Human Benchmark Dataset

Product

The 450M km AV Human Benchmark Dataset is a verification tool for assessing autonomous vehicle (AV) performance against human driving. Using real-world accident data, it highlights AV responses to complex scenarios and erratic traffic behavior, helping developers and regulators create test cases to ensure AVs meet human safety standards.

Dataset

Extensive Coverage: The dataset spans 450M km of real-world accidents, providing a solid foundation for AV verification.

Diverse Driving Environments: Includes urban and highway data for comprehensive traffic condition testing.

Regulatory-Grade Data: Originally developed for the Singapore government’s regulatory body, ensuring data quality suitable for policy and safety standards. The data is fully owned and controlled by Resembler, including annotation rights.

Expert-Led Annotations: Each incident includes two time-stamped instances for danger (with bounding boxes) five time-stamped instances for risk progression, and contextual tags, meticulously developed, reviewed and maintained by expert accidentologists to capture imminent risk factors.

Support and Integration

The 450M km AV Human Benchmark Dataset comes with comprehensive support to ensure successful migration, integration, and long-term utility:

Proprietary Abstractions and Temporal Risk Models: The dataset includes proprietary abstractions and risk models to test AVs in complex scenarios, capturing nuanced risk progression for deeper assessments.

Onsite Support for Migration & Integration: We provide onsite support to create custom scripts, ensuring seamless dataset integration and a smooth, tailored transition with your team.

Compatibility with Resembler Tools: The dataset works with Resembler’s ODD Assessment and Comparison tools (sold separately) for advanced analysis and AV safety verification.

Future Data and Updates: As part of the support package, we provide ongoing access to new incidents and accidents, ensuring continuous growth of the dataset. Our goal is to continuously expand the range and depth of scenarios available for testing.

Advanced Similarity Search with Resembler Icon Query Language (RIQL): The dataset supports advanced similarity search through the Resembler Icon Query Language (RIQL), enabling users to find incidents by danger level, tags, or risk cues. This aids in verifying algorithms like MARL and cue-based heuristics.