Real-Time View Transition between Cameras with Different Perspectives for Visual Monitoring

This paper introduces a novel method for real-time view transition between two cameras. One important feature of our method is correspondence free for the single-target case. When there are multiple targets, our approach also works with a reliable multi-camera tracker. With camera calibration data and pre-constructed background 3D models, we synthesize virtual views by first estimating the 3D position of the foreground object and then using a billboard to model the foreground. With our approach, even when the viewpoints of two cameras are quite different, it still works well as long as the monitoring areas of two cameras have intersections in the real world. The experimental results have shown natural and smooth visual effects of synthesized images, and achieved 22 frames per second on a PC with 3.4GHz Intel Pentium IV CPU and 2G RAM.