Research On Binocular Stereo Vision Based On Multi-scale And Convolutional Neural Network

Binocular stereo vision is an important method of simulating human visual perception,which has been widely used in visual navigation,automatic drive,intelligent manufacturing,virtual reality and some other fields.It calculates the binocular disparity mainly by matching the left and right image characters and uses disparity to realize the distance perception.Since the great number of complicated and indescribable characters in images causes the difficulty of precise matching,and the reconstruction quality can be easily affected by environemnt and modeling method,it is,therefore,necessary to develop relevant technology to solve thses problems.This paper focuses on research about image feature matching method and reconstruction.The main works are as follows:First,a Convolutional Neural Network(CNN)integrated with multi-scale information is proposed to calculate the similarity of image blocks.Since the image feature extracted with traditional method can easily result in great deviation in the similarity of calculation,and multi-scale image can better describe and keep featured information,an improved method is proposed in this paper.The image multi-scale information is integrated by using CNN,and the similarity between left and right image block is calculated with network model.The experimental results suggest that,due to full use of more featured information in image,the obtained matching cost is more reliable.Second,in the aggregation stage of matching cost,the image intensity and the distance between pixels are used to formulate adaptive window,so as to solve the problem of being difficult to choose the window size during the aggregation stage.In order to reduce the matching ambiguity,the smoothing constraint is added,and disparity is obtained by formulating energy function.Besides,given that it is difficult to directly solve the disparity with energy function,the problem is simplified into several combinations of one-dimensional optimization problem.At last,the disparity image is optimized through interpolation,subpixel enhancement and guided weighted-median filter.The comparison between several stereo matching algorithms verifies the effectiveness of method in this paper.Third,the local grid modeling method is proposed based on deep information segmentation.Since the scene depth reconstructed by binocular disparity has been applied in several fields,this paper researches the method of restoring depth with binocular disparity,and then conducts a scene reconstruction experiment by using this depth information.This paper analyzes and concludes the factors of influencing reconstruction in the experiments,and improves the problem of distortion caused by discontinuity of scene depth.According to the different depth of scene objects,the image is segmented into several non-overlapping regions,and then the dense feature points in these regions are selected to carry out gridding reconstruction.The experimental results show that the method mentioned in this paper can better demonstrate the reconstruction effects.