A method to segment a 3D surface point cloud for selective sensing in robotic exploration

Autonomous robotic exploration in a 3D environment requires the acquisition of 3D data to create a consistent internal model of the environment from which objects can be recognized for the robot to interact with. As the acquisition of 3D data with stereo vision or a laser range finder can be a relatively long process, selective sensing is desired to optimize the amount of data collected to accurately represent the environment in a minimal amount of time. In order to perform selective sensing, a coarse acquisition of the environment first needs to take place. Regions of interest, such as edges and other boundaries, can then be identified so that an acquisition with higher spatial resolution can occur over bounded regions. For that purpose a segmentation method of the coarse data is proposed so that regions can be efficiently distinguished from each other. The method takes a raw 3D surface profile point cloud of varying point densities, organizes it into a mesh, and then segments the surfaces present in this point cloud, producing a segmented mesh, as well as an octree of labeled voxels corresponding to the segmentation. This mesh and octree may then be used for sensory selection to drive a robot exploration task. The method is demonstrated on actual datasets collected in a laboratory environment.