A lightweight human action recognition method using binarized UWB radar input

To address the limitations of traditional vision-based human action recognition methods—such as sensitivity to lighting variations and privacy concerns—this paper proposes a lightweight UWB radar-based human action recognition method utilizing binarized input. First, an IR-UWB radar is employed to capture echo signals from various human actions. Background clutter is effectively suppressed using the Vector Mean Subtraction algorithm, enabling the extraction of discriminative dynamic features. Second, to mitigate the high computational complexity and resource consumption associated with raw double-precision input, a dynamic threshold binarization method based on a Top-K strategy is designed. This approach preserves highly responsive key features while achieving sparse encoding, significantly reducing data redundancy and computational burden. Regarding model architecture, this paper optimizes the ResNet18 structure to accommodate sparse binary input, enhancing its perception capability for key action regions. Experimental validation on a public dataset confirms the effectiveness of the proposed method: recognition accuracy remains nearly unaffected (decreasing only slightly from 96.33% to 95.62%), while model training time is drastically reduced to one-eighth of the original duration. This demonstrates excellent lightweight performance and advantages for on-device deployment.