3D Object Detection Based on Geometrical Segmentation

Objects are often occluded in cluttered real-world environments, but how to detect occluded objects effectively is seldomly studied. In this paper, we introduce an approach to identify and localize occluded objects by taking advantage of RGB-D data from a RGB-D camera, such as a Microsoft Kinect. Our approach identifies an object based on features of its geometrical surfaces obtained from segmentation rather than global features obtained from treating the object as a whole. Geometrical surfaces are segmented from the RGB-D data based on depth and surface normal continuity. Color and SIFT features are extracted to describe each surface. Experiments show that our approach can detect heavily occluded objects robustly and efficiently.