This paper introduces a new technique to automatically calibrate the extrinsic parameters of a network of cameras without using dedicated calibration objects or markers. Instead, the motion of persons walking naturally through a scene is used. Simple foreground and motion features are extracted from the individual image sequences. A Hough transform is applied in a specially defined parameter space to estimate the relative geometry between camera pairs without solving the problem of finding correct feature correspondences between views. All possible feature correspondences are examined, and a modified gradient descent algorithm is used to find the set of optimum calibration parameters within the Hough space. After calibrating each camera pair the resulting camera network is built up using a global error minimization technique. The approach is tested on several indoor scenarios and shows a high degree of robustness, especially when multiple persons enter the scene, making it difficult to resolve feature correspondences. The correctness and precision of individual camera calibrations and of the resulting camera network are thoroughly evaluated, showing that triangulation errors as low as 5cm can be reached using very little observation data.