Research On Unsupervised Optical Flow Estimation Method Based On Improved Feature Pyramid

With the continuous improvement of computer software and hardware,the research of computer vision technology have been fully developed,which are widely used in various applications.Optical flow method is an important technology for processing moving image sequence.The calculated optical flow contains the structure,position and spatial motion information of the object in the real world,which can be used for object location and recognition.Therefore,it has important applications in areas such as automatic driving,video segmentation and video semantic understanding.Optical flow can provide abundant information for other computer vision tasks,and its precision and density have an important impact on various applications.The running speed of traditional approaches cannot meet the real-time requirements in practical applications.In recent years,convolutional neural networks have been successfully applied in the field of optical flow estimation.Training process of neural network is time-consuming,but its reasoning speed is very fast,which can effectively solve the problem of poor real-time performance of traditional approaches.However,it is difficult to obtain the ground truth of dense optical flow in real scene,so the end-to-end supervised learning methods for optical flow usually use the synthetic data set to supervised train network,which often perform poorly on real-world image sequences.This paper proposes an unsupervised learning optical flow method,which is divided into three parts: improved feature pyramid network,optical flow estimation network and unsupervised training.The improved feature pyramid combines the dilated convolution with the feature pyramid network,so that the network can extract multi-scale features containing more motion information.Firstly,the multi-scale features of two images at t time and t+ 1time are extracted by two branch networks sharing network weights,and the feature images at different scales are matched.Then,the matching results are used as the input of the optical flow estimation network to infer the optical flow results at the current scale,and the positive and negative consistency check algorithm is used to calculate the occlusion area,excluding the influence of the occlusion area on the calculation accuracy.Finally,this paper uses census transform to construct unsupervised loss function to avoid the lack of a large number of real value data.Using its unsupervised training to the network,the accuracy of optical flow estimation is gradually improved through iteration.We evaluate our network on MPI-Sintel dataset and KITTI dataset to verify the effectiveness of the network.