ANALYSIS OF THE KALMAN FILTER WITH DIFFERENT INS ERROR MODELS FOR GPS/INS INTEGRATION IN AERIAL REMOTE SENSING APPLICATIONS

摘要

In the Kalman filter used for the integration of GPS/INS, the inertial sensor error model is usually considered as a random constant or random walk for both gyroscopes and accelerometers. However, the Inertial Measurement Unit (IMU) used in aerial remote sensing applications for sensor positioning and orientation is typically of tactical grade, i.e., the gyroscope drifts are on the order of 0.1 deg/h and the accelerometer biases are 100ug respectively. In this case, there is the room to improve the system performance by developing more complicated error models for the inertial sensors. In this paper, 6-state, 12-state and 15-state error models for the inertial sensors are implemented, and their performance of each in the Kalman filter is compared and analyzed. Firstly, the commonly used 6-state error model that includes three random walks for gyroscopes and three random walks for accelerometers is implemented. Then, a 12-state error model is formed by augmenting the 6-state model with three scale factors for the gyroscopes and three scale factors for the accelerometers. Thirdly, three first-order Markov procedures are considered for the gyroscopes in addition to the random walks and scale factors, thus resulting in a 15-state error model. Aerial GPS/INS data collected in the field with a tactical grade IMU and dual frequency GPS receivers is processed with these three error models. In the data processing, the loosely-coupled Kalman filter, which is the common coupling method for the aerial GPS/INS integration, is used. The 12-state and 15-state error models show obvious advantages over the 6-state error model in the test results. The accuracies of the integrated position (5cm), velocity (3cm/s) and attitude (0.002 degree for pitch and roll, 0.008 degree for heading) in the 12-state model are all better than that of the 6-state error model. However, the improvement of the 15-state error model relative to the 12-state error model is limited and insignificant.