Efficient Plane Extraction Based on Hierarchical Clustering

Plane extraction is a crucial task for many applications such as robot navigation, SLAM (simultaneous localization and mapping) and so on. Although there exists several of plane segmentation methods based on RANSAC (Random Sample Consensus), Hough transform, region growing etc. Some of these methods may not guarantee speed performance for computer vision tasks with real-time requirements. In order to improve the efficiency of the plane extraction algorithm, we propose a method based on agglomerative hierarchical clustering in this paper. Our method extracts planar surfaces in organized point clouds obtained from RGB-D sensors such as Microsoft Kinect in real time. We first divide point clouds into several groups of points as nodes. Those nodes represent point sets while the edges of the nodes represent neighborhoods. Next, we find nodes with the smallest plane fitting MSE (mean squared error) as initial nodes, and then perform agglomerative hierarchical clustering to merge nodes that belong to the same plane. We stop the step once the MSE is larger than the given threshold. Weoptimize the boundary of the extracted planes at last. We evaluate our method using the public TUM and SegComp datasets. Experiments show that the proposed approach can detect planar surfaces efficiently and correctly compared with other state-of-art methods.