Continuous Localization and Mapping of a Pan Tilt Zoom Camera for Wide Area Tracking

Pan-tilt-zoom (PTZ) cameras are powerful to support object identification and recognition in far-field scenes. Real-time detection and tracking of targets with these cameras is nevertheless complicated by the fact that the geometrical relationship between the camera view and the 3D observed scene is time-varying and, over long periods of operation, real world scenes exhibit changes either due to objects or context variations. In this paper, we present a solution that provides continuous estimation of the camera pose and mapping from scene landmarks, which is robust to rapid camera motion, changes of the environment due to illumination or objects, and scales beyond thousands of landmarks. The estimation of camera pose permits to compute the relationship between the positions of the targets in the 3D world plane and their corresponding positions in the 2D image, and to perform real-time tracking of multiple targets in world coordinates. Since camera motion is compensated, an absolute reference coordinate system can be used and targets can be tracked in the world plane with high and stable degree of accuracy also at large distances and any zooming level. We provide experimental evidence that the solution proposed is capable of supporting effective real-time tracking of multiple targets with PTZ cameras.