Research On The Computing Technology Of Variational Optical Flow For Large Displacement Motion Based On Deep Matching

Optical flow reflects the relationship between the spatiotemporal changes in the brightness of pixels in image sequence and the motion and structure of objects in the image,It not only contains the motion parameters of the observed object,but also carries affluent 3D structure information of motion scenes.Therefore,the study of optical flow calculation technology for image sequences is an important task in the realm of machine vision.The objective of researching optical flow estimation is to obtain the accurate motion fields from the image sequences,which is one of the most important foundation of advanced machine vision tasks.Since the 21 st century,with the drastic improvement of computing power of computer,image sequence optical flow calculation theory and related technologies have been widely used in robots navigation,human pose estimation,Film post-processing technology and other areas of daily life.With the rapid development of optical flow estimation technique,the accuracy and robustness of optical flow calculation technique for plain scenes such as small displacement has been vastly improved.Nevertheless,when the large displacements,complex texture structures or multiple moving objects and other difficult scenes are included by image sequence,There is still immense room for progress of the accuracy and reliability of existing optical flow calculation methods.Our paper focuses on the variational optical flow estimation method of image sequences in large displacement scenes,and aims to solve the problems of accuracy and motion blurring in the variational optical flow estimation of large displacement and complex scenes.The main work and contribution of our paper are concluded as follows:1.firstly,the research background and significance of optical flow computing technology are summarized,then the classification and existing problems of optical flow computing technology are discussed in detail.Finally,the two key issues of optical flow estimation are analyzed,in addition,the chapters and sections arrangement of this paper are summarized briefly.2.introduce our image matching calculation model based on Deep matching and Grid Motion Statistics Strategy.First construct a deep matching model for initial matching;then use the grid approximation method to divide the two consecutive frames into n×n non-overlapping image grids,and assign the matched pixels to each image grid according to the pixel coordinates;Finally,according to the motion smoothing assumption,the mismatched pixels are eliminated according to the confidence evaluation function,thereby improving the matching accuracy of the algorithm.3.In order to address the issue of accuracy and robustness of optical flow calculation under difficult motion modes such as large displacement motion and non-rigid motion,our large displacement optical flow calculation technique based on non-rigid robust matching model was proposed.First,a large displacement motion pattern of an image sequence is obtained from the aforementioned robust sparse motion field and used as a positioning clue;then a motion edge detection model based on a random forest decision tree is used to perform motion segmentation;The dense interpolation model combines the robust sparse motion field and motion segmentation information for dense interpolation.Finally,the dense motion field is brought into the variational energy function and iteratively refined to obtain the final optical flow estimation result.4.The test images provided by the Middlebury,MPI-Sintel,and KITTI databases are used to comprehensively compare and analyze the optical flow calculation methods described in this paper and the representative optical flow calculation methods such as Classic + NL,Deep Flow,Epic Flow,and Flow-Net S.The experimental results show that the proposed method achieves high optical flow estimation accuracy and robustness,especially for large motion,non-rigid motion and motion occlusion with difficult motion optical flow estimation.5.Section V summarized the work of this paper,then analyzed the limitation of our method and discussed the potential solutions in the future.