Arno Eichberger

Ji Zhou, Yilin Ding, Yongqi Zhao et al.

Reliable environmental perception remains one of the main obstacles for safe operation of automated vehicles. Safety of the Intended Functionality (SOTIF) concerns safety risks from perception insufficiencies, particularly under adverse conditions where conventional detectors often falter. While Large Vision-Language Models (LVLMs) demonstrate promising semantic reasoning, their quantitative effectiveness for safety-critical 2D object detection is underexplored. This paper presents a systematic evaluation of ten representative LVLMs using the PeSOTIF dataset, a benchmark specifically curated for long-tail traffic scenarios and environmental degradations. Performance is quantitatively compared against the classical perception approach, a YOLO-based detector. Experimental results reveal a critical trade-off: top-performing LVLMs (e.g., Gemini 3, Doubao) surpass the YOLO baseline in recall by over 25% in complex natural scenarios, exhibiting superior robustness to visual degradation. Conversely, the baseline retains an advantage in geometric precision for synthetic perturbations. These findings highlight the complementary strengths of semantic reasoning versus geometric regression, supporting the use of LVLMs as high-level safety validators in SOTIF-oriented automated driving systems.

Francesco De Cristofaro, Felix Hofbaur, Aixi Yang et al.

Lane changes of preceding vehicles have a great impact on the motion planning of automated vehicles especially in complex traffic situations. Predicting them would benefit the public in terms of safety and efficiency. While many research efforts have been made in this direction, few concentrated on predicting maneuvers within a set time interval compared to predicting at a set prediction time. In addition, there exist a lack of comparisons between different architectures to try to determine the best performing one and to assess how to correctly choose the input for such models. In this paper the structure of an LSTM, a CNN and a Transformer network are described and implemented to predict the intention of human drivers to perform a lane change. We show how the data was prepared starting from a publicly available dataset (highD), which features were used, how the networks were designed and finally we compare the results of the three networks with different configurations of input data. We found that transformer networks performed better than the other networks and was less affected by overfitting. The accuracy of the method spanned from $82.79\%$ to $96.73\%$ for different input configurations and showed overall good performances considering also precision and recall.

Francesco De Cristofaro, Cornelia Lex, Jia Hu et al.

As a result of the growing importance of lane change intention prediction for a safe and efficient driving experience in complex driving scenarios, researchers have in recent years started to train novel machine learning algorithms on available datasets with promising results. A shortcoming of this recent research effort, though, is that the vast majority of the proposed algorithms are trained on a single datasets. In doing so, researchers failed to test if their algorithm would be as effective if tested on a different dataset and, by extension, on a different population with respect to the one on which they were trained. In this article we test a transformer designed for lane change intention prediction on two datasets collected by LevelX in Germany and Hong Kong. We found that the transformer's accuracy plummeted when tested on a population different to the one it was trained on with accuracy values as low as 39.43%, but that when trained on both populations simultaneously it could achieve an accuracy as high as 86.71%. - This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible.

Julian von Schleinitz, Thomas Schwarzhuber, Lukas Wörle et al.

Professional race drivers are still superior to automated systems at controlling a vehicle at its dynamic limit. Gaining insight into race drivers' vehicle handling process might lead to further development in the areas of automated driving systems. We present a method to study and evaluate race drivers on a driver-in-the-loop simulator by analysing tire grip potential exploitation. Given initial data from a simulator run, two optimiser based on physical models maximise the horizontal vehicle acceleration or the tire forces, respectively. An overall performance score, a vehicle-trajectory score and a handling score are introduced to evaluate drivers. Our method is thereby completely track independent and can be used from one single corner up to a large data set. We apply the proposed method to a motorsport data set containing over 1200 laps from seven professional race drivers and two amateur drivers whose lap times are 10-20% slower. The difference to the professional drivers comes mainly from their inferior handling skills and not their choice of driving line. A downside of the presented method for certain applications is an extensive computation time. Therefore, we propose a Long-short-term memory (LSTM) neural network to estimate the driver evaluation scores. We show that the neural network is accurate and robust with a root-mean-square error between 2-5% and can replace the optimisation based method. The time for processing the data set considered in this work is reduced from 68 hours to 12 seconds, making the neural network suitable for real-time application.

Philip Pannagger, Demin Nalic, Faris Orucevic et al.

Virtual development and prototyping has already become an integral part in the field of automated driving systems (ADS). There are plenty of software tools that are used for the virtual development of ADS. One such tool is CarMaker from IPG Automotive, which is widely used in the scientific community and in the automotive industry. It offers a broad spectrum of implementation and modelling possibilities of the vehicle, driver behavior, control, sensors, and environmental models. Focusing on the virtual development of highly automated driving functions on the vehicle guidance level, it is essential to perceive the environment in a realistic manner. For the longitudinal and lateral path guidance line detection sensors are necessary for the determination of the relevant perceiving vehicle and for the planning of trajectories. For this purpose, a lane sensor model was developed in order to efficiently detect lanes in the simulation environment of CarMaker. The so-called advanced lane detection model (ALDM) is optimized regarding the calculation time and is for the lateral and longitudinal vehicle guidance in CarMaker.

Demin Nalic, Hexuan Li, Arno Eichberger et al.

Classical approaches and procedures for testing of automated vehicles of SAE levels 1 and 2 were based on defined scenarios with specific maneuvers, depending on the function under test. For automated driving systems (ADS) of SAE level 3+, the scenario space is infinite and calling for virtual testing and verification. However, even in simulation, the generation of safety-relevant scenarios for ADS is expensive and time-consuming. This leads to a demand for stochastic and realistic traffic simulation. Therefore, microscopic traffic flow simulation models (TFSM) are becoming a crucial part of scenario-based testing of ADS. In this paper, a co-simulation between the multi-body simulation software IPG CarMaker and the microscopic traffic flow simulation software (TFSS) PTV Vissim is used. Although the TFSS could provide realistic and stochastic behavior of the traffic participants, safety-critical scenarios (SCS) occur rarely. In order to avoid this, a novel Stress Testing Method (STM) is introduced. With this method, traffic participants are manipulated via external driver DLL interface from PTV Vissim in the vicinity of the vehicle under test in order to provoke defined critical maneuvers derived from statistical accident data on highways in Austria. These external driver models imitate human driving errors, resulting in an increase of safety-critical scenarios. As a result, the presented STM method contributes to an increase of safety-relevant scenarios for verification, testing and assessment of ADS.

Demin Nalic, Aleksa Pandurevic, Arno Eichberger et al.

Virtual testing of automated driving systems (ADS) has become an essential part of testing procedures for all automation levels. As ADS from automation level 3 and up are very complex, virtual testing for such systems is inevitable. The complexity of these levels lies in the modelling and calculation demand for the virtual environment which consists of roads, traffic, static and dynamic objects as well as the modelling of the car itself. For safety and performance analyses of ADS, the most important part is the modelling and consideration of road traffic participants. There is multiple traffic flow simulation software (TFSS) which are used to reproduce realistic traffic behavior and are integrated directly or over interfaces with vehicle simulation software (VSS). For these software environments, the possibility to manipulate traffic participants in a defined manner e.g. in the vicinity of the vehicle under test or implementing defined driver models for traffic vehicles is beneficial. In this paper, we present a software framework based on the external driver model interface provided by Vissim. This framework makes it possible to easily manipulate traffic participants for testing purposes of ADS.