Research On Object Tracking Of Siamese Neural Networks Based On Long Short-term Memory

Target tracking technology is a research hotspot in the field of computer vision.It has important research significance and has broad application prospects in many fields such as video surveillance and driverless driving.Although the object tracking technology has made great progress,due to the complexity and variety of tracking tasks,there are many factors that affect the performance of the tracking algorithm such as occlusion,background interference and appearance changes.Designing a highly accurate,real-time,and robust tracking algorithm remains a huge challenge.In recent years,the tracking algorithm based on siamese neural network has developed rapidly,and has achieved good results in VOT and other challenges.These algorithms have the advantage of real-time performance,but their generalization performance is poor and the robustness is not strong.The tracking performance of the algorithm is degraded when there is a large change in the appearance of the target.Aiming at the above problems,this thesis combines the long short-term memory network(LSTM)on the basis of the siamese neural network,and focuses on the feature learning,similarity measurement,model update,and network training of the tracking algorithm.The main works can be summarized as follows:(1)A siamese neural network tracking algorithm based on robust feature representation is proposed(SMT).The algorithm uses the similarity measure to match the target and the candidate region,and adds a region proposal layer on the basis of the siamese network to generate the candidate region.The convolutional layer uses a VGG network with stronger feature extraction capabilities,and uses hierarchical features to give the target a rich feature representation.Then add a long short-term memory layer for tracking,and introduce a confidence decision method when the template is updated.The role of the long short-term memory layer is to store and update feature information through its internal structure memory unit and gate mechanism,making the obtained feature vector more robust.The experimental results show that the proposed algorithm has better robustness and accuracy than other eight representative tracking algorithms under the attributes of fast motion,background interference,motion blur,rotation change and scale change.(2)A real-time tracking algorithm of siamese neural networks based on joint appearance and motion information is proposed(SMT-R).The network structure includes a convolution layer containing target appearance information,a long short-term memory layer that stores and updates motion information,and a regression layer that outputs target position coordinates.The algorithm extracts the target appearance information in the convolutional layer,and merges the motion information into the network model,so that the tracking algorithm can modify the target model and adapt to the new object.In long short-term memory networks,the memory unit is updated in the process of forward propagation,so that the network can continuously accept the current information and perform rapid information updates.Therefore,the movement feature of the target can be learned and updated through the long short-term memory network during the tracking process,avoiding back propagation during the tracking process,reducing the calculation amount during the tracking process,thereby improving the running speed,and making it applicable to real life scenarios.Experiments show that the proposed algorithm can run at real-time speed and has good tracking accuracy.