Experimental Evaluation of GNSS and IMU Fusion Using Gated Recurrent Unit

In this paper, a data-driven Inertial navigation systems (INS) and Global Navigation Satellite System (GNSS) fusion algorithm based on the use of the Gated Recur-rent Unit (GRU) is proposed. In this project, we trained the GRU neural network with Inertial Measurement Unit (IMU) raw data and GNSS Position, Velocity and Timing (PVT) solutions as input and the position difference between GNSS and ground truth as labels. Therefore, the trained model can estimate the rover’s positions by subtracting the predicted GNSS error from GNSS positions given IMU raw measurements and GNSS PVT solutions. To evaluate the performance of GNSS/INS fusion algorithms in realistic scenarios, we developed an experimental platform. Our experimental platform consists of a moving test rig and an external validation system. The moving test rig consists of a rover equipped with an LPMS-CU2: 9-Axis Inertial Measurement Unit (IMU) and U-Blox ZED-F9P GNSS receiver. For validation purposes, we employ an onboard real-time kinematic positioning (RTK)-GNSS receiver. The test scenarios include both open-sky and challenging conditions near buildings, which is beneficial for devolving and testing urban navigation systems. After training with collected experimental data in multiple test scenarios, the proposed algorithm is able to improve GNSS positioning accuracy by more than 60% for the open-sky environment and 30% for the urban environment.