This thesis focuses on advancing radar technology to meet the growing demands of autonomous driving, particularly regarding angular resolution and target detection. The work acknowledges the critical role of automotive radar in achieving higher levels of vehicle autonomy, where reliable detection, classification, and tracking of objects like pedestrians, vehicles, and infrastructure are critical. The primary focus of the dissertation is to explore novel techniques for enhancing radar system performance through machine learning (ML) and compressive sensing (CS) approaches, addressing the challenges faced by current radar technologies, such as low angular resolution and inefficient target detection in dynamic environments.

The dissertation begins by outlining the challenges in automotive radar systems, especially the need for improved angular resolution without increasing the physical size and complexity of the radar devices. The importance of angular resolution is emphasized for both azimuth and elevation, as modern vehicles must be able to discriminate between various objects, such as distinguishing between two vehicles at a similar distance, or identifying an object’s height to determine whether it can be driven under or must be avoided. Current methods for improving angular resolution, such as increasing the number of transmitters and receivers in multiple input multiple output (MIMO) radars, are costly and increase system complexity, thus requiring novel solutions to meet industry needs. Then, Chapter 2 briefly summarizes the theoretical background of MIMO radars and defines the terminology used in the rest of the dissertation. This is crucial since automotive radar is a multi-disciplinary topic with people from different backgrounds interacting, and often, the same concepts are named differently.

The first research chapter, Chapter 3, introduces a self-supervised learning framework designed to enhance the angular resolution of radar systems without the need for additional physical hardware. A neural network (NN) artificially expands the radar’s aperture by predicting the response of additional antenna elements based on data from radars with larger arrays. This approach leverages the correlation between antenna elements to generate a more detailed angular profile from a smaller, low-resolution radar, allowing for a more accurate estimation of incoming signals' direction of arrival (DoA). Extensive simulations and experimental results demonstrate that this method significantly enhances radar performance in separating closely spaced objects, which is critical in automotive scenarios.

Another key contribution of the dissertation is presented in Chapter 4, with the application of Bayesian compressive sensing (BCS) to automotive radar, which exploits the sparse nature of the data in the DoA domain. The BCS approach uses probabilistic models to estimate the DoA while also providing uncertainty measures, offering both accuracy and reliability in angular estimation. The research further explores how array topologies can be optimized for BCS-based DoA estimation, demonstrating that a carefully designed antenna array can achieve better performance with fewer elements, thus reducing system costs. Additionally, the work presents a computationally efficient BCS algorithm that dramatically reduces the time needed for DoA estimation without compromising accuracy. This is a crucial advancement for real-time applications, where fast processing is required for decision-making in autonomous driving.

In addition to BCS, in Chapter 5, total variation compressive sensing (TVCS) is applied to the problem of radar imaging. TVCS enforces sparsity in the gradient of the signal rather than in the signal itself, which proves particularly effective in estimating the shape of extended targets, such as vehicles or pedestrians. By applying TVCS to 2D and 3D radar data, the dissertation demonstrates that this method can reconstruct objects' shapes more accurately than traditional methods, thereby enhancing the radar's ability to classify and understand the surrounding environment. The application of TVCS marks a significant step forward in radar-based shape estimation, especially for imaging radars used in automotive systems.

The dissertation also addresses the limitations of conventional target detection methods, particularly the widely used window-based constant false alarm rate (CFAR) detectors. Window-based CFAR detectors struggle with dynamic and unpredictable environments, which are common in road traffic scenarios. Moreover, they are unsuitable for extended targets with very different sizes, such as the ones encountered in automotive radar. To overcome this, in Chapter 6, a deep learning-based detector is proposed, trained using a newly developed dataset, RaDelft, which includes synchronized radar and lidar data. This deep learning detector outperforms traditional CFAR detectors by significantly improving the probability of detection and the Chamfer distance, especially in complex and cluttered environments. The RaDelft dataset itself is another important contribution of the dissertation, providing the research community with a well-curated, large-scale, multi-sensor dataset for further exploration and development of radar-based detection and classification systems.

In conclusion, this dissertation presents a comprehensive study of methods to enhance the angular resolution, detection capabilities, and efficiency of automotive radar systems through a combination of machine learning and compressive sensing. It provides practical solutions verified with experimental data to overcome existing limitations in automotive radar technology, particularly in the areas of angular resolution, target detection, and data processing speed. These advancements contribute to the broader goal of achieving fully autonomous driving by improving the ability of radar systems to perceive and interpret complex environments.@en

In this paper, we address the limitations of traditional constant false alarm rate (CFAR) target detectors in automotive radars, particularly in complex urban environments with multiple objects that appear as extended targets. We propose a data-driven radar target detector exploiting a highly efficient 2D CNN backbone inspired by the computer vision domain. Our approach is distinguished by a unique cross-sensor supervision pipeline, enabling it to learn exclusively from unlabeled synchronized radar and lidar data, thus^eliminating the need for costly manual object annotations. Using a novel large-scale, real-life multi-sensor dataset recorded in various driving scenarios, we demonstrate that the proposed detector generates dense, lidar-like point clouds, achieving a lower Chamfer distance to the reference lidar point clouds than CFAR detectors. Overall, it significantly outperforms CFAR baselines detection accuracy.

@en

In this paper, an algorithm to generate a sparse linear antenna array for Direction of Arrival (DoA) estimation that works well in combination with Bayesian Compressive Sensing (BCS) is proposed. The proposed algorithms rely on the provided information inherent to BCS, i.e., the entropy of the recovered estimation vector, to place new sensor antenna elements in an initially empty array, so that the most additional information is gathered about the observed scene. It is shown by means of simulation and radar measurements that BCS methods for DoA estimation using sparse sensor arrays provide promising results in terms of detection probability and estimation accuracy. Furthermore, the proposed algorithms are able to generate sparse sensor arrangements which provide an improved performance when compared against randomly generated sparse arrays.

