Most of the correlation filter based tracking algorithms can achieve goodperformance and maintain fast computational speed. However, in some complicatedtracking scenes, there is a fatal defect that causes the object to be locatedinaccurately. In order to address this problem, we propose a particle filterredetection based tracking approach for accurate object localization. Duringthe tracking process, the kernelized correlation filter (KCF) based trackerlocates the object by relying on the maximum response value of the responsemap; when the response map becomes ambiguous, the KCF tracking result becomesunreliable. Our method can provide more candidates by particle resampling todetect the object accordingly. Additionally, we give a new object scaleevaluation mechanism, which merely considers the differences between themaximum response values in consecutive frames. Extensive experiments on OTB2013and OTB2015 datasets demonstrate that the proposed tracker performs favorablyin relation to the state-of-the-art methods.
展开▼
