Dual Model Manifold Regularized Correlation Filter Tracking

The tracking algorithm based on correlation filtering uses manual features to model the object appearance,which is fast but low in accuracy.The tracking algorithm based on deep learning uses convolution deep features,which is high in accuracy but slow in speed.In order to take into account the speed and accuracy of the algorithm,a dual model Manifold Regularization correlation filter tracking algorithm is proposed.The algorithm is composed of the main model and the auxiliary model.The main model combines the context sensitive framework with the correlation filtering algorithm,which effectively makes up for the background information filtered out by the cosine window in the correlation filtering learning model.Through Manifold Regularization of the context-aware samples,the purpose of punishing the context sensitive framework and optimizing the main model is achieved.The auxiliary model combines kernel correlation filtering model with convolution feature.The auxiliary model uses the depth feature to model the object appearance.Because the depth feature has a large dimension,it is time-consuming to do correlation calculation.Therefore,the principal component analysis technology is used to reduce the dimension of convolution feature to speed up the tracking speed of auxiliary model.The main model is responsible for the main tracking task.When the tracking target is occluded,deformed or beyond the line of sight,and the tracking confidence is lower than the experience threshold,the auxiliary model is used to correct and prevent the main model from drifting.The experimental results on OTB2015 and VOT2016 data sets show that the tracking speed of the main model is fast and the accuracy is low,and the tracking speed of the auxiliary model is slow and the accuracy is high.Therefore,the main model and the auxiliary model cooperate and complement each other in speed and accuracy.The proposed algorithm is tested on OTB2015 and VOT2016 data sets and achieves good results.It outperforms other related filtering algorithms in terms of accuracy and robustness.This paper has 19 figures,9 tables and 50 references.