3D Object Pose Estimation Using Local Features Based for Industrial Appliance

Object localization in the industrial field plays an important role to obtain the pose information of an object. For supporting the localization works, 3D point cloud data is used as its primacy containing the shape of an object. Despite its rapid development, the problems faced by each object used are heterogeneous. In this research, Time-of-Flight (ToF) camera is used to obtain the data considering its capability to provide a small form factor and sufficient point cloud data. As the one of most popular objects in the industrial field, fitting pipe is used in whole systems. Start from obtaining the data using a ToF camera, then followed by preprocessing which eliminate Non-measured data and point outside of the threshold range. Segmentation by separating the object with its planar, and separating each other into individual objects done by RAndom SAmple Consensus (RANSAC) and Density-based Spatial Clustering Application with Noise (DBSCAN). Before the pose estimation is applied, CNN is used to recognize the object. By using Fast Point Feature Histogram (FPFH) with SAmple Consensus Initial-Alignment (SAC-IA) as the coarse pose estimation and Iterative Closest Point (ICP) as the refinement pose estimation, average rotation error in Euclidean space (roll, pitch, yaw) as 0.47 degree, in translation is 4.17 mm. Also, we estimate the size of the pipe which has an average error of 5.30 mm.