Evidential Sensory Fusion of 2D Feature and 3D Shape Information for 3D Occluded Object Recognition in Robotics Applications

Abstract

An evidential sensory fusion using 2D feature and 3D shape information method is proposed to recognize the occluded object. For the applications of robotic object fetching, the conventional object recognition methods usually applied the algorithms separately from 2D texture matching or 3D shape fitting. It often causes the wrong recognition results when the objects are occluded. The motivation in this study is to enhance the occluded object recognition via the estimate fusion method from the RGB-D sensor, which provides both 2D image and 3D depth information. To associate the 3D shape with the 2D texture, the region of interest (ROI) is firstly captured in 3D coordinate system, and mapped onto the 2D image. The Dempster-Shafer (DS) evidence theory is applied to fuse the confidences from the recognitions of both 2D texture and 3D shape to increase the recognition rate of occluded objects. The experimental results successfully demonstrate that the proposed evidence fusion recognizes the sample object correctly where it usually has the lower confidences from 2D and 3D recognition algorithms alone, when it operates in a separate fashion.