The rapid development of Advanced Driver Assistance Systems (ADAS) necessitates enhanced performance in automotive radar systems, with Phase Modulated ContinuousWave (PMCW) radar emerging as a key technology due to its high resolution, interference resistance, and robust performance. Despite these advantages, PMCW radar faces challenges such as high computational complexity and Doppler-induced range sidelobes. This thesis addresses these challenges by proposing an adaptive block FFT correlation method to reduce computational complexity and enhance processing efficiency, ensuring reliable target detection. Additionally, we tackle Dopplerinduced range sidelobes by introducing code diversity and novel frame designs for MIMO systems, leveraging cyclic shifts and Hadamard matrices to balance sidelobe attenuation and sequence set size requirements. Through extensive analysis and simulations, the proposed methods demonstrate significant improvements in radar performance, especially in detecting weak targets behind strong reflectors. The findings contribute to developing more efficient, reliable, and scalable PMCW radar systems for advanced automotive applications.

This thesis investigates the application of multi-agent reinforcement learning (MARL) to the optimization of radar waveforms. Radar technology is crucial in fields such as aviation, maritime navigation, and defense, but faces challenges such as interference, clutter, and the need for high resolution and accuracy. Cognitive radar, which adapts to environmental changes in real-time, offers a promising solution. This research aims to explore the potential of MARL in optimizing radar waveforms and examines whether incorporating domain knowledge can enhance performance.

The radar waveform optimization problem is framed within the Decentralized Partially Observable Markov Decision Process (Dec-POMDP) framework, defining the radar environment, agents' observations and actions, and reward functions. The study experiments with different architectures, including decentralized actors with a centralized critic. The centralized critic, having access to global state information, helps stabilize the learning process and mitigate non-stationarity and credit assignment problems. The use of GNNs as a centralized critic is proposed to leverage graph data sparsity, enhancing scalability.

The proposed models are trained and tested in a radar-tracking scenario, evaluated in terms of Pareto optimality and optimization times. The results show that both Independent Actor-Critic (IAC) and Independent Actor with Centralized Critic (IACC) models outperform traditional methods in terms of probability of detection, waveform duration, and optimization speed. The findings highlight the effectiveness of MARL approaches in optimizing radar waveforms, emphasizing the benefits of centralized critics for robustness and coordination. However, the choice of architecture significantly impacts performance, and while GNNs offer potential scalability advantages, their integration of domain knowledge did not yield significant improvements in this study. This research lays a foundation for future exploration of MARL and GNNs in radar waveform optimization.

This thesis details the theoretical background and development process of a classification model for electroencephalogram-based (EEG) motor imagery (MI) signals, to be used in a brain-computer interface (BCI) system. This project was undertaken in order to demonstrate the possibility of distinguishing MI-EEG signals acquired using the g.tec Unicorn Hybrid Black EEG measurement cap. The classification model devised in this thesis is a hybrid deep neural network model, which combines a convolutional neural network (CNN) and long short-term memory (LSTM) recurrent neural network (RNN) in parallel, closing with a fully-connected (FC) layer. Much experimentation and research was needed to create this model, and this is extensively discussed within the thesis. The classification model produced for this thesis is one part of a complete end-to-end BCI system, which also entails a measurement and data processing procedure, and the development of a graphical user interface which provides visual feedback to the user.

On publicly available datasets, the classification model produced promising performance, achieving 55% on the BCI Competition IV-2a (4 class) dataset and 78% on the BCI IV-2a (2 class) dataset. The model has not yet been tested on self-collected data.

Radar technology has evolved into a versatile and robust tool for critical air traffic control, meteorology, surveillance, and defence applications. In surveillance radar, the need for continuous monitoring of large areas, often cluttered by ground or sea reflections, presents significant challenges for multitarget detection. This clutter can obscure true targets, complicating detection in environments where standard radar noise assumptions fall short.

