Research On Object Tracking Algorithm In Augmented Reality

As the continuous development of modern science and technology,augmented reality technology has been widely used and greatly improved.As the basis for the realization of augmented reality,the importance of object tracking is self-evident.It can be said that the quality of object tracking directly affects the experience of the augmented reality system.Therefore,improving the accuracy and robustness of object tracking has strong research significance and value for the application of augmented reality.In this thesis,the siamese network tracking algorithm Siam FC is used as the basis for further research.The main research contents are as follows:Aiming at the problem that the object tracking algorithm Siam FC based on siamese networks has poor feature extraction ability,and it is easy to lose the object during the tracking process,we propose a object tracking algorithm based on hard samples and mixed attention mechanism in this thesis.First of all,the edge information of the object image is used as the area with insufficient training sample size to fill in,which makes the sample image add a lot of interference information around the object to increase the difficulty of tracking to a greater extent.Meanwhile,the loss function is improved to balance the loss redundancy caused by the different ratio of positive and negative samples in the training process.Secondly,in order to enrich the information representation of the object,the residual connection is used to combine the feature maps of different stages.Then,in order to enhance the ability of the neural network to extract the main features of the object,a mixed attention mechanism is added to the search branch to highlight the useful features.At last,for the purpose of ensuring that the model tracking speed meets the real-time requirements,we use deep separable convolution to replace the traditional convolution in the network structure,which can reduce the computation amount of tracking algorithm and the parameter number of network model.Aiming at the problem that the object tracking algorithm Siam FC does not take the updating of template into account in the tracking process,which leads to the poor adaptability of the algorithm to the deformation of the object in the tracking process,a object tracking algorithm based on multi-feature fusion and template updating is proposed in this thesis.First of all,due to the strict translation invariance requirement of the siamese network,the traditional siamese network generally uses Alex Net without filling technology as the backbone network,which limits the feature extraction ability.Therefore,we adopts the modified and optimized deep neural network as the backbone network in this thesis.Then,in order to extract the object semantic information while preserving the object location information and contour information,feature pyramid is introduced to fully integrate the object features of each stage.Finally,regular update and average peak correlation energy are used to judge.When the object does not appear occlude or disappear,feature update is carried out to enhance the discrimination ability of the algorithm to deal with the deformation of the object.We test and compare our proposed algorithm with the mainstream tracking algorithm on authoritative datasets.The consequence indicates that the proposed algorithm is significantly improved contrasted with the benchmark algorithm Siam FC,and has strong competitiveness in the overall performance compared with other algorithms,and performs well in the actual tracking visualization stage.Finally,feature detection and matching are performed on the registration object to complete the registration,which shows the effectiveness of the algorithm.