Research On Stereo Matching Technology Based On Convolutional Neural Network

Binocular Stereo Vision is an important branch of computer vision.It is widely used in traditional fields like 3D scene reconstruction,intelligent robot navigation,object tracking and emerging fields like unmanned vehicles,virtual reality and mobile terminals.As the most important link in binocular stereoscopic vision,stereo matching has been a hot issue and focus of current research.In recent years,deep learning technology which got a great success in the image,voice,text analyses is developing rapidly.Convolution neural network(CNN)combines deep learning technology,artificial neural network and local correlation characteristic.It can effectively extract the features of image,so it got notable achievements recently especially in computer vision tasks such as image classification,object detection,object tracking and other tasks.However,how to combines CNN and stereo matching to get a fast and accurate algorithm is still an open question.In this paper,we summarize the achievement on CNN and stereo matching,based on classical CNN model,it is applied to stereo matching tasks.The main works in the paper list as follow:1)A cost computing algorithm based on CNN(CCNN)is proposed.On the basis of principle of traditional matching cost algorithm and constraint of dataset,this paper make use of CNN to compute the similarity of two image patches,avoiding select features manually.Experiments show that the proposed algorithm greatly reduces the initial matching error,and has good illumination robustness.Meanwhile,the influence of hyperparameter of CNN was studied.2)Proposed a CNN based stereo matching algorithm framework.Referring to classical four steps in stereo matching,in this paper,this paper uses CCNN as cost computing method,and cross-based cost aggregation and semiglobal matching algorithm were used in cost aggregation stage.After computing raw disparity,multi-level post-processing algorithm was executed to corrected the raw disparity map.Tests on open datasets show that the proposed algorithm framework has reached the advanced level in terms of accuracy and performance.