Development Process Based on Standards for Safety-Critical AI Algorithms 

To ensure the safety and reliability of AI algorithms used in autonomous driving systems, a structured development process aligned with international standards such as ISO 21448 (SOTIF) and ISO/PAS 8800 is followed. This process adopts a V-model approach that integrates both AI system/component design and scenario-based dataset development.

Scenario-Based Study for Developing Accident Avoidance Logic for Vulnerable Road Users 

Accident data involving vehicles and vulnerable road users will be analyzed using traffic accident datasets. Based on statistical analysis, scenarios will be defined according to IGLAD accident types, and additional scenarios will be specified in alignment with international standardization bodies. These scenarios will then be implemented to generate datasets for the development of decision-making algorithms. 

Parallel Architecture for Real-Time Collision Prediction and Collision Avoidance Strategy Algorithm 

This architecture enables real-time collision prediction and collision avoidance strategies through a parallel AI framework. Sensor fusion and threat assessment modules serve as inputs to three parallel components:

• Collision Prediction: Classifies collision mode based on simplified BEV(SBEV) input and CNN-based reasoning.

• Collision Avoidance Strategies: Uses the same input to determine suitable avoidance strategies such as steering, evasive maneuvers, or lane changes.

• CP-Transformer: Employs a Transformer-based architecture to derive context-aware strategies, enhancing decision-making