Uncalibrated 3D Human Tracking with A PTZ-Camera Viewing A Plane

In this paper we propose a novel method for tracking a human target based on uncalibrated video analysis techniques. The main focus here is on the observation model of sequences acquired by rotating and zooming camera sensors. In these sequences targets may appear and disappear from the field of view (FOV), their size may change very much in a few frames as a result of the camera zoom or target motion and the scene background change from frame to frame. According to this, measurements become intermittent such that a significant time lag may exist between consecutive observations of a target. Moreover targets generally change their size when they exit and then reenter the FOV of the camera. We argued that good performance in prediction (which is strongly dependent on the observation model) can improve the robustness of target tracking especially during disappearance and reappearance of target objects. The geometry of the rotating and zooming camera viewing a scene plane together with the projective geometry of imaged play-field geometric patterns (circle and lines) are exploited for computing measurement equation in the extended Kalman filter.