This thesis introduces a novel implementation based on the relevance vector machine (RVM) to address the complexities of multitarget detection in cluttered environments. Unlike conventional approaches that assume white Gaussian noise, the proposed method jointly estimates a clutter covariance matrix, allowing it to adapt to the estimated clutter model over subsequent iterations. Performance evaluations using simulated data in one-dimensional (range or angle) and two-dimensional (range-angle) settings demonstrate that the framework achieves accurate AR parameter estimation. Results indicate a marked improvement in reducing false and missed detections compared to the white-noise-based model. Notably, the framework performs multitarget detection without prior knowledge of target locations and the need for guard cells, underscoring its adaptability to real-world scenarios.

In this study, we perform human identification using accumulated radar point clouds in an outdoor scene. We employ PointNet as classification network and explore the impact of adding radars' non-spatial features as input, namely doppler velocity and radar cross section (RCS). Furthermore, we encode time as an additional time identity dimension to each point within the accumulated point cloud. We examine the effects of normalizing the RCS values, canonicalizing the spatial dimensions of the point cloud, as well as normalizing the doppler velocity with respect to this canonicalization. We examine three different PointNet configurations to understand the impact of the TransformNet blocks (T-Net) within the PointNet architecture on our six-dimensional radar data input. We have created a realistic outdoor dataset for training and evaluation purposes. Our approach of using the unnormalized six-dimensional radar data on the PointNet architecture without the two T-Net blocks achieves the highest performance of 73.4 % on our test set.

Ambiguities are an often encountered nuisance in signal processing and are the source of some of the fundamental trade-offs encountered in radar systems. The goal of this thesis is to extract unambiguous information about targets by combining a limited amount of measurements on a video integration level. A novel framework is proposed to reach this goal. At the heart of the framework lives a relevance vector machine which is extended to process the ambiguities on a video integration level and to work off-grid. The relevance vector machine is then extended to become the ambiguity aware relevance vector machine. This extension is either performed by a frequentist test or by estimating a posterior distribution. The frequentist test is used to test whether we can statistically significantly discern the returned output from ambiguities. The posterior is estimated according to Bayes’ theorem and thus allows for the incorporation of prior information. In this thesis, the framework is specifically applied to Doppler processing of a pulse-Doppler radar system. Compared to existing methods for estimating unambiguous Doppler velocity in a multi-target environment, the framework provides a general increase in performance, allows for the incorporation of prior information, and is able to give a measure of confidence in the estimates. A simulation study is set up to show the performance increase. This simulation study also highlights the utility of incorporating prior information and the quantification of uncertainty.

Gaussian process regression (GPR), a potent non-parametric data modeling tool, has gained attention but is hindered by its high com- putational load. State-of-the-art low-rank approximations like struc- tured kernel interpolation (SKI)-based methods offer efficiency, yet lack a strategy for determining the number of grid points, a pivotal factor impacting accuracy and efficiency. In this thesis, we tackle this challenge.
We explore existing low-rank approximations that facilitates the computation, dissecting their strengths and limitations, particularly SKI-based methods. Subsequently, we introduce a novel approxima- tion framework, MKISSGP, which dynamically adjusts grid points us- ing a new hyperparameter of the model: density, according to changes in the kernel hyperparameters in each training iteration.
MKISSGP exhibited consistent error levels in the reconstruction of the kernel matrix, irrespective of changes in hyperparameters. This robust performance forms the bedrock for achieving accurate approx- imations of kernel matrix-related terms. When employing our rec- ommended density value (i.e., 2.7), MKISSGP achieved a comparable level of precision to that of precise GPR, while requiring only 52% of the time compared to the current state-of-the-art method.

Machine learning models offer promising potential in precipitation nowcasting. However, a common issue faced by many of these models is the tendency to produce blurry precipitation nowcasts, which are unrealistic. Previous research on the deep learning model - TrajGRU (Shi et al., 2017) indicated that data imbalance in radar images and the double-penalty effect of pixel-wise loss functions are underlying causes for this blurriness.
In this thesis, we continue to explore various approaches to improve the predictive performance of TrajGRU. Our research has first investigated spatially and temporally enhanced loss functions to address the two remaining issues: data imbalance and double penalty. The second part of our research focuses on manifold optimizations within the model network, such as incorporating additional model inputs or increasing batch size, to understand the model’s limitations.
Our results reveal that enhanced loss functions did not lead to predictive improvements and even resulted in undesired checkerboard patterns. Changes to the model network make a difference in the image sharpness and predictive rain evolution. Our visual analysis indicates that a larger batch size generates sharper rain field edges; predictions by using multiple parameter groups exhibit more rain dynamics. The incorporation with other transformed datasets introduces finer structures within rain fields. Although the blurriness has not been completely resolved, our study recommended future work can continue exploring the optimization in the TrajGRU network.

