Validation of a Critical Driving Scenario Identification Algorithm for Automated Driving Functions Using Real and Synthetic Object Lists

Validation of automated driving functions relies on maneuver-based testing and statistical evidence, which indicates a huge amount of data. To reduce data volume and optimize data efficiency, vehicle sensor data and environmental data acquisition should focus on relevant and critical driving situations, as long-tail scenarios are of greater value for autonomous driving validation and improvement. A possible solution is to use AI-based algorithms that enable on-board systems to identify relevant and critical scenarios automatically, live, and selectively record raw data and scenario descriptions during test drives.
On the one hand, this paper presents an iterative approach for developing AI-based algorithms to detect critical scenarios in road traffic. On the other hand, the paper introduces a universal validation method for criticality algorithms to validate automated driving functions.
The algorithm was pre-trained on datasets of critical traffic scenarios, generated from both real test drives and simulated data, and labeled using rule-based metrics such as Time-to-Collision. To prove the functionality, the algorithm is integrated into the on-board system of a test vehicle, where the algorithm is fed with object lists from real test drives as well as synthetic data from a Hardware-in-the-Loop simulation with the on-board system.
... mehr
The test results show that the algorithm fulfills the desired functionality with both real and synthetic object lists. In addition, the potential to test critical driving maneuvers in the simulation and thus minimize testing efforts will be demonstrated.
The approach enables AI-based selective data recording to align with the evolving boundaries of ADAS/AD systems as automated driving advances. The resulting algorithm increases the information density of the data used for developing and testing ADAS/AD systems while significantly reducing the data volume and data protection requirements.