Differential DNA methylation patterns can serve as biomarkers for allergic diseases such as pediatric asthma and rhinitis, but age-dependent variability in epigenetic profiles undermines the reliability of predictive models. This thesis addresses that challenge by introducing a graph-based deep learning approach for \textbf{age-independent} allergy prediction from DNA methylation data. Each subject’s DNA methylation profile is represented as an individualized graph constructed via an extended Weighted Gene Co-expression Network Analysis (WGCNA) that captures global co-methylation structure and subject-specific patterns, thus balancing population-level relationships with individual epigenetic heterogeneity. Edges between CpG sites are assigned weights using a Gaussian kernel on methylation values, ensuring the graph reflects personalized similarity while maintaining biologically meaningful connections. A Graph Neural Network (GNN) with an Edge Convolution (EdgeConv) architecture is then trained on these subject-specific graphs to predict allergy outcomes. We evaluated this framework on DNA methylation data from three harmonized pediatric cohorts (PIAMA, MAKI, COPSAC) processed with the MEFFIL pipeline for cross-cohort normalization and quality control. An Epigenome-Wide Association Study (EWAS) identified key CpG features associated with asthma, rhinitis and IgE, which were used to guide feature selection for model training. Our graph-based model outperformed conventional methods like ElasticNet and XGBoost in certain cohorts and maintained robust predictive accuracy between the ages of 6 and 16, demonstrating a certain resilience to age-related methylation differences. Furthermore, we applied gradient-based saliency analysis to the trained GNN to highlight influential methylation features, providing interpretability and revealing plausible epigenetic markers of allergy. The proposed pipeline is scalable and interpretable, and its ability to deliver reliable, age-invariant risk predictions from early-life epigenetic data underscores its potential clinical utility for early allergy diagnostics in children.

Text-to-image (T2I) diffusion models have achieved remarkable image quality but still struggle to produce images that align with the compositional information from the input text prompt, especially when it comes to spatial cues. We attribute this limitation to two key factors: the lack of clear fine-grained spatial supervision in common training datasets, and the inability of the CLIP text encoder, used in the pretraining of stable diffusion models, to represent spatial semantics. While recent work has addressed object omission and attribute mismatches, accurately generating objects in the spatial locations defined in the text prompt remains an open challenge. Prior solutions typically rely on fine-tuning, which introduces computational overhead and risks degrading the pretrained model’s generative prior on other tasks unrelated to spatial reasoning. In this paper, we introduce InfSplign, a simple and training-free method that improves spatial understanding in T2I diffusion models. InfSplign leverages attention maps and a centroid-based loss to guide object placement during sampling at inference time without modifying the pretrained model. Our approach is modular, lightweight and compatible with any pretrained diffusion model. InfSplign achieves strong performance on spatial benchmarks such as VISOR, T2I-CompBench and GenEval, outperforming baselines in many scenarios.

Visual counting is an important task in computer vision with broad applications in areas such as crowd monitoring, agriculture, and environmental analysis. While deep learning has significantly advanced this field by enabling models to learn robust feature representations, deep learning approaches suffer from sensitivity to data imbalances, which occur in the distribution of object counts across counting datasets as a result of annotation effort. Most state-of-the-art counting models, categorized into clustering-, detection-, regression-, and density estimation-based methods, are built upon Convolutional Neural Networks (CNNs) and Transformers, both of which are known to be susceptible to imbalances in the training data. This study introduces a hybrid model that incorporates a programmatically guaranteed counting mechanism using the RASP language and the Tracr compiler, enabling the construction of Transformer-based models that can reliably execute predefined tasks, such as counting. By combining this exact counting mechanism with a trainable embedding module, we present a model that is capable of learning to count various tokens, even under significant data imbalance. We validate our approach on a synthetic, imbalanced dataset and compare its performance, training time, and data efficiency against standard CNN- and Transformer-based models. Results suggest that our method achieves strong generalization across the full spectrum of object counts while requiring less training data, highlighting the potential for this architecture to be further investigated and adapted to be used for robust and efficient visual counting.

