Occupancy grid mapping (OGM) is a widely adopted perception technique that represents the surrounding environment as a discretized grid of occupied and free cells. In this thesis, we investigate the OGM problem using automotive point cloud data acquired from LiDAR, radar, and their fusion. Existing OGM approaches, including pattern-coupled sparse Bayesian learning (PC-SBL) and its unrolled variants, achieve high-quality map reconstruction by exploiting the inherent sparsity and spatial correlations of occupancy grids. However, these methods rely on iterative inference, resulting in high computational overhead. To overcome this limitation, we propose an autoencoder-enhanced SBL (AE-SBL) framework that replaces iterative prior estimation with a lightweight convolutional autoencoder. The proposed framework learns a trainable sparsity-inducing prior within the SBL formulation, substantially improving computational efficiency while maintaining competitive reconstruction performance. Experimental results on the EviLOG and RADIATE datasets demonstrate that AE-SBL achieves reconstruction accuracy comparable to or better than that of state-of-the-art SBL-based methods while reducing inference time by up to two orders of magnitude. ... Read more

Optimal control of wind farms requires accurate, yet computationally efficient models to capture the complex interactions between turbines. These interactions are dependent on atmospheric turbulence, wind direction changes, but also turbine properties which can be controlled. Existing control strategies often 1) neglect atmospheric effects such as wind veer, 2) rely on steady-state simulations or 3) do not optimise turbine actuation costs. This thesis addresses these limitations through the development of an optimisation-based control algorithm in a two-step process.

Firstly, the dynamic wind farm modelling and control package FLORIDyn is extended with veer-capable wake, turbulence and power calculation models. This enables the evaluation of the farm-wide power time traces, under veered conditions, as a function of yaw control inputs. These model additions are validated with Large Eddy Simulation data in single-, double-, and farm configurations. These comparisons show reasonable agreement with the high-fidelity simulations, and dependence of prediction accuracy on turbulence level estimation is highlighted. Secondly, a sparse control parametrisation is developed and integrated, making use of the new model in FLORIDyn. This optimisation-based approach jointly solves for the optimal actuation times and turbine yaws, making use of weather forecast information. Real-time numerical performance is achieved by using an efficient genetic algorithm as optimiser. Up to 6% in power gains are observed in strongly waked dynamic scenarios. When moving to veered conditions, these gains drop to 2%, quantifying the impact of veer on yaw control for power optimisation in wind farms.

Overall, this thesis establishes a framework for real-time wind farm optimisation that supports realistic atmospheric behaviour while remaining computationally tractable ... Read more

Inspection-class autonomous underwater vehicles often operate without external positioning. Instead, they use onboard sensors to estimate their state. Because no single sensor is sufficient in isolation, they typically combine a camera, an inertial measurement unit (IMU), a Doppler velocity log (DVL), and a pressure sensor, fused within a simultaneous localization and mapping (SLAM) estimator. Coherent fusion requires extrinsic calibration so that all measurements are expressed in a common frame. Existing calibration methods address only part of the problem: they often calibrate subsets of sensors, require dedicated calibration procedures or targets, adopt continuous-time formulations not directly compatible with a discrete-time estimator, or neglect the impact of uncertainty in the motion quantities used during calibration. This thesis develops a targetless joint extrinsic calibration that recovers the full set of extrinsic parameters between the camera, IMU, DVL, and pressure sensor from onboard data. The method is formulated within a discrete-time SLAM estimator. To the best of our knowledge, no prior method estimates the pressure-sensor translation jointly with the rest of the suite. First, we show that the established baseline carries a structural bias in the IMU--DVL translation, present regardless of data quality. Our corrected formulation removes this bias, establishing a precondition for any joint calibration built on the baseline. Second, we identify a subtle bias that remains, an errors-in-variables effect from conditioning on a noisy motion measurement. This bias cannot be eliminated through tuning. We show that its severity can be predicted from the sensor data before the run, without ground truth. A further reformulation reduces it. Finally, evaluated on real underwater data, the calibration runs on demand during a mission and improves the IMU--DVL translation estimate over the baseline, in the order predicted by the analysis. ... Read more

For amateur musicians learning to improvise can be a demanding task. It can take up a lot of time and energy. The goal of this project is to simplify the learning process for jazz style pianist. This will be done by showing the musician notes that could be played in a clear manner and generating new notes when the user deviates from the proposed notes. To accomplish this goal the project is divided into 3 parts: A hardware component that will focus on displaying the notes and handling inputs, a signal processing part that focuses on extracting the frequencies from an audio stream in real time to determine what notes are being played. A machine learning part that will develop a model based on jazz solos and is able to use given chord progressions to generate jazz melodies.

