Fakultät Informatik
In recent years we have seen large language models (LLMs) demonstrating robust reasoning capabilities comparable to human performance. This makes them increasingly appealing for driver assistance, where adaptation to dynamic human context is essential. Yet, research in this area remains fragmented, often focusing on isolated applications, lacking utilization of LLM’s full potential to deliver integrated, context-specific support and action. This survey synthesizes recent advancements in LLM-driven occupant monitoring systems, focusing on their capabilities for interpreting driver states and acting appropriately, enabling a new generation of intelligent driver assistance. We critically examine pioneering frameworks, benchmarks, and foundational datasets that employ techniques like reasoning chains, multimodality, and human-in-the-loop feedback to create personalized and safe driving experiences. We lay out the current trends, limitations, and emerging patterns, in addition to a novel human-centered evaluation of the field, providing researchers with a roadmap towards transparent and trustworthy in-cabin systems that bridge safety with driver experience.
In recent years we have seen Large Language Models (LLMs) demonstrating robust reasoning capabilities comparable to human performance. This makes them increasingly appealing for driver assistance, where adaptation to dynamic human context is essential. Yet, research in this area remains fragmented, often focusing on isolated applications, lacking utilization of LLM's full potential to deliver integrated, context-specific support and action. This survey synthesizes recent advancements in LLM-driven occupant monitoring systems, focusing on their capabilities for interpreting driver states and acting appropriately, enabling a new generation of intelligent driver assistance. We critically examine pioneering frameworks, benchmarks, and foundational datasets that employ techniques like reasoning chains, multimodality, and human-in-the-loop feedback to create personalized and safe driving experiences. We lay out the current trends, limitations, emerging patterns, in addition to a novel human-centered evaluation of the field, providing researchers with a roadmap towards transparent and trustworthy in-cabin systems, that bridge safety with driver experience.
Das Vorhaben 5GoIng hatte das Ziel, 5G-basierte Anwendungen in den Bereichen Verkehrssicherheit, Nachhaltigkeit und Smart City zu demonstrieren, ein 5G Testfeld im Digitalen Testfeld "Erste Meile" in Ingolstadt aufzubauen, sowie Aspekte der Technologieakzeptanz, interkommunale Zusammenarbeit und die Förderung der Startup-Infrastruktur zu adressieren. Das 5G-Testfeld dient durch die Nutzung öffentlicher Straßenabschnitte als Schaufenster für "Vernetzte Mobilität" und eröffnet die Gelegenheit, den Nutzen der 5G-Technologien direkt zu erleben. 5GoIng adressiert sowohl die drei technischen Schwerpunkte Smart City, Verkehrssicherheit und Nachhaltigkeit, als auch die gesellschaftsrelevanten Kernpunkte Akzeptanz, interkommunale Zusammenarbeit und Förderung der Startup-Infrastruktur.
Making AI explainable requires more than algorithmic transparency: it demands understanding who needs explanations and why. In our sixth CHI workshop on Human-Centered XAI (HCXAI), we shift focus to agentic AI systems. LLM-based agents foundationally challenge existing explainability paradigms. Unlike traditional AI that produces single outputs, agents plan multi-step strategies, invoke tools with real-world consequences, and coordinate with other systems; yet current XAI approaches fail to address these complexities. Users need to understand not just what an agent might do, but the cascade of actions it could trigger, the risks involved, and why responses take time. Even our expanded HCXAI frameworks struggle with these new demands. Through our workshop series, we have built a strong community making important conceptual, methodological, and technical impact. This year, we re-examine what human-centered explainable AI means in the agentic era, bringing together researchers and practitioners to shape explainability for both users and developers of these systems.
The OPTIMAL theory of motor learning, from its introduction to its subsequent refinement, has catalyzed a substantial body of research into motivational effects on motor learning with both supportive evidence and critical debate. This paper examines the effects of goal-directed practice, provided either through autonomy-supportive practice conditions—hypothesized by the OPTIMAL theory to yield motivational benefits—or a yoked group or a low-autonomy instructor, on implicit motor sequence learning of a complex, bimanual dual task. Participants practiced a motor sequence in a virtual reality serial reaction time (SRT) task and were either given or denied control over task difficulty as a task-relevant choice. While all groups successfully acquired the target sequence, differences between groups were negligibly small or absent altogether. These results suggest that the motivational effects of autonomy support do not substantially impact the motor learning of complex tasks.