In recent years, neural networks (NNs) have seen a surge in popularity due to their ability to model complex patterns and relationships in data. One of the challenges of using NNs is the requirement for large amounts of labelled data to train the model effectively. In many real-world applications such as radar, labelled data may be scarce due to the high cost of acquiring measurements together with privacy and security concerns.

To overcome the lack of data, researchers have resorted to data augmentation (DA), a technique that aims to solve the problem at the root by generating new training samples by leveraging the available ones. In computer vision, image transformations and generative networks are used to perform DA. These techniques, however, may lead to the production of physically unfeasible samples that may hinder the generalization capabilities of classifiers in domains where the data has an underlying physical meaning.

Physics-informed machine learning aims to incorporate physical prior knowledge and governing equations of the target domain into the machine learning pipeline to improve the performance of NNs in fields with limited available data but with well-defined physical models. In this thesis, physics-informed DA in the radar domain is addressed to improve the task of classifying armed and unarmed walking individuals through micro-Doppler spectrograms.

To begin with, the usage of model-driven micro-Doppler radar simulations to improve the existing generative augmentations is investigated. After introducing several generative NN architectures, the quality and diversity of the produced synthetic spectrograms are evaluated together with their effect on the downstream classification task.


Next, the impact of visual transformations on micro-Doppler spectrograms is studied. Their effect on the underlying physics of the micro-Doppler spectrograms and their impact on feature extraction is assessed. Based on the results, a new DA technique is devised to improve the feature extraction process by informed segmentation of the input spectrogram, halving the size of the NN model and reducing the risk of overfitting.

This thesis presents a method for personnel activities observation, i.e., 3D human pose estimation and tracking, in a Catheterization Laboratory(Cath Lab). We mount five cameras from different angles in the Cath Lab, where surgeons and assistants are in similar clothes while doing surgery. Accurate 3D human pose estimation is the cornerstone of our method. Most previous 3D pose estimation methods train their models directly on a 3D pose dataset. However, these methods are not suitable for our task: 1) We do not have enough 3D pose data for training because of privacy issues and specificity 2) The model needs to be retrained in different operating rooms or the camera calibration changed. To solve these problems, we decompose the 3D human pose estimation task into two stages, avoiding the need for large amounts of 3D pose data and retraining. In the first stage, we apply YOLOX and HRNet for 2D human detection and 2D pose estimation. Simultaneously, the 2D object tracking network Bytetrack tracks person identities based on detection results. Then we use a matching algorithm to match the corresponding 2D poses from multiple views and reconstruct 3D poses. Given 3D poses and tracking identities, we at last introduce a hybrid method tracking algorithm. By feeding 2D tracking results into the matching and tracking algorithm, we increase the accuracy of the result in a scene where people are wearing similar clothes. We fine-tune and test our method with an operating room dataset. Finally, we validate the method on data from the Cath Lab.

The recent advances in wireless communication, micro-fabrication, and integration have led to the rise of multi-agent networks such as wireless sensor networks, drone swarms, and satellite arrays. These multi-agent networks comprise multiple agents, which cooperate to solve a problem collectively, beyond the individual knowledge of a single agent. Multi-agent networks are also known for improving overall system performance, especially in terms of scalability, robustness, and flexibility.

Multi-agent networks are tasked with a variety of tasks e.g., target tracking, surveillance, traffic control, and environmental monitoring. Traditionally these tasks are solved using distributed state-space filtering methods, however under the assumptions of linear models with underlying Gaussian noise. However, in general, the models could be Non-linear, and the underlying stochastic noise is not necessarily Gaussian in nature. To address this limitation, Distributed particle filters (DPF) have been investigated in recent years. Depending on the quantities exchanged in the network, DPFs can be categorized as weight-based, likelihood-based, or posterior-based.

