Abstract:This paper presents a lightweight, efficient calibration neural network model for denoising low-cost microelectromechanical system (MEMS) gyroscope and estimating the attitude of a robot in real-time. The key idea is extracting local and global features from the time window of inertial measurement units (IMU) measurements to regress the output compensation components for the gyroscope dynamically. Following a carefully deduced mathematical calibration model, LGC-Net leverages the depthwise separable convolution to capture the sectional features and reduce the network model parameters. The Large kernel attention is designed to learn the long-range dependencies and feature representation better. The proposed algorithm is evaluated in the EuRoC and TUM-VI datasets and achieves state-of-the-art on the (unseen) test sequences with a more lightweight model structure. The estimated orientation with our LGC-Net is comparable with the top-ranked visual-inertial odometry systems, although it does not adopt vision sensors. We make our method open-source at: <a class="link-external link-https" href="https://github.com/huazai665/LGC-Net" rel="external noopener nofollow">this https URL</a>

What problem does this paper attempt to address?

This paper attempts to solve the noise and error problems of low - precision micro - electro - mechanical system (MEMS) gyroscopes in robot pose estimation. Specifically, the researchers proposed a lightweight deep calibration network LGC - Net, aiming to improve the existing pose estimation methods in the following aspects: 1. **Extract local and global features**: Extract local and global features from the time window of the inertial measurement unit (IMU) to dynamically regress the compensation components of the gyroscope. 2. **Reduce model parameters**: Use depthwise separable convolution to capture local features and reduce network model parameters. 3. **Learn long - range dependencies**: Design a large - kernel attention mechanism to better learn long - range dependencies and feature representations. ### Main contributions 1. **Propose a lightweight calibration network LGC - Net**: Based on depthwise separable convolution and large - kernel attention mechanism, input the original MEMS IMU data and estimate the calibration compensation components. 2. **Achieve advanced pose estimation performance**: Using only the calibrated gyroscope data, achieve the state - of - the - art pose estimation performance with fewer model parameters. 3. **Experimental verification**: Conduct a series of experiments on the EuRoC and TUM - VI public datasets, verify the effectiveness of the proposed method, and open - source the algorithm. ### Specific problem description #### 1. Noise and error of MEMS IMU Low - precision MEMS IMUs are usually affected by complex, nonlinear, time - varying noise and random errors, resulting in inaccurate pose estimation. Therefore, developing more efficient calibration algorithms is crucial for improving the robustness and accuracy of pose estimation. #### 2. Limitations of existing methods - **Discrete calibration**: Rely on high - precision external devices (such as three - axis turntables), which are costly and cannot handle high - order coupling terms and time - varying characteristics. - **System - level calibration**: Although it reduces the dependence on external devices, it is difficult to remove nonlinear and non - Gaussian random errors. - **Traditional statistical methods**: Such as ARMA, Allan variance, Kalman filtering and wavelet transformation, etc., each has limitations and cannot comprehensively handle complex errors. #### 3. Advantages and challenges of deep learning methods In recent years, deep learning techniques have been introduced into the visual - inertial navigation field, improving the navigation and positioning accuracy. However, the existing deep - learning - based IMU calibration methods fail to consider local features and long - range dependencies simultaneously, and the models are not lightweight enough and difficult to be deployed to edge devices. ### Design ideas of LGC - Net LGC - Net solves the above problems in the following ways: - **Depthwise separable convolution**: Reduce model parameters, accelerate training and inference. - **Large - kernel attention mechanism**: Enhance the ability to learn long - range dependencies. - **Loss function design**: Adopt the Log - cosh loss function, which combines the advantages of L1 and L2 losses and reduces the fluctuation of the gradient near the minimum value. ### Experimental results The experimental results of LGC - Net on the EuRoC and TUM - VI datasets show that it performs excellently in pose estimation. Especially on unseen test sequences, the average absolute orientation error (AOE) is significantly reduced, and the number of model parameters is far less than that of existing methods. In conclusion, this paper successfully solves the noise and error problems of low - precision MEMS gyroscopes in real - time pose estimation by proposing LGC - Net, providing a new solution for low - cost and high - efficiency pose estimation.

LGC-Net: A Lightweight Gyroscope Calibration Network for Efficient Attitude Estimation

Learning to Compensate for the Drift and Error of Gyroscope in Vehicle Localization

TinyGC-Net: An Extremely Tiny Network for Calibrating MEMS Gyroscopes

Observability-Aware Intrinsic and Extrinsic Calibration of LiDAR-IMU Systems

Denoising IMU Gyroscopes with Deep Learning for Open-Loop Attitude Estimation

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

A MEMS IMU Gyroscope Calibration Method Based on Deep Learning

Robust Attitude and Heading Estimation under Dynamic Motion and Magnetic Disturbance

In-Field Gyroscope Autocalibration with Iterative Attitude Estimation

MEMS Gyroscope Multi-Feature Calibration Using Machine Learning Technique

Compounded Calibration Based on FNN and Attitude Estimation Method Using Intelligent Filtering for Low Cost MEMS Sensor Application

Rapid Gyroscope Calibration: A Deep Learning Approach

LCCNet: LiDAR and Camera Self-Calibration using Cost Volume Network

A Dual-Linear Kalman Filter for Real-Time Orientation Determination System Using Low-Cost MEMS Sensors

Robust Equipment-Free Calibration of Low-Cost Inertial Measurement Units

Low-cost MEMS gyroscope performance improvement under unknown disturbances through deep learning-based array

Attitude Estimation Using Iterative Indirect Kalman With Neural Network for Inertial Sensors

On-site scale factor linearity calibration of MEMS triaxial gyroscopes

Dynamic Attitude Estimation Improvement for Low-cost MEMS IMU by Integrating Low-cost GPS

Inertial Measurement Unit Self-Calibration by Quantization-Aware and Memory-Parsimonious Neural Networks

Optimal Compensation of MEMS Gyroscope Noise Kalman Filter Based on Conv-DAE and MultiTCN-Attention Model in Static Base Environment