Machine Learning Based Dynamic Risk Assessment for Autonomous Vehicles

Abstract

Autonomous vehicles (AVs) are complex safety-critical systems that operate in an uncertain and dynamic environment. To ensure safety, Hazard Analysis and Risk Assessment (HARA) is recommended in ISO 26262. The entire process, however, is based on the very premise that a human driver is responsible for the safety of the vehicle. On the contrary, AVs function without any human intervention. Therefore, to ensure safe behavior in all possible situations, Dynamic Risk Assessment (DRA) at runtime is a necessity to make AVs aware of themselves about the current risk and take decisions accordingly; instead of relying on static worst-case assumptions. In this paper, we present a novel approach to identify and classify the severity and controllability rating class based on the measured data from the on-board sensors. Support Vector Machine (SVM) learning technique was used to train, test, and validate the model with diverse feature sets. We illustrate the presented approach by employing an example of adaptive cruise control and discuss the case study with initial findings.