Learning-based Dynamic Risk Indicators (LADRI) for Risk Prediction in Automated Driving Systems

As Automated Driving Systems revolutionize the intelligent transportation landscape, ensuring unparalleled safety becomes increasingly essential. Moreover, as automation levels rise, the responsibility for safety shifts from the driver to the system developers. This transition necessitates a reevaluation of safety assumptions and the creation of a risk assessment framework to assess system failures within an ADS. The challenges of continuous risk assessment and early hazard identification for ADS is multifaceted. Firstly, it is crucial to achieve comprehensive scenario coverage, as current standards often focus on a limited set of predefined conditions and may not encompass the wide range of potential real-world scenarios an ADS might encounter. Secondly, integrating complex system interactions is challenging because traditional approaches may not fully account for the detailed interactions among various ADS subsystems or effectively analyze the associated risks. Thirdly, integrating and analyzing diverse data streams is essential but challenging, and current frameworks may not adequately address this aspect. Lastly, validation under various operational conditions and continuous model improvement is vital, with standard validation processes potentially not accounting adequately for the wide variability in operational conditions faced by ADS and failing to specify measures for open traffic scenarios. The LeArning-based Dynamic Risk Indicators (LADRI) framework addresses these challenges, aiding safety engineers in early hazard detection for ADS development. It employs a Plan-Do-Train-Adjust-Assess cyclic process, enabling continuous improvement in risk assessment across diverse dynamic driving conditions. This approach leverages advanced learning algorithms and incorporates risk-specific context information from both the operational environment and the ADS itself, providing runtime contextual insights for the model to accurately predict severity and controllability indicators. This strategy bridges traditional gaps in risk assessment by allowing the evaluation of severity and controllability to adapt to different dynamic environments, instead of relying on subjective judgment. By monitoring the risk spectrum during driving operations, safety engineers can identify specific risk profiles, thereby enhancing safety mechanisms for subsequent iterations of risk assessment. Iteratively applying this framework allows developers to transition from unknown-unsafe regions to known-safe regions, thus progressively enhancing safety. A use case of Highway-Lane Following with Adaptive Cruise Control (ACC) functionality is utilized, creating extensive testing across a variety of driving scenarios, on different road shapes under diverse conditions, which has demonstrated the framework's robust capability for accurate risk prediction. The analysis underscores the importance of integrating time-based, distance-based, and impact-based risk features for a comprehensive risk assessment. These findings, supported by comprehensive model performance metrics and evaluations, position the LADRI framework as an advanced tool to enhance the risk assessment process.