A cascade channel estimation of Amplify-and-Forward (AF) two-way relay networks (TWRN) was investigated, and a new estimator was proposed based on Kalman filter. Firstly, a cascade channel of AF TWRN was divided into self-interference part and transmission part functionally. Then, auto-correlation functions of these two parts (ACF) were approximated by auto-regressive (AR) model to obtain the Gauss-Markov process of the channel, and the resultant Kalman estimator was deduced according on the received training signals. After the property of convergence was proved, the bound of mean square error (MSE) was also derived in the form of Riccati equation. The final numeral simulation demonstrates that the new estimator outperforms its maximum likelihood (ML) counterpart in the merit of MSE.