Deep Learning Safety Concerns in Automated Driving Perception

Deep neural networks require specialized safety frameworks: The paper emphasizes that DNNs used in autonomous driving have "unique properties" that demand systematic safety approaches beyond traditional automotive standards, leading to the development of "safety concerns" as a structured way to evaluate these systems.

Safety concerns align with existing automotive standards: The framework is designed to work with established safety standards like ISO 21448 (SOTIF - Safety of the Intended Functionality), helping integrate AI safety into existing regulatory and development processes rather than creating entirely separate requirements.

Cross-functional collaboration is essential for AI safety: The paper stresses that addressing safety concerns in autonomous driving perception requires "cross-functional teams" to work together, suggesting that AI safety in vehicles cannot be solved by engineers alone but needs input from domain experts and safety specialists.

The field is rapidly evolving with new standards emerging: The safety concerns concept has already influenced upcoming AI safety standards such as ISO PAS 8800, indicating that this framework is becoming foundational to how the industry approaches autonomous vehicle safety governance.

Better categorization improves safety management: By refining and extending the safety concerns concept with improved categorization, the paper aims to make it easier for teams to systematically identify, understand, and address potential safety issues in DNN-based perception systems used in autonomous vehicles.