Radar-based machine learning pipelines require extensive annotated datasets. However, producing large volumes of precise labels remains prohibitively laborious and prone to inconsistency, as radar signals lack a direct visual correspondence. To address this limitation, we introduce a fully automated, multi-modal annotation pipeline built around our custom RadarBox that co-registers a FMCW MIMO radar with an Azure Kinect RGB-D camera. Precise spatial calibration and hardware-level synchronization yield exact pixel-to-radar alignment. RGB images undergo panoptic segmentation to generate per-pixel human masks, which are fused with depth measurements to reconstruct a voxelized surface mesh. We extract 3D joint positions from the Kinect Body Tracking SDK and apply a bidirectional Kalman filter to derive precise per-joint positions and velocity vectors free from sudden, non-physiological fluctuations. These enhanced labels are projected into 5D radar cube slices and target lists through robust spatio-temporal association. As a demonstration, we train a deep neural network on annotated radar target lists for indoor people localization, achieving a mean positional error of 0.31 m and 91.8% occupancy accuracy, even under occlusion. Unlike prior semi-automatic or heuristic-based methods, our approach delivers consistent 5D labels at scale, bridging spatial, temporal, and Doppler dimensions, and thus paves the way for large-scale, learning-based radar sensing in human-centered applications.

Radar-based imaging of the plantar foot surface enables contactless monitoring of deformation and swelling during gait. This work presents an inverse synthetic aperture radar (ISAR) imaging approach that integrates trajectory estimation from an Azure Kinect depth camera. As the raw depth estimates contain noise and jitter, a filtering and optimization framework is introduced to enhance reconstruction quality. A discrete Kalman filter with constant-acceleration dynamics is optimized using autofocus metrics that quantify image homogeneity and edge sharpness. The feedback loop automatically tunes filter parameters based on the reconstructed image quality rather than trajectory error. Experimental results with a moving 3D-printed foot model demonstrate that the proposed method significantly improves focus and surface consistency compared to unfiltered or root-mean-square-error (RMSE)-based filtering. These findings establish a foundation for contactless, radar-based medical imaging of the plantar foot in dynamic conditions.

Accurate simulation of near-field millimeter-wave (mmWave) multiple-input multiple-output (MIMO) radars is essential for developing and validating high-resolution imaging systems, yet remains challenging due to large virtual apertures, multi-path propagation, and polarization effects that invalidate conventional far-field assumptions. We present a Shooting and Bouncing Rays (SBR) simulation framework that formulates radar signal transport analogously to physically based rendering, replacing far-field radar cross section approximations with spatially varying bidirectional reflectance distribution functions (BRDFs) to accurately model near-field scattering. Path tracing enables tractable multi-bounce simulation, while importance sampling over antenna gain patterns and BRDFs, combined with next-event estimation for TX–RX connections, significantly improves convergence. Polarization is incorporated via a three-dimensional Jones vector formulation, enabling correct modeling of polarization-dependent scattering across multi-bounce ray paths. Quantitative and qualitative evaluation against real measurements from a 72–82 GHz high-resolution MIMO imaging radar demonstrates faithful reproduction of real radar returns, including challenging corner reflector scenes dominated by multi-bounce returns. Ablation studies confirm that multi-path modeling, polarization, and BRDF importance sampling are each essential for high-fidelity radar simulation.

The integration of inverse synthetic aperture radar (ISAR) processing with multiple-input multiple-output (MIMO) radar architectures in the near field has enabled high-resolution imaging of large objects, typically relying on auxiliary sensors to estimate the target’s trajectory. This paper presents a methodology for near-field MIMO-ISAR imaging that estimates and compensates for target motion without external sensors. Unlike previous methods that rely on external tracking or assume known motion, this approach derives the target’s trajectory directly from radar images obtained in individual MIMO acquisitions. Once the motion is estimated, these images are spatially aligned and coherently combined to synthesize a high-resolution, multi-aspect ISAR image that improves resolution and reveals features from multiple viewpoints. The method is entirely radar-based, sensor-independent, and applicable to near-field scenarios such as human-computer interaction, non-destructive testing, and security screening. Its efficacy is demonstrated through simulations and real measurements using a 70–82 GHz MIMO radar with 94 transmit and 94 receive antennas.

