Bayesian autocalibration for surveillance

In the context of visual surveillance of human activity knowledge about a camera's internal and external parameters is useful, as it allows for the establishment of a connection between image and world measurements. Unfortunately, calibration information is rarely available and difficult to obtain after a surveillance system has been installed. In this paper, a method for camera autocalibration based on information gathered by tracking people is developed. It brings two main contributions: first, we show how a foot-to-head plane homology can be used to obtain the calibration parameters and then we show an approach how to efficiently estimate initial parameter estimates from measurements; second, we present a Bayesian solution to the calibration problem that can elegantly handle measurement uncertainties, outliers, as well as prior information. It is shown how the full posterior distribution of calibration parameters given the measurements can be estimated, which allows making statements about the accuracy of both the calibration parameters and the measurements involving them