Dual neural networks for kinect camera calibration

Abstract

Kinect camera, also called as RGB-D sensors, has been widely used in robot vision control system. However, building a complex Kinect camera distortion model is still a challenging issue. In this paper, a Dual Neural Networks for Kinect camera calibration method, according to the characteristics of Kinect camera and the imaging law of point moving along the different axis in the camera coordinate, is proposed. Firstly, IR image, instead of the depth image widely used in the existing calibration method of the Kinect camera, is used to reduce the effect of noise. Then, problem of the camera calibration is considered as nonlinear mapping from 2D image coordinates to the 3D world coordinates and two different neural networks is designed respectively to realize the nonlinear mapping. Experimental results demonstrate that the proposed calibration method could provide a more reliable and accurate reconstruction results compared with popular joint calibration methods.