Abstract
This paper reports a system to track a moving object and recover its 3D shape in real time. When a camera follows the object by fixing a point at the image center, the obtained images are just equivalent to those obtained when the object rotates around the point. Then, we can recover its 3D shape from the optical flows of the object points other than the fixed point. Based on this principle, we constructed a real time system by using a pan-tilt camera and a parallel processing unit. 256 PE's in the parallel processing unit overcome the computation costs of the optical flows by effectively rearranging the image data within them, and achieve the total computation from image input to 3D reconstruction within video-frame interval (33ms).
