Research On Correlation Filter Object Tracking Algorithm Based On Dynamic Spatial-Temporal Regularization

The traditional correlation filtering object tracking algorithm uses spatial-temporal regularization method to alleviate the influence of boundary effect and filter degradation,but this method cannot update dynamically,therefore the algorithm cannot obtain spatial regularization weight with the change of the object during tracking process,and it is difficult to avoid the occurrence of tracking drift phenomenon.Aiming at these above problems,the spatial-temporal regularization method is improved,and a correlation filtering object tracking algorithm based on dynamic spatial-temporal regularization is proposed.Firstly,the content-related strategy is obtained through saliency detection algorithm,and the saliency map containing the first frame information of the object is obtained by using this strategy,after processing,spatial regularization reference weight with object content is obtained by combining it with the traditional spatial regularization weight,which can be added to the objective function of the algorithm to construct dynamic spatial regularization term.Secondly,the response judgment strategy is proposed by analyzing the change of response map,and using the strategy of the average peak-to-correlation energy and the maximum peak response to calculate reference value of temporal regularization parameter,which can be introduced into the objective function of the algorithm to construct dynamic temporal regularization term.Finally,considering dynamic spatial regularization term and dynamic temporal regularization term,the objective function of dynamic spatial-temporal regularization tracking algorithm is established,and the correlation filter,dynamic spatial regularization weight and dynamic temporal regularization parameter are respectively solved by using two alternating direction method of multipliers in the optimization process.Experiment results on OTB dataset show that the improved spatial-temporal regularization method is feasible and effective,it not only solves the problem that the spatial regularization weight cannot update dynamically with the change of the object,but also avoids the tracking drift caused by filter degradation,so that the algorithm can achieve robust object tracking under complex scenes such as deformation,rotation,occlusion and out-of-view.At the same time,the use of alternating direction method of multipliers method also reduces computational complexity and makes the algorithm meet the real-time requirements.There are 23 figures,4 tables and 58 references in this paper.