This paper presents Zwitscherkasten, a DiY, multimodal system for bird species monitoring using audio and visual data on edge devices. Deep learning models for bioacoustic and image-based classification are deployed on resource-constrained hardware, enabling real-time, non-invasive monitoring. An acoustic activity detector reduces energy consumption, while visual recognition is performed using fine-grained detection and classification pipelines. Results show that accurate bird species identification is feasible on embedded platforms, supporting scalable biodiversity monitoring and citizen science applications.
Objectives
Driving under the influence of alcohol (DUI) remains a major contributor to fatal traffic crashes worldwide. With increasing regulatory pressure, such as requirements by Euro NCAP for in-vehicle impairment detection, there is a growing need for reliable, real-time monitoring solutions. While traditional DUI detection approaches focus on driving behavior or eye movement analysis, this study explores the potential of thermal imaging as a noninvasive alternative for detecting alcohol impairment.
Methods
We conducted a large-scale experimental study with 120 participants in a high-fidelity driving simulator, capturing thermal facial data under both non-impaired and alcohol-impaired conditions. A novel temperature extraction method was developed based on facial landmarks, incorporating multiple frames to reduce noise and improve temporal stability. Ambient cabin temperature was also recorded to normalize facial temperature readings and control for environmental influences. Several machine learning classifiers, including Logistic Regression, Random Forest, Support Vector Machine (SVM), and Gradient-Boosting Models, were trained using five facial temperature features (cheek, temple, ear, forehead, and nasal tip) and evaluated via five-fold subject-wise cross-validation.
Results
Significant temperature changes were observed in specific facial regions (particularly the cheek, ear, temple, and nasal tip) under alcohol influence. Among the evaluated models, Logistic Regression achieved the highest average classification accuracy (62%), while SVM demonstrated the most stable performance across folds. The model showed a slight conservative bias toward predicting the baseline (non-impaired) class, thereby reducing the risk of false positive classifications. Environmental conditions, including cabin temperature, were verified to be stable across both driving sessions, ensuring the validity of the physiological measurements.
Conclusions
This study demonstrates the feasibility of using thermal imaging for in-vehicle DUI detection under realistic conditions. Our contributions include the development of a robust facial temperature processing pipeline, the creation of a unique dataset collected under ecologically valid conditions, and a comprehensive comparison of seven state-of-the-art classification models. Thermal imaging represents a promising complementary modality for future driver monitoring systems focused on safety and impairment detection.
Artificial intelligence (AI)-driven clinical decision support systems (CDSS) hold promise to improve diagnostic accuracy and efficiency in computational pathology. However, collaboration between human experts and AI may give rise to cognitive biases, such as automation and anchoring bias, wherein users may be inclined to blindly adopt system recommendations or be disproportionately influenced by the presence of AI predictions, even when they are inaccurate. These biases may be exacerbated under time pressure, pervasive in routine pathology diagnostics, or shaped by individual user characteristics. To investigate these effects, we conducted a web-based experiment in which trained pathology experts (n = 28) estimated tumor cell percentages twice: once independently and once with the aid of an AI. A subset of the estimates in each condition was performed under time constraints. Our findings indicate that AI integration generally enhances diagnostic performance. However, it also introduced a 7% automation bias rate, quantified as the number of accepted negative consultations, where a previously correct independent assessment gets overturned by inaccurate AI guidance. While time pressure did not increase the frequency of automation bias occurrence, it appeared to intensify its severity, as evidenced by a performance decline linked to increased automation reliance under cognitive load. A linear mixed-effects model (LMM) analysis, simulating weighted averaging, revealed a statistically significant positive coefficient for AI advice, indicating a moderate degree of anchoring on system output. This effect was further intensified under time pressure, suggesting that anchoring bias may become more pronounced when cognitive resources are limited. A secondary LMM evaluation assessing automation reliance, used as a proxy for both automation and anchoring bias, demonstrated that professional experience and self-efficacy were associated with reduced dependence on system support, whereas higher confidence during AI-assisted decision-making was linked to increased automation reliance. Together, these findings underscore the dual nature of AI integration in clinical workflows, offering performance benefits while also introducing risks of cognitive bias–driven diagnostic errors. As an initial investigation focused on a single medical specialty and diagnostic task, this study aims to lay the groundwork for future research to explore these phenomena across diverse clinical contexts, ultimately supporting the establishment of appropriate reliance on automated systems and the safe, effective integration of human–AI collaboration in medical decision-making.
