This roadmap provides a comprehensive and forward-looking perspective on the individualized application and safety of non-ionizing radiation (NIR) dosimetry in diagnostic and therapeutic medicine. Covering a wide range of frequencies, i.e. from low-frequency to terahertz, this document provides an overview of the current state of the art and anticipates future research needs in selected key topics of NIR-based medical applications. It also emphasizes the importance of personalized dosimetry, rigorous safety evaluation, and interdisciplinary collaboration to ensure safe and effective integration of NIR technologies in modern therapy and diagnosis.

The problem of contactless accurate monitoring of vital signs of patients in psychiatric settings is addressed in this work. Most state-of-the-art radar-based approaches rely on measurements acquired at short distances with the sensor positioned directly in front of the subject’s chest, typically within 1 m, to ensure a strong line-of-sight signal. In contrast, we investigate a substantially more challenging and clinically safer configuration, where multiple frequency-modulated continuous wave (FMCW) radars are mounted at ceiling height in a tilted position, remaining completely out of reach and outside the direct line of sight of patients. An experimental dataset involving 30 participants performing seven activities was collected using two of these ceiling-mounted radars and one reference radar mounted at hip height. A new processing pipeline is proposed with automatic range bin selection and fast autoregressive (AR) spectral estimation, including an approach that leverages all chirps per frame to improve performances with short observation window lengths. Despite the unfavorable sensing geometry and increased radar-to-subject distance, the results demonstrate that ceiling-mounted radars achieve mean absolute respiration rate errors comparable to the hip height reference radar. The best configuration yields a mean absolute error (MAE) of 2.18 respirations per minute (rpm) for participants in a sitting position at distances exceeding 4 m from the ceiling-mounted radars. Moreover, the proposed AR-based method significantly improves estimation accuracy for short windows of 3.1 s, achieving a MAE below 4.3 rpm for all radars and participant positions.

Conventional radar segmentation research has typically focused on learning category labels for different moving objects. Although fundamental differences between radar and optical sensors lead to differences in the reliability of predicting accurate and consistent category labels, a review of common radar perception tasks in automotive applications reveals that determining whether an object is moving or static is a prerequisite for most tasks. To fill this gap, this study proposes a neural network (NN)-based solution that can simultaneously segment static and moving objects from radar point clouds. Furthermore, since the measured radial velocity of static objects is correlated with the motion of the radar, this approach can also estimate the instantaneous 2-D velocity of the moving platform/vehicle (ego-motion). Notably, despite performing dual tasks, the proposed method employs very simple yet effective building blocks for feature extraction: multilayer perceptrons (MLPs) and recurrent NNs (RNNs). In addition to being the first of its kind in the literature, the proposed method also demonstrates the feasibility of extracting the information required for the dual tasks directly from unprocessed point clouds, without the need for cloud aggregation, Doppler compensation, motion compensation, or any other intermediate signal processing steps. To measure its performance, this study introduces a set of novel evaluation metrics and tests the proposed method using a challenging real-world radar dataset, RadarScenes. The results show that the proposed method not only performs well on the dual tasks but also has broad application potential in other radar perception tasks. More qualitative results can be viewed here: https://youtu.be/3ejS1chSvQ8?si=uGRugVA63BCyvNBV

To characterize atmospheric turbulence, the Doppler moments are estimated by weather radars. However, moment accuracy is highly sensitive to radar transmission parameters such as pulse repetition time (T_s) and number of pulses (N_p), which affect Doppler ambiguity and estimation variance. Traditional fixed-parameter radars face trade-offs between aliasing and measurement precision. This paper proposes an adaptive radar framework that dynamically adjusts T_s and N_p on a per-scan basis to improve Doppler moment estimation at a single resolution cell level. Inspired by the Fully Adaptive Radar (FAR) concept, the method also includes a novel multi-lag Doppler width estimation scheme. Results demonstrate enhanced estimation accuracy, enabling better responsiveness to localized and non-stationary weather conditions.

