IFM - Institut für Fahrerassistenz und vernetzte Mobilität
Refine
Year of publication
Document Type
- conference proceeding (article) (91)
- Article (45)
- conference proceeding (presentation) (9)
- Preprint (7)
- Doctoral Thesis (2)
- Master's Thesis (2)
- Other (2)
- Book (1)
- conference proceeding (summary, abstract) (1)
Publication reviewed
- peer review (95)
- begutachtet (42)
- nicht begutachtet (22)
Keywords
- automated driving (14)
- advanced driver assistance systems (10)
- human factor (8)
- workload (8)
- SOTIF (7)
- Simulation (7)
- driving experience (7)
- Automated Driving (6)
- lane keeping system (6)
- stress (6)
Institute
- IFM - Institut für Fahrerassistenz und vernetzte Mobilität (160)
- Fakultät Maschinenbau (53)
- Fakultät Informatik (51)
- Fakultät Elektrotechnik (40)
- Promotionszentrum (23)
- IDF – Institut für Datenoptimierte Fertigung (3)
- IFA – Institut für Innovative Fahrzeugantriebe (2)
- IMS - Institut für Maschinelles Sehen (1)
- IPI – Institut für Produktion und Informatik (1)
Assuring an AI-based perception component for highly automated driving (HAD) requires evidence from ISO 26262 (functional safety), ISO 21448 (SOTIF, safety of the intended functionality), ISO/SAE 21434 (cybersecurity), and ISO/PAS 8800 (safety of AI in road vehicles). The standards cross-reference each other normatively at the level of requirements and evidence. No standard prescribes how the claims combine under a single top-level goal in Goal Structuring Notation (GSN), and the ISO/PAS 8800 Annex B pattern does not integrate cybersecurity claims. This paper extends the Annex B pattern with those cybersecurity claims, alongside SOTIF claims from ISO 21448 and functional safety claims from ISO 26262. The pattern is positioned at the AI component level inside the encompassing system safety case and instantiated for a LiDAR-based 3D object detection component. The contributions are (i) an integrated GSN argument pattern with clause-level traceability to the cited standards, (ii) a catalogue of decision points the engineer must resolve where the standards defer to application context, and (iii) a classification of integration-induced findings that emerge only between standard scopes, distinguished from open methodological problems. Combining the cited standards reveals an evidence asymmetry at the verification and validation goal: the evidence types are disjoint by design, the cybersecurity evidence type (attack success rate) adds a fourth measurement scale to the three safety evidence types, and no standard prescribes their combination into a single sufficiency claim.
Accurate road condition assessment requires the fusion of geometric and visual information from LiDAR and camera systems. In this work, we investigate the feasibility of multimodal registration between LiDAR-derived height and intensity maps and RGB image data for road surface analysis under real-world conditions.
Our experiments show that classical feature- and intensity-based registration methods are not suitable for this application. The homogeneous nature of road surfaces, the lack of stable cross-modal correspondences, and the limited structural quality of LiDAR intensity data prevent robust and reproducible alignment.
As a result, we rely on a time- and velocity-based alignment strategy that leverages sensor synchronization and vehicle motion. This approach provides a stable coarse registration but does not allow precise multimodal fusion and is sensitive to temporal offsets and dynamic measurement conditions.
These findings indicate that reliable multimodal fusion cannot currently be achieved using classical methods in real-world road scenarios. We therefore outline learning-based approaches, such as cross-modal feature learning and self-supervised methods, as promising directions for future work.
Safety is a fundamental requirement for the acceptance of modern vehicles. When considering autonomous driving functions, this factor is crucial for the spread of this technology in the market. Human behavior plays a key role in the analysis, as these systems must be adapted and optimized to meet human needs in order to achieve a high level of market penetration. Accordingly, effect chains must be known in detail to identify influencing factors and to be able to optimize them. This paper presents a theoretical model for perceived safety - analyzing the sub-criteria of humans, vehicles, and the environment also demonstrating and linking their influences on the basis of existing investigations. These findings are intended to contribute to the understanding of human behavior in order to better adapt systems to the needs of the users.
