Doppler radar is a cost-effective Internet of Things (IoT) device widely utilized in smart homes, urban management, and health monitoring. Conventional Doppler radars, which detect targets from a single perspective, can only extract the radial information from the radar echoes and struggle to detect stationary targets or targets moving tangentially to the radar. Furthermore, the receivers commonly encounter the issue of ambiguous frequency (AF) simultaneously, making it difficult for conventional Doppler radar to track multiple targets accurately. To address these limitations, this article enhances the target detection capabilities of Doppler radars through the design of both radar hardware structure and Doppler frequency (DF) estimation algorithms. First, a multiperspective radar system is proposed to provide richer target information and substantially minimize the AF area. Second, a novel DF estimation algorithm, based on reservoir computing (RC) theory, is proposed to estimate the DFs of targets in these reduced ambiguous intervals. Lastly, an error compensation process, adapted to the characteristics of the echoes, is designed to reduce the accumulation of estimation errors. Compared to conventional Doppler radar systems, this approach reveals more precise target information and suppresses AF interference, a critical advantage in multitarget tracking environments.

Frequency-Modulated Continuous-Wave (FMCW) radar is extensively used in human activity recognition (HAR) due to its non-intrusive operation, all-weather performance, and anti-interference ability. However, FMCW signals predominantly relies on the Discrete Fourier Transform (DFT) method, which is limited by its fixed basis functions and restricted time-frequency resolution (TFR). Therefore, this paper proposes a cross-attention augmented end-to-end architecture for FMCW radar HAR called CASA-SWIN. The preprocessing module of CASA-SWIN learns the potential Range-Time (R-T), Range-Doppler (R-D), Dopper-Time(D-T) features of rawdata through multiple complex fully connected layers with preset parameters. Furthermore, it utilizes a three-branch cross-attention mechanism (TBCAM) module to extract intrinsic related information for R-T, D-T, R-D branches. For the D-T branch, proposed multi-window filtering module (MWFM) comprises multiple 1D complex convolutions with varying kernel sizes, emulating different window lengths in the Short-Time Fourier Transform (STFT). Finally, a three-branch swin-transformer is used as the HAR classifier. Moreover, proposed CASA-SWIN is evaluated on self-collected HAR, Gait recognition datasets, and open-source gesture recognition datasets. Compared to 12 advanced HAR algorithms, a series of experiments including comparative experiments, cross-environment experiments, open-source datasets experiments, and ablation experiments verify the effectiveness of the proposed module and CASA-SWIN.

Radar, with its privacy-preserving features and robust stability, shows immense promise in the field of wireless sensing. In previous research, the time-frequency analysis (TFA) methods are predominantly applied to preprocessed signals to produce spectrum for the activity recognition. Nevertheless, the limited resolution of spectrums, coupled with the constraints imposed by the Heisenberg uncertainty principle, results in frequency errors that can impact recognition accuracy. Therefore, we propose a network architecture for radar sensing called UHAR-Net, which consists of: Augmented Spectrum Network (AS-Net), Multi-Spectrum Fusion module, and Multispectral Behavior Identification Network (MBI-Net). Proposed AS-Net can enhance the time-frequency resolution (TFR) of the spectrum and suppresses the frequency errors. Multi-Spectrum Fusion module combines enhanced spectrums of different resolutions as input for the MBI-Net, which uses a complex-valued neural network (CVNN) for feature extraction and obtaining the final activity results. Moreover, proposed Frequency-Axis Phase Modulation module is used to improve the global feature extraction capability of the convolutional neural network (CNN). Subsequently, a series of experiments including human activity recognition (HAR) tasks, gait recognition tasks, and simulation signal experiments, ablation experiments, and parameter comparison experiments are performed. These experiments result suggest that the proposed UHAR-Net can be applied to a wider range of radar sensing fields.

In recent years, Doppler through-wall radar (TWR) have shown its potential in Internet of Things (IoT) applications like target positioning, smart security, and health monitoring. However, traditional time-frequency analysis (TFA) struggles with close-range or crossed instantaneous frequency (IF) from multiple targets, reducing positioning accuracy. To address this, we propose a novel algorithm, local maximum reassignment chirplet basis transform (LMRCBT). Firstly, the time-frequency characteristics of the target signal are analyzed using Short-Time Fourier Transform (STFT) and the time-frequency distribution (TFD) is chunked to ensure local consistency. Next, we introduce the chirplet basis function combined with a kurtosis-based local optimal Chirp domain theory to capture the optimal TFD and enhance signal energy concentration. Finally, we design a timefrequency reassignment operator (RO) to optimize the redistribution of time frequency coefficients, improving resolution and boosting noise suppression. The main contribution of LMRCBT is its effective handling of non-stationary signals, especially those with close-range, non-proportional, or crossed IF, improving target localization accuracy and suppressing timefrequency aliasing, as shown in experiments.

