Research On Binocular Stereo Vision Based On Deep Learning

In recent years,Binocular stereo vision is a hot topic in the field of computer vision.The binocular camera is just like the human visual system to touch the world。Stereo matching has been widely used in 3D reconstruction,public security,auto pilot and So on.Deep learning has shown its ability on computer vision research,such as object recognition,detection,segmentation.This paper mainly studies how to get the disparity from binocular pictures using Convolution Neural Network for dense prediction.Firstly Introduce the stereo matching basic concepts,the development of the stereo match algorithm,the traditional method and the deep learning method.The main contents of this paper are as follows:(1)This paper focuses on the construction of cost volume in binocular stereo matching problem,proposed a new method to construct the cost volume which has the ability to merge the left and right view features,It can save the computation and to learn the high level feature for the matching cost to help the feature learning for the matching cost.(2)We use the 3d convolution layer and 3d deconvolution layer based on the Resnet block to learn the context feature from the cost volume.It can save the computation and to learn the high level feature for the matching cost to overcome the weak texture problem.(3)We use the 1×1×1 convolution method to calculate disparity from the cost volume.Making the disparity network can be trained end to end.It is more convenient to optimize the model.(4)We use the softmax function change the cost volume to probability model to improve the performance from the 1×1×1 convolution method.We use the expect of the cost volume as the final predict,with the ground truth to supervised learning make the disparity network can be trained end to end.Finally we use the transfer learning for model parameter initialize and we ranked 10 th on the KITTI Stereo 2015 leaderboard.