Through-Wall Multi-Person Action Recognition Using Enhanced YOLOv5 and IR-UWB Radar

Impulse radio ultrawideband (IR-UWB) radar has been extensively utilized in recognizing human activities due to its high resolution, independence from light and occlusion, and nonintrusive nature. However, current research is mainly limited to single-person and line-of-sight conditions, and many challenges exist in multi-person action recognition under non-line-of-sight (NLOS) conditions. This article proposes a through-wall, multi-person action recognition framework based on the improved YOLOv5 network. First, we collected the through-wall multi-person motion echoes using the IR-UWB radar. After processing them with algorithms such as clutter suppression and adaptive amplitude balancing, we successfully constructed a dataset containing time-range images of multi-person actions. Then, to improve the detection and classification performance, we improved the YOLOv5 network and introduced the bi-directional feature pyramid network (BiFPN) module, SCYLLA-intersection over union (SIoU) loss function, and coordinate attention (CA) mechanism. The BiFPN module enhances the model’s learning ability for multiscale features. The SIoU loss function optimizes the positioning accuracy of the prediction box. The CA attention mechanism finely weights feature coordinates, enhancing the feature discrimination ability. The experimental results show that the improved YOLOv5 network achieves good detection and classification results on our dataset. Moreover, the proposed method has verified good generalization ability on the moving and stationary target acquisition and recognition (MSTAR) public dataset.