Low-Cost MEMS IMU Calibration Using Deep Learning and Visual-Inertial Odometry

In the past decade, machine learning algorithms have shown remarkable potential in comprehending and interpreting errors and their origins in Inertial Measurement Units (IMUs), Various strategies have been developed for IMU calibration and error compensation, assuming reliable ground truth data for training. The main challenge lies in creating a suitable training dataset, which often requires precise equipment that contradicts cost-effective calibration, especially for commercial use. This study introduces a pragmatic calibration framework for low-cost MEMS inertial measurement units, utilizing deep convolutional neural networks (CNNs). The lack of ground truth during learning is addressed by incorporating preliminary visual-inertial odometry results through a loosely coupled fusion algorithm based on the error-state extended Kalman Alter (ES- EKF). The proposed framework is validated for stationary, dynamic errors, and path dependency. With a precisely designed CNN and ample training and validation data, the simulation results confirm the reliability of the initial 6 seconds of ES-EKF flight navigation as a dependable CNN training reference. The scheme significantly reduces errors in a low-grade IMU attached to a quadcopter, coupled with a pre-calibrated camera, by an impressive 93.22%. The aim is to establish a highly accurate and robust offline calibration process, harnessing only quadcopter components. Moreover, the loosely coupled fusion approach empowers the trained calibration module to provide real-time online functionality within the Visual-Inertial Navigation System (VINS) by optimizing memory, CPU, and GPU usage.