Automotive radar has shown promising developments in environment perception due to its cost-effectiveness and robustness in adverse weather conditions. However, the limited availability of annotated radar data poses a significant challenge for advancing radar-based perception systems. To address this limitation, we propose a framework to generate 4D radar point clouds for training and evaluating object detectors. Specifically, we apply diffusion to a point-structured latent representation of radar point clouds. Our proposed 4DRad-Diffusion generates foreground and background points separately, conditioned on 3D bounding boxes and LiDAR data, respectively. The generated foreground points can be used as an effective synthetic data augmentation strategy or combined with generated background points to pre-train models on fully synthetic data. We demonstrate that augmenting real radar data with our synthetic data improves object detection performance on both the View-of-Delft and TruckScenes datasets, even outperforming existing augmentation methods. We also show that pre-training on synthetic data enhances performance, highlighting the potential of generative models to advance radar perception.

This thesis investigates whether traffic patterns captured by a cyclist's camera coincide with moments of rider-reported workload increases, and whether a simple, scalable pipeline using a single forward-facing camera can extract useful signals for workload modelling. We developed an end-to-end system that detects and tracks nearby road users in cyclist point-of-view video, computing traffic features from 984 temporal windows across three urban rides in the city of Delft.

Using a compact set of eight traffic features and logistic regression modelling, we evaluated performance through three complementary approaches. Individual feature analysis revealed consistent directional differences between low and high workload episodes, with six of eight features maintaining the same directional relationships in multivariate analysis. Ranking performance exceeded baseline expectations, achieving within-route ROC-AUC of 0.570 (0.070 above chance) and PR-AUC of 0.450 (0.080 above baseline), meeting established criteria for small effect sizes. Cross-route generalization proved more challenging, with performance dropping to ROC-AUC of 0.517 and PR-AUC of 0.401, though consistently exceeding route-specific baselines across all test routes.

Binary classification demonstrated above-baseline performance, with within-route F1 scores of 0.478 representing a 0.108 improvement over our baseline expectations. Threshold optimization increased F1 to 0.517 by improving recall from 55.0\% to 78.3\%, though cross-route binary performance remained more limited. The traffic patterns associated with high workload episodes point toward defensive cycling scenarios characterized by increased object density, complex motion patterns, and reduced cyclist velocity in dense urban environments.

These findings establish that camera-derived traffic metrics contain detectable workload-related information while highlighting significant constraints for practical deployment. The modest magnitude of detected relationships and limited cross-route transferability indicate that while the approach demonstrates feasibility, substantial development is needed before robust real-world application. This work provides quantitative evidence for traffic-workload relationships in cycling and establishes methodological foundations for future research in automated cycling safety monitoring.

Radar-based perception has been gaining traction in recent years, supported by improvements in deep learning techniques. Low-level radar perception focuses on utilizing the denser radar signal data instead of the conventional point-cloud. Despite the recent focus on this data representation, a lack of public datasets has limited the scope of research, especially for scene segmentation. In this paper, we address this challenge by recording a novel low-level radar dataset that includes diverse environments, sensors and complex scenarios. We propose Swin-FFM, a Swin transformer based network for free-road segmentation using the complex-valued range-Doppler signal. On our dataset, Swin-FFM achieves an IoU of 86.8%, demonstrating its ability to successfully output accurate free-road boundaries even in challenging settings. In addition to this, we compare our network with baselines for both low-level and point-cloud formats. Finally, we demonstrate the network’s ability to work with any low-level radar representation, highlighting its benefit for low-level radar perception.

