Motivated by atomic norms in gridless sparse signal recovery, we introduce a gridless direction-of-arrival (DOA) estimator capable of handling scenarios with multiple frequencies and multiple snapshots. Within the atomic norm framework, we consider a multi-frequency model analogous to a multi-snapshot model. Subsequently, we extend the model to accommodate both multi-snapshot and multi-frequency data. The proposed method is illustrated through numerical simulations and validated using ocean acoustic experimental data.
Protecting Vulnerable Road Users (VRUs) is a critical concern, particularly in the context of autonomous driving at intersections, where the environment becomes complex and challenging. Various considerations, including accuracy, reliability, and system processing time, are crucial for VRUs classification models. This paper aims to assess the feasibility of state-of-the-art models in real-world scenarios, comparing two VRU classification methods using a stationary radar: a single-frame approach and a multi-frame micro-doppler method. The single-frame method employs a Convolutional Neural Network (CNN) to a cropped Range-Doppler map (RD map) around the target position, while the micro-doppler classification considers both a CNN network and a CNN-LSTM network and uses the micro-doppler spectrogram as an input. Focused on pedestrians, vehicles, and cyclists. The results, validated through a 5-fold cross-validation and F1 score, highlight the superiority of the single-frame classification in real-world scenarios, offering instantaneous and effective multi-class multi-object object classification.
Toeplitz-rectification of a Sample Covariance Matrix (SCM) is used to enhance the performance of array processing algorithms, particularly those relying on obtaining signal or noise subspace estimates, such as MUSIC Direction of Arrival (DoA) estimation and subspace-based adaptive beamforming. A Toeplitz-Rectified SCM (TR-SCM) is obtained by averaging along the diagonals of the SCM. This paper characterizes subspace estimates obtained using TR-SCMs in environments with a wide range of signal powers. Analysis of a single source in noise reveals a deterministic structure that provides valuable insights regarding the estimated subspaces. This paper derives an analytical prediction of the principal eigenvector of TR-SCMs. The eigenvector has reduced errors in directions orthogonal to the true source, but unlike the SCM, they are non-uniformly distributed. The consequence of this is observed in the noise subspace where eigenvectors corresponding to higher noise eigen-values tend to cluster around the true source. The paper presents a geometric picture of subspace estimates obtained through TR-SCMs. It shows that when estimating DoAs for multiple sources with diverse signal strengths, cross terms between a loud source and noise can significantly increase noise levels around the true source, thereby leading to inaccurate estimations of quieter sources.
We propose a physics-informed neural network (PINN) based approach that can recover the spatially-varying acoustic properties including sound speed and attenuation via partial differential equation (PDE) recovery from noisy and incomplete wave field measurements. We encode the knowledge of the assumed PDE, i.e., the wave equation, into the loss function to be minimized during training, and formulate the coefficients of the wave equation within the spatially two-dimensional (2D) region of interest as matrices of low ranks. The method is validated using datasets of 2D wave propagation.
In this paper we develop a novel learning-based approach for mobile distributed beamforming without channel state information. We consider narrowband beamforming between a mobile UAV group and a base station under limited feedback, and propose a graph recurrent neural network (GRNN) approach to leverage local collaboration among the UAVs. The GRNN method is shown to be robust to variations in UAV speeds and group heading, and scales with the UAV group size. We compare to codebook and binary feedback methods and show that better performance is achieved with the proposed GRNN method.
Spatial frequency estimation from a superposition of impinging waveforms in the presence of noise is important in many applications. While subspace-based methods offer high-resolution parameter estimation at a low computational cost, they heavily rely on precise array calibration with a synchronized clock, posing challenges for large distributed antenna arrays. In this study, we focus on direction-of-arrival (DoA) estimation within sparse partly calibrated rectangular arrays. These arrays consist of multiple perfectly calibrated sub arrays with unknown phase-offsets among them. We present a gridless sparse formulation for DoA estimation leveraging the multiple shift-invariance properties in the partly calibrated array. Additionally, an efficient blind calibration technique is proposed based on semidefinite relaxation to estimate the intersubarray phase-offsets accurately.
In big data analytics, the collected data may be contaminated by heavy-tailed noises or outliers, and the sample size may be insufficient. In this paper, we study robust sparse regression under the presence of asymmetric heavy-tailed errors within a high-dimensional setting, where the ambient dimension can exceed the sample size. The estimation problem is formulated as an l(1) constrained regression with Huber loss function. We propose a simple projected gradient descent algorithm to solve the problem and establish its convergence properties, accounting for both computational and statistical errors. Under mild conditions, we demonstrate that the successive iterates converge at a linear rate to an estimate within the statistical precision of the model. Numerical experiments validate the robust estimation performance of the proposed method across various heavy-tail distribution settings.
