Research On Siamese Trackers Based On Attention Mechanism

In the field of computer vision,visual object tracking has important application value in actual production and life,attracting a large number of scholars to continue to work in this field.With the development of deep learning and the generation of large amounts of data,the development path of visual object tracking is gradually divided into object tracking based on traditional machine learning algorithms and deep learning.Despite the many excellent algorithms are proposed,the current visual object tracking is still facing many challenges,such as sheltering and complex background,and scale changes,in this thesis,based on the Siamese online tracking framework,from the perspective of the feature extraction and feature fusion module of object tracking algorithm,is improved,and the thesis main work is as follows:(1)The Siamese network object tracking algorithm determines whether it is the same object according to the similarity between images.When similar semantic information appears around the object,different channels in the feature map of the residual network contribute different degrees to the tracking task.Based on this,the channel attention mechanism is introduced in the feature extraction module in this thesis.The attention residual network is constructed to improve the quality of features by assigning corresponding weights to different feature channels.Experimental results show that the improved algorithm improves the tracking success rate of the model under background clutters.(2)Features at different levels in the deep neural network contain different information,and multi-layer feature fusion is helpful to improve the tracking accuracy of the algorithm.However for Siamese online tracking algorithm for multilayer features usually adopt the method of linear weighted fusion,the fusion methods for different scale information directly together,can cause the loss of information,according to the problem in this thesis,the attention mechanism of the two fusion method and multiple input fusion method,the residual network last three-phase features of fusion,two methods through the local focus and global attention mechanism to obtain the passage between the different layers attention weight,avoid the semantic inconsistency multi-layer fusion,which further solves the inconsistent on the scale information,enables the model to capture the key features of different layers,ascending scale change and rotation under the challenge of tracking effect.Based on the above improved method,experimental analysis is carried out in OTB50,OTB100,UAV123,and VOT data sets respectively.The results show that the improved method proposed in this thesis can improve the robustness and tracking accuracy of the tracking algorithm.