Single Object Tracking Algorithm Based On Siamese Network

Siamese trackers with online training strategies have recently drawn great attention because of their balanced accuracy and speed.However,some limitations remain to overcome,i.e.,trackers cannot deal with appearance variations well and can hardly discriminate target from similar background so far.In this thesis,we analyze the current development of trackers,especially the trackers based on Siamese network.Treating the Siam FC tracker as the baseline,we propose our novel real-time Siamese network trackers Siam Fusion,DCANet and DCANet++.The proposed trackers are modified from Siam FC to enhance feature representation capability of Siamese network in several ways,like changing the backbone network,proposing Co-channel and Crspatial attention modules,designing double Siamese architecture,studying template update algorithms,etc.Our contributions can be divided into three parts.First,according to the current statue of single GPU,we select the modified VGG16 network as the backbone network.The design of backbone based on two parts.One is the feature size matching of Siamese network,another is padding operation will harm the network translation invariance.We propose double Siamese network tracker Siam Fusion by using the strategy of module fusion.Experiments show that the double Siamese architecture is effective in improving the tracking performance.Second,we propose two attention modules,Co-channel attention module and Cr-spatial attention module and add the two modules into our tracker to enhance the feature representation capability of our tracker network.We design a novel real-time Siamese network(DCANet)based on the two attention modules.We perform extensive experiments on four benchmark datasets,including OTB-2015,VOT-2017,La SOT and GOT-10 k,which demonstrate that our DCANet gains a competitive tracking performance,with a running speed of more than 60 frames per second.Third,most trackers based on Siamese network always take the first frame as the template frame.However,in the long-term tracking scene,the initial template frame cannot match the current state of target well due to the deformation of the target.The template update algorithms can improve the robustness of the tracker in long-term tracking.Taking both double Siamese architecture and the two new attention modules into considerations,we design a better tracker DCANet++.Template update methods are also been discussed.Accordding to experiments results,the DCANet++ tracker can get better tracking performance than DCANet.Finally,this thesis summarizes the strengths and weaknesses of the proposed tracker and discusses the possible improvement directions in the future.