@en

A novel ensemble prediction technique is introduced to enhance the accuracy of far-field embedded element pattern (EEP) prediction under mutual coupling (MC) effects, while relaxing the training data size challenge in neural network (NN)-based algorithms. The proposed method integrates a two-stage NN for direct EEP prediction from full-wave simulated pattern data in spherical coordinates with a fully connected NN for the prediction of excitation coefficients of an array of infinitesimal dipoles, approximating the full-wave simulated EEPs via constrained infinitesimal dipole modeling (IDM). Quasi-randomly distributed five-element pin-fed S-band patch antenna arrays are used for demonstration purpose. It is shown that, for a large-sized (3500 topologies) and relatively small-sized (1500 topologies) dataset, incorporating IDM-NN with the benchmarked direct EEP-NN in an ensemble technique increases the pattern prediction accuracy by 11% and 60% on average, respectively.

@en

Efficient prediction of embedded element patterns (EEPs) is including the mutual coupling (MC) effects in the optimization of irregular planar arrays is studied for the first time in the literature. An ANN-based methodology is used to predict the pattern of each element in the whole visible space for a flexible planar array topology in milliseconds. The technique is proposed is validated on a 4-element planar non-uniform sub-array structure. Excellent accuracy on the EEP prediction while providing great efficiency in computational time and load in comparison to the full-wave simulations is demonstrated.@en

The problem of single-snapshot direction of arrival (DoA) estimation with antenna arrays has been considered. A sectorized approach based on Bayesian Compressive Sensing (BCS) has been proposed. In this method, the angular space is discretized, defining many non-overlapping small grids which cover the desired large angular space. First, a BCS estimation is run in each of the sectors to estimate the DoA of the signals. Then, a second stage is performed to correct the inconsistencies at the edges due to signal leaking between sectors. The performance of the method has been analyzed via extensive Monte-Carlo simulations in which the number of targets, their Radar Cross Section (RCS), and their location have been varied in a large extent, and the targets were observed by a Frequency Modulated Continuous Wave (FMCW) radar with an 86-element Uniform Linear Array (ULA). The results are compared with state-of-the-art methods in terms of estimation accuracy and resolution. Moreover, an analysis of the computational time, critical for many real-time applications, is presented, which shows a reduction of 20 times in the computational time compared with the standard BCS. Finally, the method has also been validated using experimental data collected with a commercial automotive radar.@en

A novel framework to enhance the angular resolution of automotive radars is proposed. An approach to enlarge the antenna aperture using artificial neural networks is developed using a self-supervised learning scheme. Data from a high angular resolution radar, i.e., a radar with a large antenna aperture, is used to train a deep neural network to extrapolate the antenna element's response. Afterward, the trained network is used to enhance the angular resolution of compact, low-cost radars. One million scenarios are simulated in a Monte-Carlo fashion, varying the number of targets, their Radar Cross Section (RCS), and location to evaluate the method's performance. Finally, the method is tested in real automotive data collected outdoors with a commercial radar system. A significant increase in the ability to resolve targets is demonstrated, which can translate to more accurate and faster responses from the planning and decision-making system of the vehicle.

@en

Linked to the increasing availability of datasets for radar-based human activity recognition (HAR), in this Student Highlights contribution, we report on a classification project that a group of 23 graduate students performed at TU Delft. The students were asked to work in groups of 2-3 members and to use the publicly available University of Glasgow dataset to develop the best classification pipeline as possible. This involved development and justification of both choices for the preprocessing techniques on the radar data (e.g., time-frequency distributions and cleaning of the signatures), and for the classification algorithms (e.g., the type of the algorithm, the hyperparameters' selection, the training-validation-testing split). While this student activity was performed at a small scale and with educational rather than research aims, we are happy to report it to the AESS readership, as we believe that such initiatives with open datasets sharing and classification algorithm benchmarking are beneficial for the wider radar research community. Furthermore, a list of publicly available datasets for radar-based HAR that can be used for similar initiatives is also reported in this article.

@en

Poor angular resolution is one of the main disadvantages of automotive radars, and the reason why lidar technology is widely used in the automotive industry. For a fixed frequency, the angular resolution of a conventional Multiple-Input Multiple-Output (MIMO) radar is limited by the number of physical antennas, and therefore improve the resolution involves increasing the size and the cost of the system, critical constraints in the automotive industry. In this work, a novel approach is presented to overcome this limitation, where a Neural Network (NN) is used to enhance the angle resolution of a MIMO radar without increasing the number of physical elements, but extrapolating the antennas signals in a teacher-student fashion. The method was validated using real data of stationary pedestrians captured outdoors, demonstrating an effective increase of three times the antenna array size. To the best knowledge of the authors, this is the first method that includes an evaluation metric in the final stages of the processing pipeline, enforcing the conservation of the target's angular shape, key for subsequent object classification.

@en

The problem of extended target cross-section estimation has been considered. A two-step method based on the Total Variation Compressive Sensing theory has been proposed to solve it. First, a coarse estimation of the target cross-section is performed with classical beamforming methods, and then Compressive Sensing algorithms have been applied to refine it. To the best of the authors' knowledge, this is the first time this approach has been applied to automotive radar signals. The method has been verified simulating extended targets as scatter point clouds and computing the response in a uniform rectangular array. Two metrics have been used, the Intersection over Union and a pseudo Integrated Sidelobe Level. Significant improvements in both metrics compared with classical beamforming methods have been demonstrated.

@en