Vehicle Segmentation and Linear Interpolation-based Ground Occlusion Reconstruction for 3D Map Improvement

Simultaneous localization and mapping (SLAM) is crucial in robotics and autonomous navigation. In real-world environments such as parking lots, parked vehicles often appear in maps created through SLAM. However, these vehicles are dynamic objects that may disappear over time, making them unsuitable for map-based localization and navigation. Therefore, the SLAM-based map creation process typically removes them. However, this removal results in an occlusion area on the ground where vehicles were previously located. Thus, we propose a ground occlusion reconstruction method using linear interpolation. First, 3D semantic segmentation is employed to segment the vehicle point cloud. Object segmentation is then performed to identify individual vehicles, and their coordinates are estimated using an L-shape fitting method. The estimated vehicle coordinates transform the corresponding vehicle point cloud, allowing for accurate occlusion area determination. Finally, the point cloud inside the vehicle boundary on the ground is generated using linear interpolation. Experiments conducted in an underground parking lot with numerous parked vehicles demonstrate that the proposed method generates more accurate maps.