Safety-driven data labelling platform to enable safe and responsible AI

Important results from the project

AI algorithms are data-driven techniques and therefore it is crucial that the safety specifications are encoded into the dataset that the AI algorithms are based upon. In this project, we aim to create such a safety-driven benchmark dataset that is ready-to-use for AI algorithm training and validation. The resulting data set contains sensor data logged in Sweden, together with ground truth labels encoded with safety specifications. This initiative aims to boost the effort of building reliable AI systems and contribute to the roadmap towards Sweden´s Vision Zero.

Expected long term effects

There are three work packages: 1) collecting data using the REVERE research vehicle; 2) developing safety encoded annotation guidelines; 3) automatically labelling the collected data. As the outcome, we have collected about 38 hours of perception data that amount to 23 TB, and we have created safety-driven annotations using our AI data platform SnapXS. The result will be used in several subsequent and related projects for applications such as road infrastructure analysis, energy efficient urban planning and data quality validation for automotive perception systems.

Approach and implementation

The project is implemented in four phases: 1) data collection, 2) data preparation and pre-annotation to enable random access and searchability, 3) annotation guideline development based on safety specifications, and 4) scalable automated labelling given the safety-driven annotation guidelines. The data is collected using the REVERE research vehicle. The data preparation and annotation steps are implemented using our AI data platform SnapXS. The end result is a set of safety-driven annotation guidelines and a ready-to-use annotated dataset for AI training and validation.