GNSS/INS sensor fusion using Kalman filter with covariance adaptation

This paper proposes a new method which estimates the Kalman filter's measurement noise covariance matrix in a loosely coupled GNSS/INS sensor fusion system. The signals received from the satellites are heavily influenced by the environmental conditions, and will cause the noise variance to change dramatically. Thus, estimating the measurement noise to improve the fusion quality is very important. Most of the proposed methods use the moving average method to estimate the measurement noise online from the innovation sequence of the Kalman filter in the window interval. However, those methods can only have a good performance in a quasi-static environment, e.g., freeway where the signals are rarely blocked. In dynamic environments such as the urban area, the satellite signals are often blocked by the buildings, resulting in a large range of the noise variance. This paper proposes to use Savitzky-Golay filter with degree adaptation (ADSG) to estimate the measurement noise. The ADSG filter uses local polynomial regression method to fit the delayed-sample innovation sequence in the window interval, and adjust the polynomial order by the statistics F-test. Experiments show that the proposed method's horizontal position error is about 0.3m, which is better than several existing methods.