We propose GMP3, a multiphase global path planning framework that generates dynamically feasible three-dimensional trajectories for unmanned aerial vehicles (UAVs) operating in cluttered environments. The framework extends traditional path planning from Euclidean position spaces to the Lie group SE(3), allowing joint learning of translational motion and rotational dynamics. A modified Bellman-based operator is introduced to support reinforcement learning (RL) policy updates while leveraging prior trajectory information for improved convergence. GMP3 is designed as a distributed framework in which agents influence each other and share policy information along the trajectory: each agent refines its assigned segment and shares with its neighbors via a consensus-based scheme, enabling cooperative policy updates and convergence toward a path shaped globally even under kinematic constraints.We also propose DroneManager, a modular ground control software that interfaces the planner with real UAVs platforms via the Micro Air Vehicle Link (MAVLink) protocol, supporting real-time deployment and feedback. Simulation studies and indoor flight experiments validate the effectiveness of the proposed method in constrained 3D environments, demonstrating reliable obstacle avoidance and smooth, feasible trajectories across both position and orientation.
SELFY envisions an agnostic toolbox for the self-management of security and resilience of the CCAM (Connected, Cooperative and Automated Mobility) ecosystem, which can be easily deployed to extend the current Operational Design Domain (ODD), providing self-awareness, self-resilience and self-healing mechanisms and enhancing trust between stakeholders. SELFY is based on four pillars: Situational awareness, Resilience, Secure Data Sharing and Trust and provides three groups of tools. SACP (Situational Awareness and Collaborative Perception) tools aim at providing all CCAM actors with a comprehensive understanding of their environment, i.e., the perception of objects, such as other traffic participants and stationary objects. CRHS (Cooperative Resilience and Healing System) tools enable self-protection actions whenever a compromising situation is detected in relation to assets, vehicles, operations, or the system itself. TDMS (Trust and Data Management System) tools establish a secure and trusted environment for data in a collaborative and cooperative context, both for infrastructure and assets, as well as for citizen’s data, such as drivers or pedestrians with special attention to privacy considerations. By defining a collaborative environment between the different tools to respond to new threats, risks and attacks SELFY facilitates the comprehension of new challenges in the cybersecurity aspect of CCAMs.
Toward Robust Agents: A Survey of Adversarial Attacks and Defenses in Deep Reinforcement Learning
(2026)
Deep Reinforcement Learning (DRL) has demonstrated remarkable success in autonomous decision-making across diverse domains, including robotics, autonomous driving, and game playing. However, recent studies have uncovered a critical vulnerability: DRL agents are highly susceptible to adversarial attacks that can significantly degrade their performance or lead to catastrophic failure. These attacks exploit different components of the learning pipeline observations, actions, rewards, and policies exposing new challenges unique to DRL compared to supervised learning. This survey provides a comprehensive examination of adversarial threats and corresponding defense mechanisms within the DRL paradigm. This also aims to serve as a foundational reference for researchers and practitioners seeking to understand and mitigate adversarial vulnerabilities in DRL.
The performance of deep learning models is known to scale with data quantity and diversity. In pathology, as in many other medical imaging domains, the availability of labeled images for a specific task is often limited. Self-supervised learning techniques have enabled the use of vast amounts of unlabeled data to train large-scale neural networks, i.e., foundation models, that can address the limited data problem by providing semantically rich feature vectors that can generalize well to new tasks with minimal training effort increasing model performance and robustness. In this work, we investigate the use of foundation models for mitotic figure classification. The mitotic count, which can be derived from this classification task, is an independent prognostic marker for specific tumors and part of certain tumor grading systems. In particular, we investigate the data scaling laws on multiple current foundation models and evaluate their robustness to unseen tumor domains. Next to the commonly used linear probing paradigm, we also adapt the models using low-rank adaptation (LoRA) of their attention mechanisms. We compare all models against end-to-end-trained baselines, both CNNs and Vision Transformers. Our results demonstrate that LoRA-adapted foundation models provide superior performance to those adapted with standard linear probing, reaching performance levels close to 100 % data availability with only 10 % of training data. Furthermore, LoRA-adaptation of the most recent foundation models almost closes the out-of-domain performance gap when evaluated on unseen tumor domains. However, full fine-tuning of traditional architectures still yields competitive performance.
