J. Marius Zöllner

Faris Janjoš, Marcel Hallgarten, Anthony Knittel et al.

The CVAE is one of the most widely-used models in trajectory prediction for AD. It captures the interplay between a driving context and its ground-truth future into a probabilistic latent space and uses it to produce predictions. In this paper, we challenge key components of the CVAE. We leverage recent advances in the space of the VAE, the foundation of the CVAE, which show that a simple change in the sampling procedure can greatly benefit performance. We find that unscented sampling, which draws samples from any learned distribution in a deterministic manner, can naturally be better suited to trajectory prediction than potentially dangerous random sampling. We go further and offer additional improvements including a more structured Gaussian mixture latent space, as well as a novel, potentially more expressive way to do inference with CVAEs. We show wide applicability of our models by evaluating them on the INTERACTION prediction dataset, outperforming the state of the art, as well as at the task of image modeling on the CelebA dataset, outperforming the baseline vanilla CVAE. Code is available at https://github.com/boschresearch/cuae-prediction.

Nikolai Polley, Svetlana Pavlitska, Yacin Boualili et al.

Effective traffic light detection is a critical component of the perception stack in autonomous vehicles. This work introduces a novel deep-learning detection system while addressing the challenges of previous work. Utilizing a comprehensive dataset amalgamation, including the Bosch Small Traffic Lights Dataset, LISA, the DriveU Traffic Light Dataset, and a proprietary dataset from Karlsruhe, we ensure a robust evaluation across varied scenarios. Furthermore, we propose a relevance estimation system that innovatively uses directional arrow markings on the road, eliminating the need for prior map creation. On the DriveU dataset, this approach results in 96% accuracy in relevance estimation. Finally, a real-world evaluation is performed to evaluate the deployment and generalizing abilities of these models. For reproducibility and to facilitate further research, we provide the model weights and code: https://github.com/KASTEL-MobilityLab/traffic-light-detection.

Svetlana Pavlitska, Malte Stüven, Beyza Keskin et al.

Mixture-of-Experts (MoE) models provide a structured approach to combining specialized neural networks and offer greater interpretability than conventional ensembles. While MoEs have been successfully applied to image classification and semantic segmentation, their use in object detection remains limited due to challenges in merging dense and structured predictions. In this work, we investigate model-level mixtures of object detectors and analyze their suitability for improving performance and interpretability in object detection. We propose an MoE architecture that combines YOLO-based detectors trained on semantically disjoint data subsets, with a learned gating network that dynamically weights expert contributions. We study different strategies for fusing detection outputs and for training the gating mechanism, including balancing losses to prevent expert collapse. Experiments on the BDD100K dataset demonstrate that the proposed MoE consistently outperforms standard ensemble approaches and provides insights into expert specialization across domains, highlighting model-level MoEs as a viable alternative to traditional ensembling for object detection. Our code is available at https://github.com/KASTEL-MobilityLab/mixtures-of-experts/.

Svetlana Pavlitska, Haixi Fan, Konstantin Ditschuneit et al.

Sparse mixture-of-experts (MoE) layers have been shown to substantially increase model capacity without a proportional increase in computational cost and are widely used in transformer architectures, where they typically replace feed-forward network blocks. In contrast, integrating sparse MoE layers into convolutional neural networks (CNNs) remains inconsistent, with most prior work focusing on fine-grained MoEs operating at the filter or channel levels. In this work, we investigate a coarser, patch-wise formulation of sparse MoE layers for semantic segmentation, where local regions are routed to a small subset of convolutional experts. Through experiments on the Cityscapes and BDD100K datasets using encoder-decoder and backbone-based CNNs, we conduct a design analysis to assess how architectural choices affect routing dynamics and expert specialization. Our results demonstrate consistent, architecture-dependent improvements (up to +3.9 mIoU) with little computational overhead, while revealing strong design sensitivity. Our work provides empirical insights into the design and internal dynamics of sparse MoE layers in CNN-based dense prediction. Our code is available at https://github.com/KASTEL-MobilityLab/moe-layers/.

Abhishek Vivekanandan, Ahmed Abouelazm, J. Marius Zöllner

Motion forecasting often requires trading interpretability for predictive accuracy. Standard anchor-based architectures rely on opaque latent queries that are highly prone to latent collapse, or naive trajectory sampling that limits multi-modal diversity. We propose an end-to-end differentiable framework that grounds predictions in a comprehensive "motion bank", a structured embedding space of physically realizable trajectories constructed via contrastive learning. Rather than regressing paths from a blank slate, our architecture dynamically retrieves explicit motion priors using a novel Anchor Retrieval Layer. This module adapts orthogonally initialized queries via a Dual-Level Gated Cross-Attention mechanism and executes discrete trajectory selection using a Straight-Through Gumbel-Softmax estimator to preserve continuous gradient flow. The retrieved semantically grounded anchors are then geometrically refined by a DETR-style decoder, optimized jointly with a Winner-Takes-All (WTA) kinematic Gaussian Mixture Model (GMM), a latent diversity penalty, and a soft-min weighted endpoint loss. By strictly conditioning the decoding phase on diverse, interpretable motion primitives, our approach eliminates the "black box" of standard latent queries while achieving competitive multi-modal accuracy on the Argoverse 2 and Waymo Open Motion datasets. Code is available at: https://github.com/abviv/recall2predict

