Continuous 3D Scan Update for Resilient IMU Preintegration in Tightly Coupled Lidar-Inertial SLAM

Abstract

Simultaneous localization and mapping, which is one of the most important techniques that allow the autonomy of mobile robots, has evolved in the last decades to provide high-performance localization and mapping. In particular, the algorithms that employ factor graph optimization have utilized a multi-sensor configuration to construct a complementary system to overcome the drawbacks of each sensor. In this thesis, a resilient IMU preintegration technique through continuous 3D scan update and adaptive noise covariance is proposed to improve the estimation accuracy of tightly coupled Lidar-Inertial SLAM. First, a method to continuously update point cloud data obtained from Lidar is introduced by defining a fixed size window along with the range measurement and moving it. In this way, Lidar measurement is updated at a higher frequency, producing a short time interval for the IMU model estimate that prevents a divergence in the preintegration step. Second, a noise covariance of the estimated key poses is configured with an adaptive uncertainty prior inserted as the preintegrated measurements into the factor graph. Providing an expected uncertainty of IMU propagation as a noise covariance in the pose domain, the uncertain state estimate can be revised with the other states and observations in confidence during the global optimization process. Then, a series of tests are conducted in normal and aggressive maneuvering conditions to evaluate the proposed method. The qualitative and quantitative analyses of the results are presented, comparing generated point cloud map and the trajectory to verify the improved accuracy and stability of the proposed framework.