A Learning-based Noise Tracking Method of Adaptive Kalman Filter for UAV Positioning

The Unmanned Aerial Vehicle (UAV) is widely used in transportation tasks, disaster rescue, reconnaissance and so on, all of which require high-accuracy positioning. The Global Positioning System (GPS) and the Inertial Navigation System (INS) are two commonly used positioning systems, but both have their shortcomings. When using GPS alone, it will cause large errors due to environmental interference. While when using INS alone, there will be cumulative errors increase with time. Kalman Filter (KF) can overcome the shortcomings of a single GPS or INS system by fusing them together. We propose a learning-based noise tracking method for KF (which is simply called RLKF). The RLKF learns the rule of noise change based on reinforcement learning. The output action of RLKF adjusts noise covariance matrix to the most suitable value, thereby strengthening the adaptive ability of KF. We make strategies output from continuous space by constructing the reinforcement learning model by Deep Determined Policy Gradient (DDPG). By comparing with the adaptive extended KF, experimental results show that the RLKF has better performance, and navigation accuracy is improved.