Neural Network-Based Real-Time Odometry Using IMU for Crane System and Its Application to Large-Scale 3D Mapping

Abstract

This paper introduces a method for neural network-based real-time odometry using an IMU (inertial measurement unit) and its application in large-scale 3D mapping using a slowly rotating 2-D LiDAR. In this method, a neural network consisting of a convolutional neural network (CNN) and long short-term memory (LSTM) is employed to estimate the change in pose. Firstly, online pre-filtering using a low-pass filter is implemented on the time windows of IMU measurements before feeding them as the input to the neural network to estimate the change in position and rotation of the sensor. After that, the estimated sensor pose is used to register the scans of 2D-rotating LiDAR to build a large-scale 3D map. The proposed method is tested in a gazebo environment by attaching the sensors to a crane boom. In this study, we also investigate the impact of different time windows of IMU measurements on the accuracy of pose estimation by the neural network.