3D Model-Based 6D Object Pose Tracking on RGB Images

Abstract

In this paper, we present a 3D-model based algorithm for 6D object pose estimation and tracking on segmented RGB images. The object of interest is segmented by U-Net neural network trained on a set of manually delineated images. A Particle Swarm Optimization is used to estimate the 6D object pose by projecting the 3D object model and then matching the projected image with the image acquired by the camera. The tracking of 6D object pose is formulated as a dynamic optimization problem. In order to keep necessary human intervention minimal, we use an automated turntable setup to prepare a 3D object model and to determine the ground-truth poses. We compare the experimental results obtained by our algorithm with results achieved by PWP3D algorithm.