A modified Kalman real-time adaptive filter method was proposed for micro-electro-mechanical system gyro which is easy to be susceptible to environmental influences and stability, and lead to the problems that the established random drift model parameter and the noise statistical property change. The order of random drift ARMA model was determined by analyzing the measured data. On this basis, the recursive least square method was used to update the model parameters in real time. According to the characteristics of gyro noise parameters, a real-time filtering method of Allan variance analysis based on fading memory factor and Sage-Husa adaptive filter algorithm was proposed to estimate the parameters of Q and R at the same time. The coupling and restriction of system state estimation and measurement noise parameter estimation were avoided. The experimental results show that, compared with the standard Kalman filter compensation method, the proposed method in this paper can compensate the random drift error of MEMS gyro more effectively in real time, and has better adaptability and stability.