One of the key challenges for effective implementation of the DPF is the need for approximation strategies to reduce the burden on communication and computation resources to suit engineering applications. To this end, we review two state-of-the-art DPF, (a) one based on applying the Graph Laplacian (GL) approximation to weights, and (b) the other using the Gaussian mixture model (GMM) to represent the local posteriors. These two algorithms significantly reduce the communication burden; however, the main limitation is the synchronization and heavy computational burden, which places high demands on hardware and software implementation of agents.

To improve on these existing methods, we propose the use of Gaussian process (GP) enhanced resampling algorithm, which reduces the particle set size, while ensuring an acceptable filter performance. Based on this GP-enhanced resampling scheme, we propose a novel Distributed particle filtering algorithm. Secondly, based on the proposed GP-enhanced sampling, we propose several metaheuristic optimization algorithms (i.e., genetic algorithm and firefly algorithm) to fuse the information across the whole network to improve the estimation performance (i.e., to seek the global optimal particle set as we can treat particles as “parents” or “fireflies”).

We compare our proposed algorithms and study their performance against the state-of-the-art DPFs in terms of time, space, and communication complexity, for a target tracking application. We use time-averaged root mean square error (ARMSE), which is used to quantify the performance of each DPF. Our numerical results demonstrate the superiority of our proposed algorithms under the condition of limited communication and computational resources, and still maintaining the optimality of particle filter.

For example, to achieve a certain ARMSE for a target tracking application, the traffic for existing GL and GMM based DPFs, is 12000 scalars/second/agent and 600 scalars/second/agent respectively, while our proposed GP-enhanced resampling and firefly algorithm uses only 900scalars/second/agent. In terms of the computation runtime, the GL and GMM based approaches take 40 and 300 seconds (about 5 minutes) respectively, while our proposed algorithm takes just 1.5 seconds, and thus shows significant improvement. Finally, we propose research directions to further improve our proposed algorithms in the future.

Accurate short term rain predictions are important for flood early warning systems, emergency services, energy management and other services that that make weather dependent decisions. Recently introduced machine learning models suffer from blurry and unrealistic predictions at longer lead times, causing poor performance on the rarer heavy rainfall events. The objective of this research was to explore how the loss function in a recurrent, convolutional neural network (TrajGRU, X. Shi et al. (2017)) can be modified, to get sharper and more realistic predictions, without worsening its performance. Six perceptive loss functions, which should better represent how an image is perceived, are thoroughly analysed to understand the reasons behind their functioning in the context of precipitation nowcasting. It was shown that these perceptive losses can lead to an improvement in sharpness for the first lead time, but that the blurriness for longer lead times arises due to the imbalance in the dataset and the uncertainty of the model with respect to the location of the rain. This thesis gets concluded with recommendations on how to deal with these two problems.

Nowadays, the aging problem is shaking the root of the healthcare system in many countries, an automatic human activity recognition (HAR) is seen as a promising solution to that problem. In particular, radar-based HAR attracts people’s attention thanks to its respect for privacy and functionality in poor lighting conditions. With a lot of research paying attention to this topic, there is still a lack of conclusive and practical methods. In particular, it is realized that dynamic motions at large aspect angles close to 90◦ or static postures have not been investigated in-depth as a part of the radar-based HAR problem. To extensively investigate this type of problem, we propose to use mm-wave FMCWMIMO radar to obtain accurate information of the human subject.

