Error analysis and experiments of 3D reconstruction using a RGB-D sensor

摘要

In this paper, we investigate the performance of KinectFusion algorithm for 3D reconstruction using a Kinect sensor. A sensor model is applied to generate depth image to evaluate accuracy of the algorithm. To obtain ground truth of depth image as well as camera pose, we generate depth data based on a CAD model in a simulation program. In the error analysis, deferent types of noise, including depth image noise and pose-predict noise are added to examine the errors of 3D reconstruction and camera localization. It is found that the KinectFusion algorithm is more robust to depth image noise, but lack robustness against pose-predict noise. Experiments of 3D reconstruction have been carried out on a Kuka 6-DOF robot manipulator with an eye-in-hand configuration. Experimental results validate the simulation results of error-analysis.