Autonomous Surface Vehicles (ASVs) must operate safely in dynamic and often cluttered maritime environments such as ports and inland waterways. Achieving reliable situational awareness in these settings remains challenging due to the limited availability of annotated datasets, the complexity of multi-sensor alignment, and the degradation of monocular depth estimation at long range. This thesis addresses these challenges by developing a multi-modal 3D object detection framework that combines radar and camera data to improve perception robustness and depth accuracy.
To support supervised learning and evaluation, a custom dataset was constructed using real-world maritime sensor data, including radar, camera, and navigation information. Three-dimensional object annotations were manually created to capture static and moving targets across a wide range of distances.
Other core contributions include a probabilistic radar association strategy that models uncertainty in sensor measurements, a learnable Radar Gaussian Parameter Network (RGPNet) for dynamic estimation of radar association parameters, and a depth-aware mechanism for dynamically adjusting the region used for radar–camera fusion.
These methods are designed to improve object localisation performance, particularly in conditions where sensor data is sparse, noisy, or spatially misaligned.
The proposed models were evaluated through a series of controlled experiments examining detection performance across different depth intervals and clutter conditions. The results show that probabilistic radar fusion improves robustness to noise and depth estimation errors, while depth-aware association strategies enhance localisation at long range without sacrificing near-field precision. Incorporating adaptive mechanisms into the fusion process was shown to be particularly effective in scenarios involving large depth variation or limited visual information.
Overall, the findings demonstrate that uncertainty-aware, adaptive fusion methods can significantly improve 3D object detection performance for ASVs. The approaches developed in this work offer a robust foundation for future research on scalable, data-efficient maritime perception systems for autonomous navigation in complex environments.

Autonomous vessels offer potential benefits in safety, operational efficiency, and environmental impact, but require reliable perception systems to function independently. This thesis investigates the reliability of camera-based detection of maritime docks, a class of static but visually diverse objects that are often difficult to distinguish from their surroundings. While extensive research exists for detecting ships and buoys, docks remain underexplored, with no publicly available datasets containing a significant number of labeled instances.
To address this gap, the Dordrecht Dock Dataset was developed, containing 30,761 frames recorded under real-world maritime conditions across eight distinct docks. A custom interpolation-based annotation tool enabled efficient labeling, achieving annotation speeds of up to 3,000 frames per hour. Two deep learning models—Faster R-CNN and YOLO11n—were trained and evaluated on this dataset using a consistent pipeline. Evaluation followed a leave-one-dock-out cross-validation strategy with fixed test sets and standard performance metrics, including mAP50, mAP50–95, F1 score, and inference speed.
Faster R-CNN outperformed YOLO11n across nearly all accuracy metrics (mAP50 of 0.85 vs. 0.69), particularly at shorter ranges. YOLO11n demonstrated over twice the inference speed (37 FPS vs. 17 FPS), but exhibited weaker generalization and reduced reliability on visually different docks. Both models showed significant performance drops beyond 100–110 meters, emphasizing the need for high-resolution input or focus on short-range detection.

The vulnerability of cyclists, exacerbated by the rising popularity of faster e-bikes, motivates adapting automotive perception technologies for bicycle safety. This paper presents an online 3D LiDAR semantic segmentation pipeline developed using our multi-sensor ‘SenseBike’ research platform. To bridge the automotive-to-bicycle domain gap, we introduce the novel BikeScenes LidarSeg Dataset, comprising 3021 consecutive LiDAR scans around the TU Delft campus, semantically annotated for 29 dynamic and static classes. By evaluating model performance, we demonstrate that fine-tuning on our BikeScenes dataset achieves a mean Intersection-over-Union (mIoU) of 63.6%, significantly outperforming the 13.8% obtained with SemanticKITTI pre-training alone. This result underscores the necessity and effectiveness of domain-specific training. We highlight key challenges specific to bicycle-mounted, hardware-constrained perception systems and contribute the BikeScenes dataset as a resource for advancing research in cyclist-centric LiDAR segmentation.

Trustworthiness in Multilingual Large Language Mod- els (MLLMs) varies across languages, often explained by differences in pretraining resources. While this associa- tion with pretraining data is well established, we hypoth- esize that typological differences between languages also influence trustworthiness. We evaluate four similarly sized MLLMs (Aya Expanse 8B, Gemma-2 9B, Ministral 7B, and Llama-3.1 8B) on trustworthiness and analyze its correla- tions with typological features such as nominal, word or- der, verbal, negation, clause structure, phonology, token overlap, language family, and script. We find moderate to very strong correlations between specific typological fea- tures and performance, though these vary across feature- criterion pairs and are present only in English language pairs. Modeling these Z-scores confirms the importance of typology, significantly reducing error in five of eight tasks. Nominal, verbal, token overlap and word order emerge as the most consistent correlators with trustworthiness differ- ences. This exploratory work reveals an expansion of estab- lished cross-lingual typological influences and underscores the complexity of multilingual language modeling.

