidence that vulnerable road users suffered was up to 35.29% (6/17 injured), though they were less involved in crashes relating to AVs. As presented in table 2, AV testers accounted for a largest proportion of the injured (17/24 injured), followed by vulnerable road users (6/24 injured) and conventional motor vehicle occupants (1/24 injured). Most of the victims reported injuries of head and neck (9/24), trunk (3/24) and limbs (3/24) coming second. Only half of the injured (12/24) sought medical treatment on site or the next day, whereas the proportion was probably underestimated due to the healthcare and absence of other road users not being followed up and descripted except for AV operators. As a result, the conservatively estimated injury incidence was 9.02 per 100 crashes according to the definition of injury epidemiology.
Epidemiological patterns of autonomous vehicle-involved road traffic injuries during field test (N=133)
Status of the injured in AV-involved crashes (N=24)
Influencing factors of AV-involved injuries
It was noted that only 19 individuals injured were reported in detailed format and ultimately included in χ2 analysis and multivariate regression. Those reports were submitted by nine manufacturers, of which GM Cruise and Waymo accounted for 64 and 32 collisions, respectively, with no significance of injury incidence (shown in table 3). More crashes occurred during non-commute periods, while the injury incidence during commute periods was significantly higher (26.32% vs 9.47%, p=0.012). Collisions with diverse objects caused notably different injury presence (p=0.037), 11.93% (13/109 injured) for motor vehicles, 35.29% (6/17 injured) for vulnerable users and none for fixed objects. Crashes in poor lighting even with street lights on led to a notably greater number of victims than those in daylight (31.58% vs 6.76%, p=0.010). Autonomous and conventional driving modes had similar rates (16.05% vs 11.54%). The movement preceding collision of AVs and other vehicles made no difference to the population distribution of injuries. Moreover, road type, collision type, weather, roadway condition and severity of AV damage did not significantly affect injury distributions among road users.
Univariable analysis of road traffic injuries caused by AV-involved crashes during field test
Significant variables, including collision time, collision object and lighting, together with driving mode, which was the key variable to investigate, were included in the binary logistic regression. As shown in table 4, driving during commute periods was 3.41 times the likelihood of non-commute periods to suffer injuries (OR 3.41, 95% CI 1.26 to 9.24). Collisions with vulnerable road users were more likely to result in injuries than that with motor vehicles (OR 4.03, 95% CI 1.27 to 12.74). Compared with crashes in daylight, those in poor lighting were at greater risk of traffic injuries (OR 6.37, 95% CI 1.69 to 24.00). After model adjustment, poor lighting was verified as a risk factor in case of same driving mode, collision time and collision object. The likelihood was basically the same as the unadjusted model (AOR 6.37, 95% CI 1.47 to 27.54).
Multivariate logistic regression analysis on road traffic injuries caused by autonomous vehicle-involved crashes during field test
Discussions
Our analysis using field test data on autonomous driving to present the injury patterns and influencing factors yielded three key findings. First, AV occupants replaced vulnerable road users as the leading victims, whereas two-wheeled vehicle users still suffered a high risk of traffic injuries. Second, driving in poor lighting was identified as a risk factor which may result in over sixfold likelihood of traffic injuries. Collision time and collision object were contributing factors of the injury occurrence imparity as well. Finally, the significant variation between autonomous and convention modes was not found in present study, which seemed to signify that autonomous driving cannot perform better than manual driving in road traffic safety to date.
From the perspective of humans involved, different from conventional crashes where vulnerable road users accounted for 54% of the fatalities, AV-involved crashes were inclined to cause injuries to AV occupants.1 As of June 2019, self-driving cars led to six deaths, including 5 AV occupants and one pedestrian, which coincided with the rank order of injured road users in our study.16 That aroused the huge controversy over whether self-driving cars were intended to protect their consumers or other road users when collisions were inevitable in certain circumstances.22 According to Moral Machine, a famous platform gathering human perspectives on moral decisions made by machine intelligence, the public do not reach a consensus on protection preference of different species and human groups especially when the number of car occupants equate with that of other road users to affect.23 Head and neck became the most common injured locations, which would place vehicle occupants into greater odds of injuries and fatalities than other body parts like limbs.24 25 Vulnerable road users as unprotected road sharers are three to four times more likely to be injured compared with vehicle users, which similarly affected AV-related RTIs.26 Assuming that mature self-driving can really eliminate vehicle-related and driver-related road traffic hazards, crashes induced by two-wheelers and three-wheelers remain to be supervised and decreased due to around 60% of AVs being rear ended even if they having come to a complete stop.18 In practice, AVs were not always at fault in the past fatal crashes according to police investigations. Hence, unsafe driving behaviours of two-wheeler and three-wheeler riders like drunk driving, non-use of helmets and speeding should be regulated as usual while developing novel driving technologies.27
Though AVs with partial automation features (levels 1 and 2) have been on the market for years and achieved great progress in technologies concerned (eg, making left turn without traffic supervision), the public seems to be not confident in the current AVs. Around three quarters of Americans were afraid to ride in fully self-driving vehicles after the fatal crashes according to the annual automated vehicle survey by American Automobile Association in early 2019.28 On the one hand, human drivers are still expected to take over the steering wheel and throttle when AVs below level 4 disengaged.13 Consumers chose to purchase AVs for the very automated features, whereas system failure initiated 52% of the disengagements.17 Additionally, more likelihood to undertake non-driving tasks (eg, napping and texting) of AV occupants may impair their susceptibility to system alerts and traffic signals and may extend the reaction time of disengagement response.29 30 On the other hand, potential travel increases and more dangerous behaviours (eg, fatigue driving and drug driving) also sparked extensive concerns.10 Predictably, increasing vehicles and more complicated occupants will share roads with the driving threshold being lowered, and there will be greater challenges on the prevention and control of traffic injuries.
