Source-linked AI summary
Ontology based Scene Creation for the Development of Automated Vehicles
Gerrit Bagschik, Till Menzel, Markus Maurer
TL;DR
The paper addresses the difficulty of identifying the many scenarios needed for automated-vehicle safety analysis and economical testing. It proposes ontology-based, knowledge-driven traffic-scene creation, producing systematically combined scenes while allowing engineers to validate the modeled knowledge base rather than review the entire catalog.
Problem
Highly automated driving requires many scenarios for safety analysis and testing, but expert-created scenarios are more creative than systematic and may not cover all possibilities.
Method
The paper models road layouts, infrastructure, traffic participants, interactions, and maneuvers in ontologies, then derives assertional traffic scenes from this formalized knowledge.
Results
1016 traffic scenes were generated for a three-lane motorway with three vehicles, while semantic duplication removal reduced the scene count by more than threefold compared with unadjusted permutations.
Takeaways & Limitations
Ontology-based creation provides broadly varied, traceable scenes and lets engineers check modeled knowledge and constraints instead of reviewing the resulting scene catalog.
Takeaways & Limitations
The scene catalog still requires valid abstraction concepts to support human-expert analysis such as hazard analysis and risk assessment.
Abstract
from arXiv · showhide
The introduction of automated vehicles without permanent human supervision demands a functional system description, including functional system boundaries and a comprehensive safety analysis. These inputs to the technical development can be identified and analyzed by a scenario-based approach. Furthermore, to establish an economical test and release process, a large number of scenarios must be identified to obtain meaningful test results. Experts are doing well to identify scenarios that are difficult to handle or unlikely to happen. However, experts are unlikely to identify all scenarios possible based on the knowledge they have on hand. Expert knowledge modeled for computer aided processing may help for the purpose of providing a wide range of scenarios. This contribution reviews ontologies as knowledge-based systems in the field of automated vehicles, and proposes a generation of traffic scenes in natural language as a basis for a scenario creation.
I. INTRODUCTION
Automated vehicles require systematic safety assessment across many operating scenarios, but expert-generated catalogs cannot guarantee completeness. The paper proposes knowledge-based systems and ontologies to support systematic scenario identification, safety analysis, and simulation-based testing.
- Safety assessment of automated driving functions is an emerging automotive-industry topic, while ISO 26262 defines a state-of-the-art functional-safety development process for series production.
- Automated vehicles must fulfill safe driving tasks across many operating scenarios, including scenarios where malfunctioning behavior could create hazards.
- Expert-generated scenario catalogs may represent critical situations, but cannot guarantee completeness across all possible combinations.
- Levels 3 and 4 require investigating and defining more scenarios than assisted or partially automated driving.
- A scenario-based process can shift test execution to simulation, but experts need a consolidated vocabulary and shared understanding of scenario organization.
- The proposed ontology-based approach models traffic infrastructure, participant behavior, and interactions in layers to support systematic scenario creation.
II. ONTOLOGIES
Ontologies formally represent shared conceptualizations through concepts, axioms, properties, instances, and facts. Their logical representation enables human-readable modeling and computer-based inference over domain and situational knowledge.
- Ontologies are formal, explicit specifications of shared conceptualizations intended to represent an abstract view of the world for a purpose.
- The ontology architecture combines terminological and assertional boxes into a knowledge base.
- Terminological boxes describe domain concepts through hierarchical classes, axioms, and properties, while assertional boxes represent class instances and observed situational facts.
- Ontologies express domain concepts in description logic, translating them into first-order logic for computer processing.
- OWL is a W3C-standardized declarative ontology language whose expressivity increases from OWL Lite through OWL DL to OWL Full.
- Reasoners infer knowledge, identify missing hierarchical concepts, check conflicts and consistency, and associate situational assertions with terminological concepts.