Marius Dähling, Sebastian Krebs, J. Marius Zöllner

In current research, Bird's-Eye-View (BEV)-based transformers are increasingly utilized for multi-camera 3D object detection. Traditional models often employ random queries as anchors, optimizing them successively. Recent advancements complement or replace these random queries with detections from auxiliary networks. We propose a more intuitive and efficient approach by using BEV feature cells directly as anchors. This end-to-end approach leverages the dense grid of BEV queries, considering each cell as a potential object for the final detection task. As a result, we introduce a novel two-stage anchor generation method specifically designed for multi-camera 3D object detection. To address the scaling issues of attention with a large number of queries, we apply BEV-based Non-Maximum Suppression, allowing gradients to flow only through non-suppressed objects. This ensures efficient training without the need for post-processing. By using BEV features from encoders such as BEVFormer directly as object queries, temporal BEV information is inherently embedded. Building on the temporal BEV information already embedded in our object queries, we introduce a hybrid temporal modeling approach by integrating prior detections to further enhance detection performance. Evaluating our method on the nuScenes dataset shows consistent and significant improvements in NDS and mAP over the baseline, even with sparser BEV grids and therefore fewer initial anchors. It is particularly effective for small objects, enhancing pedestrian detection with a 3.8% mAP increase on nuScenes and an 8% increase in LET-mAP on Waymo. Applying our method, named DenseBEV, to the challenging Waymo Open dataset yields state-of-the-art performance, achieving a LET-mAP of 60.7%, surpassing the previous best by 5.4%. Code is available at https://github.com/mdaehl/DenseBEV.

Barbara Schütt, Marc Heinrich, Sonja Marahrens et al.

Verification and validation are major challenges for developing automated driving systems. A concept that gets more and more recognized for testing in automated driving is scenario-based testing. However, it introduces the problem of what scenarios are relevant for testing and which are not. This work aims to find relevant, interesting, or critical parameter sets within logical scenarios by utilizing Bayes optimization and Gaussian processes. The parameter optimization is done by comparing and evaluating six different metrics in two urban intersection scenarios. Finally, a list of ideas this work leads to and should be investigated further is presented.

Ahmed Abouelazm, Felix Klingebiel, Philip Schörner et al.

Exploration in reinforcement learning for autonomous driving is inherently unsafe: agents must experience novel behaviors to learn, yet exploration can lead to collisions or off-road driving. We propose an uncertainty-aware framework that leverages expert advice to guide exploration while avoiding long-term dependence. Advice is triggered when epistemic or aleatoric uncertainty exceeds adaptive thresholds derived from rolling buffers, ensuring advice evolves with the agent's confidence. A commitment-cooldown strategy with a stochastic early-stop heuristic regulates the duration and frequency of guidance, exposing the agent to coherent maneuvers without exhausting the advice budget. Expert and agent experiences are combined in a shared replay buffer within an off-policy implicit quantile network (IQN) backbone, enabling efficient reuse of expert trajectories. Experiments in CARLA show that our method outperforms the IQN baseline, improving success by 5-7% and reducing failures, demonstrating that risk-sensitive uncertainty coupled with regulated expert integration enables safer and more efficient exploration for sensor-based RL policy learning in unsignalized intersection navigation.

Svetlana Pavlitska, Nico Lambing, J. Marius Zöllner

Traffic sign recognition is an essential component of perception in autonomous vehicles, which is currently performed almost exclusively with deep neural networks (DNNs). However, DNNs are known to be vulnerable to adversarial attacks. Several previous works have demonstrated the feasibility of adversarial attacks on traffic sign recognition models. Traffic signs are particularly promising for adversarial attack research due to the ease of performing real-world attacks using printed signs or stickers. In this work, we survey existing works performing either digital or real-world attacks on traffic sign detection and classification models. We provide an overview of the latest advancements and highlight the existing research areas that require further investigation.

Daniel Grimm, Maximilian Zipfl, Felix Hertlein et al.

Precisely predicting the future trajectories of surrounding traffic participants is a crucial but challenging problem in autonomous driving, due to complex interactions between traffic agents, map context and traffic rules. Vector-based approaches have recently shown to achieve among the best performances on trajectory prediction benchmarks. These methods model simple interactions between traffic agents but don't distinguish between relation-type and attributes like their distance along the road. Furthermore, they represent lanes only by sequences of vectors representing center lines and ignore context information like lane dividers and other road elements. We present a novel approach for vector-based trajectory prediction that addresses these shortcomings by leveraging three crucial sources of information: First, we model interactions between traffic agents by a semantic scene graph, that accounts for the nature and important features of their relation. Second, we extract agent-centric image-based map features to model the local map context. Finally, we generate anchor paths to enforce the policy in multi-modal prediction to permitted trajectories only. Each of these three enhancements shows advantages over the baseline model HoliGraph.

Svetlana Pavlitska, Christian Hubschneider, Lukas Struppek et al.

Sparsely-gated Mixture of Expert (MoE) layers have been recently successfully applied for scaling large transformers, especially for language modeling tasks. An intriguing side effect of sparse MoE layers is that they convey inherent interpretability to a model via natural expert specialization. In this work, we apply sparse MoE layers to CNNs for computer vision tasks and analyze the resulting effect on model interpretability. To stabilize MoE training, we present both soft and hard constraint-based approaches. With hard constraints, the weights of certain experts are allowed to become zero, while soft constraints balance the contribution of experts with an additional auxiliary loss. As a result, soft constraints handle expert utilization better and support the expert specialization process, while hard constraints maintain more generalized experts and increase overall model performance. Our findings demonstrate that experts can implicitly focus on individual sub-domains of the input space. For example, experts trained for CIFAR-100 image classification specialize in recognizing different domains such as flowers or animals without previous data clustering. Experiments with RetinaNet and the COCO dataset further indicate that object detection experts can also specialize in detecting objects of distinct sizes.

Daniel Bogdoll, Yitian Yang, Tim Joseph et al.

