Research On Binocular Stereo Matching Algorithm Based On SoC

Reconstructing scenes by using visual information is one of the most important research topics in computer vision,and has also been applied in the fields of automatic driving,intelligent robots and so on.Using of binocular stereo vision technology to obtain depth information is an important means of environment perception for automatic driving and intelligent robots.The accuracy and reliability of depth information acquisition depends on the efficiency of stereo matching algorithm.Therefore,the research in this paper focuses on improving the accuracy of stereo matching algorithm and build a binocular vision system based on SoC.Firstly,the model of binocular vision camera is studied,and the camera is calibrated offline to obtain the internal and external parameters.Based on the calibration results,the camera model is transformed into a linear model,and the corresponding points of the image are of the same height,which is conducive to improving the accuracy and efficiency of stereo matching.For binocular vision,stereo matching is one of the key points,which is usually divided into four steps: matching cost computation,cost aggregation,disparity computation,and disparity refinement.Census transform is used to extract image features,and Hamming distance is used to represent the similarity of pixels between left and right views,that is,the cost value.It's difficult to match well for low texture region because of its unclear features.In this paper,the multi-scale cost aggregation method is deeply studied,and it is combined with the minimum spanning tree as the final cost aggregation algorithm.Cross-scale cost aggregation takes into account inter-scale information,which can effectively reduce the mismatches of low-texture regions.Considering that there are still many mismatches in the initial disparity,especially in the disparity discontinuity,this paper a disparity refinement algorithm is proposed based on superpixel segmentation.Pixels in the same super-pixel always have similar characteristics,and the superpixel can closely adhere to the image edge.By combining super-pixel weights with weighted median filtering,the weight of similar pixels can be increased,and the matching accuracy can be improved.The experimental results show that the proposed method is more accurate than the bilateral weighted median filter,especially in the disparity discontinuous region.The complete binocular vision system includes image acquisition,camera calibration,stereo matching,three-dimensional reconstruction and so on.This paper builds a binocular vision system based on Zynq SoC,realizing image acquisition and Census feature extraction,and calculates matching cost,completes cost aggregation and disparity calculation,and finally obtains disparity map.The experimental results based on data sets and real scene images show that the binocular vision system in this paper has high accuracy in standard data sets and get a good effect in real scene.