Updated state-of-the-art:
During the execution of the European H2020 programme (2014-2022), many projects brought together stakeholders of the automotive industry for relevant safety and validation thematic areas. Many of them included particular developments or focus actions, but in general, they all contributed to the establishment of baseline know-how and de-facto standards in testing and validations methods for autonomous driving. To name a few, advances in data sharing and AI development were proposed and explored in ADAS&ME, AutoMate, AUTOPILOT, AVENUE, SHOW, BRAVE, CARAMEL, CoEXist, Dream4Cars, ENABLE-S3, ENSEMBLE, FABULOS, HEADSTART, ICT4CART, INFRAMIX, InLane, L3Pilot, MAVEn, Interact, SLAIN, TransAID, VI-DAS, 5G-Croco or 5G-MOBIX.
In parallel, at European level, other public and private initiatives have started to consolidate the research outcomes of these projects, including the EU expert group on Connected, Cooperative and Automated Mobility (CCAM), which, among other topics, aims to provide a Common Evaluation Methodology (CEM) and Data Sharing Framework (DSF) for Connected Automated Driving (CAD). Recently finished and ongoing projects are now pivoting around these concepts, such as the ARCADE Coordination and Support Action, which developed an online Knowledge Base (KB) gathering up to date information of CCAM related R&I projects, testing and piloting activities in Europe, regulation and policies, standards and impact assessment and data sharing methodologies. Currently, the FAME project aims to take over the results of ARCADE, and give continuity to the published KB. Focusing on scenario-based testing approaches, the Horizon Europe SUNRISE project aims to provide the first European-level Federated Scenario database for CAD testing.
Considering scientific initiatives, the most relevant contributions on AI development and testing are about the publication or massive multi-sensor datasets that can be used to train and validate models for environment perception (cameras, LIDAR, RADAR), and general scene understanding (e.g. maneuver analysis). In the last few years (2020-2022), it is noteworthy that large private companies are releasing professional-level datasets open for research purposes, such as the 3DHD CityScenes (Volkswagen), NuPlan (extension of NuScenes), MIT-AVT, Waymo Motion, Uber Pit30M, Ford Autonomous Vehicle Dataset, Drive&Act, DMD (Driver Monitoring Dataset), etc.
In the railway sector, two main datasets were recently released: FRSign, a dataset for traffic light detection and recognition, was published in 2020 as part of the TAS (Safe Autonomous Land Transport) project; Railsem19, a dataset for semantic rail scene segmentation, was published in 2019 on the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW).
Since AI is the building block of most advances in CAD, there are also important trends on AI certification under the scope of Europe's AI strategy. Since its publication in 2018, several aspects of AI are being analyzed and recommendations are produced, such as the EU Open Data Directive, European Data Governance Act, EU AI Act, etc. Bodies such as the European AI Alliance have produced Ethics guidelines for trustworthy AI (2019), with operational tool assessment list for trustworthy AI (ALTAI, 2020), and general definition of best practices for AI practitioners.
