Research On Video Object Tracking Algorithm Based On KCF And Anchor-free Siamese Network

The research of target tracking algorithm with certain robustness and accuracy has very important practical guiding significance for the theoretical development of tracking technology and engineering practice.In recent years,the correlation algorithm based on kernel correlation filter and deep twin network has become increasingly popular in the tracking field.The research achievements of correlation algorithm have been remarkable,which has effectively promoted the landmark development of the tracking field.However,with the diversity and complexity of the application environment,this field also highlights the tracking difficulties in the corresponding scene.In this thesis,the tracking algorithm is studied for the diversified scenes in the video.The research content is described as follows:1.A scale offset target tracking algorithm based on kernel correlation filter is proposed.Against the original nuclear related filtering algorithm the problem of insufficient sensitivity to scale changes,this article under the basic framework of the nuclear related filtering algorithm study,by introducing a scale offset to establish the link between the scale layer samples,and then according to the proposed box selection strategy determine the scale layer corresponding proposal box,guided by proposed boxes to conform to the current actual tracking target size box.The cyclic shift theory is adopted in the scale updating stage of the algorithm,and the corresponding sample information of the scale layer is represented by the cyclic matrix,which can be used to diagonalize and further reduce the computation.In order to illustrate the robustness and feasibility of the algorithm,this paper takes drift and occlusion events as examples.Through evaluation and test on official data sets,it is verified that the proposed algorithm has good tracking performance,which is superior to other algorithms in tracking accuracy and success rate,and achieves significant tracking effect in scaling.2.Proposed anchor-free Siamese network tracking algorithm.In view of the traditional way of Anchor box pooling single,super sensitive parameters and sample imbalances,inspired by the Anchor-free related algorithm,this paper combines the keypoint and siamese framework effectively,using two corners and center to represent the target frame size position,and consider the angle point boundary effect,the introduction of angular point of cascade pooling is sample points to determine the target.Through testing in multiple data sets and self-built simulation environment,it is verified that the algorithm can maintain good tracking performance in the video scene corresponding to multiple data sets.3.In view of the impact of low-quality sample points on network training,two-stage Hourglass network is adopted in the backbone of the network,global pooling is adopted in the initial sampling on the back end of the network,and elliptical local offset structure is introduced to optimize the design of sample labels.In addition,the attenuation penalty coefficient is introduced into the classification branch and regression branch respectively,which further improves the classification performance and regression accuracy.