LGC-Net: Lightweight Gyroscope Errors Compensation Network for Effective Attitude Estimation

Abstract

This paper presents a lightweight, effective neural network model for compensating errors of low-cost microelec-tromechanical system (MEMS) gyroscope and estimating the attitude of a robot in real time. Our proposed LGC-Net captures both local and global features from the time window of inertial measurement units (IMU) measurements to dynamically regress the output compensation components of the gyroscope. LGC-Net leverages the Depthwise Separable Convolution to capture various scale features and minimize the network model parameters after carefully deriving a mathematical calibration model. Moreover, the Large Kernel Attention is employed to improve feature representation and learn long-range dependencies. We evaluate our algorithm on two public datasets, EuRoC and TUM-VI, and demonstrate that LGC-Net can effectively and accu-rately estimate the orientation from raw IMU data. The experiment results show that the estimated attitude using LGC-Net is comparable to top-ranked visual-inertial odometry systems, even though our approach does not use vision sensors.