The drift signal of fiber optic gyroscope (FOG) is often buried in noise. It is difficult to compensate drift directly, and a novel threshold filtering method based on empirical mode decomposition (EMD) (designated as EMD-T) is proposed as a pre-processing tool. Based on the noise spread model of EMD, a bounded noise assist analysis method is introduced to improve the decomposition accuracy of EMD. The noises with low magnitude and high frequency are compressed into early intrinsic mode functions. The static output of interferometric FOG is adopted to verify the effectiveness of EMD-T. Comparison analysis with filtering methods based on wavelet packet transform (WPT) and conventional EMD (CEMD) is done. Experimental and Allan variance analysis results show that EMD-T outperforms denoising method based on WPT and CEMD. The quantization noise (Q) and angle random walk (N) are decreased from 0.7862 μrad and 4.58×10-3（°）·h-1/2 to 0.1340 μrad and 9.03×10-4（°）·h-1/2, respectively, after applying EMD-T.