Research On Target Tracking Method Based On Twin Networ

Visual object tracking is one of the most basic research topics in the field of computer vision,which has a wide range of applications such as autonomous driving,video surveillance,human-computer interaction,and ocean exploration.Given the location of the interested object in the first video frame as a reference,object tracking aims to estimate the location of the object in subsequent video sequences.Recently,thanks to the rapid development of deep learning,object tracking has made great progress.However,in complex real-world scenes,due to factors like occlusion,deformation,illumination variation,scale variation,and fast motion,the appearance features of the object change dynamically,making it still a challenge to achieve high-quality tracking methods.Therefore,this thesis studies the Siamese network based tracking methods and proposes corresponding improvements for the shortcomings of the existing methods.The specific research contents are as follows:(1)Considering the problem that traditional channel attention can lead to severe spatial information loss,a triplet attention mechanism based anchor-free Siamese network method for object tracking is proposed.The triplet attention mechanism consists of two parts,in which cross attention is used to model the dependencies between channel and spatial dimensions,explore the fine-grained features of the object,and improve the model’s ability to distinguish similar objects.Spatial attention,as a supplement to cross attention,is used to focus on spatial regions where key features are located and suppress invalid features.Different from existing attention mechanisms,the triplet attention mechanism proposed in this thesis does not require a large number of additional learnable parameters and does not increase the computational complexity of the model.Experimental results on multiple public datasets show that the proposed algorithm has better performance and robustness in complex scenes such as illumination variation and similar object interference.(2)Considering the problem that the existing Siamese network based tracking methods do not perform online model update,a gradient-weighted object template update method is proposed.This method uses the rich discriminative information in the gradient to carry out feature weighting,strengthens the new features while suppressing the background noise that may pollute the object template,and fuses with the current object template in a non-linear way to enhance the model’s ability to adapt to complex environmental changes.Different from the existing object template update methods,this thesis concentrates on the real-time performance of the tracker when designing the update strategy,so as to avoid greatly reducing the tracking speed of the model.In addition,this thesis also introduces a dual attention mechanism combined in series to improve the feature representation capacity by aggregating contextual information.The algorithm proposed in this thesis has effectively improved the performance and robustness of the model on multiple large-scale datasets.