Establishing the model of FOG’s random drift and compensating in the filter is an effective method to improve the output precision of FOG. For traditional random drift of fiber optical gyroscope has some shortages like off-line, needing pre-process and the off-line models, which are usually not universal for environmental changing, a new modeling and filtering way is put forward. First, based on a large amount of measured data, the traditional off-line AR model is improved, and a new method to build the model of FOG’s random drift is studied. Then, the comparison is made between the traditional Kalman filter and H∞ filter in real time. The result demonstrates that improved AR model has much applicability and the performance of H∞ filter is better than Kalman filter. The minimum value of fitting accuracy is 91.6% and H∞ filter can improve the performance of filtering by almost 38.5% when analyzing ingle noise.