Lightweight Similar Human Action Recognition Through Walls via IR-UWB Radar and Multi-Domain Feature Fusion

Focusing on the challenge of balancing computational efficiency and recognition performance in through-wall human activity recognition (HAR), this paper proposes a lightweight recognition framework based on IR-UWB radar. The framework first introduces an improved rank-adaptive robust principal component analysis with distance compensation (IRPCA-DC) method to suppress clutter and enhance echo quality, thereby improving the resolution of input radar maps. The core contribution lies in the construction of a lightweight multi-domain fusion network, which includes: 1) an EG2-MobileNetV3 backbone network incorporating GhostV2 and ECA attention mechanisms, significantly reducing parameter count and FLOPs; 2) a Siamese architecture with shared weights using TDM and TRM datasets as input, reducing the parameter count of the fusion network by 50% compared to traditional parallel structures; 3) integration of Efficient Multi-scale Attention (EMAF) within an Attention Feature Fusion (AFF) framework to enhance feature fusion at low computational cost. Additionally, a Dynamic Focus-Similarity Aware Hybrid Loss (DFSAL) is proposed to distinguish similar activities. The method achieves accuracies of 95.95% and 92.72% on a self-built dataset and a public fall dataset, respectively. Experiments demonstrate that the proposed approach maintains excellent performance while significantly reducing computational complexity, making it suitable for resource-constrained embedded through-wall perception applications.