A Method Of Removing False Features Of Lucas-Kanade Flow Pyramidal Implementation

Computer vision is the eye of computers to learn the word. Computer vision has become one of the most popular research fields with the popularization of automation. Motion tracking is an important research topic of computer vision, which has a decisive position in the application of UAV, AHV, traffic monitoring and etc.The most widely used computer vision motion tracking algorithm is based on tracking corners and optical flow estimation with iterations of pyramid algorithm. However, that the three hypotheses of the optical flow estimation, brightness constancy, temporal regularity and spatial consistency, can be hardly satisfied in the actual situation, which limits the application scope of optical flow estimation algorithm. Other, optical flow estimation requires large computing resource, so application of motion tracking on small embedded computing units.In this thesis an adaptive dynamic false feature removing method is proposed, which can remove false features those may lead to error results and reduce efficiency of motion tracking. The main content is as below:Chapter one makes an introduction to computer vision, application scope and the main problems of motion tracking algorithm.In chapter two, the basic conceptions and common algorithms are presented. Particularly, the definition of corner, thinking of Harris feature tracker, derivation of optical flow estimation and background of pyramidal implementation have been described in detail.The third chapter provides a false feature removing algorithm. Firstly, correction term of optical flow estimation has been analyzed, a reasonable scope of correction term in pyramidal implementation has been pointed out according to the error introduced by the three hypotheses of the optical flow estimation. To reduce error influence on optical flow estimation caused by brightness constancy, local brightness constancy compensation algorithm has been proposed. The inequality, determined by least square solution of optical flow estimation, leads to the reasonable scope of pyramidal implementation correction term. The results of every stage of every level of pyramidal implementation are compared with the scope, and the false features can be removed if the results are beyond the scope. To the result of the first stage of the first level, it can be limited by temporal regularity, if the result is larger than size of neighborhood window, it can be removed as false feature.The fourth chapter verifies false feature removing algorithm with experiments. Grid JS and pilkft.js vision library are used in the experiments. The experiments including 14 image sequences with different scenes and frame numbers record the changes of number of tracking features, accuracy of the results and proportion of false features. The experiment results of the 14 image sequences show that false features can be removed fast and completely by the false feature removing algorithm. The proportion of false features in the entire tracking process basically keeps under 0.3, and keeps at 0.15 in most cases. Almost all false features can be removed when the number of tracking features remains stable.The last chapter summaries the whole work of the thesis.The false feature removing algorithm presented in the thesis, which can remove false features fast and adaptively with preserving most effective features in different scenes, raises accuracy of the result and decreases computing strength, and make it possible to applicate vision motion tracking algorithm on small embedded computing units.