Αυτόματη εξαγωγή σεναρίων από δεδομένα οδήγησης αυτόνομων οχημάτων
Automated Driving Systems (ADS) require extensive evaluation to assure acceptable levels of safety before they can operate in real-world traffic. Although many tools are available to generate a virtual scenarios library for simulation testing, virtual testing cannot replace real-world testing. Hence a new applied research field has recently emerged on the subject of how to efficiently combine real-world testing (field tests, naturalistic driving data) and virtual testing in order to cover as many scenarios as possible.
Mixed reality test drive is a novel approach for combining the pros of reality and simulation. It uses real environmental conditions of real test drives and the accuracy and reproducibility of virtual test drive vehicle behavior descriptions (scenarios). As the scenario is a core element in AD behavior description and evaluation, automatic scenario extraction from massive volumes of recorded driving data is of key importance for modern autonomous vehicles’ testing. In this thesis we would like to i) investigate the feasibility of automatic scenario detection and classification based on logged driving data and using SoA machine learning algorithms but ii) taking into account scenarios’ classes and attributes from a formalized scenario representation (such as ASAM openScenario). This latter representation will render our results (i.e. data-driven generated scenarios) ready to be consumed from a simulation framework for future deployment of our method in a mixed reality testing framework. Exemplary scenario classes are “free driving”, “lane change”, “following a lead object”, “overtaking”. Nonetheless, different scenario decomposition could be also applied if the selected formalized representation requires it. A small set of extracted scenarios can be sufficient in order to create a small proof of concept of our proposed idea.
To facilitate the experiment set up and in order to enable fast starting point establishment, the following resources are available:
- Logged driving data (synchronized sensors + video) from motorways in the format used in H2020 EU project L3Pilot.
- Matlab scripts for data metrics derivation (already used in L3PIlot).
- Annotated set of driving scenarios from driving data (extracted manually by L3PIlot).
- Matlab tool for scenario annotation (already used in L3PIlot).
- Experience in data classification using several open source C++, Python libraries.
- Experience in smart vehicles’ perception system development and testing
- Theoretical background on time-series data analysis / video analysis.
- Experience with Matlab or other data analysis tool
- Some experience with a programming language such as Python, Java, C++. Desire to play around/experiment with open source SW for autonomous vehicle testing.
Optional: Any hands-on experience with data classification / computer vision will be considered an asset.
What you will learn: Familiarity with autonomous driving systems validation processes, hand-on experience with classification algorithms on recorded driving data, implementing and testing an automatic scenario extraction system with our help, experience on formalized driving scenario representation used in simulation environments.
 Driving scenarios describe the development of a situation within a traffic context in which at least one actor performs a (pre-) defined action and or is influenced by a (predefined) event. The action or event is specified without the definition of concrete parameters. The influenced actor may either be the ego vehicle (e.g. performing a lane change or a minimum risk maneuver) or another traffic participant (e.g. a lane change in front of the ego vehicle).
 The ASAM OpenSCENARIO 2.0 standard is meant to support the definition of tests and scenarios for the full development process of autonomous vehicles, and the full complexity of real-world scenarios, including complex inner-city traffic. Required use cases span from pure software-based simulation, through SIL, HIL, VIL hybrid testing models, up to test tracks and street driving. It will also ensure a migration path from ASAM OpenSCENARIO 1.x, with execution compatibility. The ASAM OpenSCENARIO 2.0 concept document has been published and is available on ASAM's website here.
Παναγιώτης Τσανάκας (firstname.lastname@example.org)
Αναστασία Μπολοβίνου (email@example.com)