Scenario-Based Development and Validation Framework for Autonomous Driving Decision-Making

This diagram illustrates the integrated development and validation process of decision-making algorithms required for autonomous driving—such as collision risk assessment and avoidance strategy—using the scenario-based development framework METIS.
First, the scenario generation process applies the METIS methodology, which includes approaches aligned with SOTIF (Safety of the Intended Functionality) guidelines.
Second, the scenario database is continuously updated within the METIS framework and is actively utilized in this research.
Lastly, the study also explores data analytics and learning-based techniques to enhance the overall decision-making performance.

Collision Prediction Framework for Enhanced Autonomous Driving Safety

We developing collision prediction algorithms with the goal of enhancing the safety of autonomous vehicles. By fusing data from various sensors such as radar, camera, and corner radar, we accurately recognize surrounding vehicles and predict potential future collisions. We utilize visually abstracted driving data to determine both the occurrence and location of collisions, and adopt a temporal processing structure to ensure fast and reliable predictions even in complex traffic scenarios. By integrating physics-based risk estimation with learning-based algorithms, we have built a robust prediction framework that maintains high performance under real-world driving conditions. This research plays a critical role in improving the safety of ADAS and autonomous driving systems. 

Robust Collision Prediction via Learning method

We study methods to improve collision prediction by analyzing risk scenarios from simulation and real-world driving data. To handle data imbalance and noise, we apply an active-curriculum learning framework that combines active and curriculum learning. This helps the model gradually learn difficult cases and stay robust in real environments. We also analyze misjudgment cases to refine the training set and enhance system reliability. Our approach aligns with safety standards like ISO 21448 (SOTIF) and supports data-driven learning for autonomous driving. 

Real-Time Embedded Validation of Collision Prediction Algorithms

We utilize an experimental vehicle in our lab to collect data and embed algorithms such as LiDAR detection and collision prediction into the vehicle, processing real-time inputs from various sensors. Through this setup, we verify whether the algorithms operate reliably under real-time constraints, and comprehensively analyze factors such as computation time, response delay, and processing accuracy to evaluate their effectiveness and applicability in real-world environments.