World models for autonomous driving have the potential to dramatically improve the reasoning capabilities of today's systems. However, most works focus on camera data, with only a few that leverage lidar data or combine both to better represent autonomous vehicle sensor setups. In addition, raw sensor predictions are less actionable than 3D occupancy predictions, but there are no works examining the effects of combining both multimodal sensor data and 3D occupancy prediction. In this work, we perform a set of experiments with a MUltimodal World Model with Geometric VOxel representations (MUVO) to evaluate different sensor fusion strategies to better understand the effects on sensor data prediction. We also analyze potential weaknesses of current sensor fusion approaches and examine the benefits of additionally predicting 3D occupancy.

Daniel Bogdoll, Enrico Eisen, Maximilian Nitsche et al.

Tremendous progress in deep learning over the last years has led towards a future with autonomous vehicles on our roads. Nevertheless, the performance of their perception systems is strongly dependent on the quality of the utilized training data. As these usually only cover a fraction of all object classes an autonomous driving system will face, such systems struggle with handling the unexpected. In order to safely operate on public roads, the identification of objects from unknown classes remains a crucial task. In this paper, we propose a novel pipeline to detect unknown objects. Instead of focusing on a single sensor modality, we make use of lidar and camera data by combining state-of-the art detection models in a sequential manner. We evaluate our approach on the Waymo Open Perception Dataset and point out current research gaps in anomaly detection.

Svetlana Pavlitskaya, Bianca-Marina Codău, J. Marius Zöllner

Standard approaches for adversarial patch generation lead to noisy conspicuous patterns, which are easily recognizable by humans. Recent research has proposed several approaches to generate naturalistic patches using generative adversarial networks (GANs), yet only a few of them were evaluated on the object detection use case. Moreover, the state of the art mostly focuses on suppressing a single large bounding box in input by overlapping it with the patch directly. Suppressing objects near the patch is a different, more complex task. In this work, we have evaluated the existing approaches to generate inconspicuous patches. We have adapted methods, originally developed for different computer vision tasks, to the object detection use case with YOLOv3 and the COCO dataset. We have evaluated two approaches to generate naturalistic patches: by incorporating patch generation into the GAN training process and by using the pretrained GAN. For both cases, we have assessed a trade-off between performance and naturalistic patch appearance. Our experiments have shown, that using a pre-trained GAN helps to gain realistic-looking patches while preserving the performance similar to conventional adversarial patches.

Daniel Bogdoll, Svenja Uhlemeyer, Kamil Kowol et al.

Deep neural networks (DNN) which are employed in perception systems for autonomous driving require a huge amount of data to train on, as they must reliably achieve high performance in all kinds of situations. However, these DNN are usually restricted to a closed set of semantic classes available in their training data, and are therefore unreliable when confronted with previously unseen instances. Thus, multiple perception datasets have been created for the evaluation of anomaly detection methods, which can be categorized into three groups: real anomalies in real-world, synthetic anomalies augmented into real-world and completely synthetic scenes. This survey provides a structured and, to the best of our knowledge, complete overview and comparison of perception datasets for anomaly detection in autonomous driving. Each chapter provides information about tasks and ground truth, context information, and licenses. Additionally, we discuss current weaknesses and gaps in existing datasets to underline the importance of developing further data.

Larissa T. Triess, Christoph B. Rist, David Peter et al.

A considerable amount of research is concerned with the generation of realistic sensor data. LiDAR point clouds are generated by complex simulations or learned generative models. The generated data is usually exploited to enable or improve downstream perception algorithms. Two major questions arise from these procedures: First, how to evaluate the realism of the generated data? Second, does more realistic data also lead to better perception performance? This paper addresses both questions and presents a novel metric to quantify the realism of LiDAR point clouds. Relevant features are learned from real-world and synthetic point clouds by training on a proxy classification task. In a series of experiments, we demonstrate the application of our metric to determine the realism of generated LiDAR data and compare the realism estimation of our metric to the performance of a segmentation model. We confirm that our metric provides an indication for the downstream segmentation performance.

Marc Uecker, Tobias Fleck, Marcel Pflugfelder et al.

LiDAR sensors are an integral part of modern autonomous vehicles as they provide an accurate, high-resolution 3D representation of the vehicle's surroundings. However, it is computationally difficult to make use of the ever-increasing amounts of data from multiple high-resolution LiDAR sensors. As frame-rates, point cloud sizes and sensor resolutions increase, real-time processing of these point clouds must still extract semantics from this increasingly precise picture of the vehicle's environment. One deciding factor of the run-time performance and accuracy of deep neural networks operating on these point clouds is the underlying data representation and the way it is computed. In this work, we examine the relationship between the computational representations used in neural networks and their performance characteristics. To this end, we propose a novel computational taxonomy of LiDAR point cloud representations used in modern deep neural networks for 3D point cloud processing. Using this taxonomy, we perform a structured analysis of different families of approaches. Thereby, we uncover common advantages and limitations in terms of computational efficiency, memory requirements, and representational capacity as measured by semantic segmentation performance. Finally, we provide some insights and guidance for future developments in neural point cloud processing methods.

Faris Janjoš, Max Keller, Maxim Dolgov et al.

Accurate vehicle trajectory prediction is an unsolved problem in autonomous driving with various open research questions. State-of-the-art approaches regress trajectories either in a one-shot or step-wise manner. Although one-shot approaches are usually preferred for their simplicity, they relinquish powerful self-supervision schemes that can be constructed by chaining multiple time-steps. We address this issue by proposing a middle-ground where multiple trajectory segments are chained together. Our proposed Multi-Branch Self-Supervised Predictor receives additional training on new predictions starting at intermediate future segments. In addition, the model 'imagines' the latent context and 'predicts the past' while combining multi-modal trajectories in a tree-like manner. We deliberately keep aspects such as interaction and environment modeling simplistic and nevertheless achieve competitive results on the INTERACTION dataset. Furthermore, we investigate the sparsely explored uncertainty estimation of deterministic predictors. We find positive correlations between the prediction error and two proposed metrics, which might pave way for determining prediction confidence.

