Object Tracking Based On KLT And Mean-Shift Algorithm

Real-Time Object Track in Image Sequences is a classic problem in Computer-Vision, it refers knowledge in many fields such as digital image processing, computer application, pattern recognition, mathematics, et al. It has been widely applied to human-computer interaction, television monitoring systems, video compression solutions, intelligent traffic systems and so on, how to improve the robust and accuracy of tracking algorithm is the focus of the current work. In this paper, we focused on object track which is based on KLT and Mean-Shift algorithm, after improving the limitations of tracking algorithm, we design a dual tracking system that can track objects accurately.The major aspects presented in this thesis are:1) Firstly, describing harris corner detection and epipolar geometry in detail. Analyzing the estimation of fundamental matrix, By combining with RANSAC algorithm, we get a robust estimation of fundamental matrix. Because Harris corner detection lacks necessary scale information and KLT tracking becomes powerless when cameras are in relatively fast motion, we presented an improved method, which projected them to multi-scale space for tracking and detection.2) Introducing the basic theory of Mean-Shift algorithm and proving it to be convergence, applying Mean-Shift algorithm into object tracking field successfully. Because of its limitations in object tracking field, we proposed corresponding improved methods: (a) when the color of object and background are similar, spatiograms are proposed; (b) a method based on affine structure is proposed to solve kernel's bandwidth; (c) part of model components updating strategy is proposed to update object's template.3) Designing a dual tracking system. Introducing the basic software and hardware environment which tracking system runs on. Describing algorithm flow, experimental procedures and experimental result in detail.