Enhancing UX in Automated Vehicles through Biophilic Interfaces: Insights from Prospective End Users
(2025)
Maneuver coordination enables connected and automated vehicles (CAVs) to collaboratively plan, negotiate, and execute driving maneuvers, aiming to enhance safety, traffic flow, and energy efficiency. As part of the Vehicle-to-Everything (V2X) communication system, it operates as a dedicated service that relies on detailed, bidirectional interactions between vehicles, unlike traditional broadcast-based V2X services. While maneuver coordination is the focus of ongoing research and standardization, the rules that govern when and how vehicles generate Maneuver Coordination Messages (MCMs) within the Maneuver Coordination Service (MCS) are still not fully defined. These rules are essential for ensuring timely and reliable coordination. This paper proposes three adaptive MCM generation strategies based on the operation mode of the MCS, maneuver priority, and current channel load. The first approach defines MCM rules for intent sharing, maneuver negotiation, and execution phases, factoring in maneuver importance. The second adjusts the frequency of intent MCMs when a negotiation message is received from another vehicle. The third adapts frequency based on real-time channel load during intent sharing. For comparison, two baseline strategies are also evaluated. The proposed rules are tested for highway merging and lane change use cases in traffic scenarios with increasing vehicle density and high channel loads. A comprehensive evaluation is conducted using metrics related to the network, message generation frequency, congestion control, and maneuver coordination. The results demonstrate significantly enhanced channel efficiency and communication reliability across all metrics. For example, under the final approach, maneuver negotiation time is reduced by a factor of five in congested V2X environments.
In the context of Connected, Cooperative, and Automated Mobility (CCAM), precise ego-vehicle positioning and environmental status assessment are crucial. However, these tasks can be susceptible to sensor failures, misuse, and cyberattacks. Automation disengagements and system redundancy are common strategies to achieve Minimum Risk Conditions when failures occur. This paper presents a Fail-Safe decision architecture formulated within the framework of the SELFY project (https://selfy-project.eu/). The main aim is to reduce inaccuracies in GNSS-derived positioning through the incorporation of sensor fusion, AI-guided situational assessment, trajectory planning, and mode decision components. Additionally, the architecture has been designed to enable real-time updates and communication with external entities, including the Vehicle Security Operations Centre.
Assessing the quality of audio-visual (AV) content is essential for understanding how users perceive the overall experience of combined audio and video in the modern multimedia applications. Even though there has been a lot of work in assessing the quality of audio and video separately, the integration of these channels has not been well explored especially using advanced deep learning techniques. Current literature tends to use non-industry-standard datasets and simple fusion strategies, thus hindering the real-world relevance and advancement of the field. To address these gaps, this work conducts a series of experiments starting from simple approaches and baseline models, progressively advancing to more sophisticated methods such as integrating cross-modality fusion via cross and self-attention.
Our method incorporates a deep neural network based audio model along with a dedicated video quality model where the features of both modalities are fused using attention mechanisms. We extend the feature set of one modality when needed in order to preserve temporal correspondence and relevance. One of the novelties of the proposed approach is the usage of the Concordance Correlation Coefficient (CCC) loss, which is a measure that has been employed in emotion recognition but not in AV quality modeling before. This loss function increases the stability of our quality estimations since it forces the model to output quality scores that are closer to the actual human ratings.
We also make a unique contribution by using an AV dataset that is closer to the industry practices, which contains high quality audio and video content with realistic distortions. To the best of our knowledge, this is the first work that employs cross-attention for AV feature fusion in an intrusive AV quality assessment setting. The experimental results show that this approach can enhance the prediction performance and yield high Pearson and Spearman correlations as well as low RMSE. This work sets a new baseline for AV quality assessment and demonstrates how cross-modal fusion can be useful in real-life multimedia applications.
For expert users to accept Generative AI (GenAI) as a true collaborative partner, it must move beyond simple task-awareness to an understanding of their workflow’s underlying structural rules. This paper introduces a paradigm for AI collaborators that moves beyond simple task awareness to an understanding of the semantic and hierarchical relationships within a component-based system. We investigate this concept within the context of the design-to-code workflow, where inefficiencies arise from the modification of components within design systems. Through two empirical studies with designers and developers, we found that GenAI output was often rejected because it violated the component hierarchy. Designers required granular and visual control for refinements, whereas developers valued automated setup but required transparent validation of the generated code’s logic. Based on these findings, we contribute design guidelines for achieving Component-Structure Awareness (CSA), with two core principles: the Atomic Recommender, which provides assistance that respects the component hierarchy, and Communication Archetypes, which allow GenAI to adapt its interaction style to the user’s role and the atomic nature of their task. This work provides a new, higher-level concept for designing the next generation of truly collaborative GenAI agents.
