Fast Point Cloud Registration for Urban Scenes via Pillar-Point Representation

Efficient and robust point cloud registration is an essential task for real-time applications in urban scenes. Most methods introduce keypoint sampling or detection to achieve real-time registration of large-scale point clouds. Recent advances in keypoint-free methods have succeeded in alleviating the bias and error introduced by keypoint detection via coarse-to-fine dense matching strategies. Nevertheless, the running time performance of such a strategy turns out to be far inferior to keypoint methods. This paper proposes a novel framework that adopts a pillar-point representation based feature extraction pipeline and a three-stage semi-dense keypoint matching scheme. The scheme includes global coarse matching, anchor generation and local dense matching for efficient correspondence matching. Experiments on large-scale outdoor datasets, including KITTI and NuScenes, demonstrate that the proposed feature representation and matching framework achieve real-time inference and high registration recall.