In recent years, Human Activity Recognition (HAR) using Frequency module continuous wave (FMCW) radar is an effective tool that has been widely used in the fields of healthcare, smart driving, and smart living due to its convenience, inexpensiveness, and accuracy. Past studies have mainly investigated the improvement of the accuracy of HAR models while neglecting the deployment of the models. Therefore, we propose a model named FMCW Lightweight vision transformer (FML-Vit) for human activity recognition, primarily consisting of the FML-Vit Block and FML-Vit subsample modules. The FML-Vit Block, by incorporating a cascaded linear self-attention mechanism in place of the traditional multi-head attention mechanism, can transform the time complexity from O ( k 2 ) to O ( k ). The FML-Vit subsampling modules perform dimension reduction and feature reallocation, while the context broadcasting (CB) module is used to reduce the density in the original attention maps, thereby increasing both the capacity and generalizability of the ViT. The proposed algorithm is compared with 9 different states of the art methods on self-datasets and open-source datasets. The results demonstrate that FML-Vit outperforms other current lightweight networks with the fastest inference.
2 September 2024 in IEEE Transactions on Instrumentation and Measurement
Doppler through-wall radar (TWR) is a cost-effective device to track targets in complex environments. However, when there are multiple targets in the detection area and their instantaneous frequencies (IFs) are closely arranged, mutual interference between echo components can severely degrade radar detection performance. This phenomenon is called frequency ambiguity issue. To address it, this paper proposes a time-frequency enhancement algorithm based on a complex convolutional neural network (CNN) and designs two kinds of training datasets and an improved loss function for the model. The proposed algorithm takes complex spectrograms obtained through Short-Time Fourier Transform (STFT) as input, and the final optimized spectrogram is primarily generated through the proposed energy rearrangement module (ERM) and Deep Jump Convolution Module (DJCM). The ERM, similar to the Synchronized Squeezing Transform (SST), condenses the energy along the frequency axis of the spectrogram using 1D complex convolutions. Through comparative analysis of both simulation signal and TWR target tracking experiments, the algorithm is shown to effectively improve the time-frequency resolution and suppress frequency ambiguity issue. It achieves a reduction of 57.46% in localization deviation and a decrease of 66.17% in IF deviation compared to the current state-of-the-art.
Currently, human gait recognition has emerged as an effective solution for person identification. Low-cost millimeter-wave radar, with its non-intrusive nature and high accuracy, can be effectively utilized in a wide range of scenarios. However, most current research primarily focuses on radar datasets with radial walking and limited samples, resulting in models with poor scalability and robustness. Therefore, this paper introduces a stable gait recognition algorithm that achieves commendable results in a dataset comprising 121 individuals across eight different viewpoints and three clothing variations. The proposed model integrates ResNet18 with a multi-scale temporal extraction structure as the backbone for feature extraction, effectively capturing gait characteristics over different time intervals. The model is optimized using a combination of gait loss, triplet loss, and center loss, significantly enhancing its stability. In the experiments, the proposed algorithm achieved an average accuracy of 87.9% and 75.0% under multi-view and cross-view conditions, respectively, surpassing current state-of-the-art methods.
18 February 2025 in IEEE Internet of Things Journal
Dual-frequency continuous wave radar, as a promising Internet of Things (IoT), has been used for indoor human tracking, activity detection, and smart homes. Previous indoor tracking primarily used Short-Time Fourier Transform (STFT) for IFs extraction. However, STFT has low resolution and suffers from Heisenberg’s uncertainty principle, which limits the positioning accuracy. Therefore, the paper introduces an improved Time-Frequency Analysis (TFA) algorithm called UTFA-Net, capable of significantly enhancing time-frequency resolution (TFR), potentially outperforming traditional principles, and thus boosting the precision of TWR target tracking. The proposed framework is founded on an end-to-end self-supervised neural network architecture, integrating novel Rényi and consistency loss mechanisms. Meanwhile, we propose a Multidimensional Spectrogram Generation (MSG) module, Short Time Fusion Attention (STFA), and a non-local module to boost the model’s spectrogram generation capabilities. In order to validate proposed UTFA-Net, a comprehensive set of experiments is undertaken, including simulation tests, ablation studies, and indoor tracking experiments. The results indicate that UTFA-Net surpasses current advanced algorithms in four performance metrics, showcasing the highest tracking accuracy.
The use of Doppler through-wall radar in the field of indoor target localization has attracted extensive attention due to its advantages in medium penetration, positioning accuracy, and strong anti-interference ability. However, in multi-target positioning, the echo signal consists of multiple components, leading to time-frequency aliasing in the time-frequency distribution (TFD) when using short-time Fourier transform (STFT) for time-frequency analysis. This aliasing complicates the accurate estimation of the target’s instantaneous frequency (IF) curve, thus affecting the target positioning accuracy. To tackle these problems, we propose a time-frequency enhancement algorithm that obtains a highly concentrated TFD, suppressing time-frequency aliasing and noise, particularly for signals with intersecting or adjacent IF curves. The proposed method employs the time-frequency enhancement network (TFE-Net) to learn the time-frequency features of the real and imaginary parts of STFT results under different window lengths, resulting in a TFD without time-frequency aliasing and with concentrated time-frequency energy. Subsequently, the target IF curve is estimated from the enhanced TFD to achieve target tracking. Multi-target tracking experiments demonstrate that the proposed method, compared with existing methods, not only effectively suppresses time-frequency aliasing and improves time-frequency energy concentration but also enhances the radar system’s positioning accuracy.
Doppler through-wall radar (TWR) excels in indoor target localization. The traditional localization method employs Short-Time Fourier Transform (STFT) for time-frequency analysis (TFA), but an error occurs when multiple targets’ instantaneous frequencies (IFs) cross or are close. This paper presents an algorithm using a data fusion network (DF-Net) to enhance the Wigner-Ville Distribution (WVD) by eliminating cross-terms. In DF-Net, both the WVD spectrogram and complex signals are inputs to the model, which uses complex convolutions for encoding. A channel weight reassignment module (CWR) and multilayer residual down-up sampling (MRDUS) module are employed to refine the WVD spectrogram and remove cross-terms. Target IFs extracted from enhanced spectrogram enable accurate localization. The DF-Net has been validated through simulations and real-world experiments, demonstrating its superiority. It not only performs well when dealing with IFs crossings but also exhibits superior performance in high-noise environments. As a result, the target localization error and IFs error of the proposed algorithm are reduced by approximately 59.2% and 68.8%, respectively, compared to the state-of-the-art methods.