Area-wide real-world test scenarios of poor visibility for safe development of automated vehicles
Introduction: Automated vehicles in everyday real-world traffic are predicted to be developed soon (Gasser et al., Rechtsfolgen zunehmender Fahrzeugautomatisierung, Wirtschaftsverlag NW, Berichte der Bundesanstalt für Straßenwesen F83, 2012). New technologies such as advanced object detection and artificial intelligence (AI) that use machine or deep-learning algorithms will support meeting all the maneuvering challenges involved in different degrees of automation (Society of Automotive Engineers - SAE international, Levels of driving automation for on road vehicles, Warrendale, PA., 2014; National Highway Traffic Safety Administration - NHTSA, Preliminary statement of policy concerning automated vehicles, Washington, DC, 2018). For automated series production, these vehicles of course must be safe in real-world traffic under all weather conditions. Therefore, system validation, ethical aspects and testing of automated vehicle functions are fundamental basics for successfully developing, market launching, ethical and social acceptance. Method: In order to test and validate critical poor visibility detection challenges of automated vehicles with reasonable expenditure, a first area-wide analysis has been conducted. Because poor visibility restricts human perception similar corresponding to machine perception it was based on a text analysis of 1.28 million area-wide police accident reports - followed by an in-depth case-by-case analysis of 374 identified cases concerning bad weather conditions (see chap. 1.3). For this purpose the first time ever a nationwide analysis included all police reports in the whole area within the state of Saxony from the year 2004 until 2014. Results: Within this large database, 374 accidents were found due to perception limitations - caused by ""rain"", ""fog"", ""snow"", ""glare""/""blinding"" and ""visual obstruction"" - for the detailed case-by-case investigation. All those challenging traffic scenarios are relevant for automated driving. They will form a key aspect for safe development, validation and testing of machine perception within automated driving functions. Conclusions: This first area-wide analysis does not only rely on samples as in previous in-depth analyses. It provides relevant real-world traffic scenarios for testing of automated vehicles. For the first time this analysis is carried out knowing the place, time and context of each accident over the total investigated area of an entire federal state. Thus, the accidents that have been analyzed include all kinds of representative situations that can occur on motorways, highways, main roads, side streets or urban traffic. The scenarios can be extrapolated to include similar road networks worldwide. These results additionally will be taken into account for developing standards regarding early simulations as well as for the subsequent real-life testing. In the future, vehicle operation data and traffic simulations could be included as well. Based on these relevant real-world accidents culled from the federal accident database for Saxony, the authors recommend further development of internationally valid guidelines based on ethical, legal requirements and social acceptance.