Svetlana Pavlitska, Nico Lambing, Ashok Kumar Bangaru et al.

Real-time traffic light recognition is essential for autonomous driving. Yet, a cohesive overview of the underlying model architectures for this task is currently missing. In this work, we conduct a comprehensive survey and analysis of traffic light recognition methods that use convolutional neural networks (CNNs). We focus on two essential aspects: datasets and CNN architectures. Based on an underlying architecture, we cluster methods into three major groups: (1) modifications of generic object detectors which compensate for specific task characteristics, (2) multi-stage approaches involving both rule-based and CNN components, and (3) task-specific single-stage methods. We describe the most important works in each cluster, discuss the usage of the datasets, and identify research gaps.

Svetlana Pavlitskaya, Joël Oswald, J. Marius Zöllner

Although numerous methods to reduce the overfitting of convolutional neural networks (CNNs) exist, it is still not clear how to confidently measure the degree of overfitting. A metric reflecting the overfitting level might be, however, extremely helpful for the comparison of different architectures and for the evaluation of various techniques to tackle overfitting. Motivated by the fact that overfitted neural networks tend to rather memorize noise in the training data than generalize to unseen data, we examine how the training accuracy changes in the presence of increasing data perturbations and study the connection to overfitting. While previous work focused on label noise only, we examine a spectrum of techniques to inject noise into the training data, including adversarial perturbations and input corruptions. Based on this, we define two new metrics that can confidently distinguish between correct and overfitted models. For the evaluation, we derive a pool of models for which the overfitting behavior is known beforehand. To test the effect of various factors, we introduce several anti-overfitting measures in architectures based on VGG and ResNet and study their impact, including regularization techniques, training set size, and the number of parameters. Finally, we assess the applicability of the proposed metrics by measuring the overfitting degree of several CNN architectures outside of our model pool.

Svetlana Pavlitskaya, Nikolai Polley, Michael Weber et al.

Taking into account information across the temporal domain helps to improve environment perception in autonomous driving. However, it has not been studied so far whether temporally fused neural networks are vulnerable to deliberately generated perturbations, i.e. adversarial attacks, or whether temporal history is an inherent defense against them. In this work, we study whether temporal feature networks for object detection are vulnerable to universal adversarial attacks. We evaluate attacks of two types: imperceptible noise for the whole image and locally-bound adversarial patch. In both cases, perturbations are generated in a white-box manner using PGD. Our experiments confirm, that attacking even a portion of a temporal input suffices to fool the network. We visually assess generated perturbations to gain insights into the functioning of attacks. To enhance the robustness, we apply adversarial training using 5-PGD. Our experiments on KITTI and nuScenes datasets demonstrate, that a model robustified via K-PGD is able to withstand the studied attacks while keeping the mAP-based performance comparable to that of an unattacked model.

Svetlana Pavlitska, Hannes Grolig, J. Marius Zöllner

Increasing the model capacity is a known approach to enhance the adversarial robustness of deep learning networks. On the other hand, various model compression techniques, including pruning and quantization, can reduce the size of the network while preserving its accuracy. Several recent studies have addressed the relationship between model compression and adversarial robustness, while some experiments have reported contradictory results. This work summarizes available evidence and discusses possible explanations for the observed effects.

Abhishek Vivekanandan, Ahmed Abouelazm, Philip Schörner et al.

Accurately forecasting the motion of traffic actors is crucial for the deployment of autonomous vehicles at a large scale. Current trajectory forecasting approaches primarily concentrate on optimizing a loss function with a specific metric, which can result in predictions that do not adhere to physical laws or violate external constraints. Our objective is to incorporate explicit knowledge priors that allow a network to forecast future trajectories in compliance with both the kinematic constraints of a vehicle and the geometry of the driving environment. To achieve this, we introduce a non-parametric pruning layer and attention layers to integrate the defined knowledge priors. Our proposed method is designed to ensure reachability guarantees for traffic actors in both complex and dynamic situations. By conditioning the network to follow physical laws, we can obtain accurate and safe predictions, essential for maintaining autonomous vehicles' safety and efficiency in real-world settings.In summary, this paper presents concepts that prevent off-road predictions for safe and reliable motion forecasting by incorporating knowledge priors into the training process.

Maximilian Zipfl, Moritz Jarosch, J. Marius Zöllner

Ensuring validation for highly automated driving poses significant obstacles to the widespread adoption of highly automated vehicles. Scenario-based testing offers a potential solution by reducing the homologation effort required for these systems. However, a crucial prerequisite, yet unresolved, is the definition and reduction of the test space to a finite number of scenarios. To tackle this challenge, we propose an extension to a contrastive learning approach utilizing graphs to construct a meaningful embedding space. Our approach demonstrates the continuous mapping of scenes using scene-specific features and the formation of thematically similar clusters based on the resulting embeddings. Based on the found clusters, similar scenes could be identified in the subsequent test process, which can lead to a reduction in redundant test runs.

Svetlana Pavlitskaya, Jonas Hendl, Sebastian Kleim et al.

Adversarial patch-based attacks aim to fool a neural network with an intentionally generated noise, which is concentrated in a particular region of an input image. In this work, we perform an in-depth analysis of different patch generation parameters, including initialization, patch size, and especially positioning a patch in an image during training. We focus on the object vanishing attack and run experiments with YOLOv3 as a model under attack in a white-box setting and use images from the COCO dataset. Our experiments have shown, that inserting a patch inside a window of increasing size during training leads to a significant increase in attack strength compared to a fixed position. The best results were obtained when a patch was positioned randomly during training, while patch position additionally varied within a batch.