This paper tackles the challenge of decentralised, nonconvex optimisation in situations where agents work asynchronously. Our main contribution is a new algorithm, partially asynchronous ADMM, designed to solve decentralised optimisation problems like phase retrieval. Importantly, it does not require a central coordinator and can work with arbitrary connected network setups. We also prove that our algorithm is equivalent to the randomised block coordinate Douglas-Rachford Splitting method. To illustrate the algorithm's effectiveness, we provide numerical results for the distributed phase retrieval problem, demonstrating its correctness and performance.
Crowdsourcing deals with combining and aggregating labels from crowds of annotators of unknown reliability. While most works on label aggregation operate under the assumption of independent and identically distributed data, the present work introduces an algorithm that operates under known data dependencies or correlations. To exploit these dependencies, a novel graph autoencoder-based algorithm is developed that fuses annotator labels for crowdsourced classification tasks. Numerical tests on real data showcase the potential of the proposed approach.
In this paper, we investigate the transmit signal design problem for a dual-functional radar-communication (DFRC) system equipped with one-bit digital-to-analog converters (DACs). Specifically, the one-bit DFRC waveform is designed to minimize the difference between the transmitted beampattern and a desired one, while ensuring constructive interference (CI)-based QoS constraints for communication users. The formulated problem is a discrete optimization problem with a nonconvex objective function and many linear constraints. To solve it, we first propose a penalty model to transform the discrete problem into a continuous one. Then, we propose an inexact augmented Lagrangian method (ALM) framework to solve the penalty model. In particular, the ALM subproblem at each iteration is solved by a custom-designed block successive upper-bound minimization (BSUM) algorithm, which admits closed-form updates and thus makes the proposed approach computationally efficient. Simulation results verify the superiority of the proposed approach over the existing ones in both the radar and communication performance.
We consider a terahertz (THz)-band joint radar-communications (JRC) scenario, wherein the radar targets and the communication user are in the near-field. In THz wireless systems, the transmission range is short so that the signal wavefront is spherical in near-field. Further- more, large THz antenna arrays suffer from beam-squint arising from its ultra-wide bandwidth. To compensate the loss due to beam-squint, we devise a technique using spatial path index modulation (SPIM) that also improves the spectral efficiency (SE) performance of the overall system. Specifically, SPIM allows the transmission of additional information bits to the receiver via modulating the indices of the spatial paths. To this end, we design hybrid analog/digital beam- formers for the near-field JRC scenario by generating multiple beams toward both radar targets and the communications users. Numerical experiments demonstrate that the proposed approach exhibits signifi- cant SE performance even higher than that of the use of fully digital beamformers without SPIM in the presence of near-field beam-squint.
Machine learning over graphs (MLoG) has attracted growing attention due to its effectiveness in processing relational data from complex systems such as social networks, financial markets, and the brain. However, MLoG algorithms that use the graph topology for information aggregation have been shown to amplify the already existing bias towards certain under-represented groups, often leading to discriminatory results in downstream tasks. In this context, here we consider the prob-lem of topology-induced algorithmic bias mitigation by cross-pollinating tools from MLoG and graph signal processing. Specifi-cally, we argue that application of a tunable debiasing graph filter can be reinterpreted as a graph rewiring process, thus offering an explicit handle to manipulate the utility versus topological bias tradeoff. Building on this insight, we formulate a fairness-aware network topology inference problem to obtain a rewired graph minimizing a correlation-based, unsupervised bias metric. Node classification experiments on several real-world datasets demonstrate that the proposed approach typically outperforms state-of-the-art baselines in terms of fairness metrics, and without a degradation in classification accuracy.
Crowdsourcing algorithms often work under the assumption that the data samples are independent. Recent work has shown that data dependence, such as temporal correlations in sequential data, can be leveraged to improve the label quality. Existing methods that exploit this special structure rely on thirdorder statistics of the annotator outputs to ensure the identifiability of key latent parameters, which are costly to acquire. This work proposes an approach for integrating crowdsourced annotations under the Dawid-Skene/Hidden Markov Model (DS-HMM) for sequential data based on second-order statistics, which naturally enjoys a lower sample complexity. An effective algorithm is proposed to tackle the challenging optimization problem associated with the proposed estimator. Numerical experiments showcase the effectiveness of the data labeling paradigm.
Optical Coherence Tomography (OCT) is a non-invasive technique for obtaining detailed, cross-sectional images of coronary arteries. However, cost-effective OCT systems produce only low-resolution (LR) images. Unsupervised OCT super-resolution (OCT-SR) presents a cost-effective solution, eliminating the need for high-resolution (HR) systems or co-registered LR-HR image pairs. Existing unsupervised OCT-SR methods formulate the SR task as an image-to-image translation problem, and use CycleGAN as their backbone. However, CycleGAN is known to lack translation identifiability that can result in incorrect SR results. Existing methods often empirically combat this issue by using multiple regularization terms to incorporate expert-annotated side information, resulting in complicated learning losses and extensive annotations. This work proposes a translation identifiability-guided framework based on recent advances in unsupervised domain translation. Employing a diversified distribution matching module, our approach guarantees OCT translation identifiability under reasonable conditions using a simple and succinct learning loss. Numerical results indicate that our framework matches or surpasses the state-of-the-art (SOTA) baseline's performance while requiring considerably fewer resources, e.g., annotations, computation time, and memory.
