A Radar-Based Activity Recognition Approach Utilizing an Adaptive Masking Mechanism and Gated Units

Developing efficient and robust millimeter-wave (mmWave) radar human activity recognition (HAR) networks for real-world applications presents significant challenges. Key difficulties include managing network complexity and maintaining stable performance across datasets with substantial variations in scale or imbalanced class distributions. Additionally, many current networks fail to fully leverage the latent information contained in micro-Doppler (MD) signatures. To address these issues, this article proposes the masked concatenation with gated convolution and linear unit (MC-GCLU) network, which aims to improve the accuracy and robustness of radar-based HAR. The MC-GCLU network utilizes MD signatures as input. The network consists of the masked concatenation (MC) module and the gated convolution and linear unit (GCLU) module, which features the pyramid structure. This design achieves computational efficiency, requiring only 0.06 G floating-point operations (FLOPs). Through comparative experiments on the public datasets IMG848, Ci4R, and the self-constructed RadarMD dataset, the MC-GCLU network achieved accuracies of 96.33%, 82.93%, and 97.40%, respectively. To further assess the effectiveness of the individual modules within the MC-GCLU network, ablation experiments are conducted, and hyperparameters, including the mask rate and batch size, are optimized. The experimental results showcase that the MC-GCLU network achieves competitive performance without pretraining, exhibiting strong generalization capabilities and demonstrating a favorable balance between accuracy and efficiency.