4 November 2024 in IEEE Internet of Things Journal
In the realm of human activity recognition (HAR) based on radar, the prevailing methods have been characterized by excessive complexity and a singular focus on a specific motion direction, posing challenges for practical deployment. This paper introduces a lightweight parameter estimation and deep learning hybrid extraction network (PDHE-Net) for achieving multi-directional HAR based on mmWave radar. The network consists of a lightweight deep learning feature extraction (LDE) module, a parameter estimation module, and a classification module. Specifically, the LDE module consists of multiple layers of group convolution and Ghost module, aimed at extracting deep features from the time-Doppler (TD) maps. Parameter estimation module is employed to capture direction-independent features from the TD map. Ultimately, multi-dimensional features extracted by the LDE module and parameter estimation module are classified to realize multi-directional HAR. To verify the performance of the proposed method, experimental data was collected, comprising six categories of activities wherein targets moved in multi-directions. The experimental results demonstrate that the proposed PDHE-Net achieved a recognition accuracy of 96.67% on the dataset, outperforming the state-of-the-art methods by 2.92%, while significantly reducing complexity.
13 June 2025 in IEEE Transactions on Aerospace and Electronic Systems
In this paper, we investigate a lightweight human activity recognition (HAR) method for non-line-of-sight (NLOS) scenarios. This approach constructs a lightweight multi-spectrogram parallel feature extraction model to extract time-range (TR) map and time-Doppler (TD) map features from stepped-frequency continuous wave (SFCW) radar, enabling NLOS HAR in L-shaped corner environments. Specifically, the model incorporates a multiscale depthwise separable convolution to extract multipath TR map features at different scales in NLOS scenarios. Simultaneously, it combines depthwise separable convolution with Ghost convolution to extract features from the TD map. Finally, the extracted multispectrogram features are fused and classified, achieving robust HAR in NLOS environments. To validate the effectiveness of the proposed method, we conducted experimental verification using real-world data from eight types of human activities in an L-shaped corner NLOS scenario. The results show that the proposed model achieves a recognition accuracy of 87.12%, which is a 5.82% improvement compared to the state-of-the-art recognition methods.
1 August 2026 in IEEE Sensors Journal