Multi-Scale SE-SK Network for Human Falling Detection using IR-UWB Radars

Impulse radio ultra-wideband (IR-UWB) radars are increasingly utilized for human fall detection owing to their non-contact, high precision, and high privacy attributes. However, the existing single-branch networks-based methods on falling detection only extract single feature maps, and fail to capture complete fall information. Furthermore, processing multiple spectra without the target unique feature results in feature loss after fusion, leading to decreased accuracy. This paper proposes a novel multi-scale network, which combines squeeze-and-excitation (SE) and selective kernel (SK) (SE-SK) networks, for human falling detection using IR-UWB Radars. Specifically, three features spectra of IR-UWB Radar data, that is, the time-Doppler, range-Doppler, and range-time feature spectra, are firstly obtained through the pre-processing of the radar echoes. Then, a multiscale SE-SK network is developed for human falling detection. Experimental results demonstrate improved performance with detection accuracy of 98.2% compared to the existing singlebranch networks-based method.