Gyro-Net: IMU Gyroscopes Random Errors Compensation Method Based on Deep Learning

Abstract

To solve the problem of inaccurate orientation estimation after long-term operations of the Inertial Measurement Unit (IMU), we present a learning-based method (called Gyro-Net) to estimate and compensate for IMU gyroscope random errors. We firstly introduce a semi-dense network structure, which extracts different scale features by IFES Block (IMU Feature Extraction & Selection Block), and adopts skip-connections and transition layers to adjust the feature pipeline. In this way, we can reuse features between different blocks before and after feature extraction, selection and compression. Driven by a proposed absolute and relative loss, the network can be trained and achieve the reduction of cumulative estimated orientation errors. The experimental results in public datasets show that our method can effectively and accurately estimate the orientation from raw IMU data. Moreover, we apply the network output directly to Open-VINS, and the results show that Gyro-Net can improve the accuracy of pose estimation for Open-VINS, especially in scenarios where camera-based estimation often struggles (e.g., fast motion, drastic lighting, viewpoint changes and motion blur).