Real-time plane segmentation for scene understanding in robot navigation

Plane segmentation is one of the most fundamental procedures for scene understanding in robot navigation. However, the methods based on curvature and other higher level derivations often lead to over segmentation when used to segment point cloud with high noise and outliers. To overcome these problems, we proposed a real-time plane segmentation algorithm based on region growing which can be used to segment different planes for scene understanding in indoor environment. The proposed method contains three stages: data processing, normal estimation and region growing. In the data processing step, a pass-through filter was used to remove unnecessary points. When estimating the normal for each point, the proposed method uses the combination of octree and KD-tree search method to search the k nearest neighbors of each point. This search method can also be used in region growing to select neighboring points of the current seed which makes the proposed method more efficient. To keep balance between efficiency and over-segmentation, we presented a segment smoothness parameter according to the points curvature. What's more, the proposed method can also detect different obstacles in the scene at the same time. The experiment using the public datasets show that the proposed algorithm can be used in robot navigation for fast planes segmentation and obstacles detection, while achieving precision, recall and accuracy as much as 99.88%.