Automated registration of surveillance data for multi-camera fusion

The fusion of tracking and classification information in multi-camera surveillance environments will result in greater robustness, accuracy and temporal extent of interpretation of activity within the monitored scene. Crucial to such fusion is the recovery of the camera calibration which allows such information to be expressed in a common coordinate system. Rather than relying on the traditional time-consuming, labour-intensive and expert-dependent calibration procedures to recover the camera calibration, extensible plug-and-play surveillance components should employ simple learning calibration procedures by merely watching objects entering, passing through and leaving the monitored scene. In this work we present such a two stage calibration procedure. In the first stage, a linear model of the projected height of objects in the scene is used in conjunction with world knowledge about the average person height to recover the image-plane to local-ground-plane transformation of each camera. In the second stage, a Hough transform technique is used to recover the transformations between these local ground planes.