Automatic Calibration of Stationary Surveillance Cameras in the Wild

We present a fully automatic camera calibration algorithm for monocular stationary surveillance cameras. We exploit only information from pedestrians tracks and generate a full camera calibration matrix based on vanishing-point geometry. This paper presents the first combination of several existing components of calibration systems from literature. The algorithm introduces novel pre- and post-processing stages that improve estimation of the horizon line and the vertical vanishing point. The scale factor is determined using an average body height, enabling extraction of metric information without manual measurement in the scene. Instead of evaluating performance on a limited number of camera configurations (video seq.) as in literature, we have performed extensive simulations of the calibration algorithm for a large range of camera configurations. Simulations reveal that metric information can be extracted with an average error of 1.95 % and the derived focal length is more accurate than the reported systems in literature. Calibration experiments with real-world surveillance datasets in which no restrictions are made on pedestrian movement and position, show that the performance is comparable (max. error 3.7 %) to the simulations, thereby confirming feasibility of the system.