Superparamagnetic iron oxide nanoparticles (SPIONs) have shown promise across a wide range of biomedical applications, including targeted drug delivery, magnetic hyperthermia, magnetic resonance imaging, and regenerative medicine. In the context of local tumor therapy (Magnetic Drug Targeting, MDT) SPIONs can be functionalized with chemotherapeutic agents and accumulated at tumor sites using an externally applied magnetic field. To achieve effective drug accumulation and therapeutic efficacy, precise positioning of the accumulation magnet relative to the tumor is essential. To address this need, we propose a dual-modality ultrasound imaging approach combining Magnetomotive Ultrasound (MMUS) and Passive Cavitation Mapping (PCM). MMUS detects magnetically induced tissue displacements, allowing localization of SPIONs embedded in tissue. However, MMUS is ineffective in vascular regions where no tissue displacement occurs. To overcome this, we use PCM to image circulating SPIONs, which are engineered to emit cavitation signals under focused ultrasound exposure. This complementary method enables SPIONs monitoring within tissue and flow. Validation was performed using standard phantoms and a custom-designed carotid bifurcation tumor flow phantom fabricated with 3D printing. Experimental results confirm that this hybrid strategy effectively images SPIONs in both tumor tissue and vasculature. This demonstrates the strong potential of complementary MMUS and PCM imaging for monitoring in preclinical and clinical MDT settings.

Ultrasound-induced cavitation is a fundamental physical phenomenon for histotripsy, lithotripsy, and local drug delivery. Magnetic nanoparticles (MNPs) enable magnetically assisted cancer therapy (Magnetic Drug Targeting, MDT) and can be used as cavitation nuclei. While MDT has shown promising results in preclinical studies, there remains a need for a real-time monitoring system. Here, we introduce a GPU-accelerated passive cavitation mapping (PCM) method that pairs a third-degree delay-multiply-and-sum beamformer (DMAS3) with coherence-factor weighting (CFwDMAS3) to enhance spatial localization and suppress off-axis incoherent emissions. We evaluated the approach in a flow-channel phantom perfused with lauric-acid-coated MNPs under pulsed focused ultrasound, and complemented the experiments with simulations based on Vokurka's bubble-dynamics model. Reconstructions using Delay and Sum (DAS), Delay Multiply and Sum (DMAS), and DMAS3, each with and without CF weighting, were assessed using ROC metrics (AUROC, sensitivity, specificity) and precision measures (PPV, F1-score). CF weighting consistently produced sharper maps and higher specificity; CFwDMAS3 achieved the best overall balance and supported real-time visualization (approximately 0.1 s per frame) on a GPU. These results indicate that CFwDMAS3 PCM is a practical solution for monitoring MNP-mediated cavitation during MDT and related therapeutic ultrasound procedures.

Passive cavitation mapping (PCM) is widely used to localize ultrasound-induced cavitation but typically assumes a homogeneous speed of sound (SoS), an assumption often violated in biological tissue. We integrate the Multistencil Fast Marching Method (MSFMM), an accurate Eikonal solver, into PCM to account for arbitrary, spatially varying SoS. For each array element, MSFMM precomputes travel-time maps that are then used as time delays in time-domain PCM or as phase offsets in frequency-domain PCM, enabling use with common beamformers. We benchmark this FMM-PCM against the heterogeneous Angular Spectrum Method (HASM) using in-silico experiments: a single cavitation source is simulated with a cavitation signal model based on Vokurka’s formulation, across three SoS scenarios—homogeneous, a layered medium with a high-SoS barrier, and a fully heterogeneous map. In the homogeneous case, FMM-PCM and standard PCM yield indistinguishable reconstructions, validating the integration. In layered and fully heterogeneous media, conventional PCM exhibits source mislocalization and elevated artifacts, whereas FMM-PCM consistently restores the true source position and suppresses aberration-induced errors; HASM improves over standard ASM but remains less accurate than FMM-PCM. These results indicate that embedding MSFMM into PCM provides accurate aberration correction for cavitation imaging in heterogeneous tissue.