This thesis is focused on the signal processing subgroup and their task to create the algorithm that will determine what notes are played and when they are played. To accomplish this task, the signal from the Hardware group is put through pre-processing consisting of a bandpass filter. Then using spectral flux to compute onset detection to determine when a new note is played. Followed by a harmonic product spectrum algorithm to determine the frequency that is played and finally outputting the corresponding MIDI note to the Machine Learning subgroup. Various tests were done on a real piano and the results were promising, with low latency, reliable note detection and a high accuracy. ... Read more

For amateur musicians, learning to improvise can be a demanding task. It can take a lot of time and energy. The goal of this project is to simplify the learning process for jazz pianists. This is done by showing the musician notes that could be played in a clear manner and generating new notes when the user deviates from the proposed notes. To accomplish this goal, the project is divided into three parts: a hardware component that focuses on displaying the notes and handling inputs, a signal processing part that focuses on extracting frequencies from an audio stream in real time to determine which notes are being played, and a machine learning part that develops a model based on jazz solos and is able to use given chord progressions to generate jazz melodies.

This thesis focuses on the machine learning subgroup and its task of creating the model that generates the notes the user needs to play. To accomplish this task, the Weimar Jazz Database was chosen to train a neural network called a Long Short-Term Memory (LSTM) model. This model takes the beat timings and pitches from the dataset and processes the data to train the model. The trained model determines the most likely next note and sends this to the display. Several model outputs were converted into MIDI format to evaluate the generated melodies. The results showed that the model can generate jazz melodies within certain pitch ranges for a given chord progression, but performs better when the chords are given in the key of C. To ensure the model works on the hardware, its calculation timings were measured and shown to be fast enough to compute within a 16th note at 80 bpm.
... Read more

For amateur musicians learning to improvise can be a demanding task. It can take up a lot of time and energy. The goal of this project is to simplify the learning process for jazz style pianist. This will be done by showing the musician notes that could be played in a clear manner and generating new notes when the user deviates from the proposed notes. To accomplish this goal the project is divided into 3 parts: A hardware component that will focus on displaying the notes and handling inputs, a signal processing part that focuses on extracting the frequencies from an audio stream in real time to determine what notes are being played. A machine learning part that will develop a model based on jazz solos and is able to use given chord progressions to generate jazz melodies.
The goal of this project is to simplify the process of learning musical improvisation in jazz style on the piano through the use of dynamic, real-time feedback. Through this tool, users can more easily grasp the fundamentals of jazz improvisation, hence the name Jazzify. This report focuses on the hardware aspect of Jazzify. The system utilizes a microphone, a Raspberry Pi, a display, a signal processing pipeline, and a machine learning model to analyze user input and deliver immediate, meaningful feedback. \\
The hardware component specifically uses observational results as it is easy to see whether notes are displayed correctly.
The constraints of the hardware component of the Raspberry Pi are leading. This combined with Jazzify's real-time implementation makes it important to find components that fit the requirements and constraints. This leads to several important high level design choices that will be covered in this document.\\
Overall, this project has succeeded in delivering a low-cost, portable interactive system that provides real-time dynamic feedback for learning piano improvisation. ... Read more

This dissertation focuses on Kronecker compressed sensing, recovering multidimensional sparse signals from their linear projections on Kronecker product measurement matrices. Multidimensional signals are functions of different dimensions, each conveying a specific physical quantity and they arise in applications such as wireless communications and image processing. Kronecker product matrix naturally captures the multidimensional nature, making Kronecker compressed sensing a powerful framework for the recovery. Beyond the standard sparsity, practical signals typically have additional structures.We examine three structured sparsity models: hierarchical, Kronecker-supported, and Kronecker-structured. We start with algorithms and guarantees for the Kronecker-supported and Kronecker structured patterns, and then proceed to a unified algorithmic and theoretical framework, showing how leveraging structure in measurement matrices and sparsity patterns yields gains in accuracy and efficiency.... ... Read more

