Study On Stereo Matching Of Binocular Stereo Vision

Stereo matching is one of the most active research areas in computer vision.Stereo matching as the most important and most challenging step in stereo vision is always restricting the development of stereo vision.Simultaneously,stereo matching has more and more value of application in robotic vision,three dimensional perception,unmanned,3D reconstruction fields and so on.How to obtain the high accuracy disparity map by stereo matching technology is the key of binocular stereo vision.In order to get a high quality matching disparity map,all kinds of new stereo matching algorithms are emerging.The existing dense binocular stereo matching algorithms can be divided into two categories:Local Stereo Matching Algorithm and global stereo matching algorithm.Global based stereo matching algorithm can obtain good result of stereo matching with the help of many kinds of constraint conditions and global optimization strategy,but the disadvantage of this kind of algorithm is the high complexity and large amount of computation.Compared with global based algorithm,local based stereo matching algorithm is much more difficult to get a high quality disparity map,but it is an integral method in stereo matching field because of its low complexity and small computation.In this paper,we propose a stereo matching algorithm based on the combination of deep feature stereo cost function and the matching cost filtering by guided filtering.Traditional matching cost calculation use conventional image features such as color,intensity texture and gradient.In this paper we introduce the convolutional neural network and construct convolutional neural network to get deep feature image of the original image.Then we construct matching cost function by truncated pixel color,gradient and deep feature.To the cost aggregation we filter the matching cost images by guided filter.After finishing the computation of matching cost and cost aggregation we select the optimal disparity value by the WTA for each pixel.Then we perform the the occlusion detection by left-right consistency check and filter the disparity map by bilateral filter to obtain the final disparity map.In this paper,we perform the experiment with industrial robot and four stereo image pairs of Middlebury data set.Then we make comparison with other five local based stereo matching algorithms.By the objective quantitative evaluation of the rate of error matching and the subjective qualitative evaluation of the comparison of disparity map with the ground truth we find the local based stereo matching algorithm we proposed has higher matching accuracy than other six stereo matching algorithm.