Lightweight Multiattention Enhanced Fusion Network for Omnidirectional Human Activity Recognition With FMCW Radar

Human activity recognition (HAR) based on radar has garnered wide interest due to its privacy-friendly and lighting-independent nature. Despite advancements, challenges like response speed and real-world applicability persist. To address these issues, data from different orientations utilizing frequency-modulated continuous-wave (FMCW) radar was collected to better reflect real-world conditions. In terms of data preprocessing, a sliding time-window nonlocal means filter is proposed to effectively remove the noise while retaining the micro-Doppler features. Subsequently, a multiattention enhanced fusion network (MAEF-Net) is proposed for HAR. MAEF-Net, an efficient and lightweight architecture, integrates multiple attention mechanisms to enhance robustness in omnidirectional HAR. It achieves high-accuracy omnidirectional activity recognition by simultaneously focusing on local and global features. Experimental results showcase MAEF-Net achieving 98.3% accuracy in testing, surpassing existing methods, thereby demonstrating its effectiveness in real-world scenarios.