Disparity Estimation Algorithm Based On Binocular Stereo Vision

Binocular stereo vision simulates the human visual system to deal with real scene.The disparity information is obtained by the disparity estimation of left and right imagepairs(i.e. stereo matching) and used to get the depth information of 3d scene, so thedisparity estimation about binocular stereo vision is always the focus of research in thecomputer vision. Based on binocular stereo vision model, the paper improved threedifferent algorithms of the disparity estimation.Firstly, belief propagation as a kind of algorithm based on Markov Random Fieldhas been used to solve the early vision problems (such as: stereo matching, etc). In thispaper, the message delivery channel of belief propagation is changed to inter-blockmessage delivery channel in a dual network. The initial disparity of images is calculatedby local relevant method and the reliable areas and unreliable areas are identifiedaccording to the initial results. Then BP algorithm is implemented in the created Markovdual network. The result of experiments shows that our algorithm can get better disparitymap. According to the test on the Middlebury website, the effectiveness of the proposedalgorithm is proved.Secondly, search range of the disparity is in the neighbor field of center pixels basedon the assumption of polar-line constraint and smooth constraint, so the search range canbe limited by the line segment which is based on the color and distance constraint. Thedisparity is propagated between the seed-pixels which get from the processed initialresult, and some optimization strategy is used to get the final disparity map. By thetesting results of Middlebury website, we can know that the error rate has reduced 3.59%compare with the original algorithm, and the real disparity map can be obtained.Finally, we provided a kind of effective algorithm which is used to solve theproblem about disparity continuity for the curved surface of objects in real world. Eachpixel is assigned a random plane, then the parameter of the assigned plane is propagatedand refined in support-windows. The optimization is iterated several times to eliminate the mismatch what the randomly assigned bring. Certain follow-up processingtechnology further improve the precision of disparity. By the experimental results, it canbe known that the disparity of curved surface improved. And when we lower the errorthreshold, the algorithm has better performance.