From the perspective of developing vehicular automation, decision-making of AVs that relies on the real-time capture and processing of complex traffic conditions and continuous response for hazard avoidance may result in occupant discomfort, which was estimated to account for 89% of the human factors triggering disengagements.17 Motion sickness has been verified in previous studies and has been proposed to negatively affect user acceptance, technology uptake and, ultimately, the assumed positive socioeconomic impact.31 Inadequate radar and Light Detection and Ranging (LiDAR) sensors and planning issues were considered playing a role in AV disengagements based on another dataset of CA DMV.20 Driver–car interaction and machine learning algorithms are the most crucial dilemmas at present. It was predicted that it would take 12.5 years of crash-free driving to demonstrate AV reliability and to reach an average ‘steady state’.32 Even if full autopilot can be completely ready and widely available before 2040, it is still far from a dramatic reduction of traffic injuries due to the regional inconsistency with affordability.33 Low-income and middle-income countries only possessed 60% of the global motor vehicles, of which only the registered two-wheelers and three-wheelers accounted for one-third, whereas 96% of the deaths occurred in those countries.1 In the regions having more two-wheeled and three-wheeled motor vehicles, especially Southeast Asia, with a proportion of 74.5%, crashes tend to happen among vulnerable road users and brought about 40% of the injured, over that initiated by automobile.27 Hence, it was misleadingly optimistic to save 10 million lives per decade via larger-scale replacement of conventional vehicles with private AVs globally in the short term.11
In terms of traffic environment, recognition errors account for 41% of the human driver errors, far higher than those of decision or performance.2 Poor visibility was confirmed as a critical hazard, significantly delaying driver reaction to the moving target.34 35 Unfortunately, such defects have been hardly overcome in AVs so far. Failure to perceive a sanitation vehicle and slam on the brakes while on autopilot contributed to the first fatal crash of self-driving in Hebei, China. It was estimated that the risk of traffic injuries from crashes in poor lighting was greater than that in daylight in present study. AVs have difficulty in comprehensively processing complicated traffic behaviours of higher-density road sharers during commute periods as well. The chain reaction of unmanageable traffic conditions, autopilot disengagement and delaying takeover may ultimately give rise to RTIs. As a result, the intelligent transportation system to diversify the vehicle–road communications is fairly necessary to improve the perception towards road users, vehicles and fixed objects.36 Argument that AVs should be designed to reduce light pollution appeared to go against the improvement of road safety involving AVs at present.37 All in all, only smart cars on smart roads can maximise the public health implications of unmanned vehicles.38
It was supposed that all RTIs and fatalities are preventable if appropriate measures are taken.39 Now that AVs are regarded as a revolutionary intervention for RTIs and AVs of levels 1 and 2 are on the market, unified safety assessment standards and clear legislative responsibility affirmation of crashes are urgent to advance the development.40 Improved urban lighting and easily perceived protection facilities for road users should be taken into consideration. Objective promotion and extensive knowledge dissemination campaigns are favourable to help the public better understand AVs and build their confidence in safety.41 In short, cautious optimism about AVs is more advisable, and multifaceted efforts are quite necessary for self-driving safety.
Our study provides initial insights into the traffic injuries resulting from crashes involving AVs. It may help the practitioners to perfect technical improvement and legislative frameworks. However, the injury inclusion was based on self-reports instead of defined injury epidemiology. Accordingly, we failed to make reliable comparisons with conventional crashes owing to diverse dimension, definition and potential greater reporting bias of other sources. Limited samples, injured cases, undersigned variables and multiple comparisons also restricted the stability and value. Miles driven, the extremely crucial compounding factor, cannot be controlled in the multivariate regression model.20 More selective variables like level of automation, speed and detailed medical information should be considered in future research. It should be noted that the characteristics of AV-involved crashes during field testing may be different from that during actual driving. Existing drawbacks of AVs were likely to be eliminated with technological advancement, and certain veiled defects probably appear when AVs are driven on a large scale; that is, the results were more applicable to the amelioration of technologies concerned and road safety condition rather than on-site intervention.
Conclusions
Poor lighting improvement and the regulation of commute-period traffic and vulnerable road users should be strengthened for AV-related road safety. So far, AVs have not demonstrated the potential of a dramatic reduction in traffic injuries. Cautious optimism about AVs is more advisable, and multifaceted efforts, including legislation, smarter roads and knowledge dissemination campaigns, are fairly required to accelerate the development and acceptance.
What is already known on the subject
Autonomous vehicles (AVs) are expected as a revolutionary intervention for road traffic injuries (RTIs). However, there are substantial challenges for vehicular automation, which are not only technical but also ethical and public health-related issues.
Previous studies mainly focused on automation technologies and ethical issues, and limited articles were published to discuss the public health implications and challenges, especially RTIs which have aroused worldwide controversy over fatal crashes.
What this study adds
Driving AVs in poor lighting is associated with RTIs in autonomous mode versus in conventional mode, probably resulting in higher odds of injuries. Consistent with conventional crashes, AV-involved collisions with vulnerable road users or incidents happening during commute periods contribute towards a greater number of victims. Our study provides initial insights into the patterns of crashes and traffic injuries concerned as well.
References
Footnotes
Twitter @no
Funding The authors have not declared a specific grant for this research from any funding agency in the public, commercial or not-for-profit sectors.
Competing interests None declared.
Patient consent for publication Not required.
Provenance and peer review Not commissioned; externally peer reviewed.
Data availability statement Data are available in a public, open-access repository. https://www.dmv.ca.gov/portal/dmv/detail/vr/autonomous/autonomousveh_ol316