Aiming at the problem of attitude calculation drift caused by the error of the MEMS based attitude measurement system, an attitude algorithm based on variable-step-size momentum gradient descent is proposed. The algorithm improves the attitude calculation accuracy from two aspects of sensor data denoising and attitude calculating. In sensor data denoising, an improved recursive amplitude-limited mean filtering algorithm is proposed to reduce the computational complexity of microprocessors. In attitude calculating, an adaptive step size is designed based on gradient descent method, and momentum gradient is used to optimize the direction of each iteration so as to minimize the error after each iteration, and dynamic limiting filter is used to suppress angle oscillation. Results of the static and dynamic experiments show the effectiveness and superiority of the proposed algorithm.