III. RELATED WORK
Prior automated-driving research uses ontologies mainly for situation assessment, behavior planning, environmental modeling, terminology, and simulation orchestration. The paper distinguishes its work by deriving initial scenes from modeled knowledge rather than interpreting sensor data.
- Ontologies have been applied in automated driving mainly to situation assessment and behavior planning.
- Scene-understanding approaches use sensor data, object tracks, maps, host-vehicle state, and spatio-temporal relations to infer traffic interactions.
- Ulbrich et al. used hierarchical classes and entity relations to derive an environmental model updated by sensor data for online vehicle decisions.
- Geyer et al. proposed an ontology ordering terms such as ego vehicle, scene, situation, scenario, mission, and route for unified use-case and scenario representation.
- Xiong’s scenario-orchestration framework combines an ontology, virtual drivers, supporting modules, and a scenario observer for simulation test cases.
- The paper reports no prior contribution using ontologies for initial scene creation from modeled terminological knowledge.
IV. ONTOLOGY-BASED SCENE CREATION
The paper motivates ontology-based scene creation as a semantic alternative to database combinations, aiming to generate meaningful initial scenes systematically. It frames scenes as the starting point of temporal scenario sequences and uses layered knowledge to organize traffic information.
- A scenario consists of a temporal sequence of scenes beginning with an initial scene, and the paper focuses on creating those initial scenes.
- Database-based scene generation produced physically impossible or unreasonable combinations because data entries lacked explicit semantics.
- The earlier database process was developed for a limited unmanned protective-vehicle use case and did not scale to wider use cases or more complex scenes.
- Explicitly representing assumptions and knowledge in a knowledge base supports more accurate and useful scene combinations than implicit exception handling.
- The proposed process uses semantically enriched information and ontology-based knowledge to combine meaningful real-world traffic scenes.
A. Knowledge acquisition
Knowledge acquisition combines traffic-domain guidelines, the system’s functional description, and expert knowledge to define relevant, traceable scene elements and relations.
- Knowledge acquisition: Traffic guidelines provide knowledge about how infrastructure is organized, named, and related across motorway, rural-road, urban-road, and safety domains.The guidelines are periodically updated to reflect new regulations.
- Knowledge acquisition: The system’s functional description restricts the knowledge base to scenes relevant to the intended operating range.For a vehicle operating only on the hard shoulder, scenes without a hard shoulder are unnecessary for evaluation.
- Knowledge acquisition: Functional-scope decisions about irrelevant entities and relations should be represented explicitly to preserve traceability of analysis assumptions.The paper links explicit representation to safety arguments and development traceability.
- Knowledge acquisition: Expert knowledge supplies dependencies and interactions not fully specified by individual traffic guidelines.Examples include lane-count dependencies for motorway speed limits and real-world relations among infrastructure components.
B. Layered model for knowledge representation
The proposed representation adapts a five-layer driving-scene model to separate road layout, infrastructure, temporary changes, traffic participants, maneuvers, and environmental effects.
- Layered model: Layer 1 represents road geometry and topology, while Layer 2 adds traffic infrastructure and models traffic rules as top-level semantic classes.Geometry can use straights, curves, and clothoids; infrastructure instantiates rules through signs and markings.
- Layered model: Layer 3 represents temporary infrastructure manipulations, such as construction sites, and records their changes to the original layout.The stated temporal frame for these manipulations is one day.
- Layered model: Layer 4 models stationary and movable non-infrastructure objects, including categorized traffic participants and disjoint atomic maneuvers.The maneuver set includes driving, following, approaching, passing, lane changing, turning, turning back, and safe stopping.
- Layered model: Maneuvers are described through relations, semantic rules, and parameters such as distance, relative speed, time-to-collision, and time-headway.The approach extends the maneuver set with fall back as the opposite of approach.
- Layered model: Layer 5 models environmental effects that alter infrastructure and traffic-participant interactions through different parameter ranges.Weather conditions are included as environmental influences on scene behavior.
