Background: In 2015, over 5300 pedestrians were fatally injured in U.S. traffic crashes. One benefit of emerging autonomous vehicles is that this technology may not only eliminate many driver errors, but could also eliminate or mitigate many of these pedestrian collisions. However, to achieve this goal, the vehicle must have sufficient time to detect and respond to the many ways in which vehicle-pedestrian collisions can occur. Depending on how early a potential collision is detected, countermeasures could include automated emergency braking (AEB), and deployment of an external airbag. The objective of this study was to determine the potential reduction in pedestrian crashes which could be achieved by a fully autonomous vehicle (AV).
Methods: This study was based upon 523 in-depth vehicle-to-pedestrian crash investigations extracted from the NHTSA Pedestrian Crash Data Study (PCDS). The approach was to codify AV performance as one of two comprehensive rule-based algorithms describing ideal AV driving behavior. The first algorithm was comprised of 12 rules which would constrain the AV to never violate traffic rules, e.g. failure to yield to a pedestrian in a crosswalk. The second algorithm was comprised of 13 additional rules which constrained the AV to drive cautiously in situations which could contain potential pedestrian conflicts, e.g., children darting out from between parked cars. Both algorithms were applied to each of the 523 cases assuming that the striking vehicle was an AV rather than the original car. We then reconstructed the earliest opportunity for an AV to detect the pedestrian, and potentially avoid the crash in each case.
Results: A total of 40% of the crashes in our dataset were the result of a driver violation, i.e. cases which the AV under Algorithm 1 would avoid. In the balance of the Algorithm 1 cases in which there was no driver violation, nearly 80% of the pedestrian were visible for over 1 second – allowing activation of AEB. For an AV equipped with Algorithm 2, all but 27 of 523 pedestrian conflicts would have been avoided. In most of these cases, there was sufficient time to activate AEB. However, there was one case in which a pedestrian would still be struck despite application of the rigorous Algorithm rules.
Discussion: This study found that even with idealized performance, perfect sensors and ideal weather, not all pedestrian conflicts could be avoided. Limitations in this study included the inability to account for pedestrian occlusion, the assumption of ideal weather, and no sensor degradation. Accounting for these factors would likely decrease the number of pedestrian crashes which could be avoided.
Conclusions: Even under best case conditions it is unlikely that an ideal AV could avoid every pedestrian to vehicle crash. Therefore, an AV will require safety features, such as a pedestrian-friendly front structure or an external airbag, to protect pedestrians. This study is the first of its kind to estimate the potential safety benefits of an AV in pedestrian crashes and has important implications for both automakers and regulatory bodies.