A Research On Object Tracking In Video Sequence

With the rapid development of various aspects of modern networks, the relationship between social network is more and more intimate. Visual object tracking becomes an important part of computer vision. Visual object tracking can be used in a large range of life, such as the interaction between humans and computers, video surveillance, the production of interactive video and the aspect of medical image. The procedure of object tracking is given the initial parameters of a target, which exists in the first frame of video sequence(for instance, the position of the target and its size), then according to the obtained conditions about the target and the existing tracking methods estimate the parameters of the target, which exists in the following sequences. Object tracking has highly challenges because of many affect factors, such as illumination, translucent/opaque or complete occlusions, rotations or deformation in two-dimension or three-dimension, transformation, low resolution, fast motion, blur, and confusing background and similar objects in the scene. The existing tracking methods have all kinds of restrictions that can’t completely hold in various scenarios.On the basis of the analysis and the summary of the existing tracking methods, this thesis mainly studies on the fusion of the existing object tracking methods to generate a new tracking method. This thesis mainly involves:(1) Using gradient ascent to fuse multiple existing object tracking methods to be a new method. Let these methods track multiple video sequences with specific attributes(affect factors). The tracking results of these methods are in a form of a bounding box, which is the input of the newly obtained method. In order to find the fusion result box that gets the greatest attraction for a frame we first test all tracking result boxes for how much attraction they get and keep the one with the greatest attraction. Then we perform gradient ascent starting from that box to determine the fusion result box. In this thesis, we mainly adjust parameters in attraction function to get a better tracking performance.(2) We provide a new method using an improved k-means algorithm to fuse multiple tracking methods to be a new one. We take the multiple result boxes as the input of the new method, by setting a threshold not to limit classification number and adding a weight to improve the effect of k-means and finally get a result box.This method improved the tracking performance and the running time.These two methods both evaluate the performance by a success plot and a precision plot. Precision measures the center location error in pixel space between the tracking box and the ground truth data. Success measures the overlap between a tracking and a ground truth box. The overlap is the intersection divided by the union of the two box regions. The video sequences in this paper have many attributions.