Deep Learning with RGB-D Imaging: Object Classification and Pose Estimation

Abstract

As part of the doctoral thesis, the objective is to develop a Human-Machine Interface to control a assistive robotic arm with more than 6 degrees of freedom. Deep learning techniques for object recognition and pose estimation in order to be able to interact with them is presented. Three multimodal convolutional neural network models were implemented using RGB-D images from the BigBIRD database. Each model have three classification outputs:  22 Objects - 5 Cameras - 8 Rotation. The best of model achieved 96% accuracy for objects, 98% for camera and 56% for rotation.