AWV-MOS-LIO: Adaptive Window Visibility Based Moving Object Segmentation With LiDAR Inertial Odometry

Segmenting points on dynamic objects and static objects is essential for performing various tasks in autonomous driving, such as simultaneous localization and mapping (SLAM), navigation, and collision avoidance. Conventional visibility-based moving object segmentation (MOS) methods that check free space by using LiDAR beams have been facing issues related to false predictions owing to the large incidence angle of the beam, inaccurate sensor pose, and overlapping views of slowly moving objects. To address these issues, we propose the AWV-MOS-LIO system, which performs adaptive window visibility-based MOS with LiDAR inertial odometry (LIO). The AWV-MOS-LIO consists of AWV-MOS with Object Scale Test (OST) and AWV-LIO. The AWV-MOS with OST, which utilizes adaptive window-based point-to-window comparison and object-level false prediction check, improves the MOS performance by preventing the false predictions resulting from large incidence angle and inaccurate sensor pose. In addition, AWV-MOS utilizes keyframes to check free space in various views. AWV-LIO minimizes pose ambiguity through scan-to-submap matching by utilizing static scans and static weights, and it corrects point distortion by using IMU odometry. We evaluate the online MOS performance of the proposed system on the KITTI Raw dataset and its odometry estimation performance on the UrbanLoco dataset. The proposed method achieves 6.3% higher online MOS performance (IoU) than the deep learning-based method, and 14.4% higher odometry estimation performance (ATE) than the baseline LIO method. The code will be released at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/KimSeongJun-kr/AWV-MOS</uri>.