Human Respiration and Motion Detection Based on Deep Learning and Signal Processing Techniques to Support Search and Rescue Teams

Abstract

The quick and effective detection of humans trapped under debris is crucial in search and rescue operations. This study explores the use of antennas operating within the 150-650 MHz frequency range to identify human respiration and movement under building wreckage. A debris model consisting of construction materials was generated at the laboratory, and attenuation characteristics were observed to set ideal operating frequencies. Time-dependent transmission coefficient data were collected over 20 s and processed using short-time Fourier transform, wavelet transform, and empirical mode decomposition for time-frequency analysis. To enhance signal clarity, denoising techniques were applied before the radar signals were categorized into three classes: empty debris, human respiration, and human movement. Generative adversarial networks augmented environmental noise data to enrich training datasets comprising nine subsets. Deep learning models, including temporal convolutional networks, long short-term memory, and convolutional neural networks, were employed for classification. Hyperparameter optimization via random search further refined model performance. Results indicate that the convolutional neural networks using short-time Fourier transform data consistently achieved the highest classification accuracy across subsets. These findings demonstrate the potential of combining radar with deep learning for reliable human detection under debris, advancing rescue efforts in disaster scenarios.