Accurate sensor calibration is a critical challenge in the development of automated vehicles, especially in dynamic and modular sensor configurations. Traditional target-based methods, while precise, are limited in scalability and adaptability. In this work, we propose a modular, targetless, ego-motion-based calibration framework for multi-modal sensors, including a monocular camera, LiDAR, and 4D radar. The framework leverages odometry trajectories for extrinsic calibration, incorporating temporal alignment, trajectory scaling, and both pairwise and joint optimization techniques to achieve robust and accurate sensor alignment. Experimental validation using the View-of-Delft (VoD) dataset demonstrates the framework’s
robustness across diverse sensor setups, adaptability to real-world conditions. Our results underscore the potential of scalable, targetless calibration approaches to enhance the reliability and flexibility of automated systems, supporting implementation in real-world scenarios.

To ensure safe operation of autonomous vehicles (AVs), trajectory planners should account for occlusions. These are areas invisible to the AV that might contain vehicles. Set-based methods can guarantee safety by calculating the reachable set, which is the set of possible states, for each potentially hidden vehicle. A recently published method proved in simulation experiments to reduce the cautiousness by reasoning about these occluded areas over time, assuming perfect input data. We present a novel algorithm that uses this reasoning and is applicable on a real AV with its accompanying uncertainties and imperfect sensor data. The uncertainties include sensor errors and noise, computation and communication delays and control errors in the trajectory following. This is achieved by modelling the error distributions and accounting for them in the calculations, where the confidence interval for each error is exposed as a setting. Experiments indicate that our algorithm can reduce the traversal time through an intersection by 2.2 seconds with reasoning. An ablation study of the different error measures shows that the errors in the construction of the field of view (FOV) limit the performance the most. Reducing the errors in the FOV construction is therefore the most important recommendation, besides making the method interaction-aware.

3D semantic understanding is essential for a wide range of robotics applications. Availability of datasets is a strong driver for research, and whilst obtaining unlabeled data is straightforward, manually annotating this data with semantic labels is time-consuming and costly. Recently, foundation models have facilitated open-set semantic segmentation, potentially aiding automatic labeling. However, these models have largely been limited to 2D images. This work introduces Label Any Pointcloud (LeAP), which leverages 2D Vision Foundation Models (VFMs) to automatically label 3D data with any set of classes in any kind of application. VFMs are used to create image labels for the desired classes which are then projected to 3D points. Using the Bayesian update, point-wise labels are combined into voxels to improve label consistency, and label points outside the camera fustrum. A novel Cross-Modal Self-Training (CM-ST) approach further enhances label quality. Through extensive experiments, we demonstrate that our method can generate high-quality 3D semantic labels across diverse fields without any manual 3D labeling. Models adapted to new application domains using our labels show up to 3.7× (12.9 → 47.1) mIoU improvement compared to the unadapted baselines. This ability to provide labels for any domain can help accelerate 3D perception research.

The training process of machine learning models for self-driving applications suffers from bottlenecks during loading and processing of LiDAR point clouds with large storage complexity.
Many studies aim to remedy this problem from an implementation perspective by developing efficient data loading and processing pipelines.
This study, on the other hand, explores an alternative approach by augmenting data representations to achieve lower storage complexity known as point cloud compression.

A broad analysis is presented on novel point cloud compression codecs using tensor decomposition methods.
Several point cloud representations and tensor decomposition methods are considered over a range of hyperparameter choices and compression values.
In order to assess the performance of the presented codecs: the compression rate, quality of the reconstruction, and time complexity is compared to the octree-based baseline model: TMC13.
Compared to the baseline model, the performance of the presented tensor decomposition-based codecs falls short. One of the presented codecs does notably outperform the others. This codec uses synthetic tensorization followed by sorting using z-location and decomposition using the TT-SVD algorithm.
Sorting by z-value isolates the ground plane, which is a dominant low-rank feature, which can effectively be decomposed using the TT-SVD algorithm yielding adequate results.

