Rank detection is crucial in array processing applications, as many algorithms rely on accurately estimating the rank of the data matrix to ensure optimal performance. Under Gaussian white noise, rank can be detected through eigenvalue analysis. However, in arbitrary noise, prewhitening the data matrix with the noise covariance matrix is necessary, and rank detection is achieved by examining the generalized eigenvalues. Existing methods often assume the noise covariance structure or require a large number of noise samples. This thesis focuses on addressing the rank detection problem in scenarios with limited noise samples and arbitrary noise environments.

Firstly, we investigate the largest generalized eigenvalue threshold for the prewhitened data sample covariance matrix according to the random matrix theory. We develop a rank detection algorithm based on the threshold via a sequential test, and provide the performance analysis. A series of simulations demonstrate its superiority over conventional methods such as Minimum Description Length (MDL) and Akaike's Information Criterion (AIC).

Secondly, since the Short-time Fourier Transform (STFT) is commonly used for non-stationary signal analysis, we extend our rank detection method to the STFT domain. The correlations introduced by the STFT have a significant impact on the distribution of the noise. Therefore, we develop a technique to remove correlations among time-frequency bins based on exact expressions of these correlations. After successfully eliminating these correlations, our proposed rank detection method achieves enhanced reliability and performance in the STFT domain.

Lastly, we evaluate the effectiveness of our rank detection method in speech enhancement applications. Simulations confirm that utilizing the estimated rank improves speech quality compared to using the known number of sources.

Over the years in the automotive industry, the use of radar sensors in advanced driver assistant systems (ADAS) has been increasing, shining a light on various safety and security concerns to be dealt, that come along with it. The performances of existing sensor technologies such as camera, LIDAR and GPS can be very well complemented by the use of millimeter-wave radar due to its robust performance in various conditions. In assisted or autonomous driving tasks, localization of targets and their height information is important to be determined. Since height information is lacking from automotive radar systems without elevation measurements capabilities, this remains a challenge to be worked on. In the application discussed in this thesis, using the radar specifications providing a high range and velocity resolution, the multi-paths can be resolved in the range-Doppler domain and be used for height estimation techniques. The automotive applications for which short and ultra-short range radar systems are used, require a high range and velocity resolution. In this thesis, a low complexity and fast target height classification algorithm is presented using a FMCW radar.

In the process of having a quick response automotive radar system, low latency DoA estimation technique was employed. To estimate the height properly, knowledge of the ego-velocity is needed. A histogram method can be used to roughly estimate the ego-velocity. The ego-velocity estimate can be refined using an Extended Kalman filter that tracks targets that are supposed to be annotated with height information. Using the vertical doppler beam sharpening technique, the improved ego-motion estimate helps in estimating the elevation angles of the targets and therefore, their corresponding height. An overview of the height estimation process is explained to provide a real time mapping of the target in 3-Dimensions, showcasing the improvement in ego-motion estimation leading to an elevation angle estimation. Simulations and experiments are carried out to validate and verify the algorithm. The results show promising results and also reveal potential improvements and the need for more simulations and experiments to mature the concept further.

Rainfall is increasing in frequency and intensity due to climate change. Hydrological models exist that can report bottlenecks in urban infrastructures. However, these require accurate rainfall estimations with high temporal and spatial resolution. The fulfillment of these requirements is challenged due to high costs. This can be solved with cost-effective acoustic rainfall sensors, that can be densely distributed over a city. A (machine learning) prediction model is then needed to infer rainfall from acoustic data. It is infeasible to create or calibrate a seperate model to the specific characteristics of every acoustic sensor. This research focuses on domain adaptation techniques that transfer a model designed for one sensor, called the source, to a characteristically different one, called the target. A two-stage regressor model is devised that estimates rainfall intensity from acoustic data. In a supervised learning setting, this gave an average $R^2$ score of 0.67. Transfer Component Analysis (TCA) has been used as a semi-supervised domain adaptation technique, where no annotations from the target sensor are needed. This resulted in an average $R^2$ of 0.59. Finally, a supervised domain adaptation technique, where some annotations are used, is proposed using TCA and a mean matching method. Two flavours are researched. In the batched approach a small selection of target annotations are gathered, where after the rainfall is predicted with a delay. The online approach predicts in real-time, but adjusts the prediction model when a new target annotation is available. These resulted in an average $R^2$ of 0.52 and 0.32, respectively.