Compressive sensing has allowed for the improve-ment of angular resolution in radar technology, which involves two aspects: sparse signal recovery and measurement matrix design. Assuming a sparse target scene, compressive sensing radar depends solely on the design of the measurement matrix to possess certain properties, such as satisfying the restricted isometry property (RIP) and low coherence. The design of the measurement matrix depends on the location of the antennas. In this work, we consider the antenna placement problem in compressive sensing radar. The problem is interpreted as a binary program, where we propose to solve it directly using a heuristic binary optimization algorithm. The proposed binary differen-tial evolution (BDE) algorithm is able to navigate the search space with relatively high diversity while still refining promising candidates. Results illustrate the superiority of approaching the problem directly using BDE rather than resorting to relaxation approaches in the literature.
In this paper, we propose a new structured Grass-mannian constellation for noncoherent communications over multiple-input multiple-output (MIMO) Rayleigh block-fading channels with two transmit antennas. The constellation, which we call Grass-Lattice, is based on a measure preserving mapping from the unit hypercube to the Grassmann manifold. The constellation structure allows for on-the-fly symbol generation and low-complexity decoding. Simulation results show that Grass-Lattice offers a superior bit error rate performance than other structured Grassmannian constellations such as Exp-Map.
This paper addresses the problem of calibration for antenna arrays with multi-port polarimetric elements. Model-based array signal processing techniques require an accurate model of the complex array response and model errors can, for example, cause significant systematic direction finding errors. Modeling the response of polarimetric antenna arrays can be particularly challenging due to cross-polarization and mutual coupling effects in the multi-port antenna elements. This work proposes a new calibration technique for polarimetric antenna arrays using neural networks that learn any mismatches between the modeled and the actual array response. The technique is evaluated based on the measured response of a five-element dual-polarized antenna array and outperforms conventional calibration techniques like mutual coupling calibration or local polynomial approximation. Its performance is studied exemplarily for the direction finding problem.
This paper addresses the problem of spatial waveform design for collocated multiple-input multiple-output (MIMO) radar systems with sparse antenna arrays. The use of sparse arrays allows to obtain narrower beams and therefore higher angular resolution and accuracy. However, if the spatial waveform is not designed properly, the resulting transmit-receive beam-pattern may suffer from significant sidelobes or ambiguity, which can strongly degrade the estimation performance. The Bayesian Cramer-Rao bound (BCRB), which is commonly used for waveform design, may produce inappropriate results as it considers only local errors and ignores the effect of sidelobes and ambiguity. To overcome this limitation, we propose using the arbitrary test-point transformation Weiss-Weinstein bound (AT-WWB) that was recently proposed, as an optimization criterion. This bound is a simpler and tighter version of the Weiss-Weinstein bound (WWB). This bound is derived for collocated MIMO radar and is minimized with respect to the ratio between coherent and orthogonal signals. The proposed method is demonstrated via simulations, and compared to optimization schemes using the BCRB and the WWB. It is shown that the spatial waveform optimized by AT- WWB exhibits superior performance in terms of direction-of-arrival estimation accuracy.
In this paper, we develop a scheme to partition a one- or multi-dimensional consecutive integer number set into multiple identical, possibly rotated, subsets. The proposed technique first exploits one-dimensional nested subsets, and the results are extended to achieve two- and multi-subset partitioning as well as in two- and multi-dimensional spaces. The number of consecutive lags in each case is examined. The results are useful to various sensing and communication applications, and sparse step-frequency waveform design for range estimation in automotive radar is demonstrated as an example.
Non-line-of-sight (NLOS) propagation could severely degrade the performance of wireless localization systems. Thus, algorithms that are robust to NLOS error are valuable. In this paper, based on hybrid angle-of-arrival (AOA) and time-of-arrival (TOA) measurements, we propose a localization approach that is robust against NLOS errors. It starts with a modified maximum likelihood estimation (MLE) problem, where range NLOS errors and angle NLOS errors are included in the objective function. Then three observations are exploited to develop an effective iteration algorithm: 1) the range NLOS error is always positive; 2) the angle NLOS error could be either positive or negative, making it more difficult to address than range NLOS errors; 3) the angle NLOS error is highly correlated with the range NLOS error for each link. In the iterative process, range NLOS errors, the weight matrix of the weighted least-squares (WLS) formulated from AOA measurements, and the source location are alternatively updated. The performance of the proposed algorithm is compared with that of existing methods in simulation.