Radar-based human activity recognition (RadHAR) is an attractive alternative to wearables and cameras because it preserves privacy, is contactless, and is robust to occlusions. However, dominant convolutional neural network (CNN)- and recurrent neural network (RNN)-based solutions are computationally intensive at deployment, and recent lightweight vision transformer (ViT) and state-space model (SSM) variants still exhibit substantial complexity. In this article, we present RadMamba, a parameter-efficient, micro-Doppler-oriented Mamba SSM tailored to radar HAR under on-sensor compute, latency, and energy constraints typical of distributed radar systems. RadMamba combines 1) channel fusion with downsampling; 2) Doppler-aligned segmentation that preserves the physical continuity of Doppler over time; and 3) convolutional token projections that better capture Doppler-span variations, thereby retaining temporal–Doppler structure while reducing the number of Floating-point Operations/Inference (\# FLOP/Inf.). Evaluated across three datasets with different radars and types of activities, RadMamba matches the prior best 99.8% accuracy of a recent SSM-based model on the continuous wave (CW) radar dataset, while requiring only 1/400 of its parameters. On a dataset of non-continuous activities with frequency-modulated continuous wave (FMCW) radar, RadMamba remains competitive with leading 92.0% results using about 1/10 of the parameters, and on a continuous FMCW radar dataset it surpasses methods with far more parameters by at least 3%, using only 6.7k parameters.

The problem of estimating the mounting angle of millimeter-wave automotive radars installed on moving vehicles is investigated. We address this angle estimation problem during normal driving, without relying on controlled environments, dedicated radar targets, or specially designed driving routes. To achieve this, we propose a signal processing pipeline that combines radar and inertial measurement unit (IMU) data to enable accurate and reliable estimation under realistic driving conditions. Unlike previous studies, the method employs neural networks to process sparse and noisy radar measurements, reject detections from moving objects, and estimate radar motion. In addition, a measurement model is introduced to correct IMU bias and scale factor errors. Using vehicle kinematics, the radar mounting angle is then computed from the estimated radar motion and the vehicle’s yaw rate. To benchmark performance, the proposed approach is comprehensively compared with two alternative problem formulations and four estimation techniques reported in the literature. Validation is carried out on the challenging RadarScenes dataset, covering over 79 km of real-world driving with different velocities and trajectories. Results show that stable and accurate mounting angle estimates are obtained within approximately 25 seconds of driving. To the best of our knowledge, this is the first study to demonstrate that automotive radar mounting angles can be estimated during complex, real traffic conditions using only onboard sensor data.

The problem of estimating instantaneous distributed target velocity using noisy measurements by multiple asynchronous automotive radar sensors is investigated. Two novel neural networks (NNs)-based approaches are proposed to address the problem. Both NNs use the point cloud with radar detections as an input. In the first approach, a hybrid NN is designed to take a set of points inside a cluster as its input, extract spatial-dynamic features to be used as weights for each input point, and apply them to obtain a weighted least square (WLS) solution for instantaneous velocity estimation. To this end, dedicated loss functions are proposed to allow the model to predict weights that can follow a velocity profile curve satisfying target constraints. Moreover, a small offset in the radial velocity value of each point is applied to adjust errors in the sensor measurements. In the second approach, a deep NN is proposed that takes a set of points inside a cluster as its input and directly outputs velocity estimates. Both approaches have been verified experimentally using the large open-source automotive RadarScenes dataset. The results show a significant improvement in terms of mean absolute error in velocity estimation over the state-of-the-art alternative techniques. Moreover, the estimated velocity is used as an additional measurement value inside a target tracker. Results show that this can increase the performance of the tracker, especially during challenging scenarios such as abrupt changes in the velocity of the target.

The problem of reconstructing 3D signatures of human activities for monitoring and classification is considered in this work. A method based on data fusion from distributed MIMO (multiple-input multiple-output) radar nodes is developed in order to generate 3D intensity maps and related voxel-wise velocity vectors. The proposed method was evaluated with a dataset collected using three 60 GHz radars and including 7 activities performed by 30 participants. The results show that both static postures and dynamic activities can be captured effectively: consecutive phases of activities/movements can be identified by combining spatial intensity and velocity vectors, and the participant can be localized in the area under test. Furthermore, initial promising classification results of 98.3% macro F1-score are demonstrated in a three-class problem using the proposed 3D intensity maps and velocity vectors as inputs to a Convolutional Neural Network classifier.

In this paper, an automatic labelling process is presented for automotive datasets, leveraging on complementary information from LiDAR and camera. The generated labels are then used as ground truth with the corresponding 4D radar data as inputs to a proposed semantic segmentation network, to associate a class label to each spatial voxel. Promising results are shown by applying both approaches to the publicly shared RaDelft dataset, with the proposed network achieving over 65% of the LiDAR detection performance, improving 13.2% in vehicle detection probability, and reducing 0.54 m in terms of Chamfer distance, compared to variants inspired from the literature.

