Research On Stereo Matching Algorithms

Binocular vision is a well-known technique to obtain depth information from digital images. One of its key problems is to find corresponding points in two or more images of the same scene, referred to as stereo matching. Nowadays stereo matching has recently experienced significant progress as many new algorithms are presented.In the stereo matching field, numberless algorithms are presented every year. Although these algorithms present complicated features, we can easily make a comparative study of various algorithms using the classify methods and the Middlebury testing benchmark presented by Daniel Scharstein and Richard Szeliski. Through their research, stereo matching algorithms can usually be classified into two groups:local and global methods. Global methods regard the stereo matching problem as an energy minimization problem and many of them achieve excellent results. However, global methods are always hard to implement and computationally expensive. As a result, area-based local methods still play an important part in real-time applications as they are often simple and fast.In this paper, we focus on the local matching methods, which yield a dense disparity map by matching small image patches as a whole, relying on the assumption that nearby points usually have similar displacements. Generally, these methods could be parted into four steps:matching cost computation, cost aggregation, disparity computation and disparity refinement. Most researchers choose the WTA (Winner-Takes-All) method in the part of disparity computation and improve the algorithms in the remaining three parts. After doing the initial matching which usually contains the first three parts, some authors lastly introduced some methods to refine the disparity.In particular, we mainly focus on the improvement of cost aggregation methods. We present a new window-based method for stereo matching in this paper. Our algorithm follows the simple process. Firstly, we improve the traditional window-based matching by proposing a new aggressive method using data dividing means. Then, we present a method to improve the disparity. In particular we also introduce the median filter to reduce noise with disparity image. Our method has a much simpler structure. Actually the early local matching algorithms usually have much simper structure and we can easily accomplish fast implementation. Actually, the question that has been on our minds during recent years is how to achieve accuracy without increasing complexity. In order to get a better comparative study, we also give a capsule description of another published algorithm in which we focus on the invalid regions which usually happen at depth discontinuities and weakly-texture regions.Through lots of experiments about the new approach on the standard images, it can be concluded that our method proposed in this paper is effective. Another virtue resides in its clarity and brevity.