3D shape recovery from images viewed by multiple video cameras

点击翻译按钮获取中文摘要

The proposed research "3D Shape Recovery from Images Viewed by Multiple Cameras" will attempt to estimate the 3 dimensional coordinates of all pixels in a 2 dimensional image with the help of additional images viewed from multiple cameras. 3D modeling of objects and scenes contained in multiple 2D views is the conceptual objective. This research goes beyond still images and works on video images viewed by a multi-camera assembly of 4 (or 5) video cameras interfaced to a personal computer. The real time multi-camera vision system continuously produces and displays a dense stereo disparity map that shows the distance to every point of the observed scene by the multiple cameras. The system is also capable of displaying a virtual reality image viewed from an arbitrary viewpoint separate from the positions of real cameras. Established mathematical theories of multiple view geometry are extensively applied to the epipolar (cross-eye stereo) configuration as well as to the parallel view axis configuration. Potential applications include (1) robotic applications, such that the multi-camera vision system continuously provides distance information to a self-navigating robot for collision avoidance, (2) video surveillance capable of detecting 3D motions to indicate whether a video object is moving closer to or away from the cameras, and (3) synthesis of 3D stereo views for projection on stereoscopic 3D LCD (Liquid Crystal Display) equipped with the parallax barrier, which will set a new standard for stereoscopic video, video games and TV.