Research On Real-time Target Tracking Method Based On Siamese Network

Target tracking has always been a hot issue in computer vision,which is applied widely in many fields,such as missile positioning,video surveillance,and UAV reconnaissance.Although target tracking algorithms based on siamese networks have flourished in recent years,it's speed and accuracy are still limited by some complex scenes,so that it cannot be effectively applied.The main problems are that the feature extraction framework of the algorithm based siamese network is too shallow and cannot be learned online,which lead to the lack of discriminability of the model.In order to solve the above problems,this thesis mainly studies the target tracking algorithm based on siamese network.The main contributions of the thesis are as follows.In order to solve the problem of tracking failure caused by tracking target drift,which is due to the similar semantic information interferers in the fully convolutional siamese object tracking network(Siam FC),this thesis proposes a new Multi-level Features Enhanced Siamese network(MFESiam)to improve representation capabilities by enhancing the high-level and shallow-level features respectively.Firstly,for shallow features,a lightweight and effective feature fusion strategy is adopted to simulate some changes in complex scenes through a data enhancement technology,such as occlusion,similarity interference and fast motion and so on,to enhance the texture characteristics of shallow features.Secondly,for high-level features,a Pixel-wisely global Contextual Attention Module(PCAM)is proposed to improve a long-term localization ability of the target.Finally,a large number of experiments are carried out on three challenging tracking benchmarks: OTB2015,GOT-10 k and 2018 Visual-Object-Tracking real-time challenge(VOT2018).The results show that the proposed algorithm's success rate index on OTB2015 and GOT-10 k are 6.3% and 4.1% better than the benchmark respectively and runs at 45 frames per second for the real-time tracking.In the VOT2018 real-time challenge,it surpasses the champion-high performance visual tracking with Siamese Region Proposal Network(Siam RPN)and verifies the effectiveness of proposed algorithm.In order to solve the problem of tracking accuracy degradation caused by the insufficient stability of the tracker,which is due to the violent target deformation and background chaos in the fully convolutional offline training siamese tracking network,this thesis proposes a graph attention mechanism siamese network target tracking algorithm based on online learning.First of all,this thesis proposes an online learning sub-network,which uses the template updating strategy and discriminant classifier to learn the characteristics of specific targets,so as to improve the performance of the tracker.Second,a graph attention mechanism is to establish the local topological correspondence between the target and the search region to deal with the lack of information between the target partial levels caused by global matching.Experiments show that the proposed algorithm has a great improvement in the OTB2015 and VOT2018 tracking benchmarks,which surpasses the baseline by 8.8% and 6.7% respectively,and achieves real-time tracking at a speed of 50 frames per second.It also surpasses the same type of Graph Convolution Tracking(GCT)by 2.2% in the OTB2015 target tracking library.