Use Cases
Use Case 1
Trustworthy Perception Systems for CCAM
UC-1 Trustworthy Perception Systems for CCAM
Rationale
Rationale
Perception components receive raw data from sensors (e.g., cameras, LIDAR, ultrasonics), and apply AI methods to extract objective information (e.g., presence, distance and classes of surrounding objects, driver status, etc.). Trusted AI is needed to understand which objects (classes) are perceived, how different sensor information is used (redundancy, fusion), and how possible discrepancies are solved.
Objectives
Objectives
Enable use in safety-critical applications where pedestrian detection is in the safety-critical path. Enabling the system to be used in an open context.
Demonstrator
Demonstrator
Reliable Pedestrian Detection in Urban Environments
Aim: Enable use in safety-critical applications where pedestrian detection is in the safety-critical path. Enabling the system to be used in an open context.
Approach towards trustworthy AI:
- Reliable pedestrian detection based on multiple sensor modalities (redundancy) and by using safety mechanisms. Time Coherent multisensor fusion
- Explainable Layers: Neural Network Intermediate Representations
- Visualization of conflicting perception and training on data with bad lighting/bad weather
Perception
– Trustworthy Perception Systems for CCAM
Trusted AI is needed to understand which objects are perceived? How different sensor information is used? How possible discrepancies are solved?
UC-1 Leader:
Partners involved:
Use Case 2
AI extended Situational Awareness / Understanding
UC-2 AI extended Situational Awareness / Understanding
Rationale
Rationale
Demonstrator 1
Demonstrator 1
Collision prediction with Hybrid AI data fusion models
Aim: Test robustness and trustworthiness in difficult perception situations such as occlusion and bad lighting/weather conditions, with specific predictive collision detection situations.
Approach towards trustworthy AI:
- Implementing data fusion for robustness against occlusions – handle continuity / time consistency (hybrid AI models of Work Package 3)
- Training on data with bad lighting/bad weather and other edge-cases (real and synthetic data from Work Package 2 activities)
Demonstrator 2
Demonstrator 2
Robust Prediction modules for Robo-taxi in urban environment
Aim: Trustworthy and robust prediction of traffic participant’s indented motion enabling a safe and predictable operation of robo-taxis.
Approach towards trustworthy AI:
- Optimal combination of AI and physics-based approaches enabling large-prediction horizons together with computational efficient implementations
- Validation for trustworthiness of AI through the usage of XiL testing methodologies together with test data collection and analysis
- Seamless transition from virtual and simulation environments to the embedded, real-time platforms using container-based development approach
- Safety SW platform capable of embedded real-time container execution and fault-tolerant decision-making subsystem concept
Understanding
– AI extended situation awareness and understanding
AI needs to understand communications, map information and perception layers. This includes understanding local dynamic maps (LDM) with accurate and complete knowledge about the scene.
UC-2 Demonstrator 1 Leader:
Partners involved:
Use Case 3
Trustworthy and Human understandable Decision-making
UC-3 Trustworthy and Human understandable Decision-making
Rationale
Rationale
Objectives
Objectives
Demonstrator
Demonstrator
Explainable and robust decision making (manoeuvre and trajectory)
Aim: Combine ML decision-making with human understandable definitions of traffic rules encoded in the HD Maps. Decisions are visualized (if possible) such that human occupants get the chance to understand them before they are executed by the AI.
Approach towards trustworthy AI:
- Develop fusion models for decision-making using perception, localisation, HD Maps and external information via V2X communication
- Improve situation awareness using Hybrid AI system (knowledge and data-driven AI)
- Human-aligned agent
Decision
– Trustworthy and human understandable decision making
AI path planning and manoeuvring execution should maximize safety, comfort, and eco-driving. The user understands why, when, and how a decision is taken.
Perception
– Trustworthy Perception Systems for CCAM
Trusted AI is needed to understand which objects are perceived? How different sensor information is used? How possible discrepancies are solved?
UC-3 Leader:
Partners involved:
Use Case 4
AI based Traffic management
UC-4 AI based Traffic management
Rationale
Rationale
Objectives
Objectives
Define a benchmark/reference for desired/acceptable behaviour and/or minimal driving performance expectations.
Analyse the response of AI-models to traffic management data, at strategic, tactical, and operational levels.
Demonstrator
Demonstrator
AI Models and Traffic Management
Aim: the capabilities of the AI models will be tested and evaluated in use cases particularly relevant for traffic management and road authorities. The decisions of AI models in response to the data are of special interest. To assess this, the behaviour of the vehicle will be benchmarked against what is considered good behaviour as defined by the rules of the road.
Approach towards trustworthy AI:
- Prepare data sets with real traffic management data
- Real-world data will serve as a starting point, for example from deployed traffic systems, data from national access points, weather data and databases of road operators
- Define good behaviour patterns according to road rules
- Create causality, transparency and predictability indicators
Traffic
– AI based traffic management
AI models are analysed in macroscopic scenarios with interaction with other AI and non-AI systems. Trustworthiness is gained by benchmarking “good behaviour” of AI models in transport level contexts.
UC-2 Leader:
Partners involved:
