Research On Visual Location Technology Based On Optical Flow

With the development of robotics and Artificial Intelligence,Visual Navigation Systems come to the fore in many fields,such as Smart Robots Systems,Automated Driving Systems and VR(Virtual Reality)/ AR(Augmented Reality)Systems.As an important subfield of Computer Vision,Optical Flow is widely applied to Vision Localization.There are vast motion information and structure information in optical flow field.The motion estimation is performed with the calculation of optical flow,following the fulfillment of navigation and position.In this paper,the method of calculating optical flow field and the applications of optical flow in Vision Localization are studied.Firstly,a great deal of literature researchs on optical flow are carried out.then,an improved method of calculating optical flow field based on TV-L1 algorithm and CLG algorithm is proposed.The implementation of the improved algorithm is accompanied with the structure-texture decomposition approach,the multiscale pyramid image algorithm and the image filtering methods.In addition,a lot of comparative tests on Middle Berry dataset are conducted.The results of tests show that the improved approach improves the accuracy and robustness of the calculation of optical flow with smoother optical flow field.The Average Angle Error(AAE)of the optical flow field computed by the improved algorithm on dataset is 3 degrees and the Average End-Point Error(AEPE)is0.25 pixels.After that,the key techniques of binocular stereo vision based on the optical flow method are studied,such as the camera model,camera calibration,depth estimation and pose estimation.Then,a novel binocular vision odometer method is put forward based on a state-of-the-art monocular direct vision odometer approach called Direct Sparse Odometry(DSO).Finally,the evaluation of improved binocular vision odometer method is carried out on two group tests.The first one is based on KITTI Dataset,and the average translation error of our method is 1.6%,the average rotation error is 0.03 degrees per meter.Two campus experimental platform are built in another,numbers of experiments draw a conclusion that the accuracy of location of the improved method is better than original one,and the improved method is more robust than ORB-SLAM2 algorithm.