Online Auto-Calibration in Man-Made Worlds

This paper proposes a method to auto-calibrate a static camera in man-made worlds. The calibration is done with vanishing points which are estimated with line segments. Instead of a single image, the method uses constantly acquired images from a camera. This online behavior of the method has two advantages: A continuous and adaptive estimation of the vanishing points over thousands of images suppresses the amount of noise in their positions. Newly detected line segments can further improve the estimated positions of vanishing points. This method is of practical importance, because simple installation and maintenance of cameras in video surveillance is a competitive advantage. Furthermore, applications like tracking can profit, because knowledge about the scene geometry can be a potential clue. The method was evaluated in a broad range of indoor and outdoor environments. The relative error between the internal camera parameter estimates and the true values was on average 5%.