Several limitations of the presented tensor decomposition-based codecs are: the omission of bitwise compression on the factor matrices, and the trade-off between bitwise precision and truncation due to tensor decomposition.
Future work could improve in these areas along with considering the use of different heuristics and optimizing the tensor network topology.

The ever-increasing complexity of Artificial Intelligence (AI) models has led to environmental challenges due to high computation and energy demands. This thesis explores the application of tensor decomposition methods—CP, Tucker, and TT—to improve the energy efficiency of large Convolutional Neural Networks (CNNs) during inference by reducing energy consumption. The energy consumption of several convolution layers was measured using a watt meter across various CNN configurations and different hardware architectures (Central Processing Unit(CPU) and Graphics Processing Unit (GPU)). In addition, several regression models were fitted to estimate energy savings, incorporating memory usage. It was found that TT decomposition consistently provided the most significant energy savings across various compression ratios, influenced by CNN hyperparameters such as input/output channels, feature sizes, and kernel sizes, whereas CP decomposition was the least effective in reducing energy. The GPU implementations generally resulted in additional energy consumption, and the GPU regression models suggested a need for more complex relationships. The thesis also revealed that the efficiency of tensor decompositions might be highly dependent on the implementation details of software libraries, such as TensorlyTorch, which can significantly impact the computation and memory complexities. These findings underscore the importance of both hardware specific considerations and careful software implementation in achieving energy efficient CNNs, providing a foundation for further research in energy-constrained environments.

Humans are our best example of the ability to learn a structure of the world through observation of environmental regularities. Specifically, humans can learn about different objects, different classes of objects, and different class-specific behaviors. Fundamental to these human abilities is evolved sensory hardware and automatic pattern recognition systems thought to be powered in part by the leading neuroscience theory of predictive coding. Artificial intelligence research is often inspired by neuroscience and algorithms already exist that implement predictive coding. In this paper, we seek to evaluate a leading predictive coding video-prediction algorithm, PredNet, for its ability to perform human-like learning of the types mentioned. By successfully training PredNet on a custom Simple Shape Motion (SSM) video dataset that explicitly requires structure learning to occur in order to accurately predict the next frame, we establish that PredNet is capable of rudimentary structure learning. We investigate PredNet filters and feature maps but find scant evidence of truly symbolic knowledge, and propose instead that PredNet performs semi-symbolic learning. We perform ablation studies that reveal the aspects of PredNet that critically contribute to its structure learning ability. Finally, we detail a set of modifications made to PredNet to allow object-centric processing as a promising step change towards human-like structure learning. Evaluation results and investigations are provided. Performance was slightly worse than Baseline, likely due to a noted implementation flaw. Code and instructions to reproduce dataset creation and model training / evaluation are available at https://github.com/ofSingularMind/parallel_prednet.

Human Mesh Recovery (HMR) frameworks predict a comprehensive 3D mesh of an observed human based on sensor measurements. The majority of these frameworks are purely image-based. Despite the richness of this data, image-based HMR frameworks are vulnerable to depth ambiguity, resulting in mesh inaccuracies in 3D space. Several HMR frameworks in the literature use LiDAR data to avoid this depth ambiguity. However, these frameworks are either only LiDAR-based, which limits performance due to LiDAR sparseness and limited training data, or they are model-free, making the resulting meshes vulnerable to artifacts and limited detail. Therefore, this work introduces SMPLify-3D, an optimization-based HMR framework that combines the richness of image data and the depth information within sparse LiDAR data to improve the 3D mesh inaccuracies of image-based HMR frameworks. The proposed framework consists of three main steps: 1) a body part visibility filter based on the 2D detected keypoints, 2) rough alignment between the mesh and the observed LiDAR point cloud using the ICP algorithm, and 3) an optimization scheme, inspired by SMPLify, that modifies the actual pose and shape of the mesh to improve both the 3D and image alignment. SMPLify-3D is versatile compared to other methods and outperforms image-based and SMPL-compatible LiDAR-based HMR frameworks by improving the Per-Vertex-Error (PVE) with 45% and 26% on the 3DPW and HumanM3 datasets respectively. Multiple quantitative experiments are conducted to show the effects of LiDAR noise and sparsity on the framework’s performance. Additionally, qualitative results illustrate how the proposed method achieves superior results on out-of-sample data recorded by a mobile robot. The source code for this work is available at: https://github.com/guidodumont/SMPLify-3D

