LIO-DOR: a novel LiDAR/inertial odometry system with real-time dynamic object removal

Simultaneous localization and mapping (SLAM) technology has attracted significant attention for its ability to enable intelligent autonomous vehicles to perform localization and mapping in unknown environments. Traditional LiDAR SLAM techniques largely rely on static assumptions, but in real-world scenarios, the presence of dynamic objects often renders these assumptions invalid, thereby affecting localization accuracy and mapping quality. To address this issue, we propose a LiDAR/Inertial SLAM dynamic object removal system, LIO-DOR. Specifically, we propose a dynamic point cloud region identification algorithm that combines point cloud volume and intensity information, using different descriptor selection strategies to identify regions where dynamic objects appear. Furthermore, based on groundtruth fitting results, we propose a dynamic point cloud clustering removal algorithm, designing a dynamic adjustable threshold factor to efficiently remove dynamic point clouds while preserving valid static point clouds. Finally, we validated the proposed LIO-DOR system on the open-source KITTI and UrbanNav datasets, as well as our self-collected real-world dataset. Compared to baseline methods, our LIO-DOR algorithm achieves a 3.52% improvement in overall dynamic object filtering performance on the KITTI dataset. In terms of mean positioning accuracy, the system achieves maximum improvements of 21.88% and 16.37% on the UrbanNav dataset and self-collected dataset, respectively. The results demonstrate that our method LIO-DOR achieves higher dynamic point cloud removal rates and lower absolute trajectory error.