3D model estimation using a single RGB-D image

Model estimation is important for robotic grasping. Since in order to make a decision about how to grasp the object, the robot should know where the target is and what it looks like. In this paper we propose a method of 3D model estimation using a single RGB-D image. The target object is segmented out from background using RANSAC and convex hull algorithm. And a clustering method is designed to separate different objects. After recognizing the types of the objects, the parameters of the model of objects are estimated according partial observed information and the symmetrical property of objects. Although the parameter estimation process may vary for different kinds of model, yet the key of the method are RANSAC and Ordinary Least Squares (OLS). Experimental results show that our method is effective in model analyzing.