Safe Deployment of AI and ML based Software and Algorithms in ADAS Systems

As the autonomy of ADAS features are moving from SAE level 0 autonomy to SAE level 5 autonomy of operation, reliance on AI/ML based algorithms in ADAS critical functions like perception, fusion and path planning are increasing predominantly. AI/ML based algorithms offer exceptional performance of the ADAS features, at the same time these advanced algorithms also bring in safety challenges as well. This paper explores the functional safety aspects of AI/ ML based systems in ADAS functions like perception, object fusion and path planning, by discussing the safety requirements development for AI/ML systems, dataset safety life cycle, verification and validation of AI systems, and safety analysis used for AI systems. Among all the safety aspects listed above, emphasis is put on dataset safety lifecycle as that is not only the most important element for training ML based algorithms for ADAS usage, but also the most cumbersome and expensive. The safety characteristics associated with dataset lifecycle are dataset safety analysis, dataset requirements development, dataset design and implementation, dataset verification and validation and then finally dataset maintenance. All these dataset life cycle characteristics are discussed in detail. Holistically, this paper outlines the process flow on what is needed from safety point of view to evaluate AI/ML based systems to claim the vehicles with advanced AI/ML systems are free from unreasonable risk. Also, considering perception system as an example, Key Performance Indicators (KPI) from safety perspective are defined to explain the acceptance and rejection criteria of the AI/ML based perception system.