Research On Tracking Of Moving Objects

Object tracking has been widely applied to various fields such as intelligenttransportation, video surveillance, video editing, It is an important area in computervision research. There are many effective algorithms have been proposed, but all ofthem are useful in specific situations. Actually, the illumination changes, occlusion, thescale transformation, rotation and other factors that affect the tracking effect areuncertain. So it is still a challenge to propose an algorithm which reliable, robust anduseful in complex environments. In order to solve these problems, we studied the targetrepresentation and location progress of tracking algorithm, the main work of this paperis as follows:1. We studied the target tracking algorithm based on mean shift, which representthe target with color histogram. The target representation based on RGB color isunstable, because the different picture can produce the same color histogram, this paperproposed a novel way to represent the target which combine the color and local texture.The experimental results show that this novel representation is insensitive to brightnessand illumination variation. Most of all, the target itself may be experience sometransformation, it is necessary to use the new information update the target model, so weproposed a update strategy which improved the adaptability. The experimental resultsshow that the increased computation of novel algorithm doesn’t affect its real-timeperformance; it can track the target object robustly in real time.2. We studied the target tracking algorithm under particle filter framework, thetarget objects are represented by sparse representation in this section. The representationbased on holistic templates can’t distinguish between the target object and background.So we divide the target and candidate objects into local patches to solve this problem.That is to say, the patches are used to create a histogram by their spatial relationship,then we can find the patches which occluded by other disrupting things. These patchesare neglect when conduct histogram matching, this representation method is morerobust than holistic representation. What’s more, we proposed model update strategy tocalculate the appearance change caused by target move or the camera displacement and the background change. In holistic model, only background atoms are updated, while inlocal patches, the patches without occlusion are updated. Numerous experiments showthat the proposed algorithm can tack the target object robustly in complex environments,such as drastically illumination change, occlusion, etc.