Automatic Stereo Camera Calibration in Real-World Environments Without Defined Calibration Objects

Abstract

Using the images of a stereo camera, a 3D reconstruction of the real environment can be calculated, enabling a machine to see three-dimensionally. To calculate the 3D reconstruction, precise geometric dimensions of the stereo camera, called stereo camera parameters, are required. In the course of this work, three algorithms were realised that can calculate the stereo camera parameters from the images and motion of a stereo camera. None of the three is dependent on defined calibration objects. The algorithms are the calibration using epipolar geometry, relative world point error, and absolute world point error. These allow the calibration of a stereo camera with only the help of a measuring tape and the data sheet of the cameras. Continuous recalibration is also possible, where changes in the physical structure of the stereo camera are compensated during operation. The accuracy of the algorithms was determined by measuring the deviation of the 3D reconstruction from the real scene. The algorithms have been designed for use on a mobile robot to autonomously calibrate and recalibrate its stereo camera, constituting its visual sense.