Optimization And Mplementation Of Optical Flow Estmation Algorithm Based On Image Sequence

Optical flow is an important research direction in the field of computer vision.It is a method to descr:ibe the movement of pixels between images over time.It is widely used in the field of image processing such as motion detection,motion estimation and self-driving.The traditional optical flow estimation algorithm uses a feature pixel or all pixels of an image to be associated with velocity,and then finds a corresponding feature pixel in the next frame.Although the traditional algorithm has better results in dealing with some scenes,there are bottlenecks in calculation time and estimation accuracy in dealing with scenes with large pixel motion.With the increasing amount of information,traditional image processing algorithms can not meet the characteristics of high resolution,real-time and so on.Compared with traditional image algorithms,CNN has great advantages in calculating speed and optical flow estimation accuracy,but the regression value of optical flow estimation for key parts of pixels still needs to be improved,especially in the edge part of image frame,the predicted value deviation of regression it is bigger.This thesis focused on the enhancement of the edge regression value of optical flow prediction,built a deep neural network model,and complete a series of research and related work.The main contributions of this paper are as follows:(1)Designing an optical flow estimation algorithm based on deep learning,using the TensorFlow framework to construct an optical flow prediction FPN(feature pyramid networks)network model.Using theFlying Chairs optical flow public dataset to train the network to convergence,and built an end-to-end optical flow estimation network.The image sequence runs on the model,the deviation of the optical flow estimation result is less than 1%,and the processing speed is higher than 30fps,which can achieve real-time effects.The optical flow estimation for the edge enhancement innovative experiments laid the foundation as the baseline.(2)For the problem of edge contour blurring of optical flow estimation results,based on the established baseline network model,two methods are proposed to train the edge enhanced optical flow network.The first group is the edge map of adding images in the input model.The other group is to add the foreground obj ect contour to the input module.The two groups of experiments also use the same training dataset training network in the baseline network.Compared with the baseline experiment,the two sets of enhanced experiments on the EPE(End Point Error)index have a good result of 3.71%and 7.54%.