Research On Methods For Long-term Object Tracking Based On Correlation Filters

In this article,based on object tracking methods of the correlation filters for long-term are researched.Aiming at the problems of target scale variation,target occlusion and target disappearance and reappearance,the main contributions of this paper are as follows:(1)Long-term object tracking method using background context-aware modelAiming at the problem of object appearance changes caused by external environment change and self-changes under long-term object tracking,in this paper,a model of online tracking method based on background-aware is proposed,this method is a combination of context awareness model,kalman filtering,occlusion detection based on response figure and model updating based on the average peak correlation energy.Compared with other algorithms,the proposed algorithm performs better in the complicated scenarios of occlusion,fast movement,motion blur and scale variation,and achieves high performance on data sets.Experimental results show that the overall performance of the improved algorithm is 10.6% higher than that of the context-aware STAPLE＿CA,and the tracking success rate of the improved algorithm is 23.2% higher than that of the long-time target tracking algorithm LCT without the formula.In the case of occlusion,the success rate of the improved algorithm is 3.3%,24.8% and 7.3% higher than that of benchmark algorithm BACF,long time tracking algorithm LCT and context aware algorithm STAPLE＿CA,respectively.(2)Long-term object tracking method with scale adaptation and online re-detectionAiming at the problem that the short-time tracking method cannot solve the object scale variation and target disappearance effectively in the long time tracking process,a correlation filtering method with scale adaptive and online re-detection is proposed to reduce the drift problem effectively,and the target can be recaptured from the tracking failure.Firstly,a multi-feature kernel correlation filter is used to train the position filter to identify the target position in each frame.Secondly,in order to solve the problem of scale variation,a detection scheme using correlation filtering method is used.Finally,a new online re-detection strategy is introduced to relocate the target when tracking fails.Experimental results show that the overall success rate of the improved algorithm is 7.5%and 69.9% higher than that of LCT and TLD algorithms,respectively.Compared with KCF,DSST and SAMF,the success rate of the improved algorithm under the object scale variation is improved by 64.6%,13.4% and 14.1%,respectively.(3)Long-term object tracking method based on reliability evaluation and re-detectionFor long-term tracking results lack of reliable and effective update mechanism,presents a reliability evaluation method to generate a reliable tracking results,and the introduction of low-rank sparse method for re-detection,this paper proposes a long-term tracking model composed by detection module and re-detection module.The model can check the reliability of the target and update model through multiple aspects adaptively,Finally,the performance of the redetection method and the long-term tracking model is verified by a large number of experiments on OTB-2013 and OTB-2015 data sets.Experimental results show that the overall success rate of the improved algorithm on dataset OTB-2015 is 13.8%,5.16% and 18.9% higher than that of Staple,Siamfc3 s and LCT algorithms,respectively.In the case of occlusion,the success rate of the improved algorithm on dataset OTB-2013 is 2.8%,6.3% and 6.7% higher than SRDCF,HCF and Siam FC,respectively.