Research On Target Tracking Method Based On Dense Siamese Network

Target tracking is one of the hot research issues in computer vision,and it has broad application prospects in unmanned driving,smart transportation,smart surveillance,and military reconnaissance.In recent years,the target tracking technology has made great progress.However,due to the complexity of the environment and various uncertainties in real life scenarios,the target tracking technology still faces great challenges.On the basis of the siamese network framework,this thesis conducts in-depth research and exploration for the existence of background clutter,target deformation and target scale change during target tracking.The main content and results of this thesis are as follows:(1)A dense siamese network tracking algorithm based on global context is studied.In order to improve the feature extraction ability of the network and enhance the network’s discrimination of similar target backgrounds,based on the feasibility analysis of the Dense Net,the shallow Alex Net network is replaced with a more powerful dense network.The convolutional layers in the dense convolutional network are connected in a densely connected manner,and the features between the layers are spliced on the channel to achieve feature reuse and improve the generalization ability of features.In order to further enhance the network’s robustness to changes in the appearance of the target,a global context module is added after the siamese network template branches to aggregate the global context information of the target,so that the network can output feature maps with a lot of semantic information.(2)A dual-template dense siamese network tracking algorithm based on spatial-temporal attention mechanism is studied.When the target is deformed,a single fixed template in the siamese network is not enough to express the current state of the target.In response to this problem,this paper designs a new template to complement the original template to achieve the dynamic update of the target template.The new template is selected based on the average peak-related energy value of the historical prediction result score map.After selection,the original template and the new template are input into the siamese network to obtain their respective feature maps,and then the two feature maps are passed through a spatial-temporal attention module to obtain their respective weight maps,weighted fusion of the weight maps with the corresponding feature maps to obtain the fused template feature map,and then cross-correlation with the current frame feature map obtained by the current frame through the dense network to obtain the final response map,which can determine the target location.This algorithm for real-time updating of the target template can also play a good target tracking effect even when the target is distorted.(3)A dense siamese network tracking algorithm based on anchor-free classification and regression is studied.In order to solve the problem of target scale change in target tracking,a target tracking task is decomposed into a classification task and a regression task.The classification task separates the target from the background to get the rough position of the target,the regression task predicts the target state to get the target bounding box,and the two combine to get a precise target location.The network classifies and predicts the target frame pixel by pixel,abandoning the previous method of relying on anchors for classification and regression,and avoiding the uncertainty and high calculation amount caused by the hyperparameter setting of the anchor.In order to solve the inconsistency between the quality evaluation branch and the classification branch in the training and testing of the network,the classification branch and the quality evaluation branch are integrated,the meaning of the training label is redefined,and a new loss function is set to solve the goal.With regard to scale issues,good tracking results have been achieved.The algorithm in this thesis is experimentally verified on the OTB2015 dataset.The experimental results show that the dense siamese network tracking algorithm based on the global context still achieves a good tracking effect under the complex and changeable background.The tracking algorithm of the dense saimese network based on dual templates still has high accuracy and robustness in target tracking when the target undergoes huge deformation.The dense saimese network tracking algorithm based on anchor-free classification regression can effectively estimate the target position and bounding box size,and the accuracy is still high when the target scale changes.