Split and Merge for Accurate Plane Segmentation in RGB-D Images

In this paper, we propose an accurate and efficient method to detect planar surfaces indoors based on an RGB-D camera. First, we segment the RGB image using a graph-based segmentation approach because of its efficiency and capability in preserving sharp region borders. The graph-based color segmentation methods usually result in over-segmentation or under-segmentation. Then to achieve better plane segmentation results, we propose a split-andmerge strategy. We first segment the planes in the split step by applying a random sampling and consensus (RANSAC) approach to each graph-derived point cloud based on a plane-fitting mean squared error (MSE). In the merge step, we can simultaneously merge some over-segmented regions obtained from the split step by a maximal clique clustering approach. Experiment demonstrates that our plane segmentation algorithm can detect planes indoors at a frame rate of 10Hz, and can achieve very promising performance.