Millimeter-wave imaging is a promising technique for non-destructive testing (NDT). It is a contactless sensing modality, it can penetrate many optically opaque materials, and millimeter-waves are not harmful to the operator. To enable automated defect recognition in millimeter-wave images, artificial intelligence (AI), such as deep learning, has been investigated in the research community. One problem associated with deep learning, however, is that the network’s decisions are not based on analytical criteria, but on a non-linear optimization process during training. Because of this, the bases for the network’s predictions are not transparent. This “black box” behavior limits confidence, which restricts the use of deep learning in NDT so far. To make deep learning networks more transparent, explainable AI (XAI) techniques have been introduced in the machine learning community. In this study, we introduce explainable AI to millimeter-wave based defect detection. It enables an understanding of how the deep-learning network classifies defects. At the same time, shortcomings of the deep learning network can be made visible. Since explainable AI provides visual maps of important image regions, it is possible to perform a basic defect localization without the need for a specific localization network. This approach is termed “weakly supervised object localization”. In this paper we discuss both defect classification and defect localization based on XAI techniques. Experimental demonstration is performed by means of a millimeter-wave imaging dataset of PVC objects with artificial defects.

Near-field multiple-input multiple-output (MIMO) radar systems allow for high-resolution spatial imaging by leveraging multiple antennas to transmit and receive signals across multiple perspectives. This capability is particularly advantageous in challenging environments, where optical imaging techniques struggle. We present a novel approach to inverse rendering for near-field MIMO radar systems, aimed at reconstructing material properties such as surface roughness, dielectric constants, and conductivity from radar and ground-truth mesh data, for example obtained from multi-view stereo. Drawing inspiration from physically based rendering techniques in computer graphics, we formalize an advanced inverse rendering algorithm that integrates electromagnetic wave propagation models directly into the optimization process. To avoid bias from conventional radar image reconstruction algorithms in the optimization process, we directly derive gradients from raw radar outputs, resulting in more accurate material characterization. We validate our approach through extensive experiments on both synthetic and real radar datasets, demonstrating its effectiveness in a multitude of scenarios.

Spine-related diseases such as scoliosis can be diagnosed with X-ray. As a less harmful alternative, optical scanning has been established for spine posture estimation. From a scan of the human’s back topology, characteristic points can be extracted to derive the spine’s posture. While this modality is less expensive and harmful than X-ray, it comes with the disadvantage that the patient has to undress, which takes time and may cause discomfort. Radar imaging can be an alternative that can capture the back’s topology without the necessity to undress. Therefore, in this paper, we investigate the potentials of millimeter-wave radar imaging to capture the surface topology of the back for spine posture estimation. We will present measurement results with a mannequin and compare the results to measurements with an optical reference system.

Superparamagnetic iron oxide nanoparticles (SPIONs) have shown promise across a wide range of biomedical applications, including targeted drug delivery, magnetic hyperthermia, magnetic resonance imaging, and regenerative medicine. In the context of local tumor therapy (Magnetic Drug Targeting, MDT) SPIONs can be functionalized with chemotherapeutic agents and accumulated at tumor sites using an externally applied magnetic field. To achieve effective drug accumulation and therapeutic efficacy, precise positioning of the accumulation magnet relative to the tumor is essential. To address this need, we propose a dual-modality ultrasound imaging approach combining magnetomotive ultrasound (MMUS) and passive cavitation mapping (PCM). MMUS detects magnetically induced displacements to localize SPIONs embedded in elastic tissue, while PCM monitors cavitation emissions from circulating SPIONs under focused ultrasound exposure. In addition to detection, PCM has the potential to enable feedback-based control of cavitation exposure, allowing cavitation parameters to be kept within a safe regime. The dual imaging modality approach was validated using standard phantoms and a complex carotid bifurcation tumor flow phantom fabricated via 3D printing. Experimental results demonstrate the first coordinated spatiotemporal imaging of MMUS and PCM within the same anatomical model, resolving the key bottleneck of SPIONs monitoring in blood vessels/tissue. This demonstrates the strong potential of complementary MMUS and PCM imaging for monitoring in preclinical and clinical MDT settings.

The integration of radar sensing and imaging capabilities into future integrated sensing and communication (ISAC) networks enables advanced use cases, including autonomous vehicle navigation, real-time health monitoring, and smart city management. However, ultraprecise time and frequency synchronization is crucial for unlocking the full potential of such networked ISAC systems. In this article, a novel real-time wireless time and frequency synchronization scheme is developed and fully implemented on a high-end radio frequency system-on-chip field-programmable gate array (FPGA) platform. The excellent performance and robustness of the proposed solution in practical applications are demonstrated. It is evidenced that the recursive nature of the Kalman filter is well suited to the dynamic capabilities of FPGA-based simultaneous synchronization. Observed values obtained through the precision time protocol (PTP) are iteratively refined, thus effectively compensating for uncertainties encountered during a synchronization packet exchange. Due to the deterministic processing time inherent in the FPGA, the proposed synchronization method achieves exceptional precision, with clock offset deviations in the nanosecond range and clock rate deviations limited to only a few parts per billion, even across considerable distances between the network nodes.

