Contextual Global Registration of Point Clouds in Urban Scenes

This paper presents an innovative and powerful approach for contextual global registration of partially overlapping point clouds in urban scenes. First, a fast and robust strategy for extraction of feature points from large-scale scanning scenes is proposed and used to represent the scanning scene. Correspondences are then established using a contextual rule-based method at the two-dimensional and three-dimensional levels. A penalization strategy is then introduced to generate dense corresponding sets between two overlapping point clouds. Finally, a three-stage optimal strategy is implemented to efficiently match these point clouds. The proposed approach highlights the following aspects: (1) The designed voting-based feature extraction method can efficiently and robustly represent a scanning scene in a complex environment with a huge database; (2) The contextual information enables the generation of more reliable correspondences beyond the geometric features; (3) The novel penalization strategy and the three-stage optimal method benefit the approach to achieve tight alignment at a lower computational cost. State-of-art baseline algorithms from the field of photogrammetiy are used to evaluate the performance of the proposed method in the comprehensive experiments. The presented approach outperforms other methods in registration accuracy and computational cost.