IFM - Institut für Fahrerassistenz und vernetzte Mobilität
Refine
Year of publication
Document Type
- conference proceeding (article) (85)
- Article (42)
- conference proceeding (presentation) (9)
- Preprint (4)
- Master's Thesis (2)
- Book (1)
- conference proceeding (summary, abstract) (1)
- Doctoral Thesis (1)
- Other (1)
- Report (1)
Publication reviewed
- peer review (83)
- begutachtet (43)
- nicht begutachtet (19)
Keywords
- automated driving (14)
- advanced driver assistance systems (10)
- human factor (8)
- workload (8)
- Simulation (7)
- driving experience (7)
- Automated Driving (6)
- lane keeping system (6)
- stress (6)
- Advanced Driver Assistance Systems (5)
Institute
- IFM - Institut für Fahrerassistenz und vernetzte Mobilität (147)
- Fakultät Maschinenbau (45)
- Fakultät Elektrotechnik (39)
- Fakultät Informatik (35)
- IDF – Institut für Datenoptimierte Fertigung (2)
- IFA – Institut für Innovative Fahrzeugantriebe (1)
- IMS - Institut für Maschinelles Sehen (1)
- IPI – Institut für Produktion und Informatik (1)
This paper focuses on developing a high-fidelity model of a frequency-modulated continuous wave (FMCW) radio detection and ranging (RADAR) sensor for automated train systems. The model uses ray tracing for virtual environmental perception in railway scenarios. It includes a multiple input multiple output (MIMO) antenna array and a complete signal processing toolchain of real RADAR sensors. The model outputs raw data, including range maps (RMs), range-Doppler maps (RDMs), and detection lists, including distance, relative radial velocity, and signal-to-noise ratio (SNR), radar-cross section (RCS), azimuth, and elevation angles. Results show a strong correlation with real measurements with a mean absolute percentage error (MAPE) below 4.8% for all the parameters defined at the detection level. To the author's knowledge, these error levels are among the lowest reported for RADAR sensor model validation. This finding allows for a cost-effective perception of virtual environments, facilitating simulation-based testing of automated railway systems.
Virtual radar for the railway of tomorrow – testing automated trains in a virtual environment
(2025)
This article presents a method for developing and validating a virtual radar sensor model designed to simulate automated train operations as part of the Digitale Schiene Deutschland (DSD) sector initiative. The model generates realistic sensor data in a digital environment and supports the evaluation and future validation of radar-based object detection systems used in automated trains. An initial validation has been carried out under controlled laboratory conditions so as to ensure the accuracy and reliability of the virtual sensor. A static test setup was used in which a radar corner reflector served as a reference target for comparing the real and simulated sensor data. The use of simulation instead of physical testing on tracks means that complex railway scenarios can be evaluated more flexibly, cost-effectively and without any safety risks. This accelerates the development process and reduces the reliance on time consuming field trials.
Vision-guided robotic systems play a key role in industrial automation, particularly in flexible pick-and-place tasks. We present a modular approach that integrates automatic data extraction of 2D assembly drawings, object recognition, automatic robotic part placement, and augmented reality for quality inspection. The system combines a YOLOv8-based segmentation pipeline with classical shape-matching techniques to detect, identify, and align components during assembly. We apply a hybrid method to interpret assembly drawings that merge traditional computer vision with neural segmentation masks. The robotic setup features a 6-DOF arm controlled via Robot Operating System (ROS), with overhead cameras for object localization. This paper presents ongoing work. Preliminary results indicate improved segmentation accuracy using AI-based methods for extracting the foundational data from the 2D assembly drawings. Future work will focus on enhancing grasp and path planning, increasing placement precision, and improving the robustness of part detection. Additionally, we aim to generalize the system to previously unseen assembly plans and expand the feature extraction capabilities for complex assembly drawings.
There is evidence that the driving style of an autonomous vehicle is important to increase the acceptance and trust of the passengers. The driving situation has been found to have a significant influence on human driving behavior. However, current driving style models only partially incorporate driving environment information, limiting the alignment between an agent and the given situation. Therefore, we propose a situation-aware driving style model based on different visual feature encoders pretrained on fleet data, as well as driving behavior predictors, which are adapted to the driving style of a specific driver. Our experiments show that the proposed method outperforms all evaluated baselines significantly and forms plausible situation clusters. Furthermore, we found that feature encoders pretrained on our dataset lead to more precise driving behavior modeling. In contrast, feature encoders pretrained supervised and unsupervised on different data sources lead to more specific situation clusters, which can be utilized to constrain and control the driving style adaptation for specific situations. Moreover, in a real-world setting, where driving style adaptation is happening iteratively, we found the MLP-based behavior predictors achieve good performance initially but suffer from catastrophic forgetting. In contrast, behavior predictors based on situation-dependent statistics can learn iteratively from continuous data streams by design. Overall, our experiments show that important information for driving behavior prediction is contained within the visual feature encoder. The dataset is publicly available at https://huggingface.co/datasets/jHaselberger/SADC-Situation-Awareness-for-Driver-Centric-Driving-Style-Adaptation.
