According to the signal characteristics of inertial sensors in optoelectronic tracking system, an improved Kalman filtering method was designed, by which the AR model of denoised signals was built up and parameters were estimated after the signals were filtered by wavelet. Because the signals filtered by wavelet is approximative to actual ones, the variance of system noises was obtained by the variance of observation noises multiplying a coefficient less than 1. Finally, the experiments about real signals of some inertial sensors verify that the algorithm is efficient and has some catholicity.