RISK ESTIMATION FOR PERCEPTION FAILURES IN AUTOMATED DRIVING

Abstract

The failure of sensors to perceive the environment correctly is one of the primary sources of risk that needs to be quantified in the development of active safety features for autonomous vehicles. By extracting training data from the CARLA simulator, an object detector was trained to simulate a perception system of an autonomous vehicle. Using the detection model and gathering data for incorrect detections, various extreme value models were created and compared to investigate if extreme value theory is a viable option for estimating the risk of sensor failures of the perception system. An analysis of the extreme value's dependency on the velocity of the vehicle is performed and a risk measure is presented.