Pose Estimation with Mismatching Region Detection in Robot Bin Picking

Abstract

3D object detection and pose estimation based on 3D sensor have been widely studied for its applications in robotics. In this paper, we propose a new clustering strategy in Point Pair Feature (PPF) based 3D object detection and pose estimation framework to further improve the pose hypothesis result. Our main contribution is using Density Based Spatial Clustering of Applications with Noise (DBSCAN) and Principle Component Analysis (PCA) in PPF method. It was recently shown that point pair feature combined with a voting framework was able to obtain a fast and robust pose estimation result in heavily cluttered scenes with occlusions. However, this method may fail in the mismatching region caused by false features or features with insufficient information. Our experimental results show that the proposed method can detect mismatching region and false pose hypotheses in PPF method, which improves the performance in robot bin picking application.