A novel Kalman filter for dead reckoning of biped robots is proposed. In order to reflect the change of the foot contact condition, its observation error model is adaptively varied in accordance with the reaction force. For the rolling or rotation happened on the supporting foot, the minimum velocity point (MVP) which proposed in our previous work is introduced into the kinematics computation. MVP estimation is sensitive to the attitude error, so that we lower that influence by taking the ground reaction force at the contact point into consideration. The efficacy of the proposed method is verified through simulations.