Research On Object Tracking Algorithm Based On Siamese Networks

Computer vision uses computers to simulate human vision mechanisms,and then models the real world with the help of image processing,pattern recognition and other technologies.A popular yet challenging research task in the field of computer vision is to track the visual object,that is to predict and calibrate the spatial position of the visual object according to the context information of the image sequence.It is heavily used in vision applications including vehicle assisted driving system,video monitoring system and human-computer interaction,to name a few.In the past few decades,object tracking algorithms have developed rapidly.Among them,Siamese networks based trackers to a certain extent take into account both the tracking accuracy and the running speed,and have drawn attention of a great many scholars to conduct researches.In this paper,the representative fully convolutional siamese network(Siam FC)and the siamese region propose network(Siam RPN)are improved and two different tracking schemes are proposed.The specific research content of this paper is arranged as follows:First of all,this article introduces the research significance of target tracking algorithms and the challenges faced by target tracking in complex real-life scenarios.The research status of tracking technology at home and abroad is further introduced.Subsequently,a brief overview of the relevant basic theories of neural networks is carried out,leading to the siamese neural network structure used in this article and its advantages over traditional algorithms in solving tracking tasks.At the same time,the overall framework of target tracking is introduced and the corresponding data set is derived.The evaluation indicators are introduced in detail so that the follow-up tracking algorithms can be compared clearly and accurately.Secondly,on the basis of Siam FC algorithm,a deeper siamese network target tracking algorithm based on multi-level feature fusion is designed.The algorithm explores the influence of various parameters of deep backbone network on tracking performance,and modifies Res Net50,which has stronger feature extraction ability.To some extent,it solves the problem that the padding in the deep network will make the tracking algorithm learn the center position deviation in the training process.Then two fusion methods of high-level semantic and low-level spatial information are adopted and verified respectively.At the same time,attention mechanism is introduced to achieve the lowest degree of goal adaptation.The algorithm is tested on OTB2013,OTB2015,VOT2016,UAV123 and other tracking benchmarks,and the corresponding indicators are evaluated.The results show that the proposed algorithm is effective.Finally,on the basis of Siam RPN algorithm,a siamese network real-time target tracking algorithm based on Distance-Io U loss is designed.The main idea of the algorithm design is based on the more accurate boundary box regression of Distance-Io U loss,and combined with the advantage of multi-layer feature fusion to construct the siamese network with stronger representation ability.The algorithm is tested on OTB2015,VOT2016,VOT2018,UAV123 and other tracking benchmarks,and the corresponding indicators are evaluated.It is proved that the algorithm can ensure the tracking accuracy on the premise of fully meeting the real-time tracking requirements.