Research On Single Target Tracking Algorithm Based On Fully Convolutional Twin Networ

The fully convolutional siamese network object tracking algorithm has been rapidly developed due to its well-balanced between accuracy and speed.However,there are many difficulties in object tracking restricting its performance,such as background clutters,occlusion and deformation,etc..Especially when the siamese network object tracker encounters interference with similar background information,the tracking is prone to drift,and the existing anchor-based trackers have many hyper-parameters,which brings extra complexity and computational consumption.In addition,only using a single feature extracted by the network for tracking prediction will lead to underutilize image information and cannot form robust tracking,and the effect of simply adaptive fusion multiple features is not ideal,so,it is necessary to apply a more reasonable way to make full use of the corresponding advantages of different features to track.Aiming at the above problems,this paper proposes two methods to solve them.The specific research contents are as follows:(1)In order to avoid many hyper-parameters and suppress the influence of distractors near the target,by redesigning the classification and regression of siamese tracker,and changing sampling strategy,a tracking algorithm based on center pixel weighted anchor-free siamese tracker is proposed(CPWSiam).First,a regression method with anchor-free is introduced to directly classify and predict the target bounding box on each pixel,avoiding the impact of excessive hyper-parameters and increased computational complexity;secondly,the exact division of samples makes the pixels of positive samples contain less background and improves the ability of feature discrimination;finally,through the method of center weighted sampling,it reflects the different attention of positive pixels,reducing the influence of adjacent similar background information enables the tracker to make more precise predictions.Experimental results on VOT2018 benchmark shows that the EAO of CPWSiam is2.9 higher than Siam RPN++,achieving advanced performance.(2)In order to make full use of the features at different levels of the image and give full play to the feature extraction ability of the Res Net50.On the basis of the CPWSiam only using a single level of features,combined with the weighted voting method,the features extracted at different stages of the Res Net50 are imported into the classification and regression network.The classification score map and regression map are divided into weighted voting,a multi-layers voting siamese network object tracking algorithm(MVCPWSiam)is proposed.Through a multi-layers voting design,avoiding the simple processing which adaptively fused prediction of multi-layers features,accordingly achieve more accurate target prediction,and further improve the tracking performance.Experimental results on VOT2018 benchmark shows that the EAO of MVCPWSiam is 1.6 higher than CPWSiam,and the advantages of multiple attribute challenges are further improved,which shows outstanding generalization and robustness performance.