This thesis work aims to fully exploit the six dimensions of information provided by an imaging radar: range, azimuth, elevation, velocity, power and time. Two complementary data representations- point cloud and spectrogram- are utilized to represent these dimensions of information. A signal processing flow is implemented to generate the desired data representations. A hierarchical pipeline consisting of three cascaded deep learning-based classification modules is proposed to process the input data. Particularly, human orientation classification is achieved through the so-called "T-Net" network learning the geometric distribution of point clouds. The positive contribution of each module in the proposed pipeline is validated via an ablation study. The superior performances of the proposed pipeline are also established by comparing with those of the state-of-the-art baselines. The robustness of the proposed pipeline concerning a noisy environment is also discussed. It is also presented that the size

With the development of machine learning techniques, more and more classification models have been designed for seizure detection. The creation of these models has dramatically improved the convenience of epilepsy detection and made seizure labeling automation possible. However, many of the current researches in this field use EEG datasets with small data volumes and are mainly designed for scientific purposes, which do not have a good performance of actual medical data. Besides, most models require complex time-frequency domain transformation and feature extraction process, which result in low classification speed and makes it difficult to achieve real-time monitoring. Moreover, the excessive complexity also means higher power consumption, so most of these models cannot be implemented with wearable EEG devices.
This thesis proposed a new seizure detection algorithm based on the bidirectional long short-term memory(BiLSTM) technique. The seizure detection function is achieved using time-domain features and LSTM networks. The preprocessing steps of this model are simple, and the complexity is low. Thus its operation speed is significantly improved compared to other traditional models. Also, this model is developed and tested based on TUH EEG corpus, which is an open-access dataset. Therefore, the results are directly comparable to others in the literature.

When a patient is in a hospital, it is very important to monitor their vital signs. Doctors and nurses use this information to assess the condition of the patient. Most of the existing vital signs measurement devices need physical contact with the patient. This thesis focuses on a non-contact vital signs estimation method. Using a mmWave radar, one or more persons in the view of the sensor are being monitored. This monitoring consists of finding the chest region of a person, and monitoring this chest for vibrations. These vibrations are caused by breathing in and out, and the beating of the heart. Using signal processing, these vibrations can be converted to a heart rate and a respiration rate.

This thesis is about getting insight in the already available options regarding vital signs monitoring, programming the Texas Instruments IWR6843ISK mmWave radar module to estimate the vital signs of multiple persons and validating this project against trusted vital signs monitors.

The implemented solution which followed from this project is able to track multiple persons inside the radar view, and is able to measure the vital signs for up to four persons in real-time. The mean accuracy gained for one person heart rate estimation is 10.8%, the mean accuracy gained for one person respiration rate estimation is 7.6%. The mean observed accuracy for multiple person heart rate estimation is 13.4%, the multiple person respiration rate mean accuracy is 10.6%.

THz time-domain systems driven by photoconductive antennas (PCAs) promise a bandwidth of the order of hundreds of GHz. This characteristic can be utilized to build see-through radars with sub-mm resolution and a field of view comparable to the current mm-wave radars at a relatively low cost. These radars can solve the issues in the field of public safety and security. However, the analysis and design of a THz time-domain radar is a challenging and time-consuming task even at the preliminary level. In this work, a Matlab based GUI is built that serves as a one-stop solution for the THz time-domain radar design and analysis. The GUI has three interfaces, the first one is the “Transmitter interface” which implements the Norton equivalent circuit of the PCAs and quantifies its radiated THz power along with other important parameters of the PCA semiconductor gap such as the conductance and current across the gap, and the resistance of the gap. The second interface, known as the “QO Channel Analyzer interface”, of the GUI tool analyzes a pre-defined QO channel model to estimate the power loss in the channel and the power received at the receiver. The third interface is the “Main interface” which analyzes the detection and imaging-related parameters of the radar such as the resolution, field-of-view, detection bandwidth and integration time.

