Target Tracking Method Based On Correlation Filtering

Correlation filter has shown good performance in common video target tracking and UAV tracking.Many methods based on Discriminative Correlation Filter(DCF)successfully use the model to alleviate the degradation problems of boundary effect and time filtering in video target tracking.These methods mainly depend on the explicit prior regularization item,with a framework to control objective function to update the loss of degradation,but often ignore the data fidelity term losses,and these methods are usually also be affected by the boundary effect and background noise caused by the distortion problem of limitation.The restrictions on UAV tracking is more apparent.To solve these problems,we propose a bilateral weighted regression ranking model with spatial-temporal regularized correlation filter.Here,we approach the above problem in two ways.Firstly,a bilateral constraint is introduced into the data fidelity item to control the row and column loss of the filter learning data item.In order to avoid the problem of tracking offset and model degradation,the weighting matrix can be used as an adaptive penalty for mass data loss during the learning process.Secondly,in each iteration,the updated weighting matrix is updated by ranking and numerical transformation to obtain a new weighting product matrix to update the filter.In terms of UAV tracking,we propose an aberrance Kullback-Leibler divergence correlation filter model by introducing Kullback-Leibler divergence,which improves the Euclidean distance in the original model and reduces the tracking offset caused by noise distortion in the filter.At the same time,the Jensen-Shannon divergence is introduced to replace the Kullback-Leibler divergence,so as to study the influence of the measurement method on the tracking effect and the model performance.In addition,we propose a dual-domain Jensen-Shannon divergence correlation filter model by combining with the bilateral constraint method.In the iterative solution,the above model is approximated as a linear equation constraint problem and is solved iteratively by the Alternating Directions Method of Multipliers(ADMM).Through qualitative and quantitative evaluation on some widely used video target tracking datasets,the effectiveness and superiority of our proposed method are proved.