Abstract: Borehole rescue technology, as a novel approach for rapidly establishing life channels, has been widely applied in practice. Ultra-Wideband (UWB) radar enables the detection of human vital signs through boreholes in obstructed environments. However, its echo signals are susceptible to complex underground noise, background clutter, and the target’s own motion states, making it difficult to directly extract effective life features of trapped individuals. Based on the principle of UWB radar detection, this paper constructs a mathematical model of human target recognition suitable for coal-rock complex medium environment. Combined with Doppler effect and two-dimensional Fourier transform, the time-frequency distribution characteristics of human vital signs in radar echo are systematically analyzed. Combined with the practical application of mine rescue, the human body motion signal is taken as the main detection target. Aiming at the problems of strong clutter, multipath effect and signal aliasing in the mine environment, a radar signal preprocessing method combining exponential weighting, adaptive gain and slow time domain energy average elimination is proposed, which significantly suppresses the interference caused by background noise and shelter. In order to further improve the ability of human body motion signal extraction, an improved complete ensemble empirical mode decomposition (ICEEMDAN) algorithm is proposed. By constructing noise-containing simulation signals, the signal-to-noise ratio and root mean square error are used as evaluation indicators. It is verified that the algorithm is superior to EMD and EEMD methods in signal decomposition and reconstruction, which effectively improves the accuracy and robustness of body motion signal extraction. Furthermore, a dedicated UWB radar detection device was developed, and single moving target detection experiments were conducted in a simulated obstructed environment composed of 50 cm crushed coal and a 30 cm brick-concrete wall. The results show that after processing with the ICEEMDAN algorithm, the fundamental and harmonic features of human targets can be clearly extracted within a distance range of 3.8–10.8 m, with harmonic frequencies exhibiting integer multiples of the fundamental frequency. This verifies the algorithm’s enhanced detection capability for moving targets under complex obstructed conditions. This study provides an effective technical method and equipment reference for single moving target recognition in strong occlusion and high noise scenes such as mine rescue.