Research On Binocular Stereo Matching Technology For Complex Scenes

The basic principle of binocular stereo vision is to capture the same scene using two different positions or a moving camera.The depth information is obtained by calculating the disparity of the spatial point in the two images.Because binocular stereo vision has the advantages of low cost and high performance in non-contact measurement,binocular stereo vision technology research is a hot research field in computer vision,image processing,and pattern recognition.Since the 21 st century,with the rapid development of computer technology,binocular stereo vision technology has shown an increasingly important position in the fields of robotics,military,aerospace,medical information and other fields,such as traffic video surveillance,drone navigation and control.unmanned autopilot system,appearance inspection of industrial assembly lines,restoration and reconstruction of cultural relics,etc.In recent years,with the continuous deep research on binocular stereo vision technology,new technologies for computational models and optimization methods of binocular stereo matching continue to emerge.Stereo matching technology has made breakthrough progress in calculating disparity accuracy and computational efficiency.However,when the image pair contains complex scenes such as complicated edges,occlusion areas,similar areas of color backgrounds,and discontinuous texture areas,the accuracy and robustness of binocular stereo matching calculations need to be further studied.This dissertation focuses on the research of disparity calculation in the context of complex scenes using stereo matching technology.By designing a stereo matching algorithm model combining image edge and brightness information,it overcomes the inaccuracy of weight structure in the edge region of images with similar color background.In addition,the method of learning the similarity measure model by using the optimized convolutional neural network instead of the gray similarity measure method improves the accuracy and robustness of binocular stereo matching disparity calculation.The main work of this article includes the following points:1.Examine the existing typical stereo matching algorithms and summarize them.Summarize the key problems in the current stereo matching,and briefly introduce and analyze them.In particular,the stereo matching algorithm based on minimum spanning tree is deeply discussed.The existing problems of existing stereo matchingalgorithm based on minimum spanning tree are analyzed in detail.2.For the traditional minimum spanning tree stereo matching method in the image with similar color background pixels between the weight of the problem is not accurate,this paper proposes a minimum spanning tree stereo matching method that combines image edge and brightness information,first treat the matching image Bilateral filtering removes noise while preserving image edge information,then uses Canny operator to extract image edge information,and replaces the traditional color information with pixel edge color and edge information to reconstruct the MST weight function.Finally,the cost aggregation is performed using the weight function that introduces the edge and color information and the final disparity result is calculated and refined.3.For the gray similarity measurement,the problem of low initial cost precision is calculated in the weak texture region and the Census transform in the structural repeat region.This paper uses convolutional neural network to learn the similarity measure and use it to initialize the matching cost calculation to replace it,which effectively improves the performance of the algorithm.In this paper,a fast parallax estimation optimization method is proposed,which uses Akaze and Surf feature detection algorithm to perform fast disparity estimation on the left and right images to be matched,and reduces the range of disparity search,thereby improving the computational efficiency of disparity without reducing the accuracy.Aiming at the problem that the traditional aggregation strategy can easily lead to void or partial region calculation errors in the result of the disparity refinement,our paper proposes a secondary refinement method based on image segmentation,which effectively improves the accuracy of the disparity refinement,and the accuracy of the disparity calculation.4.Using the standard test image set and extended image set provided by the Middlebury and KITTI databases,the methods described in this paper were comprehensively compared and analyzed.The experimental results show that the two stereo matching methods proposed in this paper have better disparity calculation results in complex scenes,effectively improve the precision of disparity estimation,and have good robustness and wide application prospects.