Machine Learning (ML)-based LiDAR 3D object detectors in automated driving produce false detections, missed detections, and localisation errors under adverse weather and reduced visibility. Detection errors arising without hardware or software faults constitute performance insufficiencies under ISO 21448, Safety of the Intended Functionality (SOTIF), and the standard requires identification of the triggering conditions responsible. The prescribed analysis methods assume a design specification, but ML-based LiDAR object detectors have no design specification because the mapping from point clouds to bounding boxes is learned from training data. This paper proposes an uncertainty evaluation methodology that uses disagreement among deep ensemble members to separate correct from incorrect detections. Ensemble disagreement and performance insufficiencies arise from insufficient training data coverage of the operating condition. The methodology evaluates whether three uncertainty indicators derived from ensemble disagreement (mean confidence, confidence variance, and geometric disagreement) separate correct from incorrect detections. The evaluation produces outputs mapped to ISO 21448 analysis activities: discrimination metrics, triggering condition rankings by false positive share, frames flagged for investigation, and acceptance gates reporting coverage and false acceptance rate. A case study using simulated ensemble predictions across 22 weather configurations shows that geometric disagreement achieves the strongest separation, with acceptance gates that retain only true detections at reduced coverage. The observed separation arises because false detections produce spatially inconsistent bounding boxes across ensemble members where no physical object constrains the predicted position, while true detections remain spatially consistent.
Uncertainty in LiDAR sensor-based object detection arises from environmental variability and sensor performance limitations. Representing these uncertainties is essential for ensuring the Safety of the Intended Functionality (SOTIF), which focuses on preventing hazards in automated driving scenarios. This paper presents a systematic approach to identifying, classifying, and representing uncertainties in LiDAR-based object detection within a SOTIF-related scenario. Dempster-Shafer Theory (DST) is employed to construct a Frame of Discernment (FoD) to represent detection outcomes. Conditional Basic Probability Assignments (BPAs) are applied based on dependencies among identified uncertainty sources. Yager's Rule of Combination is used to resolve conflicting evidence from multiple sources, providing a structured framework to evaluate uncertainties' effects on detection accuracy. The study applies variance-based sensitivity analysis (VBSA) to quantify and prioritize uncertainties, detailing their specific impact on detection performance.
Das Ziel dieser Arbeit ist die subjektive Bewertung des Einflusses von Schäden an Achsbauteilen auf die Fahrzeugbeherrschbarkeit. Ein besonderer Fokus liegt dabei zum einen auf der Entwicklung von Methoden, um Gesamtfahrzeugmodelle mit Achsbauteilschäden in Echtzeit berechnen zu können. Zum anderen wird ein dynamischer Fahrsimulator verwendet, um die Gesamtfahrzeugmodelle subjektiv zu erleben und deren Funktionalität zu validieren. Für die echtzeitfähige Berechnung von Gesamtfahrzeugmodellen können kennfeldbasierte Zweispurmodelle sowie adaptierte Mehrkörpermodelle verwendet werden. Beide Ansätze werden zunächst an einem dynamischen Fahrsimulator appliziert und das grundlegende fahrdynamische Verhalten mit dem eines Realfahrzeuges verglichen. Anschließend werden als erste Schadenskonfiguration Bauteildeformationen untersucht. Dabei werden verschiedene Ansätze entwickelt, um Bauteildeformationen im Echtzeitmodell zu integrieren. Darauf aufbauend wird mit einer umfangreichen Validierungsstudie untersucht, inwiefern sich die subjektive Wahrnehmung der Schadenskritikalität von deformierten Bauteilen im Realfahrzeug mit der Wahrnehmung im dynamischen Fahrsimulator deckt. Dabei kann für ein breites Spektrum an Bauteildeformationen absolute Verhaltensvalidität nachgewiesen werden. Danach wird untersucht, wie Risse in Bauteilen auf echtzeitfähige Gesamtfahrzeugmodelle übertragen werden können. Ein physikalisches Ersatzmodell, welches die Nichtlinearitäten aus dem Material, der Geometrie und dem Kontakt der Rissufer erfasst, ermöglicht die echtzeitfähige Berechnung angerissener Bauteile. Ebenfalls untersucht wird der vollständige Bauteilabriss. Da infolge eines Lenkerabrisses die Radebene im Allgemeinen nicht mehr statisch definiert ist, sind kennfeldbasierte Modelle für derartige Anwendungsfälle ungeeignet. Durch eine Optimierung des Mehrkörpermodells sowie des zugehörigen numerischen Integrationsverfahrens kann jedoch eine echtzeitfähige Berechnung von spontanen Lenkerabrissen realisiert werden. Bisher ist die Bewertung des Einflusses von Achsbauteilschäden auf die Fahrzeugbeherrschbarkeit auf subjektive Fahreindrücke und somit (virtuelle) Fahrversuche angewiesen. Daher wird abschließend untersucht, ob Korrelationen zwischen objektiven fahrphysikalischen Größen und den Subjektivbewertungen gefunden werden können. Diese können teilweise identifiziert werden und ermöglichen somit eine Prädiktion des Subjektiveindrucks basierend auf Gesamtfahrzeugsimulationen in Kombination mit den Korrelationsmodellen.
