Abstract: The mine roadways are narrow in space and complex in environment, where trapped personnel are prone to being buried once a disaster accident occurs. Owing to its advantages of non-contact detection and anti-interference capability, millimeter-wave radar-based vital sign detection plays an irreplaceable role in mine disaster rescue. Aiming at the key problems of strong background noise interference in radar signals and reduced heartbeat extraction accuracy caused by frequency coupling between respiratory harmonics and heartbeat signals, this paper proposes a vital sign detection method integrating the Whale Optimization Algorithm (WOA) and Successive Variational Mode Decomposition (SVMD). The core innovations are as follows: An adaptive parameter optimization model is constructed, with the Maximal Information Coefficient (MIC) as the fitness function. Leveraging the global optimization capability of WOA, the model realizes adaptive solution of the SVMD balance parameter, which avoids decomposition deviations caused by empirical parameter settings and adapts to the dynamic changes of signals in complex mine environments. A two-stage Intrinsic Mode Function (IMF) screening mechanism of "energy ratio screening - correlation coefficient screening" is designed. The IMFs containing heartbeat information are initially filtered by energy ratio, and then the optimal IMF is accurately identified by correlation coefficient, which improves the reconstruction accuracy of heartbeat signals and meets the demand for rapid identification of survivors in mine rescue. The adaptability and stability of the algorithm are verified from the dimensions of measurement angle, measurement distance and physiological state, providing support for the application in actual mine rescue scenarios. Experimental results show that the Mean Absolute Error (MAE) of the proposed algorithm is as low as 2.71%. Compared with traditional methods including band-pass filtering, Empirical Mode Decomposition (EMD), Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and unoptimized SVMD, the similarity between the separated heartbeat signal and the real reference signal is improved by 20.61% ～ 77.01%, enabling high-precision and highly stable heart rate detection under complex conditions.