This paper investigates the fusion of radar and depth camera for maintaining high-quality radar images in walk-through security scanners with significantly reduced array geometries. Therefore, after sensor calibration, the reconstruction volume is initially divided into sub-volumes and linked to the closest body part of a walking person being scanned. The depth camera's motion tracking provides the location and orientation of each body joint during the gait, allowing us to determine their trajectory and velocity. This information is then used to compensate for movement in the image reconstruction of each recorded radar frame. Afterwards, the motion compensated images are then superimposed to one ISAR image, considering the complete trajectory of the person. We conducted measurements with a $3.6 \,\mathrm{G}\mathrm{Hz}$-to-$10.6 \,\mathrm{G}\mathrm{Hz}$ walk-through security scanner and a real person carrying various objects. Motion tracking with a depth camera shows promising results for both imaging approaches. In the motion-compensated images, small details became visible and focused compared to reconstruction results without applied motion compensation. The ISAR approach demonstrated effective alignment of individual frames into one superimposed image, producing a full-body image even with a reduced array dimension. While these results prove the benefit of the fusion of radar and depth camera, further investigations into array design are necessary, as object visibility is strongly affected by the selected array geometry for ISAR imaging. This fact could complicate the reliable detection of potential threat objects with a downsized imaging array.

The 3D reconstruction of faces gains wide attention in computer vision and is used in many fields of application, for example, animation, virtual reality, and even forensics. This work is motivated by monitoring patients in sleep laboratories. Due to their unique characteristics, sensors from the radar domain have advantages compared to optical sensors, namely penetration of electrically non-conductive materials and independence of light. These advantages of radar signals unlock new applications and require adaptation of 3D reconstruction frameworks. We propose a novel model-based method for 3D reconstruction from radar images. We generate a dataset of synthetic radar images with a physics-based but non-differentiable radar renderer. This dataset is used to train a CNN-based encoder to estimate the parameters of a 3D morphable face model. Whilst the encoder alone already leads to strong reconstructions of synthetic data, we extend our reconstruction in an Analysis-by-Synthesis fashion to a model-based autoencoder. This is enabled by learning the rendering process in the decoder, which acts as an object-specific differentiable radar renderer. Subsequently, the combination of both network parts is trained to minimize both, the loss of the parameters and the loss of the resulting reconstructed radar image. This leads to the additional benefit, that at test time the parameters can be further optimized by finetuning the autoencoder unsupervised on the image loss. We evaluated our framework on generated synthetic face images as well as on real radar images with 3D ground truth of four individuals.

Magnetic Drug Targeting (MDT) is a promising technique for local chemotherapy, employing magnetic nanoparticles (MNPs) and an external electromagnet to concentrate chemotherapeutic agents at tumor sites. Achieving complete tumor perfusion with the drug requires precise control over MNP distribution, which can be optimized by repositioning the electromagnet. However, this process needs a therapy monitoring system capable of 3D mapping of MNP distribution within the tumor. While ultrasound-based imaging techniques, particularly magnetomotive ultrasound (MMUS), have been shown to effectively track MNP accumulation in 2D, a full 3D characterization remains challenging. In this study, we employ a matrix transducer combined with a sliceby- slice 2D global ultrasound elastography (GLUE) approach to reconstruct a 3D magnetomotive displacement map, enabling improved visualization of the MNP distribution.

This paper presents a deep learning-enabled method for human pose estimation using radar target lists, obtained through a low-cost radar system with three transmitters and four receivers in a multiple-input multiple-output setup. We address challenges in previous research that often relied on extracting ground truth poses from RGB data, which are constrained by the need for 3D mapping and vulnerability to occlusions. To overcome these limitations, we utilized optical motion capture, which is widely recognized as the gold standard for precise human motion analysis. We conducted an extensive optical motion capture study involving various recorded movement activities, which resulted in mmRadPose, a new dataset that enhances existing benchmarks for radar-based pose estimation. This dataset has been made publicly accessible. Building on this approach, we designed an application-tailored radar signal processing chain to generate suitable input for the machine learning algorithm. We further developed an attentional recurrent-based deep learning model, PntPoseAT, which predicts 24 keypoints of human poses using radar target lists. We employed cross validation to thoroughly evaluate the model. This model surpasses previous approaches and achieves an average mean per-joint position error of $6.49 \,\mathrm{c}\mathrm{m}$ with a standard deviation of $3.74 \,\mathrm{c}\mathrm{m}$ on totally unseen test data. This excellent accuracy of the reconstructed keypoint positions is particularly remarkable when you consider that a very simple radar was used for the measurements. Additionally, we conducted a comprehensive analysis of the model's performance by exploring aspects such as network architecture, the use of long short-term memory versus gated recurrent units, input data selection, and the integration of multi-head self-attention mechanisms.

