By studying geometric properties of arbitrarily placed cameras and their spatio-temporal relation to moving people in a surveillance system, we have elaborated several possibilities of self-calibration. Using purely visual information we can fully calibrate a network of non- or slightly-overlapping cameras in an automated manner. Another patent-protected solution of great practical interest uses Structure from Motion (SfM) along with the support of an external image sequence acquired by a moving camera to infer the 3D poses of stationary surveillance cameras within a large-scale network. These auto-calibration and localization capabilities are especially relevant in large camera networks (airports, shopping malls, etc.).
Application Example: Vision-Based Localization
Indoor localization is in general an unsolved problem, where satellite and radio based solutions fail or are too inaccurate. We have elaborated on an alternative passive solution which is based on the fact that environments exhibit unique visual appearance. This allows indoor localization to be cast as a purely visual problem. We have demonstrated that a surveillance camera or a mobile phone can be localized in real-time just by analyzing the content of its image and by matching it to a previously constructed visual map. The accuracy of the solution is far better than existing solutions.