Research On Binocular Stereo Vision Based On Convolutional Neural Network

Binocular stereo vision is an important branch in the field of machine vision. It simulates the visual system of the human being to perception objective world, which is widely applied into many fields, such as reverse engineering, test and measurement, cultural industries, public safety, visual Navigation, map generation, aviation surver and so on. Therefore, it has a certain theoretical value and very important practical significance to study on some issues of stereo matching and 3D reconstructions from binocular stereo vision. Focusing on stereo matching and 3D reconstructions, the author had done amount of researches ong these aspect. The followings are the main researching content in the paper.(1)This paper presents a method for extracting depth infromation from a rectified iamge pair. Our approach focuses on the first stage of many stereo algorithm:the matching cost computation. We approach the problem by learning a similarity measure on small image patches using a convolutional neural network. Training is carried out in a supervised manner by constructing a binary classificaiton data set with examples of similar and dissimilar pairs of patches, which is downlowned from the middlebury site. We evaluate our method on the Middlebury stereo data sets and show it outperforms other approaches, such as SAD、CT algorithm and so on.(2) The raw outputs of the convolutional neural network are not enough to produce accurate disparity maps, with errors particulary apparent in low-texture regions and occluded areas. This paper presents a series of post-processing steps follow:cross-based cost arregation, semiglobal matching, a left-right consistency check, a guided weighted median filter,a median filter and subpixel enchancement. Experiments show that these method greatly improves the accuracy of the disparity map.(3)In the paper, we use octree data structure to storing 3D point cloud data and projection greedy triangulation algorithm to 3d reconstruction. The reconstruction results demonstrate the effectiveness of this method.