When the UAV flies in the attitude mode, the attitude angle error fluctuates greatly.According to the complementary characteristics of magnetometer, accelerometer and gyroscope, an Adaptive Unscented Kalman Filter (AUKF) algorithm is proposed to optimize the MEMS sensor data.The attitude quaternion and gyro drift are taken as state variables, and the output of accelerator and magnetometer is taken as measurement variables.The gradient descent algorithm is used to optimize the key parameter of Unscented Kalman Filter, namely, process noise covariance, so as to improve the accuracy of attitude calculation.The analysis of actual flight data shows that the proposed method has the highest accuracy compared with conventional Kalman filter and traditional unscented Kalman filter, and can ensure flight stability of small UAVs in various situations.