When optimising the placement of sensors on an Autonomous Vehicle (AV), research often uses evolutionary algorithms, offering a flexible way to explore complex solution spaces with multiple candidate configurations. However, this approach limits the ability to optimise one particular configuration directly. Sensor placement optimisation methods generally aim to optimise sensor-related aspects, such as visibility or coverage. When sensor placement is optimised based on the performance of a downstream task, this task performance is generally implicit, using surrogate metrics such as important task elements or probabilistic-related metrics. To address this, we propose to use gradient-descent-based optimisation in combination with a differentiable renderer. Making rendering differentiable allows gradients to flow between the initial rendering settings and the output rendered image, enabling gradient-based optimisation methods to optimise parameters based on the gradients of the objective function. This combination allows us to more directly optimise the placement of sensors on an associated downstream task. For this approach, we optimise the position of cameras on an AV based on 2D object detection performance. We create triangle mesh representations of traffic scenes from an AV simulator to use with the differentiable renderer. We use an objective function that combines optimisation losses related to in-frame rotational differences and visibility based on the detected object area of non-ego vehicles. This optimisation also considers constraints such as a minimum distance between each camera pair and camera positions that do not extend across the positioning bounds. The objective formulation results in an improved mean Average Precision (mAP) compared to a random sampling strategy and an intuitive baseline represented by cameras around the highest point on the ego vehicle.

This paper presents a comprehensive and quantitative framework for enhancing the safety of autonomous vehicles by integrating sensor detection performance with braking dynamics under extreme weather conditions and variable road slopes. Using high-fidelity simulations in CARLA and PyChrono, extreme weather datasets are generated, sensor detection distances are evaluated and braking distances are simulated under varying friction coefficients and slopes. This study quantifies the relationship between detection distance and safe braking performance, establishing objective metrics for the dynamic Operational Design Domain of AVs. Furthermore, an adaptive speed control strategy based on reinforcement learning — implemented via the Soft Actor-Critic algorithm — is proposed, which continuously adjusts vehicle speed to ensure that braking distances remain within sensor detection limits. Experimental results demonstrate that models trained with extreme weather data significantly outperform general-purpose detectors and that integrating real-time braking distance predictions into the control strategy improves safety margins and operational efficiency across diverse driving scenarios.

The recent explosion in research on 3D generative AI has shown that learning based editing methods can generate a wide range of shapes in an intuitive and low effort manner. However, current methods show limited local editing control due to the coarse shape representations they are based on, as well as shared or propagated geometry information propagating deformations to a larger region. Methods for theoretically unlimited fine editing exist, however they generally lack shape understanding, which is essential for reducing the effort involved in 3D shape editing. We propose a flexible shape representation in combination with an editing procedure to produce theoretically arbitrary fine local edits with learned shape understanding. Our work builds upon the learning based shape representation and generation method SPAGHETTI, which enables meaningful part mixing and interpolation. Our method enables the selection of a local surface region, which can be transferred through mixing or interpolation to a target shape guided by learned semantics. To empirically assess the locality of edits, we propose a new metric, which we use to conduct experiments on two baseline local shape mixing and interpolation methods. Our locality metric and surface displacement visualizations show that our method achieves more localized edits than the baseline methods, which yield significant deformation propagation.

Building registry updates are essential for urban planning but remain a labor-intensive process. This thesis introduces PolyChange, an adaptation of the PolyBuilding model, to automate mutation delineation by integrating aerial imagery with reference maps to produce precise, vectorized building mutations. PolyChange combines raster-based and direct polygonal representation methods, augmented with angle loss, double-thresholding, and label smoothing to enhance delineation accuracy. Experiments on our synthetic dataset demonstrate that raster-based methods achieve superior delineation, with an AR of 89.3\% and AP of 85.2\%. However, real-world datasets reveal challenges in generalization due to overfitting, semantic vagueness, and annotation inconsistencies, resulting in reduced detection and delineation capabilities, indicated by an AR of 30\% and AP of 13.7\% on the test set. Future work should address these limitations by improving dataset quality, refining model architectures, and leveraging reference maps more effectively to achieve fully automated and accurate building registry updates.

