Human Activity Recognition with UWB Radar-Integrated UAV in Cluttered Low-Attitude Space: a Video Swin-Transformer Approach

Benefiting from high resolution and capability to operate in visually constrained environments, Ultra-Wideband (UWB) radar has garnered significant attention in Human Activity Recognition (HAR). Radar-integrated Unmanned Aerial Vehicles (UAV) plays a critical role in rescue and military scenarios because of its mobility and flexibility, while the movement of UAV platform introduces severe interference to radar data, posing a major challenge of significantly degraded accuracy for HAR systems. To address this challenge, in this work, we propose a Video Swin-Transformer based HAR (VSTH) scheme for UAV-borne UWB radar, which has HAR capability with high accuracy and efficiency in complex environments. Specifically, a Phase-Correlation-based Feature Reconstruction (PCFR) method is developed to estimate UAV movement characteristics, thereby facilitating the recovery of human features. Besides, a UWB radar Video Swin Transformer (UV-SwinT) model is designed to efficiently extract activity-related features. Experiments on real-world datasets demonstrate that the proposed VSTH can achieve a recognition accuracy exceeding 90% while exhibiting strong robustness against external factors including UAV flight modes, angle, environmental noise, and operational distance.