The sensors in the inertial measurement unit have strong nonlinearity and noise uncertainty, resulting in large errors and divergence when using the conventional Kalman filtering. In order to solve this problem, an improved adaptive incremental Kalman filter (AIKF) algorithm is proposed in this paper. The algorithm uses complementary filtering to fuse the data from accelerometers, magnetometers and gyroscopes, and the filtered data increment is used as the observation of the Kalman filter. At the same time, the adaptive online estimation of the system noise is carried out to obtain accurate attitude output. The experimental results show that the algorithm can achieve accurate measurement of attitude, the errors of the pitch angle and roll angle are less than 0.05° and the heading error is less than 0.15° with the rolling table test, and it has a better ability to suppress the noise.