Forensic microtrace investigation relies on a time- and labour-intensive process of manually analysing samples via microscopy. To aid forensic experts in their investigations, an image recognition model for microtrace localisation and classification is needed. This work investigates the trace recognition accuracy that can be achieved by analysing images captured with automated microscopy through deep learning. Fibres, hairs, skin, glass and sand are pixel-wise classified in microscopy scans of tape-lift samples. As deep learning requires extensive amounts of annotated training data, pretraining is investigated to minimise the required annotation workload. ImageNet pretraining, pretraining with self-supervised learning and a sequential application of these approaches are tested. It is found that pretrained models are able to reduce the required annotated data twofold compared to models trained from scratch, while retaining the prediction accuracy. While the ImageNet-pretrained models outperform the self-supervised-pretrained models, the highest accuracy is achieved by combining the two approaches. With this, traces are recognised with a mean intersection over union of 0.56 when training on only 2.2 dm2 of annotated tape lift scans.

Monitoring wildfires using multiple unmanned aerial vehicles (UAVs) is essential for timely intervention and management of the fire while minimising risks to human lives. A research gap was identified for practical UAV solutions integrating critical features, such as localisation, fire safety and battery management. This thesis presents a novel rule-based algorithm designed for effective wildfire monitoring using cooperative UAVs. A simulation environment with a wildfire spread model was developed to test and evaluate various strategies under dynamic conditions. The proposed algorithm operates in two phases: first, locating the fire and then monitoring its progression. Key features include a heat avoidance mechanism to maintain UAV safety and recharging schedules to maximise operational uptime. A fire detection map is centrally created using the infrared data from the UAVs. The convex hull around the fire serves as the basis for path planning to track the fire perimeter. The best strategy is determined and optimised through a series of experiments in the simulation environment. The overall best-performing algorithm makes navigational decisions based on the importance of different sections of the fire perimeter. The importance of the perimeter is constantly updated based on the fire spread, visit times, and distances of the UAVs. The assignment of various UAV tasks was reviewed to enhance UAV cooperation, which showed performance benefits across different scenarios. Other key findings
include the recommendation of a small fleet size of up to six UAVs, with a flight speed of 10 m/s at an altitude of 100 meters, balancing costs, performance and safety. The algorithm demonstrated performance equal to state-of-the-art reinforcement learning techniques
while offering advantages in explainability. Additionally, the algorithm has been successfully validated in a lab environment, demonstrating its potential as a practical and cost-effective solution for wildfire monitoring. This work brings a fire monitoring system closer to real-world implementation and will possibly help fight wildfires effectively.

Unsupervised 3D object detection methods can reduce the reliance on human-annotations by leveraging raw sensor data directly for supervision. Recent approaches combine density-based spatial clustering with motion and appearance cues to extract object proposals from the scene, which serve as pseudo-annotations. However, density-based methods struggle with the uneven data densities seen in LiDAR point clouds, and fail to distinguish between foreground and background objects effectively. To address this issue, this thesis introduces MobileClusterNet, a learnable framework designed for 3D spatial clustering. MobileClusterNet incorporates a novel loss module which utilizes appearance embeddings alongside scene flow information, thereby learning to generate high-quality clusters consisting of both static and dynamic mobile objects. Annotations generated by MobileClusterNet can be used for training any existing supervised detector, without the need for extensive self-training. Experimental results on the Waymo Open Dataset demonstrate that MobileClusterNet outperforms traditional density-based methods like HDBSCAN in clustering performance by a large margin, and provides high quality proposals for training supervised detectors.