This paper presents a method that shows how a lidar-based 3D static environment can be constructed from a driving scenario and is used to aid in the creation of a digital twin from that scenario. Built with limited computational resources in mind, the resulting 3D static background reconstruction method is lightweight and nonetheless up to par with similar works, proving itself to be a viable alternative to similar methods. It uses ground truth labels and open-source, out-of-the-box working building blocks to create the pipeline that filters the lidar frames, performs registration between those frames, and meshes the resulting point cloud into a 3D mesh of the static background. We performed experiments in the Siemens Prescan ADAS Simulator of a reenactment of the traffic scenario in combination with our 3D static background and measured the inference domain gap using the Bidirectional Chamfer Distance and a real-data trained object detector, comparing both the synthetic and original data. Using Prescan as the simulator, AD models van be evaluated closed-loop. This work opens up the possibilities of applying sensor domain shifts to existing data or creating reality-based traffic scenarios from existing traffic scenarios and their corresponding 3D static background of edge cases that do not, or too little, exist in real-world data.

Automotive radars are increasingly used in automated driving systems due to their cost effectiveness, ease of integration, and ability to withstand adverse weather conditions. Semantic segmentation for radar point clouds is a crucial step in radar pre-processing, which can be used on almost all downstream tasks of radars, such as detection and tracking. Radar point clouds are noisier compared to LiDAR point clouds due to sensor noise and the multi-path propagation, which makes the segmentation task for radar more challenging. In this paper, we address the problem of segmentation in noisy radar point clouds in terms of ghost target vs. real detection, moving vs. static objects as well as semantic segmentation of moving road users. We demonstrate how these three tasks can be performed in a single, unified pipeline using an auto-labeled radar dataset. Our approach, called Real, Moving, and Semantic Segmentation Network (RMSNet), is able to output point-wise labels for all three tasks simultaneously. On our dataset, RMSNet attains an IoU of 82.5% for real detection segmentation, 66.9% IoU for moving object segmentation, and 64.9% mIoU for semantic segmentation. We also did a live demonstration, using the model on two different radars. The result shows that the model is capable of running inference in real-time and demonstrates good ability in model generalization.

Currently, self-driving vehicles have trouble detecting partially and fully occluded objects such as pedestrians, vehicles, and static obstacles. It has been proven that a drone surveilling the area around the vehicle improves the vehicle's awareness of its surroundings. This work explores planning strategies for the drone and evaluates how much the strategies assist the vehicle. Previous work proposed a metric called PKL, which describes the awareness of a vehicle as a function of the detections of the vehicle using a planner model. Subtracting this metric calculated using detections done by the drone from the metric without these detections results in an awareness improvement metric. This metric was used to evaluate drone positions and therefore the drone planning strategies. It was found that aiming to hover directly above the autonomous vehicle improves the awareness of the vehicle measured by median PKL improvement by 1.8%. Using only the relative position of the autonomous vehicle to the drone, the velocity of the vehicle in x and y directions and the number of vehicles visible to the drone, an imitation learning-based strategy performed the best with 11.6% median PKL improvement. If the drone's planner knows the future position of the autonomous vehicle or the positions of all vehicles in the area, the median PKL improvement is 74.5% and 80.8%, respectively. Optimizing the trajectories using a genetic algorithm further improves the performance to 96.7%. From these numbers, we can conclude that aiming to stay directly above the vehicle does not benefit the autonomous vehicle the most. We show that intelligent drone trajectory planning strategies can be learned which improve the awareness of the autonomous vehicle and therefore the safety of the people in and around this vehicle.

LiDAR technology is gaining popularity for use in 3D object detection, necessary for self-driving cars. However, due to class imbalances in state-of-the-art LiDAR datasets, detection algorithms often tend to lack performance in detecting cyclists. To address this issue, we introduce the \textit{SenseBike}, a LiDAR-equipped bicycle suited for collecting novel data, including more cyclists. We have created the \textit{SenseBike dataset}, which features distinctive data from the city of Delft, The Netherlands. Recording from a bicycle brings unique challenges, and we explain and evaluate our solutions to these issues.
To evaluate the impact of this new dataset on the performance of LiDAR object detection, we adapted an existing pseudo-labeling pipeline. Despite the recommendation, we did not self-train this pipeline, which would have resulted in higher quality pseudo-labels. Nonetheless, when we train CenterPoint, a well-known and fast 3D LiDAR object detector, on these lower-quality pseudo-labels, we still achieve an 85% Average Precision for cyclists, evaluated with maximum center-distance differences of 1m.

