Loop Closure Detection Based on Geometry of Semantic Point Cloud

The spatiotemporal constraints provided by loop closure detection play a crucial role in the simultaneous localization and mapping of robots. Traditional methods based on key points information are vulnerable to point-level feature instability. In this paper, a loop closure detection method based on robust geometry of the objects extracted through semantics of point cloud is proposed. Our method is based on the fact that point cloud can robustly describe geometric properties, and the extracted local features are extremely robust and invariant to rotation, thus achieving remarkable results in loop closure detection. We conduct experiments on the KITTI odometry dataset and confirm the effectiveness of our method.