Spatial cues allow a listener to determine the direction sound is coming from. In addition, recognising spatially separated sound sources facilitate the listener to focus on specific sound sources. Because of this, preservation of spatial cues in multi-microphone hearing assistive devices is important to the listening experience and safety of the user.
A number of linearly-constrained-minimum-variance-based methods exist for this purpose. Most of these are limited in the number of interfering sources for which they can preserve the spatial cues. In this thesis, a method of selecting the most important interfering point sources using convex optimisation is proposed.
The method is presented based on two different convex relaxations, which are compared, using simulation experiments, to existing, exhaustive search and randomised methods in terms of noise suppression and localisation errors.
Both methods are shown to improve the performance of the joint binaural linearly constrained minimum variance beamformer, an existing method for simultaneous noise reduction and spatial cue preservation, by giving it more degrees of freedom for noise reduction and allowing it to handle a larger number of (virtual) sources present in the scene.

This report details the design and implementation of a subsystem within a localization system designed for the Bosch DICENTIS wireless conference system. The goal of this localization system is to determine the location of individual units. These units each contain a microphone and a loudspeaker and are used by the participants at a conference. Potential use cases for the locations of these units are beam-forming and the mapping of participants within a room.

The subsystem discussed in this report is tasked with transforming estimated propagation times of audio signals into locations of the units. To this end, a TOA-based and a TDOA-based algorithm for the self-localization of individual conference system units is presented. Experimental results show that both methods are able to recover positions, with the TOA-based method slightly outperforming the TDOA-based method. Real-life measurement results show a root mean square error in position of about 8 cm for the TOA-based method.

Hearing aids as a form of audio preprocessing is increasingly common in everyday life. The goal of this thesis is to implement a blind approach to the cocktail party problem and challenge some of the regular assumptions made in literature. We approach the problemas wideband FD-BSS. From this field of research, the common assumption of contineous activity is dropped. Instead a number of users detection is implemented as a preprocessing step and ensure the appropriate number of demixing vectors for each time frequency bin. The validity of the standard mixing model used for STFT’s is challenged by looking at the response of a linear array. Source separation is achieved by demixing vectors based on the GSVD, derived in a model-based approach. While most ermutation solvers offer an a posteriori solution for all users, we looked at finding local solutions for a single user. Combining this with the user identification called the alignment step, we conclude that the permutation problem can be reduced to selecting a demixing vector for each discrete time-frequency instance.
The correlation coefficient proves to be a sufficient metric to couple reconstructions to the original data as it selects most of the active time-frequency bins. In the far-field case, our approach performs in a comparable
but not superior manner. We did find that our method is much more robust against inaccuracies introduced when narrowband channels are assumed but not actually available. This is strongly exemplified by our experiment of a changing DFT-size.
The Frobinius norm was suggested as a measure of distance between the estimate STFT and the orignial signals time frequency domain description but it resulted in counter intuitive results which didn’t correspond with other metrics used in this thesis. It is expected that there are effects induced by changing the size of the STFT which are not accounted for.
Our demixing vectors achieve comparable inteligibility, measured by STOI, as the compared techniques and it is more robust against smaller sample sizes than the theoretically SINR optimal MVDR.

Wearable health has become a striking area in our daily life.
Electrocardiogram (ECG) is one of the biomedical signals collected by the wearable or portable devices, which is widely used in heart rate monitoring and cardiac diagnosis. However, automatic ECG signal analysis is difficult in real application because the signals are easy to be contaminated by the noise and artifacts. Thus, the quality of ECG signals is essential for the accurate analysis.
The objective of this project is to design a reliable automated ECG signal quality indicator based on the supervised learning algorithm, which intends to estimate the quality of the signals and distinguish them.

The methodology of this project is creating a classification model to indicate the quality of ECG signals based on the machine learning algorithm. The model is trained by the extracted features based on the Fourier transform, Wavelet transform, Autocorrelation function and Principal component analysis of ECG signals. Subsequently, the feature selection techniques are proposed to remove the irrelevant and redundant features and then the selected features are fed to classification algorithms. The classifier was then trained and tested on the expert-labeled data from the collected ECG signals. Particularly, we focus on the performance of classifier and use the best training model to predict the quality of new ECG signals.

The acoustics of a space contribute much to the perceived size and atmosphere of the environment. Modeling these acoustics in virtual environments can make a big improvement to the immersion of the player. There are very accurate models that can reproduce realistic acoustics in virtual spaces, but these are computationally very expensive. Cheaper algorithms are either very inaccurate or impose very limiting constraints on the environment. This thesis aims to introduce an acoustics model that can run in real time, while still providing a reasonable degree of accuracy. The perceptually most important parts of the reverberation are modeled accurately in real time, allowing for excellent sound
source localization. The remaining part of the reverberation is generated using an artificial reverberator whose input parameters are derived from the geometry, again in real time. This system is intended to give players a convincing sound in any environment without heavily taxing the available processing power.