One of the most important prerequisites for blind source separation to achieve an effective separation is to estimate the number of sources accuratly. For underdetermined problems of which the sources signal number is unkown and less than the observed signal number, this paper proposes a kind of signal source number estimation method based on Empirical Mode Decomposition(EMD), together with Singular Value Decomposition(SVD) of covariance matrix, and by means of Bayesian selection principle combining with Laplace approximation. Before procesing the observation signal with EMD, the proposed method denoised the observation signal by means of Kalman filtering algorithm in order to remove the mode mixing phenomenon of Intrinsic Mode Function(IMF). The method is validated using both simulated signals and measured signals respectively. The research on this shows that the proposed method can estimate the sources signal number exactly and provide an accurate priori information to the blind source separation.