Correlation Filter Moving Target Tracking Algorithm

The tracking of moving object is one of the important parts of Computer Visual system,it was widely used in areas like military,public security and autonomous car.Detecting techniques'development starts before tracking techniques,there are many remarkable algorithms whose theoretical foundation has already been completed.Recently,a method named Tracking by Detection showed an excellent performance.It tracks object by detecting techniques in every single frame,then achieves the tracking through the whole sequence.Kernelized Correlation Filter is the representative of Tracking by Detection,it's training a correlation filter to track object.This paper focuses on multi-scale correlation tracking and long-term tracking when drifting happens and proposes an Adaptive Correlation Filter to deal with multi-scale tracking.Currently modified algorithms are based on Minimum Output Sum of Squared Error Filter,while in this paper,we expanded it to KCF to do accurate and efficient position filtering through diagonalization of circulant structure.Then we employed scale pyramid and convolution theorem to conduct scale filtering in Fourier domain.And based on the ratio of scale changing of last frame,we chose scale priority strategy or position priority strategy suitably for a better robustness between scale and position change.In this paper,we also suggested a long-term object appearance model named CUR filter for template drifting.The main theorem of CUR filter is the matric dimension reduction,which builds CUR filter from history matrix Q which contains all the image information of successful detection by a method derived from CUR decomposition.In this way,most features of the object could be kept,which would realize the object's long-term presentation.If adapt CUR filter as the object appearance model of next frame,even though when the failure of detection happens in Adaptive Correlation Filter,the tracker can ignore the influence of the failure frame and catch up the object because of the long-term presentation character of CUR filter.