C. Process for creation of traffic scenes
Scene creation derives assertional boxes from an ontology by generating infrastructure, inferring arrangements, adding traffic participants, and applying semantic rules to infer behaviors.
- C. Process for creation of traffic scenes: The ontology represents entities in natural language, assigns their formal order, and connects each entity to parameter relations in physical state space.Data properties are proposed for modeling these parameter relations.
- C. Process for creation of traffic scenes: The process aims to derive all possible assertional boxes from the ontology rather than only infer knowledge for observed scenes.This distinguishes the proposed creation process from related approaches focused on augmenting observed assertional boxes.
- C. Process for creation of traffic scenes: Creation begins with Layer 1 and Layer 2 concepts, whose mandatory, optional, relational, and cardinality constraints define possible road layouts.The example RQ 31 layout contains two lanes and a hard shoulder, with optional barriers, embankments, and speed limitations.
- C. Process for creation of traffic scenes: The first generated infrastructure scenes encode which elements exist, after which logic reasoning infers direct-neighbor arrangements such as left, right, in front of, and behind.These directional properties are reflexive, so one inferred relation yields its converse.
- C. Process for creation of traffic scenes: The first evaluation does not implement the third layer of the layered model.Temporary infrastructure manipulations are therefore outside this evaluation stage.
- C. Process for creation of traffic scenes: For Layer 4, configurable positions are distributed across lanes, participants are permuted, and semantic-web rules infer maneuvers from their relative positioning.Semantic repetitions are eliminated when participants belong to the same semantic class; weather setups are permuted at the end.
V. PRELIMINARY RESULTS AND DISCUSSION
The evaluation demonstrates large-scale generation of natural-language traffic scenes from a motorway ontology, while discussion emphasizes reduced knowledge-base complexity and remaining real-world coverage limits.
- V. PRELIMINARY RESULTS AND DISCUSSION: 284 ontology classes, 762 logical axioms, and 75 semantic web rules generate scenes for German motorways.The number of generated scenes depends on positions per lane and traffic participants.
- V. PRELIMINARY RESULTS AND DISCUSSION: 1016 traffic scenes result from a three-lane motorway setup with three vehicles, lane positions, maneuver combinations, and good weather.Removing semantic duplications prevents the scene count from becoming more than three times larger.
- V. PRELIMINARY RESULTS AND DISCUSSION: 652 infrastructure setups are generated for two-, three-, and four-lane motorways with different traffic rules, before traffic-participant combinations.The paper uses this count to indicate the scale of combining multiple representation layers.
- V. PRELIMINARY RESULTS AND DISCUSSION: Natural-language scene descriptions provide an overview of infrastructure, vehicles, and maneuvers for generated scenes.The figures use red squares for vehicles, green squares for infrastructure positions, and gray squares for scenes.
- V. PRELIMINARY RESULTS AND DISCUSSION: The ontology reduces knowledge-base complexity relative to the resulting scene catalog and lets analysts inspect modeled constraints and rules directly.Correct ontology modeling is stated to support correct combination of elements across layers.
- V. PRELIMINARY RESULTS AND DISCUSSION: The approach does not account for vehicle dimensions in lane positioning and cannot capture every infrastructure setup allowed by real-world data.The authors propose complementing ontology-based generation with real-world data analysis.
VI. CONCLUSION AND FUTURE WORK
The paper proposes ontology-based knowledge representation for automated vehicle scene creation, deriving possible observations from modeled knowledge. Future work targets simulation-format transformation and abstractions for analyzing the resulting scene catalog.
- The approach derives possible traffic-scene observations from modeled ontological knowledge for automated vehicles.
- Automated creation from formalized knowledge reduces validation and verification to reviewing the modeled knowledge base rather than the resulting scene catalog.
- Future work will transform natural-language scenes into OpenScenario and OpenDrive formats and identify valid abstractions for scene-catalog analysis.