This article presents the first subject-specific head pose estimation approach using only one frequency-modulated continuous wave radar data frame. Specifically, the proposed method incorporates a deep learning framework to estimate head pose rotation and orientation frame-by-frame by combining a convolutional neural network operating on range-angle radar plots and a PeakConv network. The proposed method is validated with an in-house collected dataset, including annotated head movements that varied in roll, pitch, and yaw, and these were recorded in two different indoor environments. It is shown that the proposed model can estimate head poses with a relatively small error of approximately 6.7°–14.4° for all rotational axes and is capable of generalizing to unseen, new environments when trained in one scenario (e.g., lab) and tested in another (e.g., office), including in the cabin of a car.

This article studies the problem of estimating the 2-D motion state of a moving vehicle (ego motion) using millimeter-wave (mmWave) automotive radar sensors. Unlike prior single-radar or synchronized radar systems, the proposed approach (named DeepEgo+) can achieve sensor fusion and estimate ego motion using an unsynchronized radar sensor network. To achieve this goal, DeepEgo+ combines two neural network (NN)-based components (i.e., Module A for motion estimation and Module B for sensor fusion) with a decentralized processing architecture using the late fusion technique. Specifically, each radar sensor in the network has a Module A that processes its output and computes an initial motion estimate, while Module B fuses the initial estimates from all radar sensors and outputs the final estimate. This novel architecture and fusion scheme not only eliminates the synchronization requirement but also provides robustness and scalability to the system. To benchmark its performance, DeepEgo+ has been tested using a challenging real-world radar dataset, RadarScenes. The results show that DeepEgo+ provides significant performance advantages over recent state-of-the-art approaches in terms of estimation accuracy, long-term stability, and robustness against high outlier ratios and sensor failures. Furthermore, the influence of vehicle nonzero acceleration on ego-motion estimation is identified for the first time, and DeepEgo+ demonstrates the feasibility of compensating for its effect and further improving the estimation accuracy.

For the effective deployment of countermeasures against drones, information on their intent is crucial. There are several indicators for a drone's intent, for example, its size, payload and behaviour. In this paper, a method is proposed to estimate two or more of the following four indicators: a drone's wing type, its number of rotors, the presence of a payload and its mean rotor rotation rate. Specifically, three multitask learning (MTL) approaches are analysed for the simultaneous estimation of several of these indicators based on radar micro-Doppler spectrograms. MTL refers to training neural networks simultaneously for multiple related tasks. The assumption is that if tasks share features between them, an MTL model is easier to train and has improved generalisation capabilities as compared to separately trained single-task neural networks. The proposed MTL approaches are validated with experimental data and in a variety of combined classification and regression tasks. The results show that MTL approaches can provide improvement in several tasks compared with conventional approaches.

The problem of radar-based multi-target tracking for indoor human monitoring is considered. Tracking and counting the number of people moving as a group is particularly challenging as multiple individuals are close together and their radar signatures are mixed. A transformer-based classification approach for counting the number of grouped people is proposed. The Neural Network model is trained with selected features from the spatial domain and Doppler frequency domain, which are concatenated over multiple frames to form a sequence for the transformer network. Compared to statistical classifiers, the self-attention mechanism allows transformers to capture feature long-term dependencies. The proposed classifier is integrated into a tracking pipeline in order to monitor the position and number of people in the grouped targets. The method proposed is experimentally verified using a 24 GHz Frequency Modulated Continuous Wave (FMCW) radar with 250MHz bandwidth. Despite the relatively coarse range resolution, the proposed method achieves 92.5% accuracy in these initial tests. Furthermore, the method performances and related accuracy is analyzed according to various parameters.

The role of radar for building situation awareness in (semi)autonomous vehicles is severely restricted by its low angular resolution. The physical size of the radar, which determines its antenna aperture size and thus the radar angular resolution, is often a subject of stringent limitations to physically fit the system in the vehicles. Multiple input multiple output systems are used to increase the achievable angular resolution, and these are often combined in the literature with algorithms inspired by synthetic aperture radar techniques that exploit the velocity of the vehicle itself for finer resolution. Some of the most common approaches are reviewed, in this context, with a specific focus on challenges for the implementation of data collected in real driving scenarios. Key experimental results using representative algorithms and driving data collected in the city of Delft, The Netherlands, are presented and discussed.