The Object Tracking Algorithm Research Based On Mean Shift

Objecttrackingistosearchthemostsimilarpartstotargetsinvideosequences,whichisanimportant research in computer vision. In recent years, object tracking is used in manyapplications such as surveillance, perceptual user interfaces, augumented reality, object basedvideo compression, and driver assistance. Among the various tracking algorithms, Mean Shifttrackingalgorithmhasbecomepopularduetoitssimplicity,efficiencyandgoodperformance.This paper focuses on Mean Shift tracking algorithm, which is a modeling mechanismbased on statistical probabilitydensityfunction (pdf). In the tracking process,the target area andthecorrespondinghistogram areusuallyestahlished in thefirst framebyuser. In thesubsequentframes,thebesttargetcandidateareasaresearchediterativelybyMeanShiftalgorithmbasedonthe Bhattacharyya similarity function. Mean Shift method can be performed in real time and isrobust for partial occlusion and distortion. However, it has a few defects in some aspects. Forexample,whenthetargetissimilartothebackground,itisdifficulttodistinguishthetargetfromthebackgroundandleadstofailintracking.In the paper, various representative video sequences are utilized in experiments to analyzeand verify the performance of Mean Shift. And a novel Harris Mean Shift algorithm based oncorner detection is put forward to overcome the tracking defect of the Mean Shift algorithm incomplex scenes. The proposed algorithm just selects some key feature points by the Harriscorner detection algorithm to build the target model and the candidate model, which reallyenhances the performance of Mean Shift tracking. Experimental results show that Harris MeanShift algorithm could not only reduce the background's interference for target localization incomplexscenes,butalsomitigatetheinfluenceoflighteffect.