An Efficient Scene Semantic Labeling Approach for 3D Point Cloud

摘要

Using point cloud to label objects of different categories in 3D scenes is a hard task because of their complex topological structure. In this paper, we propose an efficient approach to extract point's descriptor by employing our Voxel-Neighbor Structure. Using classifier learned via Random Forest, we label the scene into semantic categories. Finally, by using Conditional Random Fields with additional contextual relationship we define at first, we build up the semantic affiliation between points and improve the performance by minimizing energy function using graph cut. Experiments based on Oakland 3-D Point Cloud Dataset demonstrate that our proposed method is effective and robust comparing to state-of-the-art.