Vision-Based Robotic System for Polyhedral Object Grasping using Kinect Sensor

Since the introduction of the Microsoft Kinect sensor, real-time 3D perception of objects in a close-range scene has become one of the major trends in the current research on robotics and computer vision domain. As part of a long-term goal to develop a robust object recognition system for industrial applications, this paper focuses on the research topic of 3D planes recognition and separation for polyhedral object grasping using industrial manipulators and Kinect as a 3D vision system. Since the data acquired by the vision sensor such as Kinect are invariably noisy, a pipeline of algorithms has been implemented to reduce the noise and clustering the data points properly. With the feature vectors computed for each data point that represent the object of interest given, a split and merge algorithm has been developed to cluster the data exploiting the flexibility and robustness of the fuzzy logic inference system and then fitting each segment to a planar model using RANSAC. This work mainly uses the eye-to-hand Kinect-based vision system to retrieve the 3D position and normal vectors of the planar faces of the object and then give instructions to a robotic arm for grasping. Experimental results indicate that the proposed pipeline of algorithms written in C# for plane segmentation can successfully segment the polyhedral object, and the Kinect-based robotic vision system developed in this work can achieve real-time automatic polyhedral object grasping with high precision.