A Multitarget Action Recognition Method Based on UWB MIMO Radar Space–Time–Frequency Fusion Features

Human Action Recognition (HAR) has broad application potential but traditional optical methods are limited by lighting, occlusion, and privacy concerns. Radar-based approaches overcome these issues through all-weather, non-contact sensing, yet existing micro-Doppler methods relying on global time–frequency analysis lack spatial resolution, causing feature aliasing in multi-target scenarios. To address this limitation, this paper proposes a novel space–time–frequency fusion framework for multi-target HAR using ultra-wideband (UWB) Multiple-Input Multiple-Output (MIMO) radar. The proposed method performs two-dimensional imaging to separate multiple targets, followed by multi-height three-dimensional imaging to extract independent micro-Doppler features of each subject in the image domain. By coupling spatial and frequency-domain representations, the framework effectively suppresses inter-target interference, enhances motion feature separability, and preserves fine-grained temporal dynamics. Finally, the extracted features are fed into a lightweight recognition network, Spatio-Temporal Decomposed Network (STD-Net), for efficient classification. Extensive experiments demonstrate that the proposed method achieves superior accuracy, robustness, and real-time performance compared with existing approaches. It can effectively distinguish actions of multiple individuals while significantly reducing Doppler cross-talk. Field experiments in real-world environments further validate the generalization ability and practicality of the system, highlighting its strong potential for intelligent sensing and human–computer-interaction applications.