118 / 2023-09-20 14:23:30

Semantics and Spatial Structures Guided Point Cloud Registration

point cloud registration,semantic-guided,spatial structure,Simultaneous Localization and Mapping

Special Sessions > SS05 - Computer Vision for Automation Science and Engineering

Point cloud registration in SLAM plays a key role in mapping and localization tasks, which involves finding the transformation between two sets of point clouds. Iterative-based algorithms in point cloud registration tend to perform well on different shapes of point clouds but can be highly dependent on those initial positions. There is often a trade-off between speed and precision in these algorithms. In response to the trade-off in iterative point cloud registration algorithms, this paper proposes a semantic and spatial structure guided point cloud registration algorithm. First, the semantic spatial structure based coarse registration algorithm uses semantic extracted in space to segment spatial centroids of various object categories. Graph matching is performed after dimension reduction to obtain the transformation relationships of corresponding point clouds, which are then applied to transform the initial point cloud, resulting in a reasonably good initial alignment. Second, in the fine registration, the Manhattan assumption and semantic information are used to filter the point cloud for registration, thereby accelerating the convergence of the iterative algorithm. Experiments shows that when compared to the state-of-art method, the proposal requires only around 25% of the iteration time, and achieves higher registration accuracy, especially when dealing with initial registration errors.