This paper presents a new method for estimating the tilt angles of endoscopic images. Disorientation is one of the major challenges during natural orifice translumenal endoscopic surgery (NOTES). Reorientation allows surgeons or gastroenterologists to work in off-axis conditions and provides an important reference for coupling a secondary image. Some published studies of angle estimation for NOTES still have the limitation under the influence of movement or vibration. This study proposes a new sensor-fusion method for reducing the shock-based error. A triaxial accelerometer measures the gravitational vector (g-components) in all static states. When motion appears, the angular velocity from a triaxial gyroscope is used to calculate the elemental changes in g-components. A so-called predict-and-choose process relies on this data to predict the future state by giving many prediction values. The relationship between these values, the newest accelerometer readings, and their variation determine the final choice. Hence, under all conditions, the gravitational components are iteratively estimated to calculate the tilt angles. The result is evaluated by being applied in a well-known application, endoscopic horizon stabilization. Compared with the reference method, the proposed method has notable advantages. The simulation and experimental results show small errors, smooth angle change, and a small delay time. The tilt angles are reliable without any cumulative error under the prolonged motion. Therefore, this study gives surgeons or gastroenterologists an improved rectified image for reorienting under off-axis conditions. Further research will identify more applications for the development of surgical instruments for NOTES.