Maximilian Zipfl, Moritz Jarosch, J. Marius Zöllner

Examining graphs for similarity is a well-known challenge, but one that is mandatory for grouping graphs together. We present a data-driven method to cluster traffic scenes that is self-supervised, i.e. without manual labelling. We leverage the semantic scene graph model to create a generic graph embedding of the traffic scene, which is then mapped to a low-dimensional embedding space using a Siamese network, in which clustering is performed. In the training process of our novel approach, we augment existing traffic scenes in the Cartesian space to generate positive similarity samples. This allows us to overcome the challenge of reconstructing a graph and at the same time obtain a representation to describe the similarity of traffic scenes. We could show, that the resulting clusters possess common semantic characteristics. The approach was evaluated on the INTERACTION dataset.

Svetlana Pavlitskaya, Şiyar Yıkmış, J. Marius Zöllner

The growing use of deep neural networks (DNNs) in safety- and security-critical areas like autonomous driving raises the need for their systematic testing. Coverage-guided testing (CGT) is an approach that applies mutation or fuzzing according to a predefined coverage metric to find inputs that cause misbehavior. With the introduction of a neuron coverage metric, CGT has also recently been applied to DNNs. In this work, we apply CGT to the task of person detection in crowded scenes. The proposed pipeline uses YOLOv3 for person detection and includes finding DNN bugs via sampling and mutation, and subsequent DNN retraining on the updated training set. To be a bug, we require a mutated image to cause a significant performance drop compared to a clean input. In accordance with the CGT, we also consider an additional requirement of increased coverage in the bug definition. In order to explore several types of robustness, our approach includes natural image transformations, corruptions, and adversarial examples generated with the Daedalus attack. The proposed framework has uncovered several thousand cases of incorrect DNN behavior. The relative change in mAP performance of the retrained models reached on average between 26.21\% and 64.24\% for different robustness types. However, we have found no evidence that the investigated coverage metrics can be advantageously used to improve robustness.

Svetlana Pavlitska, Christopher Gerking, J. Marius Zöllner

Safety and security are essential for the admission and acceptance of automated and autonomous vehicles. Deep neural networks (DNNs) are widely used for perception and further components of the autonomous driving (AD) stack. However, they possess several limitations, including lack of generalization, efficiency, explainability, plausibility, and robustness. These insufficiencies can pose significant risks to autonomous driving systems. However, hazards, threats, and risks associated with DNN limitations in this domain have not been systematically studied so far. In this work, we propose a joint workflow for risk assessment combining the hazard analysis and risk assessment (HARA) following ISO 26262 and threat analysis and risk assessment (TARA) following the ISO/SAE 21434 to identify and analyze risks arising from inherent DNN limitations in AD perception.

Daniel Bogdoll, Lukas Bosch, Tim Joseph et al.

In recent years there have been remarkable advancements in autonomous driving. While autonomous vehicles demonstrate high performance in closed-set conditions, they encounter difficulties when confronted with unexpected situations. At the same time, world models emerged in the field of model-based reinforcement learning as a way to enable agents to predict the future depending on potential actions. This led to outstanding results in sparse reward and complex control tasks. This work provides an overview of how world models can be leveraged to perform anomaly detection in the domain of autonomous driving. We provide a characterization of world models and relate individual components to previous works in anomaly detection to facilitate further research in the field.

Daniel Bogdoll, Svetlana Pavlitska, Simon Klaus et al.

Anomalies in the domain of autonomous driving are a major hindrance to the large-scale deployment of autonomous vehicles. In this work, we focus on high-resolution camera data from urban scenes that include anomalies of various types and sizes. Based on a Variational Autoencoder, we condition its latent space to classify samples as either normal data or anomalies. In order to emphasize especially small anomalies, we perform experiments where we provide the VAE with a discrepancy map as an additional input, evaluating its impact on the detection performance. Our method separates normal data and anomalies into isolated clusters while still reconstructing high-quality images, leading to meaningful latent representations.

Faris Janjoš, Lars Rosenbaum, Maxim Dolgov et al.

The Variational Autoencoder (VAE) is a seminal approach in deep generative modeling with latent variables. Interpreting its reconstruction process as a nonlinear transformation of samples from the latent posterior distribution, we apply the Unscented Transform (UT) -- a well-known distribution approximation used in the Unscented Kalman Filter (UKF) from the field of filtering. A finite set of statistics called sigma points, sampled deterministically, provides a more informative and lower-variance posterior representation than the ubiquitous noise-scaling of the reparameterization trick, while ensuring higher-quality reconstruction. We further boost the performance by replacing the Kullback-Leibler (KL) divergence with the Wasserstein distribution metric that allows for a sharper posterior. Inspired by the two components, we derive a novel, deterministic-sampling flavor of the VAE, the Unscented Autoencoder (UAE), trained purely with regularization-like terms on the per-sample posterior. We empirically show competitive performance in Fréchet Inception Distance (FID) scores over closely-related models, in addition to a lower training variance than the VAE.

Tim Joseph, Marcus Fechner, Philipp Stegmaier et al.

Intrinsic rewards for exploration in reinforcement learning condition on different contexts: lifelong rewards score each transition against accumulated experience but ignore within-rollout redundancy; episodic rewards penalize intra-trajectory repetition but discard lifetime progress. Hybrid methods combine both signals through heuristic weights or require Gaussian-process dynamics that do not scale beyond low-dimensional state spaces. Trajectory-level information gain decomposes into per-step terms that condition on the replay buffer and rollout prefix simultaneously, but remains intractable for deep models. We derive the Conditional Information Gain (CIG) reward as a tractable surrogate: a log-determinant objective over an ensemble disagreement kernel whose Cholesky factorization yields causal per-step rewards that retain both conditioning sets while scaling to high-dimensional state spaces. We instantiate CIG in a model-based setting, where rollouts are short and within-rollout corrections remain largely unexplored. Across twelve tasks spanning discrete (MiniGrid) and continuous control (OGBench), in both clean and stochastic-distractor settings, CIG outperforms or matches prior exploration methods while remaining robust to stochastic distractors.