Occupancy maps are used in automotive driving applications to understand the scene around the vehicle using data from sensors like LiDAR and/or radar on vehicles. In state-of-the-art work, pattern-coupled sparse Bayesian learning (PCSBL) was used to estimate the occupancy map by leveraging spatial dependencies across grids in the map for both single modalities and the fusion of multiple modalities. The PCSBL method, however, has high computational complexity, making real-time implementation challenging for large-scale grid maps. To address this limitation, we propose several methods to improve the computational efficiency of PCSBL while maintaining mapping accuracy. First, we utilize a precomputed lookup table to accelerate selection matrix construction. Second, we implement adaptive resolution reduction based on sensor measurements. Third, we develop two novel methods that exploit the narrow angular interactions between measurements and the map regions to enhance computational efficiency. The first method partitions measurements into spatially disjoint submaps that enable parallel processing. The second method exploits the angular structure to impose a block structure on the selection matrix, reducing computational overhead. Experiments on the nuScenes and RADIATE public datasets show that the presented methods reduce computational costs compared to the benchmark PCSBL and fusionbased PCSBL methods while preserving detection accuracy. ... Read more

With the continuous advancement of autonomous driving technology, the precision and efficiency of perception systems have become increasingly critical. Among various sensors, LiDAR plays a central role, and solid-state optical phased arrays (OPAs) are widely regarded as a promising future direction. However, traditional uniform OPAs often face challenges such as high power consumption and limited scalability.

This thesis addresses the design and optimization of sparse non-uniform OPAs, aiming to balance trade-offs among the number of antennas, element spacing, beamwidth, and side lobe level. We propose a novel formulation that simultaneously considers array sparsity and performance while enforcing distance constraints, which is solved using a modified genetic algorithm. The simulation results reveal a clear trade-off between sparsity and array performance, while also offering practical solutions to the constraints faced by current LiDAR systems. Furthermore, we investigate the impact of array configuration on beam steering and introduce a mathematical transformation that reformulates the steering design problem to be compatible with our model. The comparison results demonstrate that the proposed approach significantly improves array performance across the entire steering range. ... Read more

Occupancy grid mapping represents the surrounding environment with a discretized grid, providing information about obstacles and the drivable region using sensors such as LiDAR or radar. For automotive driving applications, these maps are central to safe autonomous navigation. While both model-driven and deep learning-based approaches exist, this thesis develops a hybrid method to estimate the occupancy grid map from point cloud data. Specifically, the proposed method builds on the pattern-coupled sparse Bayesian learning (PC-SBL) algorithm, which is well suited to the block-sparse, spatially correlated structure of automotive grids. By replacing explicit parameter updates with a lightweight convolutional neural network, the spatial correlations and sparsity profile are learned directly from the data. Based on qualitative and quantitative evaluation on LiDAR point cloud data from the nuScenes dataset, we show that the proposed approach surpasses the strong PC-SBL baseline in both accuracy and runtime. Moreover, when applied without further training to LiDAR and radar point clouds from the RADIATE dataset, it marginally outperforms PC-SBL, indicating robust cross-dataset and cross-sensor generalization. ... Read more

Occupancy grid mapping is a method of representing the environment and its obstacles and drivable areas in a discretised grid that is usually constructed with point cloud data from sensors such as radars. These maps facilitate path planning and decision-making in autonomous driving applications, making them a crucial component. This thesis addressed the problem of dynamic occupancy grid mapping, which additionally models moving objects. The aim was to jointly model the spatial structure and temporal dynamics of the environment. Building on an existing sparse Bayesian learning framework exploiting sparsity and spatial structures, we proposed a dynamic extension that integrates radar range-rate measurements into a motion prediction module. This module predicted the future positions of moving objects directly from the point cloud and incorporated them into the Bayesian inference process by altering the Gamma hyperprior, enabling better tracking of moving objects. The proposed method was evaluated on the real-world View-of-Delft dataset containing urban driving scenarios and compared against three other benchmark methods. Experimental results demonstrated superior detection of moving objects and improved shape reconstruction without significantly increasing false positives. ... Read more

This thesis presents the development of a sensor fusion framework that integrates ultrasonic sensors with a rotating Light Detection and Ranging (LiDAR) system to generate an occupancy grid map. The objective is to improve spatial awareness for autonomous navigation by employing an adaptive LiDAR approach, wherein ultrasonic sensors are used to identify regions of interest for focused scanning.

