An Indoor Mobile Robot Pose Estimation Method Using Deep Learning Models for IMU Error Calibration and LiDAR Measurement Noise Covariance Tuning

Methods that fuse IMU and LiDAR data using an extended Kalman filter (EKF) are commonly employed to estimate the position and orientation of indoor mobile robots. Correcting the inherent errors of the IMU and mounting misalignments, as well as tuning the EKF measurement noise covariance according to the environmental conditions that affect the LiDAR performance, are required to improve localization and orientation estimation accuracy. This study proposes a method to address these issues by using two neural networks: one based on a long short-term memory network (LSTM) to compensate for IMU errors, and another based on a convolutional neural network (CNN) to estimate the error of the position and orientation measurements obtained from the LiDAR, which is then used to tune the EKF measurement noise covariance. To effectively train these networks, the CNN is first trained and then used to assist in training the IMU error compensation network. Experiments conducted in indoor environments, such as corridors and rooms, demonstrate that the proposed method significantly improves the position and orientation estimation performance of the robot compared to using only LiDAR or a conventional EKF.