mmCMD: Continuous Motion Detection From Visualized Radar Micro-Doppler Signatures Using Visual Object Detection Techniques
Zhimeng Xu, Junyin Ding, Shanshan Zhang, Yueming Gao, Liangqin Chen, Željka Lučev Vasić, Mario Cifrek, Zhizhang Chen
Abstract
This article presents a novel millimeter-wave (mm-wave) radar-based continuous motion detection approach called mmCMD for the long-term care of the elderly. The proposed mmCMD visualizes the micro-Doppler signatures of continuous motions generated from millimeter radar as images, which are then analyzed by an object detection network for continuous motion detection. To improve the imaging quality of micro-Doppler signatures, a dynamic feature visualization (DFV) method is proposed by selectively mapping the micro-Doppler matrix (MDM) elements with significant values, highlighting human motion to enhance subsequent detection network’s accuracy in capturing the details of the motion. Furthermore, a novel detection network is designed for the visualized micro-Doppler images by combining the specially designed fusion squeeze-and-excitation (FSE) module with the coordinate attention (CA) into the YOLOv5 architecture, which is distinct from prior works that overlook global contextual information. Experimental results demonstrate that the proposed mmCMD achieves a mean average precision (mAP) of 93% at the intersection over union (IoU) thresholds from 0.5 to 0.95 and an F1 score of 99% for 12 actions, which makes it a promising solution for remotely monitoring and detecting elderly individuals’ activities to enhance safety and risk prevention capabilities.