Taking the navigation system for unmanned aerial vehicle (UVA) as research background, the error correction method of magnetic sensor and magnetic heading error after calibration have been researched in this work. By analyzing the magnetic sensor error sources, the horizontal plane ellipse direct fitting model is established. Considering some errors existing between GPS track angle and magnetic heading angle calculated by corrected magnetic sensor, a learning function between heading error angle and the magnetic heading is given, which assisted UAV navigation by using magnetic heading learned in the case of losing GPS. Flight test results show that the elliptical model and self learning function proposed in this paper can improve the accuracy of magnetic heading, which can assist the UAV to calculate heading angle in the case of GPS losing.