We propose GMP3, a multiphase global path planning framework that generates dynamically feasible three-dimensional trajectories for unmanned aerial vehicles (UAVs) operating in cluttered environments. The framework extends traditional path planning from Euclidean position spaces to the Lie group SE(3), allowing joint learning of translational motion and rotational dynamics. A modified Bellman-based operator is introduced to support reinforcement learning (RL) policy updates while leveraging prior trajectory information for improved convergence. GMP3 is designed as a distributed framework in which agents influence each other and share policy information along the trajectory: each agent refines its assigned segment and shares with its neighbors via a consensus-based scheme, enabling cooperative policy updates and convergence toward a path shaped globally even under kinematic constraints. We also propose DroneManager, a modular ground control software that interfaces the planner with real UAV platforms via the MAVLink protocol, supporting real-time deployment and feedback. Simulation studies and indoor flight experiments validate the effectiveness of the proposed method in constrained 3D environments, demonstrating reliable obstacle avoidance and smooth, feasible trajectories across both position and orientation. The open-source implementation is available at https://github.com/Domattee/DroneManager.
As automated vehicle technology advances, explainable AI has emerged as a critical tool to enable users to understand and predict the behavior of AI systems, particularly in safety-critical applications such as automated driving. However, increased transparency in AI explanations may inadvertently contribute to an “illusion of control”, a cognitive bias in which drivers overestimate their influence or understanding of the AI’s actions. We aim to better understand how the level of detail in AI explanations affects users of automated vehicles. In a virtual reality study, N = 44 participants experienced different explanation levels (low, medium, high) in an automated ride (SAE L4) compared to a baseline condition with no explanations. The results show a significant improvement in participants’ user experience, acceptance, and explanation satisfaction, with more detailed explanations. Our findings also indicate that as AI explanations become more detailed, users’ perceived level of control increases significantly, although this perception does not correlate with actual control capabilities. At the same time, it decreased their desire to take control, indicating users’ susceptibility to the ’illusion of control’ bias in the context of automated driving. Overall, this suggests that the design of explanation interfaces should strive for a balanced level of detail that promotes AI transparency without causing cognitive overload. At the same time, explainable AI can be utilized to decrease users’ desire to intervene in the AI’s actions.
Trust is a fundamental requirement in vehicular systems, particularly in the context of autonomous and cooperative driving, where decisions often depend on data received from other vehicles or infrastructure. This paper presents a novel approach to integrating trust-based reasoning—specifically, decision-making based on subjective logic—into Attribute-Based Access Control (ABAC) systems, using the XACML policy language. We draw a parallel between trust management and access control, showing that trust assessments can be modeled as access decisions and expressed through standard access control policies. This integration allows for the reuse of mature ABAC infrastructure to represent and enforce trust-based decisions. Furthermore, we demonstrate how trust discounting mechanisms, which combine node-based and data-based trust, can be implemented within XACML policies. To evaluate this approach, we introduce a Rust-based XACML engine capable of interpreting trust-enhanced policies. Its integration in both inter- and intra-vehicle scenarios is discussed, assuming a modern zone-based architecture. Our results suggest that policy-based trust management can be efficiently implemented, enabling flexible and standardized trust reasoning in automotive environments.
Modern car insurance companies can dynamically adjust premiums with mobile sensor units that constantly record, analyze, and grade driving behavior. However, attackers can tamper with these systems to cheat the insurance company and reduce their premiums. We developed an attacker model for this threat in conjunction with a risk assessment using the ISO/SAE 21434 Risk Rating Methodology, a widely accepted standard for evaluating and managing risks in automotive systems. To support and showcase our model, we examined the static IT security of the mobile sensor unit DriveWell and its companion Android application HUK Mein Auto. In particular, we utilized digital forensic tools to analyze the stored data in the sensor’s internal storage and the app’s smartphone storage, with a focus on encryption mechanisms for locally stored data and internet communication. During this analysis, we spotted a critical vulnerability that enables attackers to manipulate their insurance scores and potentially reduce premiums by up to 30%. Our model and real-world analysis provide a solid foundation for reducing the risk of such manipulation of insurance premiums.
Quantum Key Distribution (QKD) enables two parties to generate a secret key, based on quantum physical properties. The secrecy of the exchanged key is guaranteed by these quantum physical properties, even against an attacker using a quantum computer. QKD therefore offers an Quantum Information Theoretically Secure (Q-ITS) alternative to Post Quantum Cryptography (PQC) schemes. This paper discusses advantages and challenges of deploying a QKD system in an Intelligent Transport System (ITS), by examining two exemplary use cases. Approaches to effectively implement QKD in these use cases are proposed, with a short discussion of advantages over the use of classical and PQC schemes.
Khepri: The Beetle Experience: A Culturally-Grounded Mobility System for Luxor’s Heritage Tourism
(2025)