An improved Sage-Husa adaptive extended Kalman filter algorithm is proposed to ensure the precision and stability of calculating attitude angles of a multi-rotor Unmanned Aerial Vehicle (UAV) under the actual flight conditions, such as unknown and time-varied noise statistical properties, main disturbance source in vibration and attitude angles high dynamically changed. The algorithm uses attitude angle variance estimated by a gyroscope in real time to estimate system noise variance and only adopts an adaptive filter algorithm to estimate measurement noise variance on-line to ensure the precision and stability of filtering. Meanwhile, it introduces the criterion of filter convergence to restrain the divergence of Kalman filter through combining with a strong tracking Kalman filter algorithm. A flight experiment and corresponding analysis show that the root-mean-square errors of the pinch and roll angles estimated by the improved algorithm are 1.722 and 1.182, obviously better than that of the conventional Sage-Husa adaptive Kalman filter algorithm. It concludes that the improved algorithm has strong adaptive ability, good real-time performance, high precision and reliable operation. It meets the need of multi-rotor UAV autonomous flight and can be applied to other navigation information measuring systems with high dynamic performance requirements if the parameters are modified appropriately.