Research On Cost Function Algorithm Strategy Of 3D Reconstruction Based On Stereo Matching

At present,stereo matching algorithm plays an important role in completing scene reconstruction.Stereo matching algorithm can obtain the depth information of the scene through camera shooting,and restore the three-dimensional structure of the scene according to the depth information.Three dimensional reconstruction technology based on stereo matching is widely used in military,aviation,agriculture,industry and other fields.Therefore,getting high-precision and real-time stereo matching method has always been the research goal of researchers.Researchers generally believe that it is difficult to obtain accurate and fast stereo matching algorithm in stereo vision.The most critical step in stereo matching algorithm is cost matching,which is the process of finding corresponding points in two constrained images.However,there are many problems and challenges in the process of finding corresponding points,mainly due to many uncontrollable factors in the reconstructed scene,such as optical distortion,noise,projection reduction,perspective distortion,low texture,repeated texture and so on,which makes stereo matching difficult to achieve the expected accuracy.In order to achieve highprecision stereo matching under limited realistic conditions,this paper mainly studies the selection strategy of 3D reconstruction cost function based on stereo matching.Census cost function based on non-parametric transformation has been widely used in recent years,but there are some deficiencies in the process of obtaining parallax map.In view of the deficiency of census cost function,there are many improvements of census cost function.The improved census cost function is summarized in the form of summary.Firstly,according to the basic principle of calculating parallax map with census cost function,the limitations of census cost function in calculating parallax map are analyzed;Then,according to the limitations of census cost function,this paper summarizes the calculation methods of improved census cost function in the form of classification,and divides the improved census cost function in recent years into six categories;Finally,the advantages,disadvantages and core ideas of each kind of improved census cost function are summarized,and the performance and effect pictures of six kinds of improved census cost functions are compared and analyzed,so as to find the problems and solutions in the application of census cost function.Although the census cost function algorithm based on non-parametric transformation has been widely used in recent years,the census cost function method relies too much on the gray value of the central pixel in the process of obtaining the parallax map.Aiming at the deficiency that the census cost function of non-parametric transformation depends on the central pixel,an improved census cost function based on cross domain is proposed.The algorithm finds the pixels with similar color around the reference pixel through the cross domain,and optimizes the generation value of the reference pixel by using the pixels in the cross domain,so as to reduce the matching error caused by the excessive dependence of census cost function on the central pixel and further improve the accuracy of disparity map.The experimental results show that the parallax accuracy of the scene obtained by the improved census cost function based on the cross domain is improved,and the accuracy of weak texture region and repeated texture region in the scene is also improved.In addition,under the condition of ensuring the matching accuracy,the cost optimization time of the algorithm is greatly reduced compared with the census algorithm of four path cost aggregation.