2D to 3D transfer learning has proven to be an effective method to transfer strong 2D representations into a 3D network. Recent advancements show that casting semantic segmentation of outdoor LiDAR point clouds as a 2D problem, such as through range projection, enables the processing of 3D point clouds with a Vision Transformer (ViT). For autonomous vehicles, collecting camera data alongside point cloud data is inexpensive. In this work, we investigate how we can best use these available 2D images, by exploring the benefit of camera pretraining to LiDAR semantic segmentation by using a shared ViT backbone. We propose a label generation method that generates pixel-wise pseudo labels for the camera images from the LiDAR annotations. We show that we can effectively use these for pretraining before finetuning on point cloud semantic segmentation. Besides, we compare jointly optimizing a shared ViT encoder for image and point cloud semantic segmentation simultaneously, with finetuning on point cloud semantic segmentation after pretraining on the camera domain. We show that joint optimization can lead to improved performance compared to pretraining when training from scratch. By using a shared encoder for the camera and LiDAR domain we can investigate the joint Camera-LiDAR feature space and find it is possible to create a shared feature space where LiDAR and camera features from the same class are mapped to the same location in the feature space. However, this does not contribute to a better performance on LiDAR semantic segmentation. These experiments provide valuable insight into 2D to 3D transfer learning and the creation of a shared Camera and LiDAR feature space.

Training models for autonomous vehicles (AVs) necessitates substantial volumes of high-quality data due to the strong correlation between dataset size and model performance. However, acquiring such datasets is labor-intensive and expensive, requiring significant resources for collection and labeling. To optimize the utility of available data, augmenting the dataset or generating synthetic data presents a cost-effective and efficient solution. Traditional methods that operate within the RGB domain frequently overlook crucial information, such as object frequency, scene composition, and agent trajectories. To address these limitations, a pipeline employing controllable diffusion models and vehicle simulation software is proposed. This approach involves loading collected data into a physics-based simulator, which allows for augmentation beyond the pixel space into the structural space. The augmented simulated data is subsequently transformed back into the photorealistic domain using generative artificial intelligence. This process generates high-fidelity synthetic data, thereby enabling models to train effectively on an expanded and varied dataset, enhancing robustness through the increased variation. The proposed method is evaluated in both image and video domains to assess its effectiveness.

Neural Radiance Fields (NeRFs) have showcased remarkable effectiveness in capturing complex 3D scenes and synthesizing novel viewpoints. By inherently capturing the entire scene in a compact representation, they offer a promising avenue for applications such as simulators, where efficient storage of real-world data, fast rendering and dynamic generation of new content are crucial. However, the potential for compression in NeRFs has been largely neglected in the existing literature. Moreover, the practical deployment of NeRFs in real-world scenarios, including simulators, faces significant obstacles such as constraints in training time, rendering speed, and scalability to large scenes. While recent advancements have tackled some of these hurdles individually, none have offered a comprehensive solution. In this paper, we introduce a new NeRF architecture based on a textured polygon-based method and augment this architecture by integrating encodings to expedite training. Additionally, we introduce learned pose refinement and an appearance embedding to enhance scalability to larger scenes. Through experimentation on the nuScenes dataset, we demonstrate that our method achieves competitive reconstruction performance with existing techniques while surpassing them in rendering speed. Furthermore, in terms of compression, our findings indicate that our method achieves competitive compression rates comparable to image-based compression techniques, while also enabling novel-view synthesis. This underscores its potential utility in applications like simulators.

Supercritical CO₂ (SCO₂) is a promising alternative to traditional working fluids in heat pumps and power cycles due to its high density, thermal efficiency, and stability. These properties allow for the design of more compact and efficient equipment. However, accurately modeling supercritical heat transfer, especially near its pseudocritical point, is challenging, because of extreme variations in its thermophysical properties. As a result of these inaccuracies in modeling, large fault margins must be taken into account, leading to overdesign of equipment.

Current methods for estimating heat transfer in such systems include empirical measurements, computational fluid dynamics (CFD), and Nusselt correlations. Although measurements provide accurate data, it is not a scalable solution. CFD methods offer this scalability. However, CFD cannot be applied in complex scenarios such as modeling of SCO₂ due to the trade-off between computational cost and accuracy. Nusselt correlations have low computational demand and are a scalable solution but come with low accuracy. Therefore, Nusselt correlations are not suitable for the design of equipment.