Driving simulators are increasingly used in chassis development. An application example is the subjective assessment of the controllability of the vehicle in the event of damage. As a result, an axle concept can be developed with regard to the chain of occurring damages or risk assessments of potential axle damage can be carried ou in a cost-saving, reproducible manner and without endangering the driver. A central question in the application of driving simulators is the validity of the results. In this work, the subjective evaluations and physiological stress indicators in the real vehicle and in the driving simulator are compared. For this purpose, three real vehicles are equipped with differently deformed tie rods on the rear axle and in parallel a validated real-time capable full vehicle model is used. A slalom, a double lane change and a cornering manoevre are completed by 20 semi-professional test drivers. A comparison of the subjective evaluations shows that for tie rod deformations up to 15 mm absolute validity is achieved. For the 25 mm deformed tie rod, there are statistically significant differences, which can be explained by the physical limits of the full vehicle model. There are no statistically significant differences in the physiological quantities, which indicates a comparable stress level in the vehicle and in the driving simulator.
Implementation of a Real-Time ESC Simulation Environment into a Dynamic Driving Simulator (SiL/DiL)
(2025)
The Electronic stability control (ESC) is a safety system that helps prevent vehicle skidding and loss of control. ESC systems are usually tested and calibrated on real vehicles, which is costly and time-consuming. This paper presents a novel approach to implement a serial production ESC simulation environment in a dynamic driving simulator. This allows performing objective and subjective evaluations of the ESC system at an early stage of the development process. The hardware and software components of the real-time simulation environment are described, as well as the integration of the ESC system and validation of its functionality. In a pilot study, expert drivers compared the simulated ESC system with the real one in different driving scenarios. The results show that the simulation environment and therefore the dynamic driving simulator can reproduce the behavior and performance of the real ESC system with high fidelity and accuracy, by providing useful feedback about the driving characteristics and stability of the vehicle. To conclude, the approach could reduce the development time and cost of ESC systems, as well as improve their quality and safety.
Federated Learning (FL) offers a solution to the challenges of traditional centralized machine learning by enabling decentralized training and exchanging only model updates instead of raw data. This approach addresses key issues such as privacy concerns and high data transfer costs. However, integrating FL into existing Machine Learning Operations (MLOps) pipelines presents challenges, particularly regarding model versioning, synchronization, and scalability. This paper introduces a concept for centralized model management that enables the integration of FL into existing MLOps pipelines without the need to overhaul the existing architecture. The concept is specifically developed for deployment in an industrial setting, with plans for implementing both FL and Transfer Learning (TL) in the future. The proposed approach emphasizes flexibility, ensuring that it can be easily extended to accommodate additional methods and seamlessly integrated into diverse, pre-existing infrastructure. The management of the system is facilitated using the open-source tool MLflow, which offers significant advantages over specialized FL frameworks, particularly in terms of adaptability and resource optimization.
The response time of Artificial Neural Network (ANN) inference is critical in embedded systems processing sensor data close to the source. This is particularly important in applications such as predictive maintenance, which rely on timely state change predictions. This study enables estimation of model response times based on the underlying platform, highlighting the importance of benchmarking generic ANN applications on edge devices. We analyze the impact of network parameters, activation functions, and single- versus multi-threading on response times. Additionally, potential hardware-related influences, such as clock rate variances, are discussed. The results underline the complexity of task partitioning and scheduling strategies, stressing the need for precise parameter coordination to optimise performance across platforms. This study shows that cutting-edge frameworks do not necessarily perform the required operations automatically for all configurations, which may negatively impact performance. This paper further investigates the influence of network structure on model calibration, quantified using the Expected Calibration Error (ECE), and the limits of potential optimisation opportunities. It also examines the effects of model conversion to Tensorflow Lite (TFLite), highlighting the necessity of considering both performance and calibration when deploying models on embedded systems.
Light detection and ranging (LiDAR) sensor technology for people detection
offers a significant advantage in data protection. However, to design these systems cost and energy-efficiently, the relationship between the measurement data and final object detection output with deep neural networks (DNNs) has to be elaborated. Therefore, this paper presents augmentation methods to analyze the influence of the distance, resolution,
noise, and shading parameters of a LiDAR sensor in real point clouds for people detection. Furthermore, their influence on object detection using DNNs was investigated. A significant reduction in the quality requirements for the point clouds was possible for the measurement setup with only minor degradation on the object list level. The DNNs PointVoxel-Regionbased Convolutional Neural Network (PV-RCNN) and Sparsely Embedded Convolutional Detection (SECOND) both only show a reduction in object detection of less than 5% with a reduced resolution of up to 32 factors, for an increase in distance of 4 factors, and with a Gaussian noise up to µ = 0 and σ = 0.07. In addition, both networks require an unshaded height of approx. 0.5 m from a detected person’s head downwards to ensure good people detection performance without special training for these cases. The results obtained, such as shadowing information, are transferred to a software program to determine the minimum number of sensors and their orientation based on the mounting height of the sensor, the sensor parameters, and the ground area under consideration, both for detection at the point cloud level and object detection level.