Tilt sensor is usually necessary for attitude control of a biped robot when it walks on an uneven terrain. There are many sensors to measure the tilt angles, and gyro sensor is widely used for estimation of tilt angles because it can offer the sufficient bandwidth, so it is suitable for sensing the rapid motions that create high frequency pose variations. However, its major disadvantage is the lack of accuracy and drift over time. Vision sensor can estimate an accurate attitude of robot directly from the image of camera. However, it is hard to control attitude of rapid or abrupt rotation due to the time delay and its low bandwidth. In this paper, we propose novel fusion filter frame that combines two sensor signals using Extended Kalman Filter (EKF) based on the modified track-to-track model to enhance the sensor characteristics. The simulation and experimental results show that the accurate attitude information can be obtained by the proposed fusion filter combining two sensor signals.