Autonomous robots are increasingly used in more and more applications, such as warehouse robots, search-and-rescue robots and autonomous vacuum cleaners. These applications are often in environments where the GPS signals are denied or inaccurate, which makes it difficult to localize the robot in an unknown environment. To overcome this problem the framework of Simultaneous Localization and Mapping (SLAM) is typically used. This solution constructs a map of the environment with the use of cameras or range sensors, while keeping track of the location of the robot in it. To extend the exploration time of these battery powered robots, the energy consumption of the SLAM algorithm could be reduced. It is assumed that if the computational load of an algorithm reduces, the energy consumption of the algorithm reduces as well. An existing paradigm to solve SLAM is the use of a particle filter, which tracks the trajectory of the robot and simultaneously maps the environment. The question answered in this thesis is how to make this algorithm more energy-efficient to be able to deploy this framework in more applications and make the existing robots more sustainable. In this thesis two methods are investigated. In the first method, the information about the landmarks is incorporated in the trajectory estimation as spatial constraints, to try to achieve a higher accuracy with less particles and thus subsequently a smaller computational load. The proposed method is validated by simulations on synthetic datasets. This method shows improvements in terms of the estimation accuracy. However, it is more computational complex than the existing algorithms, so it is considered less energy-efficient. The second method researched in this thesis, is the implementation of a parallelized particle filter. This method processes the observation measurements in parallel for the different particles and communicates the information between the particles efficiently. It should reduce the computational time, to enable partial computation of the algorithm to reduce the computational load. This method is validated on the same datasets as the first method using simulations. This method shows improvements on the run time and thus on the computational load, especially for a larger number of particles and is therefore more energy-efficient. The two separate methods have been analyzed and compared with state-of-the-art methods. Both methods deliver equally good or better results in terms of accuracy. However, the computation time of constrained FastSLAM does not outweigh the improvement in accuracy. On the other hand, the parallelized particle filter shows significant improvement over the existing solutions.

Accurate short-term forecasts, also known as nowcasts, of heavy precipitation are desirable for creating early warning systems for extreme weather and its consequences, e.g. urban flooding. In this research, we explore the use of machine learning for short-term prediction of heavy summer rainfall showers in the Netherlands. We explore the use of a recurrent, convolutional neural network (TrajGRU, Shi et al., 2017) with lead times of up to 100 minutes. We aim to train the model to predict the formation and dissipation of dynamic, heavy, localized rain events, a task for which traditional Lagrangian extrapolation based nowcasting methods still come up short. The network is trained, validated and tested on a 13-year archive of radar images with 5-min temporal and 1-km spatial resolution from the precipitation radars of the Royal Netherlands Meteorological Institute (KNMI). We report on different ways to optimize predictive performance for heavy precipitation events through two experiments. In the first experiment different training dataset compositions are explored. The large dataset available provides many possible configurations for training. To focus on heavy rainfall intensities, we use different subsets of the radar dataset through using different conditions for event selection and varying the ratio of light and heavy precipitation events present in the training data set. In the second experiment we change the loss function used to train the model. To assess the performance of the model, we compare our method to a current state-of-the-art deterministic Lagrangian extrapolation-based nowcasting system from the pySTEPS library, S-PROG (Seed et al, 2003). The results of the experiments are used to discuss the pros and cons of machine-learning based methods for precipitation nowcasting and possible ways to further increase performance. Model behaviour was found to be significantly influenced by the formulation of the loss function. It was concluded that there is always a trade-off between performance at low rainfall intensities and performance at high rainfall intensities: (1) If a model makes smaller errors at low rainfall intensities this results in a low total error, but also in the failure to detect high rainfall intensities. (2) If model performance is improved at detecting high rainfall intensities, this results in a decreased performance at low rainfall rates and increases the total error.

When symptoms of atrial fibrillation (AF), a common cardiac arrhythmia, are experienced, a Holter monitor or event recorder is used for official diagnosis. Apart from the fact that these devices are experienced as inconvenient, AF can already manifest damage in a pre-symptomatic phase. This thesis is aimed at developing a method for recording heart activity using a wearable device to permit convenient early detection of AF. For this, heart activity is measured continuously by means of photoplethysmography (PPG). A classification algorithm is used to detect AF episodes in the PPG recording. If the algorithm suspects AF, a limb lead I ECG recording is requested from the user. The ECG recording can be analyzed by a clinician for official diagnosis. The Maxim Integrated Max86150 chip is used for the implementation of PPG and ECG. Acceleration data is gathered by means of the Adafruit MMA8451 accelerometer to allow for detection of motion artefacts. These sensors and the data they retrieve are controlled and processed by the ARM Cortex-M7 microcontroller. From the results, PPG recordings have a higher quality when infrared light is used as compared to when red light is used. However, both types of recordings are of sufficient quality for monitoring the heart rate accurately when in stasis. Although complete functionality of the system could not be verified, the results are promising for future work.