Research On Segmentation-Based Stereo Matching Algorithms

As an important branch of computer vision, stereo vision is responsible for recovering the three-dimensional structure of natural scenarios from two-dimensional images acquired by the camera. It can be used to a wide range of applications with high commercial value, such as 3DTV, multi-view/free-view video, driver assistance systems, military, virtual reality, etc. The key problem of stereo vision is to obtain the depth information of the objects in the scene, which is mainly accomplished by the stereo matching algorithms. Stereo matching technology has been studied for quite many years, but still is the research focus in the fields of pattern recognition and artificial intelligence. In this paper, two improved stereo matching algorithms are proposed based on learning principles of stereo matching and summarizing the existing stereo matching algorithms. In general, the main ideas are included as follows:Firstly, a local stereo matching algorithm based on adaptive weight and image segmentation is proposed. Support windows are used as matching primitives for most local matching algorithms currently. According to the importance of contributions, different weights are assigned to pixels in the support window. Greater contributions correspond to higher weights, and less contributions with lower weights. A new weighting function is presented with integration of the adaptive support windows and the information of image segmentation, which can pick up the "high-quality" pixels with similar disparities among the support window. The outlier points are detected in the initial disparity map, and then the disparities in these areas are rectified by several different strategies. The specific procedures of the matching algorithm are listed, and some experiments are taken at last.Secondly, an improved global stereo matching algorithm is proposed. Global energy function typically contains two items: data item and smooth item. The data item represents the matching cost between a pixel in one image and the corresponding pixel in another image under the given disparity. In order to meet the real similarity degrees of pixels better, the distinguishable property should be enhanced, which is compatible with the main idea of the 3rd Chapter. The piecewise smooth prior of the disparity map is merged into the smooth item. The disparity of the same object surface is generally changed smoothly, while the disparity discontinuous boundaries are mostly distributed on the edge of the object. This is consistent with the segmentation hypothesis, which states that the pixels of the same segment belong to the same object, and edges of the object line with the boundaries of the segment. Thus, the disparity discontinuity within a segment should be punished with a greater value. A new energy function is constructed, using the matching cost proposed in the 3rd Chapter and the image segmentation information, and the methods of multi-label to binary-label conversion and directed graph construction are listed. Finally, outliers in the initial disparity map are detected and improved.