A flexible method for time-of-flight camera calibration using random forest

Abstract

A learning based approach is proposed to calibrate the geometric distortion of a Time-of-Flight (ToF) camera. Our method is flexible as it requires only a ToF camera and a standard camera calibration chessboard. We treat the noise model of a ToF camera as a black box, and employ random forest to automatically learn the underlying unique noise model. The geometric property of the point-cloud can be effectively restored by the learned distortion model. The method can be used in a range of computer vision applications including e.g. hand pose estimation.