Abstract

Abstract MEMS gyro is widely used in the field of autonomous system navigation, but due to the large measurement error of low-cost MEMS gyro, the orientation estimation of gyro carriers cannot be met by simple calibration only. In this paper, we propose a learning method to correct the measurement error of MEMS gyro in IMU. Our method utilizes dilated convolution to increase the receptive field, designs a lightweight attention module to extract the gyro random error, and uses channel transformation to extract deterministic errors. We also design a multi-timescale loss function, enabling the network to notice the cumulative orientation errors at different timescales. We tested our method on public datasets EUROC and TUM-VI and compared it with the Visual-Inertial Odometry (VIO) methods as well as other gyro correction methods. The experimental results show that the proposed method can have comparable or even higher accuracy of orientation estimation than the visual inertial combination method, and the gyro orientation estimation error using MGCNet correction is reduced by 15-20% compared to other advanced gyro processing methods.