Impact of Automotive System Safety Design on Machine Learning Based Perception Systems
Vasu Singh, Mandar Pitale · 2021
Perception in autonomous driving is a prominent safety-critical application of machine learning (ML). Recent attention towards ML safety has contributed to a variety of guidelines for establishing assurance in ML based systems - however these guidelines are agnostic to automotive system safety design. This paper bridges the gap between top level safety decomposition and its impact on ML based perception systems. We study system design for different driving scenarios as examples. We then explore the relationship between the decomposition of safety goals to camera based perception and its impact on model learning and deployment aspects of the perception system, namely accuracy measures, ensemble methods, and distribution shift. We show how the safety decomposition plays an integral role in creating guidelines for ML design and deployment.