First Real-World Results of a Deep Neural Network Assisted GNSS/INS Kalman-Filter with MEMS Inertial Sensors for Autonomous Vehicle

Abstract

The integration of global navigation satellite system (GNSS) and inertial navigation system (INS) is a powerful technology that provides accurate, available, and continuous navigation solutions, which is critical for autonomous vehicles (Mikhaylov et al., 2020). Due to the advancements in micro-electromechanical system (MEMS) inertial sensor technology, the use of low-cost, small size, and low power consumption MEMS inertial measurement units (IMU) becomes attractive for land vehicles (Li et al., 2019; Yang et al., 2014). However, the INS cannot operate stand-alone to provide long-term accuracy in the GNSS challenging environments because the errors in the IMU measurements are integrated into the navigation solutions (Woodman, 2007). The accumulated errors and the IMU measurement errors are usually estimated by an error-state extended Kalman filter (ES-EKF) (Madyastha et al., 2011). The performance of the integration algorithm is highly dependent on the knowledge of noise statistics and system models. The noise covariance matrices are formulated empirically under independent Gaussian noise assumptions whereas the system models are designed by linearizing the nonlinear equations of the system. Considering the highly nonlinear error propagation and the complex IMU error model of low-cost MEMS IMU, the ES-EKF based GNSS/INS integration is not sufficient for meeting the navigation requirements of land vehicles. In order to address the nonlinear issue, several advanced integration algorithms are utilized such as unscented Kalman filter (Meng et al., 2016), cubature Kalman filter (Cui et al., 2017) and factor graph (Wen et al., 2021). An alternative approach is to estimate other IMU error components other than bias (Godha, 2006). Despite advancements, these algorithms are still unable to optimally address nonlinear issues or require significant computational loads. On the other hand, external sensors such as odometer, lidar, and camera can be integrated into the system to improve the performance by providing additional measurements (Chiang et al., 2019). The use of auxiliary sensors could limit the application areas and increase costs. Given the remarkable success of deep learning (DL) in various fields and the impressive learning capability of deep neural networks (DNN) (LeCun et al., 2015), we present a DL-assisted integration algorithm in this paper.