To estimate the parameters of signal in chaotic background is very important. A novel method named minimization of relative singular value(MRSV) is proposed in this paper, which is based both on the fact that points reconstructed from the mixed signal time series typically lie off the embedded image of the manifold for the dynamics underlying the chaotic time series, and also on the characteristic of the signal in chaotic background. The parameter estimation can be achieved by minimizing the relative singular value of the output of an inverse filter of the received signal in a reconstructed phase space. Several experiments of estimating the AR model parameters and sinusoidal signal frequency are carried out to confirm the effectiveness of the method.