To overcome these limitations, this research applies a new model to predict heat transfer to SCO₂ flowing upward in a heated vertical tube. Current prediction methods use artificial neural networks (ANN). This research applies a convolutional neural network (CNN) for its ability to capture spatial context from the heat transfer trajectory. This property enables CNN to capture global and local patterns important for accurate prediction in SCO₂ heat transfer. In addition, the multiple kernels per layer enables the model to extract a large range of features from the heat transfer trajectory allowing for higher accuracy. The developed CNN model has shown superior performance, achieving an R² score of 0.970 and an MSE of 1477, outperforming existing ANNs and Nusselt correlations.

Furthermore, a feature importance study is done, to identify the nondimensional parameters that are important for heat transfer prediction The feature importance study highlighted the importance of parameters such as the Reynolds number and the Froude number in improving the model predictions. In addition, the feature importance study led to the development and identification of a new nondimensional parameter, (q_wk)/(T d), as one of the most important features influencing heat transfer. Furthermore, the experiments show that normalization is vital to enhance model stability and performance, countering issues such as exploding or vanishing gradients.

The findings in this research suggest great potential for using machine learning models to design more effective and compact heat transfer equipment. Future work will focus on a feature importance study for normalized, dimensional and nondimensional parameters for improved predictions. In addition, more synthetic data should be generated in the heat transfer detoriation and heat transfer enhancement areas for better prediction of those parts. Finally, a study into the application of machine learning models for designing heat transfer equipment should be done to show the benefits of such models.

Airlab, a collaboration between TU Delft and Ahold Delhaize, is developing Albert, a robot tailored to work in a complex supermarket environment. Key to Albert is a product detection and classification module that tells it what products to grasp and where they are located in a shelf. Albert’s existing YOLO‑based product detector a significant issue: Adding new products without re‑training the whole model is impossible. Especially in a dynamic supermarket environment with an ever‑changing stock, the latter is a major issue.

This problem will be the main focus of this paper and is addressed through few‑shot learning, which predicts similarity between query and target products. This simplifies adding new products to just supplying new target images. Few‑shot learning also requires significantly less data to train on. In supermarkets with 300.000 different products, requiring only a few images per product is a major advantage. For this reason, this paper aims to deploy a few‑shot model to classify products as either the target class or non‑target class for Albert’s picking task and defines the following research question: “What few‑shot classifier can identify products in a supermarket environment, is able to detect non‑target classes, and meets the requirements of deployment on a robotic platform like Albert best?“

This paper first analyses the potential of using TRIDENT and P>M>F, two state‑of‑the‑art few‑shot models, for deployment on Albert, and evaluates them on the requirements of this paper. P>M>F performs better on all requirements, which makes it the preferred model for Albert. However to work well, it still requires adjustments. Its inference time is still too high to work on Albert and it cannot classify query images as not the target product.

For this reason, this paper uses P>M>F’s two key ideas to construct Product‑ProtoNet, a new Albert‑suitable few‑shot model: 1) Using a good pre‑trained feature extractor; and 2) Comparing query images to a set of classes and matching only to the likeliest. P>M>F uses a ProtoNet model for classification that essentially does this; Like ProtoNet, Product‑ProtoNet constructs class prototypes from one or multiple examples of class images. Product‑ProtoNet then uses a sigmoid classifier to predict if query images have the same class as those prototypes. It compares query images to a set of similar class prototypes(helper prototypes) and classifies it as the likeliest. Product‑ProtoNet uses a ViT pre‑trained with DINO to extract image features. To bring down inference time, Product‑ProtoNet computes product prototypes before deployment.

With an accuracy of 99.1% on product classes seen during training and 99.8% on novel classes in a realistic supermarket setting, a low inference time of 2.89 ms and a memory usage lower than 4GB, Product‑ProtoNet is the only model that passes all requirements of this paper. When deployed on Albert Product‑ProtoNet successfully guides Albert to the right product in 97% of attempts. This makes Product‑ProtoNet the only few‑shot classifier that can identify products in a supermarket environment, is able to detect non‑target classes, and meets the requirements of deployment on a robotic platform like Albert.