Through-Wall Human Presence Detection Using Chip-Scale Ultra-Wideband Radar and Edge-Deployed VGG16: A Feasibility Study for Rescue Applications

After an earthquake, numerous buildings are destroyed, and the frequency of structural collapses increases significantly in regions lacking comprehensive geophysical surveys. Furthermore, the number of deaths increases dramatically in countries without disaster prevention strategies and effective search-and-rescue operations. Consequently, urban search-and-rescue (USAR) teams encounter significant challenges in promptly locating survivors buried beneath substantial rubble. This research presents a feasibility study for detecting human presence by analyzing micro-respiratory motion behind structural obstacles, using a system designed for edge deployment compatible with compact platforms. Initially, millimetric chest movements of survivors were measured using a chip-scale Ultra-Wideband Impulse Radar (IR-UWB radar). Data acquisition over three respiratory cycles was performed by an algorithm proposed in this study. Clutter suppression was accomplished through a moving average filter. A thresholding technique was then applied to enhance the signal-to-noise ratio (SNR), thereby generating clearer images suitable for training a deep neural network. Following this, a dataset comprising 680 images was created, with 340 images representing scenarios with human presence and 340 without human presence. A convolutional neural network (CNN) based on VGG16 was trained via transfer learning to identify human presence behind 28 cm drywall walls. Results indicated that the CNN achieved a validation accuracy of 93.3% [86.67%, 98.33%] with 100% precision, outperforming conventional signal-processing baselines by over 13 percentage points. With an inference time of 23.91 ms on an embedded edge device, this research demonstrated the potential effectiveness of IR-UWB radar as a crucial technology for USAR teams in locating survivors following disasters.