Application Story
The AVEAS Projekt
An Open Approach towards the acquisition of safety-critical traffic scenarios for ADAS/AD validation: The AVEAS project
Summary
The AVEAS research project (www.aveas.org), funded by the German Federal Ministry for Economic Affairs and Climate Action (BMWK), has developed methods for acquiring real-world data for use in the virtual Validation of automated vehicles. Led by understandAI GmbH, it joined the partners dSPACE, Porsche Engineering, Continental, PTV, Allianz, GOTECH, Spiegel Institut, KIT and the three Fraunhofer institutes EMI, IVI and IOSB. Associate partners include ASAM, TÜV Süd, ADAC, ALP.Lab and EDI. Between 2021 and 2024, hundreds of hours of public traffic were collected by road vehicles, airplanes and infrastructure sensors and simulator studies with hundreds of participants were conducted, to establish data-driven simulations. The outcome: A unified set of files, as a connection of ASAM OpenLABEL, ASAM OpenDRIVE and ASAM OpenSCENARIO, representing the static and dynamic state of the acquired scenarios.
Initial situation
With growing complexity and criticality of automated driving functions in road traffic and their operational design domains (ODD), there is increasing demand for covering significant proportions of development, Validation, and Verification in virtual environments and through Simulation models.
If, however, simulations are meant not only to augment real-world experiments, but to replace them, quantitative approaches are required that measure to what degree and under which preconditions Simulation models adequately represent reality, and thus, using their results accordingly. Especially in R&D areas related to the safety impact of the “open world”, there is a significant shortage of real-world data to parameterize and/or validate simulations – especially with respect to the behavior of human traffic participants, whom automated driving functions will meet in mixed traffic. For this reason, prevailing Simulation models for human behavior are mostly based on manually defined models that are calibrated against high-level aggregated data, for example traffic count data for highways. Despite the good availability of so-called “microscopic”, agent-based behavior models for road traffic participants, the amount of data is by far insufficient to parameterize or validate detailed Simulation models that enable quantitative safety tests in virtual environments.
This is despite the fact that several large and well-known international campaigns to acquire such data have been conducted in the past, including datasets from the US Next Generation Simulation (NGSIM) project (US Highway 101, Interstate 80, Lankershim Blvd.), the computer vision-oriented datasets KITTI Vision Benchmark Suite, Cityscapes, BDD100K, nuScenes, ACDC and A2D2, the US Strategic Highway Research Program Naturalistic Driving datasets SHRP 2 NDS or the highD dataset from the PEGASUS project, as well as subsequent datasets such as inD and roundD. Similarly, a large number of test areas acquires comparable data continuously, but this data is often not stored or distributed, partly for reasons of privacy protection. While each of these data sources can, in whole or in part, contribute to an extensive body of data of real-world driving scenarios, the datasets are not only based on different acquisition methods, but also follow different acquisition and processing principles and data formats. Hence, their joint consideration is difficult and no research efforts are known that leverage a larger proportion of these datasets cumulatively. Furthermore, most of these datasets originate from temporary project enterprises. No systematic large-scale acquisition effort is known that aims to provide significant data for, e.g., data-driven virtual Validation of automated driving scenarios.
Success Stategy
The AVEAS project focused on developing a systematic approach to acquire real-world data under “relevant conditions” and transforming these into Simulation models for the virtual Validation of automated vehicle functions.
First, the results comprise a considerable range of methods for acquiring and processing real-world data with a focus on opening the potential for an economically viable, systematic, long-term data acquisition. These include methods for deriving ASAM OpenDRIVE maps from lightweight vehicle sensors, the reliable, privacy-safe detection and tracking of objects to achieve consistent ASAM OpenSCENARIO and ASAM OpenLABEL scenario data, and the derivation of data-driven models. All methods have been evaluated exhaustively to establish best practices for acquisition and processing, which have been published as peer-reviewed open access papers. During the three year project duration, real-world acquisitions at traffic accident hotspots in Germany have been conducted in the Dresden area via stationary long-wave (thermal) infrared cameras by Franhofer IVI, in Paderborn via road vehicles by dSPACE and UnderstandAI, as well as in the Karlsruhe area via road vehicles by Porsche Engineering and KIT, via static LiDAR sensors by KIT, and via light aircraft-mounted cameras by Fraunhofer IOSB.
Second, a joint Metadata Standard based on ASAM OpenLABEL has been developed as part of DIN SAE SPEC 91518, which combines the flexibility of ASAM OpenLABEL with detailed specifications developed by an international consortium that harmonizes definitions, scenario parameters, principles and units of Measurement, criticality metrics and corresponding measures for uncertainty across the heterogeneous acquisition methods (airborne, road vehicle-based and infrastructure) and heterogeneous applications, including virtual testing, traffic safety and traffic analysis. This data format is intended to unify and consolidate the international efforts in data acquisitions and establish a harmonized product interface for applicable data, to support the establishment of a consistent and compatible body of data. This harmonized data format is expected to minimize efforts for assuring consistency across different data sources as well as for data format conversion while assuring an increased data quality across heterogeneous sources.
Third, the project has acquired extensive data as an initial dataset to support the newly developed data format, part of which will be provided as open data as part of the project deliverables. This dataset will enable organizations to familiarize themselves with the format, as well as developing data-driven methods based on the data and evaluating their own methods’ performances.
Challenges During The Project
During its three-year period, the project faced various challenges ranging from technology over regulation up to force majeure. But a principal challenge, already foreseen during the initial preparation, was the unification of risk concepts, data and Simulation purposes and eventual testing applications. With the goal to develop a method to acquire data across highly different sensor platforms that can be use across highly different applications in different domains, cross-compatibility and harmonization of metrics, concepts, and approaches, was a key project focus throughout.
Most automated driving-related datasets available today introduce custom data formats and schemes tailored to the specific characteristics of the dataset and acquisition methods. When designing an approach to cover a wide range of methods to be used interchangeably, instead, the choice of formats and definitions that reflect any method alike is far from obvious. In this case, the need for harmonization covers all steps of the data processing pipeline, from a basic definition about how the “position” and “scale” of an object can be determined (e.g., with or without rearview mirrors of cars), how it is filtered and matched to the road network, up to a common definition of concepts like “lane change” or “time to collision”.
The establishment and further development of standardized formats is crucial to ensure sufficient interchangeability and reusability of input data across different Simulation frameworks that serve as the basis for XIL test benches and/or driving simulators in the development and Validation process.
Tille Karoline Rupp, Senior Manager Simulation at Porsche Engineering
Business Benefits
The AVEAS project results contribute to streamlined processes in traffic safety and the virtual / data-driven Validation of automated driving primarily in the three dimensions outlined priorly.
First, the developed range of methods for acquiring and processing real-world data has the express focus on opening the potential for an economically viable, systematic, long-term data acquisition. All methods have been evaluated exhaustively to establish best practices for acquisition and processing, which have been published as peer-reviewed open access papers.
Second, the joint Metadata Standard specified in DIN SAE SPEC 91518 is expected to minimize efforts for assuring consistency across different data sources as well as for data format conversion while assuring an increased data quality across heterogeneous sources.
Third, the AVEAS dataset is intended to enable organizations to familiarize themselves with the format, as well as developing data-driven methods based on the data and evaluating their own methods’ performances.
BACK TO OVERVIEW
Follow us on
