Visual object tracking is a fundamental task in both computer vision and human-computer interaction. Its applications spread over many fields, such as vehiclenavigation, video surveillance, and intelligent robot. These applications facilitatepeople’s daily life and accelerate the development of industry.Despite the improvements in both theory and experimental performance, visualobject tracking remains a challenging work. The occlusion and drifting problems havenot been well solved. The destabilizing factors of illumination, target scale, view pointand complicated background make the tracking problem difficult. The work of thispaper focuses on the occlusion problem and the drifting problem. With the analysis ofthe previous work, two novel trackers are proposed.Occlusion problem is a common but tough problem in visual object tracking.Many trackers are trying to avoid this problem in an indirect way. In this thesis, a novelmethod is proposed to solve the occlusion problem directly by using the occlusioninformation. This new algorithm first gives an occlusion detection mechanism, whichcan detect the occlusion accurately. Second, a combined tracker of discriminative andgenerative methods adjusted by occlusion information is presented. Third, a templateupdate method with occlusion information is given. This new tracker solves theocclusion problem well and outperforms the state-of-the-art trackers.By analyzing the drifting problem of the local sparse tracker, we find the causeof the drifting problem. Local features may lead to an over-fitting problem, especiallywhen the background area is similar to the target area. In order to solve this problem, aglobal constraint on the contributions of each local part is added to the objectivefunction, after analyzing the tracking procedure of local sparse tracker. Also thisconstraint is exploited to guide the template update process. With these improvements,the local sparse tracker with global constraints is robust and stable, which alleviates the drifting problem effectively. The comparison of experiment results with several state-of-the-art trackers on the benchmark datasets shows the superiority of our tracker. |