3D Vision for Indoor Robotic Manipulation using Kinect V2 Sensor with Point Cloud Library (PCL

Abstract

In order to perform any robotic task object recognition plays a very vital role. In this research we propose a real time recognition system that is able to recognize single object using the latest Kinect v2 depth sensor. This sensor captures RGB images with depth information for each pixel and is also low cost. The recognition system consists of two domains i.e. training and testing. For training, a point cloud including the single object is retrieved, processed and its global feature named as Clustered Viewpoint Feature Histogram (CVFH) using Point Cloud Library (PCL) is calculated and for all the required objects a database is made. For testing a point cloud from Kinect V2 is obtained and after preprocessing (CVFH) is calculated for each cluster obtained and using K-Nearest Neighbor (KNN) method nearest feature match is searched from database. If a match is found, then object is considered as recognized. The system was tested successfully. CVFH has performed well than VFH in the case of partially occluded objects and noisy environments and since it considers geometry of object so its computational cost is also low as compared to other local features so it is quite suitable for the situation where computation time is main concern.