For the assessment of axle damages, real vehicle tests have mostly been used so far, but they are dangerous and difficult to reproduce. Therefore, driving simulators are becoming increasingly important for the virtual rating of vehicles. Regardless of whether a real vehicle or a driving simulator is used, the prediction of the subjective perception of axle damages requires time-consuming driving tests. A powerful dynamic driving simulator is used to obtain subjective evaluations of various axle damages. Objective vehicle quantities are logged simultaneously. Subsequently, multilinear regression (MLR) models and artificial neural networks (ANN) are used to identify correlations and predict subjective evaluations based on objective data. Furthermore, real-time capable vehicle models in CarMaker and multibody dynamic (MBD) models in ADAMS/Car are used to virtually carry out driving manoeuvres and generate synthetic data. By combining the simulated vehicle data with an ANN, subjective driver evaluations can be predicted entirely virtual.
Up to now, cracked axle components are subjectively examined in real vehicle tests. In order to save development costs and time, these tests should be carried out in a dynamic driving simulator. A necessary prerequisite for this is a real-time capable full vehicle model that correctly represents all crack-related non-linear effects. This publication develops and validates a purely virtual process chain to describe cracked axle components in real time. Rear axle tie rods with different crack lengths at the same position are considered as an example. First, the stiffness behaviour is simulated at the component level and compared with real tests. The determined stiffness curves serve as a basis for the parameterisation of a physical substitute model. The tie rods are modelled as a non-linear FE component, as a linear flexible body and via the physical substitute model. These modelling approaches are integrated and validated in a FE semi-axle model as well as in the MBS full vehicle model. It is shown that the physical substitute model provides very good results. Finally, the full vehicle model with a cracked tie rod is converted into a real-time model based on elastokinematic maps and validated based on the MBS model.
In the development process of passenger cars, various scopes are defined by subjective criteria, which have to be determined in road tests. Therefore driving simulators are increasingly used in order to improve the efficiency of the development process. In the context of vehicle dynamics, map-based models are used predominantly, as they are unconditionally real-time capable. Multi-body simulation (MBS) models have a higher complexity and are therefore more accurate. However, adherence to the real-time condition depends on the available computing power and model complexity. A main advantage of this approach is that no conversion into map-based models is required. As a result, spontaneous changes can be made to the full vehicle model, which significantly enhances the tuning process at the driving simulator. First, a theoretical comparison of the two simulation approaches is made. This shows that real-time MBS models deliver significantly better results than map models, especially at higher frequencies. Subsequently, expert drivers assess both, the full vehicle models (map-based model and real-time MBS model) on a dynamic driving simulator in direct comparison with the real vehicle on the proving ground. The vehicle's controllability and steering behaviour through lane change manoeuvres and sinusoidal steering are assessed. It turns out that the subjective assessments between the driving simulator and the real vehicle agree very well. There are hardly any differences between the two simulation approaches for the driving manoeuvres examined. As a result, the process for integrating vehicle dynamics models on a driving simulator can be significantly downsized by the use of realtime multi-body models without experiencing any loss of evaluation quality. At the same time, this opens up potential for assessing comfort issues on the driving simulator.
Steering feedback provides an essential contribution to the validity of driving simulators. Previous works have shown significant effects on steering feedback in dynamic driving simulators through steering and vehicle body excitations up to 100 Hz. Based on these findings, this work extends the discussion by addressing its implications for lateral driving performance. It presents the results from a meta-analysis of the data from three subject studies comparing participants’ lateral driving performance between a reference vehicle and different variants of its virtual representation in a dynamic driving simulator. The effects of different modifications of steering feedback through steering wheel and vehicle body excitations covering the frequency ranges of noise, vibration, and harshness are investigated. The results show a clear beneficial effect of rotational steering wheel vibrations up to 30 Hz whereas non-rotational excitations in higher frequency ranges yield less conclusive results. The influence of the participant selection and interactions between modifications is discussed. Recommendations for the use of high-fidelity driving simulators and the development of modern steering systems are developed.
