Bayesian Test Design for Reliability Assessments of Safety-Relevant Environment Sensors Considering Dependent Failures

With increasing levels of driving automation, the perception provided by automotive environment sensors becomes highly safety relevant. A correct assessment of the sensors' perception reliability is therefore crucial for ensuring the safety of the automated driving functionalities. There are currently no standardized procedures or guidelines for demonstrating the perception reliability of the sensors. Engineers therefore face the challenge of setting up test procedures and plan test drive efforts. Null Hypothesis Significance Testing has been employed previously to answer this question. In this contribution, we present an alternative method based on Bayesian parameter inference, which is easy to implement and whose interpretation is more intuitive for engineers without a profound statistical education. We show how to account for different environmental conditions with an influence on sensor performance and for statistical dependence among perception errors. Additionally, we study the impact of error dependence among several sensors on the perception reliability of a redundant multi-sensor system. To this end, we simplify the sensor data fusion with a majority voting scheme, which implies that the multi-sensor system's perception fails whenever more than half of the individual sensors commit unacceptable errors. For a redundant multi-sensor system, in which error occurrence is weakly dependent, it can be shown that empirical reliability assessments are feasible. While the presented method does not encompass entirely the full complexity of the problem, it provides an initial systematic estimate of the necessary test drive effort and facilitates the use of sound statistical methods for test effort estimation.