Fusion Object Tracking Based On Attention Mechanism And Multiscale Geometric Network

With the rapid development of modern science and technology,Object Tracking has penetrated into people’s daily life.Single Object Tracking task usually gives the size and position of a specific target in the initial frame of a video sequence.Algorithms need to track the specific target in all subsequent frames.In recent years,object tracking algorithms based on correlation filters and deep siamese networks are research hot spots,they have achieved high performance in both speed and accuracy.However,some problems such as similar object interference,occlusion,etc.all still require better solutions.Trackers based on region proposal siamese tracking networks are susceptible to interference from similar objects in the background,and have the problem of insufficient discrimination.At the same time,the scale estimation result of siamese tracking networks which based on segmentation still need to be refined.The structures of siamese tracking networks cause that the quality of the prior image in the first frame will affect the tracking result.Therefore,in view of the above-mentioned problems,this thesis mainly makes the following three improvements.First of all,in view of the poor discriminative ability of the deep siamese trackers based on the region proposal networks in the case of similar obiects,backfround interferense,etc.This thesis proposes a fusion target tracking network which is based on the attention mechanism and multi-layer features,the channel squeeze attention module improves the correlation between channels,and can effectively obtain the spatial attention distribution of the image,thereby increasing the importance of the object candidate area.At the same time,it integrates the multi-layer features of different depths and levels of the network,and combines the attention mechanism with it to improve the accuracy and discriminative ability of the network.Through experimental verification,compared to the siamese tracking networks based on region proposal,the fusion object tracking network based on the attention mechanism and multi-layer features proposed in this thesis can effectively improve the discriminative ability,the success rate and precision rate on OTB2015 dataset were improved by 1.4% and 2.3% respectively,and on OTB2013 dataset were improved by 1.8% and 1.3%respectively.The performance of the network is significantly improved when dealing with similar objects,background interference,etc.Secondly,in view of the problem that the scale estimation result of the siamese networks based on segmentation are not refined enough,this thesis proposes a tracking network based on segmentation and multi-scale geometric features.It uses image shearlet transform to obtain multi-scale image edge contour texture feature,and integrates the siamese network with multi-scale geometric features,which improves the network’s feature expression,so that the bounding box can be predicted more accurately,and a more refined scale estimation result.Through experimental verification,the proposed fusion target tracking network based on segmentation and multi-scale geometric features can more accurately estimate the target size and effectively improve the performance of the tracker compared to the siamese network based on segmentation,the success rate and precision rate on OTB2015 dataset were improved by 1.2% and 3.0% respectively,and on OTB2013 dataset were improved by 1.6%and 4.0% respectively.Finally,in view of the problem that the performance of the deep siamese network trackers are affected by the quality of the prior patch,this thesis proposes a fusion target tracking network based on Scale Invariant Feature Transform and multi-scale geometric attention.The motion model based on Scale Invariant Feature Transform can improve the problem that the prior image patch does not contain the target when the lens shakes,effectively prelocating the possible position of the target in the next frame,and improves the probability of the existence of the target in the prior image patch.At the same time,it is combined with the above two network structures,the attention mechanism is used to improve the network discriminative ability,and the image multi-scale geometric features are used for fine target scale estimation.Through experimental verification,the fusion tracking network based on Scale Invariant Feature Transform and multi-scale geometric attention proposed in this thesis can improve the effect of prior image patch quality on the tracking results and improve tracker performance compared to the siamese tracking network based on region proposal,the success rate and precision rate on OTB2015 dataset were improved by 2.5% and 3.1%respectively,and on OTB2013 dataset were improved by 2.8% and 2.3% respectively.