| Object tracking is an important task within the field of computer vision. The popularity of high-powered computers, the availability of high quality video cameras, and the increasing need for automated video analysis has generated a great deal of interest in object tracking algorithms. There are three key steps in video analysis: detection of moving objects, tracking of such objects from frame to frame, and analysis of object tracks to recognize their behavior.There are many local characteristics used for tracking, such as contour, edge, point, etc. Among them, point of interest is what extracted most easily. In this thesis, Harris corner is extracted as tracking object. Harris corner is defined as the point where window moving in any direction will cause obvious gray change. Harris corner is rotationally invariant and insensitive to illumination. However, since corner detection is based on corner response of every pixel, it will consume a large amount of computation time when processing high resolution image. A proposed method is pre-screening based on difference in gray level between center and other points in an area to extract alternative points. Instead of every pixel, computing response of alternative points will save much time. In order to detect interest points invariant to scale, a multi-scale representation for the Harris corner is computed and selecting points at which a local measure Laplacian is maximal over scales. Curvature is computed to filter out false corners and enhance adaptability to scale change.Tracking algorithm in this thesis is optical flow based on gradient. Optical flow is the 2D motion field, a project of the 3D velocities of surface points onto the image surface. Based on optical flow, the position of tracking points can be estimated precisely, except high speed moving. A proposed method is based on image pyramid to decrease the size of the image in order to satisfy the optical flow constraint of "small movements". The estimated optical flow is corrected using iteration method. Based on EFB, the estimated error between 2 trajectories of different time flow in the same frame, tracker is able to self-evaluate its reliability and detect tracking failures.Finally, experimental results of differents video sequences show that tracking algorithm based on improved corner detection and pyramidal implementation of LK tracker has better tracking performance. The estimated error, EFB, enables reliable detection of tracking failures and selection of reliable trajectories in order to estimate the position of object. |