Philipp Stegmaier, Karl Kurzer, J. Marius Zöllner

Automated vehicles require the ability to cooperate with humans for smooth integration into today's traffic. While the concept of cooperation is well known, developing a robust and efficient cooperative trajectory planning method is still a challenge. One aspect of this challenge is the uncertainty surrounding the state of the environment due to limited sensor accuracy. This uncertainty can be represented by a Partially Observable Markov Decision Process. Our work addresses this problem by extending an existing cooperative trajectory planning approach based on Monte Carlo Tree Search for continuous action spaces. It does so by explicitly modeling uncertainties in the form of a root belief state, from which start states for trees are sampled. After the trees have been constructed with Monte Carlo Tree Search, their results are aggregated into return distributions using kernel regression. We apply two risk metrics for the final selection, namely a Lower Confidence Bound and a Conditional Value at Risk. It can be demonstrated that the integration of risk metrics in the final selection policy consistently outperforms a baseline in uncertain environments, generating considerably safer trajectories.

Niklas Wagner, Felix Mätzler, Samed R. Vossberg et al.

Understanding emotions and expressions is a task of interest across multiple disciplines, especially for improving user experiences. Contrary to the common perception, it has been shown that emotions are not discrete entities but instead exist along a continuum. People understand discrete emotions differently due to a variety of factors, including cultural background, individual experiences, and cognitive biases. Therefore, most approaches to expression understanding, particularly those relying on discrete categories, are inherently biased. In this paper, we present a comparative in-depth analysis of two common datasets (AffectNet and EMOTIC) equipped with the components of the circumplex model of affect. Further, we propose a model for the prediction of facial expressions tailored for lightweight applications. Using a small-scaled MaxViT-based model architecture, we evaluate the impact of discrete expression category labels in training with the continuous valence and arousal labels. We show that considering valence and arousal in addition to discrete category labels helps to significantly improve expression inference. The proposed model outperforms the current state-of-the-art models on AffectNet, establishing it as the best-performing model for inferring valence and arousal achieving a 7% lower RMSE. Training scripts and trained weights to reproduce our results can be found here: https://github.com/wagner-niklas/CAGE_expression_inference.

Albert Schotschneider, Svetlana Pavlitska, J. Marius Zöllner

Deep neural networks (DNNs) are widely used in perception systems for safety-critical applications, such as autonomous driving and robotics. However, DNNs remain vulnerable to various safety concerns, including generalization errors, out-of-distribution (OOD) inputs, and adversarial attacks, which can lead to hazardous failures. This survey provides a comprehensive overview of runtime safety monitoring approaches, which operate in parallel to DNNs during inference to detect these safety concerns without modifying the DNN itself. We categorize existing methods into three main groups: Monitoring inputs, internal representations, and outputs. We analyze the state-of-the-art for each category, identify strengths and limitations, and map methods to the safety concerns they address. In addition, we highlight open challenges and future research directions.

Svetlana Pavlitska, Oliver Bagge, Federico Peccia et al.

Model compression and hardware acceleration are essential for the resource-efficient deployment of deep neural networks. Modern object detectors have highly interconnected convolutional layers with concatenations. In this work, we study how pruning can be applied to such architectures, exemplary for YOLOv7. We propose a method to handle concatenation layers, based on the connectivity graph of convolutional layers. By automating iterative sensitivity analysis, pruning, and subsequent model fine-tuning, we can significantly reduce model size both in terms of the number of parameters and FLOPs, while keeping comparable model accuracy. Finally, we deploy pruned models to FPGA and NVIDIA Jetson Xavier AGX. Pruned models demonstrate a 2x speedup for the convolutional layers in comparison to the unpruned counterparts and reach real-time capability with 14 FPS on FPGA. Our code is available at https://github.com/fzi-forschungszentrum-informatik/iterative-yolo-pruning.

Marc Uecker, J. Marius Zöllner

Recently, LiDAR segmentation methods for autonomous vehicles, powered by deep neural networks, have experienced steep growth in performance on classic benchmarks, such as nuScenes and SemanticKITTI. However, there are still large gaps in performance when deploying models trained on such single-sensor setups to modern vehicles with multiple high-resolution LiDAR sensors. In this work, we introduce a new metric for feature-level invariance which can serve as a proxy to measure cross-domain generalization without requiring labeled data. Additionally, we propose two application-specific data augmentations, which facilitate better transfer to multi-sensor LiDAR setups, when trained on single-sensor datasets. We provide experimental evidence on both simulated and real data, that our proposed augmentations improve invariance across LiDAR setups, leading to improved generalization.

Daniel Bogdoll, Finn Sartoris, Vincent Geppert et al.

Autonomous vehicles drive millions of miles on the road each year. Under such circumstances, deployed machine learning models are prone to failure both in seemingly normal situations and in the presence of outliers. However, in the training phase, they are only evaluated on small validation and test sets, which are unable to reveal model failures due to their limited scenario coverage. While it is difficult and expensive to acquire large and representative labeled datasets for evaluation, large-scale unlabeled datasets are typically available. In this work, we introduce label-free model failure detection for lidar-based point cloud segmentation, taking advantage of the abundance of unlabeled data available. We leverage different data characteristics by training a supervised and self-supervised stream for the same task to detect failure modes. We perform a large-scale qualitative analysis and present LidarCODA, the first publicly available dataset with labeled anomalies in real-world lidar data, for an extensive quantitative analysis.

