A Study On Anti-Occlusion Optical Flow Estimation Algorithm

Optical flow estimation is the basic research content of computer vision.It has a large number of applications in video segmentation,object tracking and three-dimensional reconstruction.After more than forty years of development,the optical flow estimation research has made great progress.With the continuous expansion of applications and the increasing demand for high-precision optical flow,optical flow estimation has become a research hotspot in the field in recent years.The optical flow records the movement of the object onto the image plane.According to the number of image frames used,the optical flow estimation can be divided into two categories based on two frames and based on multi-frame sequence images.In our dissertation,the former is mainly studied.For the scene with occlusion,a robust optical flow estimation algorithm based on adjacent frames is proposed.Firstly,this dissertation studies the generation problem of the warped image introduced by a class of optical flow estimation algorithm in its algorithm step,and proposes a warped image synthesis and correction algorithm.We know that the differential method is the most basic operation in the optical flow estimation algorithm.Although the warped image is only a phased result in the algorithm,it plays a key role in the algorithm.It is found that due to the problem of the algorithm itself,some parts of the object or object elsewhere in the scene will appear in the warped image.The existence of such ghosting artifacts will mislead the subsequent optical flow estimation and must be correct.To this end,we first propose a ghost-removal algorithm based on scene priors by statistical analysis of the foreground moving objects and backgrounds regions in the database.Then,starting from the definition of ghosting artifacts,through the pixel contrast,a ghost-removal algorithm based on correlation(correlation between warped image and pre-motion image)is proposed.Then,by analyzing and combing the motion relationship between the pixels of the warped image,a ghost-removal algorithm based on motion relationship is proposed.Finally,a pre-motion image-based padding and correction algorithm is proposed for the ghosting artifacts of the warped image.Experimental results on multiple datasets show that the proposed warped image synthesis algorithm can effectively eliminate the ghosting artifacts,thus providing support for obtaining accurate optical flow estimation results.Furthermore,this dissertation studies the occlusion problem in optical flow estimation.Through analysis,we find that there is a very high correlation between the ghosting artifacts in the warped image and the occlusion area produced by the motion of the foreground object.In this way,we can mine occlusion information from the warped image.To this end,we first propose an iterative optimization algorithm for occlusion reasoning and optical flow estimation with the warped image as the link.Then,we use multiple information to improve the accuracy of occlusion reasoning,and a stepwise optimization algorithm for occlusion reasoning and optical flow estimation is also proposed in this dissertation.In order to verify the effectiveness of the method,a related experimental study of optical flow estimation was carried out.The experimental results show that the optical flow estimation algorithms proposed in this dissertation are effective and can obtain the correct optical flow estimation result in the presence of occlusion.Finally,the optical flow estimation problem is deeply studied by using the framework of deep learning.Based on the convolutional neural network model and combining the advantages of traditional algorithms,the optical flow estimation algorithm based on prior knowledge is proposed.Optical flow is a description of motion,and motion edges generally appear in the image edges of pre-motion images.Therefore,we introduce the edge information of the pre-motion image into the network,and propose an optical flow estimation algorithm based on the edge prior.The convolutional neural network is the learning and expression of complex mapping relations,and the occlusion area as the outliers has great interference to the learning of the mapping relationship.To this end,we propose an optical flow estimation algorithm based on occlusion prior,which introduces the occlusion inference into the network input and loss function at the same time to suppress the influence of the occlusion region on the network mode'l.The experimental results show that the fusion of prior knowledge can effectively improve the accuracy of optical flow estimation.