Autonomous Driving Vehicles Using Adaptive Learning Method for Data Fusion

Abstract: This paper presents an adaptive learning method for data fusion in autonomous driving vehicles. The localization is based on the integration of Inertial Measurement Unit (IMU) with two Real-Time Kinematic (RTK) Global Positioning System (GPS) units in an adaptive Kalman filter (KF). The observability analysis reveals that $i$) integration of a single GPS with IMU does not constitute an observable system; $ii$) integration of two GPS units with IMU results in a locally observable system provided that the line connecting two GPS antennas is not parallel with the vector of the measured acceleration, i.e., the sum of inertial and gravitational accelerations. The latter case makes it possible to compensate for the error in the estimated orientation due to gyro drift and its bias without needing additional instruments for absolute orientation measurements, e.g., a magnetic compass. Moreover, in order to cope with the fact that GPS systems sometimes lose their signal and receive inaccurate position data, the self-tuning filter estimates the covariance matrix associated with the GPS measurement noise. This allows the KF to incorporate GPS measurements in the data fusion process heavily only when the information received by GPS becomes reliably available.