The combination of human activity recognition (HAR) with machine learning (ML) has been the subject of extensive research interest in recent years for a number of reasons. A significant amount of academic institutions worldwide are engaged in research activities within this field. However, only a limited number of publications consider the utilisation of networks and a multistatic radar setup. Furthermore, an even smaller number of researchers make the data acquired in their measurement campaigns publicly available. This paper presents a dataset on HAR acquired with a millimeter-wave multistatic radar network. The radars were positioned orthogonally to one another. This configuration allows for the direction-independent acquisition of Doppler data. Additionally, to our best knowledge, our dataset is the first publicly available dataset that provides azimuth information along with range and Doppler. The configuration allowed for the acquisition of direction-independent data. The objective of this paper is to make the acquired data accessible for researchers, with the measurement setup, sensor technology and subjects described, and the data and access information provided.

Radar-based human pose estimation enables privacy-preserving motion tracking for ambient intelligence, yet the noisy nature of radar sensing makes uncertainty quantification essential. We present RadProPoser, an end-to-end probabilistic framework that predicts three-dimensional body joints with per-joint uncertainties from raw radar tensor data. Using a variational encoder-decoder with spectral attention that fuses real and imaginary radar components across temporal frames, we model aleatoric uncertainty through learnable Gaussian and Laplace distributions. Trained on a new benchmark dataset with optical motion-capture ground truth, our method achieves 6.425 cm mean per-joint position error. The model outputs per-joint aleatoric uncertainties, and isotonic recalibration yields calibrated total uncertainty with expected calibration error of 0.027. Since spectral attention operates on individual radar tensor components, extending to multi-radar configurations requires only concatenating additional input streams. On the HuPR benchmark with dual orthogonal radars, this achieves 5.042 cm MPJPE. The framework runs at 89 frames per second (FPS) on an NVIDIA RTX 3090, exceeding the 15 Hz radar frame rate.

Ultrasound-induced cavitation can be used in various biomedical therapies, including localized drug delivery, sonoporation, gene transfer, noninvasive sonothrombolysis, lithotripsy, and histotripsy. It can also enhance thermal ablation of tumors and facilitate trans-blood-brain-barrier treatments. Accurate monitoring of cavitation activity, including dose and location, is essential for the safe and effective application of these therapies. Passive cavitation mapping (PCM) is a key technique used to achieve this. However, conventional Delay and Sum (DAS) beamforming methods suffer from low resolution and high side-lobe levels in standard diagnostic ultrasound transducer, limiting their effectiveness or are computationally expensive, in the case of robust capon beamformer (RCB). To address these challenges, we propose a higher-order nonlinear Delay Multiply and Sum (DMAS) beamformer for improved passive cavitation mapping. Our approach utilizes a novel implementation with linear complexity, using a determinant from symmetrical polynomials. Simulation and experimental results demonstrate that the proposed method enhances both axial and lateral point spread function, resolution and increasing image quality, while exhibiting linear complexity. These improvements suggest that higher-order nonlinear beamforming is a promising advancement for more accurate and reliable cavitation monitoring in biomedical applications.

Passive Acoustic Mapping (PAM) is a technique used to localize cavitation events in biomedical applications such as targeted drug delivery, histotripsy, and lithotripsy. Accurate cavitation mapping is essential for optimizing therapeutic efficacy and safety. However, conventional PAM assumes a homogeneous speed of sound within tissue, which does not reflect the heterogeneous acoustic properties of biological media. In this work, we introduce the Fast Marching Method (FMM) to PAM for accurately computing time delays in arbitrary speed of sound distributions. We evaluate the proposed approach through simulations, demonstrating its potential for improved cavitation mapping in realistic tissue environments.