Research On Single Object Tracking Algorithm Based On Siamese Network

Object tracking is a fundamental yet significant research field of computer vision.It possesses a wide use of scientific and technical applications such as intelligent surveillance,public safety,human-computer-interactions,unmanned vehicles and so on.The main task of visual object tracking can be simply summarized as: given the bounding box of an arbitrary target in the first frame,locate the very target in the subsequent frames as precisely as possible.In the real environment,the object will inevitably encounter scale variations,deformation,motion blur and other interference factors.Therefore,in-depth research of object tracking based on Siamese network is carried out and improving strategies are proposed in this paper.The research contents and results of this paper are as follows:(1)In this paper,a paralleled attention module Paralleled Spatial and Channel Attention(PSCA)and Adaptive Focal Loss(AFL)are proposed.Attention mechanism is introduced to enhance the feature representation ability and meanwhile suppressing the influence of irrelevant information.So the extracted features that are more robust and tracking accuracy can be improved.PSCA is composed of two parts.First,Group-wise Heterogeneous Spatial Attention(GHSA)is designed to highlight the semantic features from each space of sub-features.Next,Multi-size Channel Attention(MSCA)is proposed by referring to various sizes of 1D convolutions.It reallocates the channel weights from both global and local aspects.PSCA,which can be integrated into Siamese networks to optimize object tracking,is constructed by a parallel connection between GHSA and MSCA.In this paper,AFL is proposed through improving focal loss to reduce the negative influences of easy samples on the model during training stage.Based on the training process and importance of each sample,AFL can enable the network to adaptively focus on the hard examples that are difficult to distinguish so as to increase training efficiency.(2)On the basis of Siamese network,a feature fusion network(F-net)based on multi-level convolutional feature fusion for object tracking is proposed in this paper.As the structure of the convolutional network continues to deepen and become more complex,the convolutional features learned by the deep network do not bring obvious improvement for tracking.This paper refers to the characteristics of Siamese network to improve the classic deep network model and construct a more suitable feature extraction network.The network involved in this article first extracts the shallow convolution features in feature extraction stage.The feature of shallow convolution is not discriminative but retains the structural information.Next,deep convolution features are extracted.As the network deepens and convolution operations continue to accumulate,target boundary information will be blurred,but the features have rich semantic information and high discrimination.Finally,various levels of convolution features are fused to form complementary features,which enables the network to not only distinguish object and background accurately,but also obtain the precise size of the object during the tracking process.(3)On the basis of Siamese network,M-net,which is based on mutual learning,is proposed in this paper for object tracking.It bridges the gap between the two branches and updates the object template to a certain extent through mutual information learning.First,a series of matrix operations are applied on the features extracted from the two branches to change feature dimensions for convenience of calculation.Next,search branch learns information from template branch via operations such as matrix multiplication and vector dot product to enhance the feature representations.Finally,template branch learns contextual information from search branch to ensure that the target template is dynamically updated according to background information.Furthermore,M-net and F-net are combined to form a double Siamese network,which not only includes the intra-branch feature fusion,but also pays attention to inter-branch mutual learning,for single object tracking.In summary,strategies including PSCA,AFL,F-net and M-net are proposed in this paper to improve single object tracking algorithms based on Siamese network.Besides,abundant experimental results demonstrate the excellent performance of the proposed algorithms on several public datasets.Moreover,the proposed methods in this paper satisfy the real-time condition,which can meet the need of online tracking.