Stereo Matching Algorithm

Stereo vision is an important researching field in the computer vision. Stereo vision has been widely used in autonomous robot, human computer interaction, surveillance, intelligent control, 3D measurement, robot navigation,unmanned aerial vehicle and 3D stereo film etc. Stereo matching is an important and key researching field in the stereo vision, the result of the stereo matching has more influence on the result of the 3 dimension reconstruction. The thesis is mainly focused on the stereo matching algorithm.In real scene, there are many surfaces and planes are curved or slanted.It will lead to many errors to the scene which has many slanted planes and curved surfaces. In this thesis, we discard the hypothesis and allow there are curved surfaces and slanted planes. We optimize the disparity between segments and in segment and use the contextured geometric information of the segment as the constraint to get the optimal disparity plane.One of the many challenges in stereo matching is the low textured regions. In this thesis, we discuss the low textured regions in scenes and present the concepts on the beliable disparity and unbeliable disparity. We use the probability distribution of the beliable disparity to inferent the unbeliable disparity. Experiments give the support to the algorithm.In this thesis, we present a new stereo matching algorithm based on the adaptive supported region. We use the probability distribution information of the intensity and select the supported pixel based on the probatility distribution information of the disparity. An adaptive size and shape supported region is formed by all the supported pixels. We use the constraint of all the supported pixels to cost aggregation. At the same time, we give an adaptive support weight to the supported pixels. Experiments prove the effectiveness of the algorithm.Convational stereo matching algorithms use the gray images and do not take use of the color information. Actually, the color information is very useful to reduce the ambiguity of matching. We fully use the color image and different color channel information to establish the reasonable support weight model between the pixels which can robust to the environmental condition. Experiments prove the effectiveness of the algorithm.