Abstract.
Real-time stereo matching has many important applications in areas  such as robotic navigation and immersive teleconferencing. When processing  stereo sequences most existing real-time stereo algorithms calculate disparity  maps for different frames  independently without considering  temporal  consistency between adjacent frames. While it is known that temporal  consistency information can help to produce better results, there is no efficient  way to enforce temporal consistency in real-time applications.  In this paper the temporal correspondences between disparity maps of  adjacent frames are modeled using a new concept called disparity flow. A  disparity flow map for a given view depicts the 3D motion in the scene that is  observed from this view. An algorithm is developed to compute both disparity  maps and disparity flow maps in an integrated process. The disparity flow map  generated for the current frame is used to predict the disparity map for the next  frame and hence, the temporal consistency between the two frames is enforced.  All computations are performed in the image space of the given view, leading  to an efficient implementation. In addition, most calculations are executed on  programmable graphics hardware which further accelerates the processing  speed. The current implementation can achieve 89 million disparity estimations  per second on an ATI Radeon X800 graphic card. Experimental results on two  stereo sequences demonstrate the effectiveness of the algorithm.