Scale-adaptive Correlation Filter Tracking With Density Peak Clustering

During the process of complex scenarios and multi-scale changes,image pyramid is adopted especially when relevant filtering algorithm is used to forecast the scale.In the tracking process,tracking filter is trained based on multi-scale object samples,thus leading to more calculations and slower tracking speed.The article proposes the method of self adaptive tracking algorithm according to the scale of clustering by fast search and find of density peaks and applies it into the framework of Background-Aware Correlation Filters.Firstly,feature information about targets and background can be extracted respectively in the searching areas.Clustering by fast search and find of density peaks is adopted to conduct a clustering analysis of targets and backgrounds.The method of spaced samples is used for clustering.In doing so,it can reduce the impacts of background characteristics on model training against the context of complex scenarios while increasing the speed along the entire tracking process.Secondly,the target position is roughly predicted by a single-scale filter,and the constraint term is placed in the optimization function by the augmented Lagrange multiplier method to reduce the time complexity of the solution process.Lastly,foreground points and background points in the search areas are classified.And confidence coefficient of the scale can help get the final location and scale of targets to realize scale adaptation.Compared with Background-Aware Correlation Filters(BACF),the tracking accuracy and success rate of the method in this paper have been improved,and the tracking speed has been greatly improved.The tracking speed on the OTB2013 dataset has increased by 61.26%,and the tracking speed on the OTB2015 dataset has increased by 67.18%.The tracking speed on the DTB70 dataset is improved by 51.91%.Meanwhile,it produces better tracking outcomes and higher success rate in terms of complex situations such as rotating and blocking,meeting real time requirements.This paper has 24 figures,8 tables,and 59 references.