Motion-Based Feature Selection and Adaptive Template Update Strategy for Robust Visual Tracking

The Kanade-Lucas-Tomasi(KLT) method is a classictracking algorithm, which however suffers from a longexisting gradual template drift problem. In this paper, wepresent an improved tracking algorithm which can sustain thetracking performance for a long term against drift. In thisapproach, we formulate the tracking over a state comprising of a template with kinematic motion. Based on the sparsemotion field generated by KLT, a motion consistency basedmethod is applied to filer out the outliers which are the causeof cumulative drift errors. To sustain the tracking performance in a long term, an adaptive template update strategy monitoredby the appearance and motion continuities of the template isproposed. Finally, quantitative testing on several benchmarksequences demonstrate the advances of the proposed methodin long term tracking.