Contributing Factors to Consider While Defining Acceptance Criteria and Validation Targets for Assuring SOTIF in Autonomous Vehicles

Two major steps involved in SOTIF analysis are defining acceptance criteria and estimating the validation target. While acceptance criteria aids in determining if we have an acceptable residual risk corresponding to a hazardous scenario, the validation target specifies the amount of testing effort (in hours or representative miles) that is needed to ensure that the acceptance criteria are met. The current approaches for defining acceptance criteria heavily rely on existing fatality databases or naturalistic driving study data sets. The criterion is selected based on average number of fatalities or crashes per mile or per one hour of operation. The validation target is then calculated based on acceptance criteria. However respective validation targets., are these values really reflecting the acceptable risk criteria and targets? According to statistics, for a given data set and a random sample derived from the dataset, only the mean of population of the data set and the sample can be considered equal. In the case of autonomous vehicles, this implies we can only generalize the mean of crash statistics of all vehicles in a country to a mean of crash statistics of a sample fleet of vehicles. However, the current acceptance criteria consider arithmetic mean across states but not vehicles, and often do not take into account the operational design domain (ODD) factors, the driver characteristics, and justifications behind their selection. Similarly, while the validation target is defined from the acceptance criteria there are no minimal requirements with respect to scenarios that are stated which needs be met within the target. To overcome these limitations, in this paper, we discuss the various factors that need to be weighed in for defining acceptance criteria and validation target. We illustrate with an example how the results can vary drastically when the factors we defined are considered and compare our results with the currently adopted methods.