PREPER: A Public Dataset with Key Performance Indicators for the Safety Evaluation of Deep Learning Visual Perception Systems for Self-Driving Cars

Deep learning algorithms are being widely used in autonomous driving (AD) and advanced driver assistance systems (ADAS) due to their impressive capabilities in visual perception of the environment of a car. However, the reliability of these algorithms is known to be challenging due to their data-driven and black-box nature. This holds especially true when it comes to accurate and reliable perception of objects in edge case scenarios. So far, the focus has been on normal driving situations and there is little research on evaluating these systems in a safety-critical context like pre-crash scenarios. This article describes a project that addresses this problem and provides a publicly available dataset along with key performance indicators (KPIs) for evaluating visual perception systems under pre-crash conditions.