Daniel Bogdoll, Meng Zhang, Maximilian Nitsche et al.

Great progress has been achieved in the community of autonomous driving in the past few years. As a safety-critical problem, however, anomaly detection is a huge hurdle towards a large-scale deployment of autonomous vehicles in the real world. While many approaches, such as uncertainty estimation or segmentation-based image resynthesis, are extremely promising, there is more to be explored. Especially inspired by works on anomaly detection based on image resynthesis, we propose a novel approach for anomaly detection through style transfer. We leverage generative models to map an image from its original style domain of road traffic to an arbitrary one and back to generate pixelwise anomaly scores. However, our experiments have proven our hypothesis wrong, and we were unable to produce significant results. Nevertheless, we want to share our findings, so that others can learn from our experiments.

Svetlana Pavlitska, Beyza Keskin, Alwin Faßbender et al.

Estimating accurate and well-calibrated predictive uncertainty is important for enhancing the reliability of computer vision models, especially in safety-critical applications like traffic scene perception. While ensemble methods are commonly used to quantify uncertainty by combining multiple models, a mixture of experts (MoE) offers an efficient alternative by leveraging a gating network to dynamically weight expert predictions based on the input. Building on the promising use of MoEs for semantic segmentation in our previous works, we show that well-calibrated predictive uncertainty estimates can be extracted from MoEs without architectural modifications. We investigate three methods to extract predictive uncertainty estimates: predictive entropy, mutual information, and expert variance. We evaluate these methods for an MoE with two experts trained on a semantical split of the A2D2 dataset. Our results show that MoEs yield more reliable uncertainty estimates than ensembles in terms of conditional correctness metrics under out-of-distribution (OOD) data. Additionally, we evaluate routing uncertainty computed via gate entropy and find that simple gating mechanisms lead to better calibration of routing uncertainty estimates than more complex classwise gates. Finally, our experiments on the Cityscapes dataset suggest that increasing the number of experts can further enhance uncertainty calibration. Our code is available at https://github.com/KASTEL-MobilityLab/mixtures-of-experts/.

Svetlana Pavlitska, Enrico Eisen, J. Marius Zöllner

Vulnerability to adversarial attacks is a well-known deficiency of deep neural networks. Larger networks are generally more robust, and ensembling is one method to increase adversarial robustness: each model's weaknesses are compensated by the strengths of others. While an ensemble uses a deterministic rule to combine model outputs, a mixture of experts (MoE) includes an additional learnable gating component that predicts weights for the outputs of the expert models, thus determining their contributions to the final prediction. MoEs have been shown to outperform ensembles on specific tasks, yet their susceptibility to adversarial attacks has not been studied yet. In this work, we evaluate the adversarial vulnerability of MoEs for semantic segmentation of urban and highway traffic scenes. We show that MoEs are, in most cases, more robust to per-instance and universal white-box adversarial attacks and can better withstand transfer attacks. Our code is available at \url{https://github.com/KASTEL-MobilityLab/mixtures-of-experts/}.

Svetlana Pavlitska, Leopold Müller, J. Marius Zöllner

Adversarial attacks on traffic sign classification models were among the first successfully tried in the real world. Since then, the research in this area has been mainly restricted to repeating baseline models, such as LISA-CNN or GTSRB-CNN, and similar experiment settings, including white and black patches on traffic signs. In this work, we decouple model architectures from the datasets and evaluate on further generic models to make a fair comparison. Furthermore, we compare two attack settings, inconspicuous and visible, which are usually regarded without direct comparison. Our results show that standard baselines like LISA-CNN or GTSRB-CNN are significantly more susceptible than the generic ones. We, therefore, suggest evaluating new attacks on a broader spectrum of baselines in the future. Our code is available at \url{https://github.com/KASTEL-MobilityLab/attacks-on-traffic-sign-recognition/}.

Abhishek Vivekanandan, J. Marius Zöllner

Representing diverse and plausible future trajectories is critical for motion forecasting in autonomous driving. However, efficiently capturing these trajectories in a compact set remains challenging. This study introduces a novel approach for generating scene-specific trajectory sets tailored to different contexts, such as intersections and straight roads, by leveraging map information and actor dynamics. A deterministic goal sampling algorithm identifies relevant map regions, while our Recursive In-Distribution Subsampling (RIDS) method enhances trajectory plausibility by condensing redundant representations. Experiments on the Argoverse 2 dataset demonstrate that our method achieves up to a 10% improvement in Driving Area Compliance (DAC) compared to baseline methods while maintaining competitive displacement errors. Our work highlights the benefits of mining such scene-aware trajectory sets and how they could capture the complex and heterogeneous nature of actor behavior in real-world driving scenarios.

Annemarie Hoffsommer, Helen Schneider, Svetlana Pavlitska et al.

Accurately predicting emotions from brain signals has the potential to achieve goals such as improving mental health, human-computer interaction, and affective computing. Emotion prediction through neural signals offers a promising alternative to traditional methods, such as self-assessment and facial expression analysis, which can be subjective or ambiguous. Measurements of the brain activity via electroencephalogram (EEG) provides a more direct and unbiased data source. However, conducting a full EEG is a complex, resource-intensive process, leading to the rise of low-cost EEG devices with simplified measurement capabilities. This work examines how subsets of EEG channels from the DEAP dataset can be used for sufficiently accurate emotion prediction with low-cost EEG devices, rather than fully equipped EEG-measurements. Using Continuous Wavelet Transformation to convert EEG data into scaleograms, we trained a vision transformer (ViT) model for emotion classification. The model achieved over 91,57% accuracy in predicting 4 quadrants (high/low per arousal and valence) with only 12 measuring points (also referred to as channels). Our work shows clearly, that a significant reduction of input channels yields high results compared to state-of-the-art results of 96,9% with 32 channels. Training scripts to reproduce our code can be found here: https://gitlab.kit.edu/kit/aifb/ATKS/public/AutoSMiLeS/DEAP-DIVE.