The range-Doppler (RD) spectrum of frequency-modulated continuous-wave (FMCW) radar is widely used in various fields, making high-resolution RD spectrum estimation a critical research focus. This letter proposes HRRD-Net, an end-to-end high-resolution RD spectrum estimation algorithm for FMCW radar. HRRD-Net uses 2-D complex signals as input and achieves strong anti-noise performance with a matched filtering module and denoising loss. A coarse RD spectrum is generated using a discrete Fourier transform (DFT)-initialized module, and its resolution is enhanced by a super-resolution module. Four experiments—perception, frequency accuracy, frequency resolution, and real super-resolution RD—demonstrate HRRD-Net’s ability to generate high-resolution RD spectra, even under noisy conditions, outperforming advanced RD estimation algorithms, and its frequency accuracy has been improved by approximately 37.78% compared to the Periodogram.

Doppler through-wall radar (TWR) is a promising device for the Internet of Things (IoT), effective for indoor tracking, health monitoring, and smart homes. However, employing it to estimate the trajectories of multiple targets presents challenges associated with time-frequency analysis (TFA). In this article, a multimodal network called MWVD is proposed, which eliminates the crossterm problem of Wigner-Ville distribution (WVD) and improves the accuracy of instantaneous frequency (IF) extraction to obtain accurate localization. In the MWVD, both the WVD spectrogram and the 1-D complex signals are used as inputs to the network. The complex signals are passed through the proposed multiwindow short-time filtering (MWSTF) module followed by an adaptive wavelet attention fusion (AWAF) module to simulate the wavelet transform. Subsequently, the enhanced WVD spectrogram is obtained by the energy compression module. As a result, comprehensive experiments, including simulated signal tests, module ablation studies, fusion mode ablation analyses, and real TWR target tracking, are conducted to demonstrate the proposed algorithm’s excellence, which will be combined with more IoT applications in the future.

Fully-supervised deep learning has achieved great success in many fields, but its performance is often hindered by the abundance and quantity of available labeled training data. In the field of radar-based human motions recognition (HMR), obtaining sufficient training data is really a challenge due to the scarcity of labeled data, which causes deep learning methods to fall into an over-fitting state easily. To overcome this limitation, we propose a GAN-based semi-supervised learning model for radar-based human motions classification, which can leverage a large amount of unlabeled data to enhance classification performance. In addition, according to the characteristics of multi-classification tasks, we improve the loss function of GAN, leading the model to utilize unlabeled data more effectively. We did comparative experiments on human motions radar data measured by the doppler radar, the experimental results show that the proposed model has significant advantages in classification accuracy compared with the other models.

Accurate acquisition of human micro-Doppler (m-D) frequencies remains a challenging task, which can help us to identify the target of interest and provide valuable information about its motion dynamics. In this letter, a novel theoretical method to estimate the m-D frequencies of human specific scattering parts from continuous-wave radar echo is proposed with a united application of a modified Viterbi algorithm and a nonlinear prediction technology. Although the proposed algorithm achieves a limited improvement in the estimation accuracy and produces better human m-D frequency curves, this algorithm suppresses the frequency ambiguity effectively. The employed search method enhances upon traditional ones and improves the efficiency of finding the optimal paths considerably. Experimental results demonstrate the superior performance of the proposed method.

The micro-Doppler (m-D) modulations to radar backscattering introduced by the flexible body articulations and complicated movement patterns of human movers can provide valuable information for activity classification and help to identify the interested targets. In particular, the m-D signal of limbs, as a highly distinctive feature of human activities, can be used as an effective clue to discriminate between the armed and unarmed persons, as well as the humans against other small animals. In this article, a novel theoretical method is proposed to extract the target m-D trajectories through an integrated application of modified Viterbi algorithm and Hough transform. Through this method, multiple components corresponding to various target scattering parts and their respective m-D trajectories can be accurately extracted and estimated, even in the overlapping regions of different scattering parts in the time–frequency (TF) distribution. The employed search method enhances upon the traditional ones and improves the efficiency of finding the optimal paths considerably. Finally, a series of experiments is conducted to illustrate the validity and performance of the proposed techniques. Compared to short-time Fourier transform (STFT) peak detection and traditional Viterbi algorithm, the average error of m-D frequency estimated by the proposed algorithm is reduced by 82.1% and 71.8%, respectively. Besides, the processing time is reduced by 36.4% compared to the traditional Viterbi algorithm.