Deep Learning Based 3D Point Clouds Recognition for Robotic Manufacturing

Abstract

The present work aims to develop a light-weight deep learning algorithm for 3D vision based on point cloud perception. We first analyzed the current 3D vision models for object classification and pose estimation purposes, and proposed that the global features extracted by PointNet was able to facilitate performing both classification and pose estimation tasks by visualizing the activations of a trained PointNet model. Then, the customized point cloud datasets were produced for both training and testing through Blensor. With our datasets, the performance of our proposed algorithm was evaluated. The predicted results of our model were also compared with those predicted by the framework of [4], which generally showed similar performance but with slightly lower accuracy. Nevertheless, our framework is considered to have a much better efficiency than that of [5], due to simpler structure of our framework that avoid repetitive work of feature extraction that experienced by using the framework of [5].