Marius Dähling, Sebastian Krebs, J. Marius Zöllner

This paper introduces a novel method for end-to-end crowd detection that leverages object density information to enhance existing transformer-based detectors. We present CrowdQuery (CQ), whose core component is our CQ module that predicts and subsequently embeds an object density map. The embedded density information is then systematically integrated into the decoder. Existing density map definitions typically depend on head positions or object-based spatial statistics. Our method extends these definitions to include individual bounding box dimensions. By incorporating density information into object queries, our method utilizes density-guided queries to improve detection in crowded scenes. CQ is universally applicable to both 2D and 3D detection without requiring additional data. Consequently, we are the first to design a method that effectively bridges 2D and 3D detection in crowded environments. We demonstrate the integration of CQ into both a general 2D and 3D transformer-based object detector, introducing the architectures CQ2D and CQ3D. CQ is not limited to the specific transformer models we selected. Experiments on the STCrowd dataset for both 2D and 3D domains show significant performance improvements compared to the base models, outperforming most state-of-the-art methods. When integrated into a state-of-the-art crowd detector, CQ can further improve performance on the challenging CrowdHuman dataset, demonstrating its generalizability. The code is released at https://github.com/mdaehl/CrowdQuery.

Svetlana Pavlitska, Haixi Fan, Konstantin Ditschuneit et al.

Robustifying convolutional neural networks (CNNs) against adversarial attacks remains challenging and often requires resource-intensive countermeasures. We explore the use of sparse mixture-of-experts (MoE) layers to improve robustness by replacing selected residual blocks or convolutional layers, thereby increasing model capacity without additional inference cost. On ResNet architectures trained on CIFAR-100, we find that inserting a single MoE layer in the deeper stages leads to consistent improvements in robustness under PGD and AutoPGD attacks when combined with adversarial training. Furthermore, we discover that when switch loss is used for balancing, it causes routing to collapse onto a small set of overused experts, thereby concentrating adversarial training on these paths and inadvertently making them more robust. As a result, some individual experts outperform the gated MoE model in robustness, suggesting that robust subpaths emerge through specialization. Our code is available at https://github.com/KASTEL-MobilityLab/robust-sparse-moes.

Svetlana Pavlitska, Jamie Robb, Nikolai Polley et al.

Realistic adversarial attacks on various camera-based perception tasks of autonomous vehicles have been successfully demonstrated so far. However, only a few works considered attacks on traffic light detectors. This work shows how CNNs for traffic light detection can be attacked with printed patches. We propose a threat model, where each instance of a traffic light is attacked with a patch placed under it, and describe a training strategy. We demonstrate successful adversarial patch attacks in universal settings. Our experiments show realistic targeted red-to-green label-flipping attacks and attacks on pictogram classification. Finally, we perform a real-world evaluation with printed patches and demonstrate attacks in the lab settings with a mobile traffic light for construction sites and in a test area with stationary traffic lights. Our code is available at https://github.com/KASTEL-MobilityLab/attacks-on-traffic-light-detection.

Daniel Bogdoll, Felix Schreyer, J. Marius Zöllner

Autonomous driving is among the largest domains in which deep learning has been fundamental for progress within the last years. The rise of datasets went hand in hand with this development. All the more striking is the fact that researchers do not have a tool available that provides a quick, comprehensive and up-to-date overview of data sets and their features in the domain of autonomous driving. In this paper, we present ad-datasets, an online tool that provides such an overview for more than 150 data sets. The tool enables users to sort and filter the data sets according to currently 16 different categories. ad-datasets is an open-source project with community contributions. It is in constant development, ensuring that the content stays up-to-date.

Yin Wu, Daniel Slieter, Carl Esselborn et al.

Deploying ADAS and ADS across countries remains challenging due to differences in legislation, traffic infrastructure, and visual conventions, which introduce domain shifts that degrade perception performance. Traditional cross-country data collection relies on extensive on-road driving, making it costly and inefficient to identify representative locations. To address this, we propose a street-view-guided data acquisition strategy that leverages publicly available imagery to identify places of interest (POI). Two POI scoring methods are introduced: a KNN-based feature distance approach using a vision foundation model, and a visual-attribution approach using a vision-language model. To enable repeatable evaluation, we adopt a collect-detect protocol and construct a co-located dataset by pairing the Zenseact Open Dataset with Mapillary street-view images. Experiments on traffic sign detection, a task particularly sensitive to cross-country variations in sign appearance, show that our approach achieves performance comparable to random sampling while using only half of the target-domain data. We further provide cost estimations for full-country analysis, demonstrating that large-scale street-view processing remains economically feasible. These results highlight the potential of street-view-guided data acquisition for efficient and cost-effective cross-country model adaptation.

Daniel Bogdoll, Iramm Hamdard, Lukas Namgyu Rößler et al.

The scale-up of autonomous vehicles depends heavily on their ability to deal with anomalies, such as rare objects on the road. In order to handle such situations, it is necessary to detect anomalies in the first place. Anomaly detection for autonomous driving has made great progress in the past years but suffers from poorly designed benchmarks with a strong focus on camera data. In this work, we propose AnoVox, the largest benchmark for ANOmaly detection in autonomous driving to date. AnoVox incorporates large-scale multimodal sensor data and spatial VOXel ground truth, allowing for the comparison of methods independent of their used sensor. We propose a formal definition of normality and provide a compliant training dataset. AnoVox is the first benchmark to contain both content and temporal anomalies.