An Anchor-free Siamese Network With Template Update For Object Tracking

Object tracking is one of the basic tasks of computer vision research,and has important research and application value in tasks such as video surveillance,unmanned driving,and vehicle tracking.To improve the tracking ability of the existing Siamese network to deal with problems such as occlusion,deformation and scale variation,an Anchor-free Siamese network with template update for object tracking is proposed.The main works of this research are listed below.(1)A deep Siamese network with high-confidence template update for object trackingIn order to improve the expression ability of features and solve the problem of template not being updated,a deep Siamese network with high-confidence template update for object tracking is proposed.By changing the sampling strategy of positive samples,reducing the residual network stride and increasing the receptive field,the original shallow backbone network is replaced by the deep network to obtain more abstract features with stronger expressive ability.In order to further utilize the features of the high-level and low-level,a fusion module is designed to effectively improve the tracking accuracy.In order to reduce model drift caused by frequent updating of templates,a high-confidence template update mechanism is added to introduce high-confidence tracking result feature information,reduce the introduction of erroneous information in the background,and alleviate the impact of deformation,fast motion,background clutters and other factors on tracking.The experimental results show that the success rate and accuracy of the proposed algorithm on OTB100 are improved by 3.4%and2.6%respectively compared with the Siam FC algorithm,and the average overlap and success rate(SR_0.5)are increased by about 5.3%and 6.3%respectively when running on the dataset of GOT-10k.(2)An anchor-free Siamese network for object trackingIn order to effectively deal with the scale variation of the object and reduce the computation amount of scale testing,an anchor-free Siamese network for object tracking is proposed.The low-level feature maps are concatenated with the high-level feature maps to take full advantage of feature maps with spatial information and semantic information.and the multi-layer feature maps are used for prediction respectively.The anchor-free prediction network is used to directly obtain the object category and bounding box in a per-pixel manner,effectively avoiding the increase of parameters and calculations caused by scale testing and anchor bounding boxes.Finally,in order to obtain continuous and stable tracking results,a high-confidence template update method is used to update the template,and multiple tracking results are weighted and fused.Experimental results indicate that the success rate and accuracy of the proposed algorithm on the GOT-10k and La SOT datasets are higher than the anchor-based Siam RPN++algorithm.The success rate and accuracy of the proposed algorithm on the GOT-10k and La SOT datasets are higher than the anchor-based Siam RPN algorithm.