Loop Robots develops and operates the next generation of fully autonomous disinfection robots in hospitals and healthcare settings. Accurate localization is essential in order to navigate reliably and effectively disinfect the tight hallways and corners of a patient room, operating room, or intensive care unit in a hospital. Odometry, or dead-reckoning, is the process of estimating the robots pose with respect to a known initial pose. It forms the backbone of the robots localization strategy, as well as being critical to many other processes that run on the robot, such as control and mapping.
The current strategy of generating odometry using the wheel encoders is prone to drift due to the integration of errors into the estimate. Moreover, the estimation takes place on the horizontal plane due to the nature of the wheel encoders. To improve the odometry and extend it to the full 3D space, a solution that makes use of all of the onboard sensors in a tightly-coupled manner is required.
In this Thesis, we make the first steps towards this larger goal. The contributions of this thesis include: (1) an Extended Kalman Filter (EKF) algorithm to fuse data from the Inertial Measurement Unit (IMU) and wheel encoders to estimate position and orientation of the robot in 6-DoF; (2) a line feature extraction and tracking methodology to extract primitives from LIDAR data and (3) A Moving Horizon Estimation (MHE) scheme based on a factor-graph formulation to perform pose estimation on the horizontal plane using LIDAR data and wheel encoders.
We test the three modules individually using a combination of simulations and real-world data wherever possible. We found that the MHE scheme was able to reduce drift over the long term, but is sensitive to the effects of outliers in feature matching, motion distortion of LIDAR scans, and wheel slip. The EKF scheme is able to reduce the overall drift and correct for wheel slips.
Based on these results, promising avenues for the improvement of all the proposed modules are given, along with recommendations on how to combine them all in a tightly-coupled fashion.

Visual inspection of liquid medicine containers for contamination and defects is mandatory and crucial to ensure their safety for injection. This document presents research and development of three modules of the Visual Particle Inspection Subsystem (VPIS), an automatic inspection subsystem with the task of detecting and classifying particle contamination in liquid medicine containers. The proposed VPIS comprises three modules for which research and development is performed separately: the backgrounds subtraction and segmentation module, the classification module, and the tracking module.

For the background subtraction and segmentation module, a solution is proposed with filtered temporal background modelling and locally adaptive threshold segmentation. This solution works on the assumption that moving objects such as particles and bubbles are sparse and do not occur more than twice in a pixel in twenty frames. Therefore, a representative incomplete background cluster can be obtained by removing the two lowest or highest pixel values depending on the mode of illumination. Next, the average value of the background cluster is used for background subtraction and the standard deviation of the background cluster is used to determine a locally adaptive segmentation threshold. The solution has been positively evaluated for detection of low contrast objects, insensitivity to disturbances, processing time, and parameter configuration. Additionally, an add-on solution is proposed that makes it possible to detect objects in the challenging region near the rubber stopper of a syringe.

For the classification module, out of four candidate classification methods a Convolutional Neural Network (CNN) is chosen for the classification of single object detections. The CNN classifier achieves an accuracy of 0.93 and is used as a baseline classifier for the rest of this research. Next, with three exploratory research questions, research is performed into opportunities and pitfalls for this classification problem summarised below.
-It is found that certain handcrafted features correlate to the classification accuracy of a detection. A method is proposed to predict the classification accuracy of detections and filter out detections that cannot be classified reliably. Compared simple filters that filter out detections based on a single feature such as area or total contrast, the proposed method should achieve a higher subsequent classification accuracy and reject less detections.
-With the simplest classification strategy that is often used for inspection, a container will be rejected if a single detection is classified as a particle. It is shown through simulation that this classification strategy is not effective for this classification problem. This is as the large number of bubble detections in a clean container would result in a high false container rejection rate.
-Two multi-detection classification strategies are proposed that use multiple detections for classification. Median voting classification uses results from the tracking module to classify multiple detections of the same object. Multi-positive classification does not use tracking results but requires multiple detections to be classified as particles. Simulation results indicate that with both strategies, the classification accuracy for containers can be drastically improved. It is theorized that the best results can be achieved with median voting classification...

