Fast plane detection for SLAM from noisy range images in both structured and unstructured environments

This paper focuses on fast plane detection in noisy range images. First, two improvements to the state-of-the-art region growing algorithm are presented to make it faster without losing precision for unstructured environments. One is to add the seed selection procedure based on local shape information to avoid blind growth. The other is to simplify the plane fitting mean square error computation complex. Second, a novel algorithm called grid-based region growing is presented for structured environments. The point cloud is divided into small patches based on neighborhood information when it is viewed as a range image. The small patch is called grid. Then the grids are classified into different categories according to their local appearance, including sparse, planar, spherical and linear. Finally, the planar grids are clustered into big patches by region growing. The plane parameters are incrementally computed whenever a new grid is added. The resulting planes can be used for 3D plane simultaneous localization and mapping (SLAM). Experimental results show promising plane detecting speed for both structured and unstructured environments.