The design and implementation of a test system are described, along with the development of an occupancy grid map capable of representing data from both LiDAR and ultrasonic sensors. To enhance the accuracy and reliability of the environmental representation, the occupancy grid map incorporates an inverse sensor model in combination with Bayesian statistical methods. ... Read more

This thesis presents a novel approach based on Joint Deep Probabilistic Subsampling with Cram´er-Rao Lower Bound Integration (J-DPSC) for sparse antenna array design in distributed radar systems. The method addresses the critical challenge of achieving high angular resolution in autonomous driving applications while maintaining low hardware complexity. The proposed J-DPSC framework extends Deep Probabilistic Subsampling with theoretical foundations based on worst-case dual-target Cram´er-Rao Lower Bound analysis. This enables simultaneous optimization of both transmitter and receiver arrays across mono-static, bi-static, and joint operational modes. A Neighborhood Masking Mechanism ensures physical realizability of the designed arrays. Extensive simulations validate that our proposed sparse arrays achieve performance comparable to fully populated arrays while using significantly fewer elements. The results demonstrate peak sidelobe levels up to −10 dB with minimal resolution sacrifice (≤ 0.1 degree) for two-sensor distributed radar networks. The approach allows extension to three-sensor configurations for potential applications with multi-node networks. Monte Carlo evaluations across varying SNR levels and source (i.e., targets) separations show consistent superiority over random selection methods.

Summarizing, this work advances sparse array design methodology by bridging machine learning techniques with estimation theory, enabling a more generalized and flexible design tool for distributed radar networks in next-generation autonomous driving applications. ... Read more

This thesis investigates a method to dynamically adapt the angular resolution of a 2D spinning lidar (Light Detection and Ranging) using an ultrasound sensor. The ultrasound sensor data is used to locate areas where a higher angular resolution is desired compared to other regions. It proposes two distinct methods to locate these areas based on real problems encountered in the autonomous automotive industry and assign numerical scores to areas of interest. These scores are then taken into account when deciding the angular resolution for that area. The adaptive algorithm is meant to increase the amount of data available on object of interest for use in object avoidance and trajectory mapping systems in autonomous vehicles. The thesis will cover the design, development and testing of the whole system. ... Read more

This thesis addresses the design and optimization of sparse non-uniform optical phased arrays (OPAs) for advanced automotive LiDAR systems. As autonomous driving technologies advance, the demand for high-resolution, reliable, and compact LiDAR systems has become increasingly critical. Traditional uniform OPAs, while effective, face limitations regarding power consumption. This work introduces an innovative approach to designing sparse non-uniform OPAs that achieve desired performance metrics essential for automotive applications, including beamwidth, field of view, and sidelobe levels, while minimizing element count and, consequently, energy consumption.

Through mathematical modelling and simulation, we formulate the problem of sparse OPA design as an optimization problem, leveraging techniques from compressive sensing to identify the most efficient element arrangements. We propose using the sparse array synthesis method to formulate the sparse OPA design problem, utilizing algorithms such as LASSO, thresholding, and iterative reweighted l1-norm minimization to achieve optimal sparse configurations. Our results demonstrate substantial improvements in effectiveness, offering a practical solution to the constraints posed by current LiDAR systems. This thesis contributes to the field by providing a comprehensive framework for the design of sparse non-uniform OPAs, highlighting the trade-offs and benefits of various design strategies. The findings advance our understanding of OPA design principles. ... Read more

Occupancy grid maps are fundamental to autonomous driving algorithms, offering insights into obstacle distribution and free space within an environment. These maps are used for safe navigation and decision-making in self-driving applications, forming a crucial component of the automotive perception framework. An occupancy map is a discretized representation of a chosen environment that is constructed using point cloud information obtained from sensor modalities like LiDAR and radar. In this project, we formulate the problem of estimating the occupancy grid map using sensor point cloud data as a sparse binary occupancy value reconstruction problem. We utilize the inherent sparsity of occupancy grid maps commonly encountered in automotive scenarios. Besides, the spatial dependencies between the grid cells are exploited to provide a better reconstruction of the boundaries of the objects inside the range of the map and to suppress the false alarms emerging from the reflections coming from the road. To address sparsity and spatial correlation jointly, we propose an occupancy grid estimation method that is based on pattern-coupled sparse Bayesian learning. The proposed method shows enhanced detection capabilities compared to two benchmark methods, based on qualitative and quantitative performance evaluation with scenes from the automotive datasets nuScenes and RADIal. ... Read more