Understanding traffic participants’ behaviour is crucial for predicting their future trajectories, enabling autonomous vehicles to better assess the environment and consequently anticipate possible dangerous situations at an early stage. While the integration of cognitive processes and machine learning models has demonstrated promise in various domains, its application in trajectory forecasting of multiple traffic agents in large-scale autonomous driving datasets remains lacking. This work investigates the state-of-the-art trajectory forecasting model Trajectron++ which we enhance by incorporating a smoothing term in its attention module. This attention mechanism mimics human attention inspired by cognitive science research indicating limits to attention switching. We evaluate the performance of the resulting Smooth- Trajectron++ model and compare it to the original model on various benchmarks. Our results show improved performance on the large-scale nuScenes dataset, revealing the potential of incorporating insights from human cognition into trajectory prediction models.

Tracker-level fusion (TLF) is recognized as an effective approach to comprehensively improve visual object tracking performance by combining the capabilities of multiple baseline trackers. Although there is considerable interest in TLF, there are still challenges related to insufficient understanding, high cost, and unstable performance that make studying TLF difficult. In this thesis, I begin with an explicit summary of the overall pipeline of TLF which has significant guidance for TLF study. Additionally, I conduct a deep analysis of the positive and negative effects of baseline trackers to fully understand their influence on TLF. For visual object tracking, I propose three TLF frameworks based on three Distributed Kalman filters which are optimized for different scenarios and enable the fusion of different baseline trackers to enhance tracking performance. My TLF frameworks fuse the tracking results of different baseline trackers based on the principle of minimal trace to produce fusion results.Additionally, they exhibit superior and stable performance with general baseline tracker requirements, while also being simple, online, and real-time.Furthermore, the proposed frameworks can benefit from state-of-the-art baseline trackers over time, which will further improve their tracking performance. The proposed analysis and frameworks are studied extensively on three challenging benchmarks: generic tracking OTB2015, short-term tracking GOT-10k, and long-term tracking LaSOT. At over 240 FPS, the state-of-the-art success AUC score of 72.7% is achieved on OTB2015.

In the literature, neural network compression can significantly reduce the number of floating-point operations (FLOPs) of a neural network with limited accuracy loss. At the same time, it is common to manually design smaller networks instead of using modern compression techniques. This thesis will compare the two approaches for the object detection network YOLOv7. YOLOv7 can run in real time on a desktop GPU. For edge GPUs a smaller version, called YOLOv7-tiny, was manually designed by the authors of YOLOv7. This thesis answers the question: Can a state-of-the-art compression of YOLOv7 achieve higher accuracy than YOLOv7-tiny at the same number of floating-point operations?
First, two state-of-the-art compression methods are selected and compared on YOLOv7-tiny. Then the best performing method, GBIP, is used to compress YOLOv7 till it has the same number of FLOPs as YOLOv7-tiny. From the experiments it is determined that GBIP is not able to achieve higher accuracy than YOLOv7-tiny at the same number of FLOPs.

The rapid advancement in autonomous driving technology underscores the importance of studying the fragility of perception systems in autonomous vehicles, particularly due to their profound impact on public transportation safety. These systems are of paramount importance due to their direct impact on the lives of passengers and pedestrians. Additionally, their reliability can be easily compromised given the complexity and unpredictability of driving environments. However, current research and existing regulations often fail to adequately address the adversarial robustness of autonomous vehicle perception systems. This thesis delves into the adversarial robustness of camera-based perception systems of autonomous vehicles. Our research concentrates on developing and implementing evasion attacks that use black-box gradient estimation, as well as physical attacks in traffic sign detection and classification systems. Our findings indicate that even minor perturbations can impact the accuracy of these systems, leading to detection and classification errors. This finding highlights a critical vulnerability in the perception system's robustness against adversarial attacks. Moreover, the study extends to assess the transferability of adversarial examples across diverse perception models. Our results also expose significant gaps in the current regulatory frameworks of autonomous vehicles, necessitating the establishment of more rigorous and comprehensive safety standards.