Technology And Application Of Visual Object Tracking Based On Deep Learning

As one of the basic problems in the field of computer vision,visual object tracking has broad application prospects in many aspects such as intelligent monitoring,autonomous driving,human-computer interaction,medical diagnosis,and robot navigation.Therefore,the research on visual object tracking technology is of important practical significance,which is also a hot topic researched by many scholars.It has made great progress in the past ten years,especially in recent years,the continuous development of deep learning-based visual object tracking methods has achieved surprising performance improvements and promoted the progress in this field.However,due to the continuous emergence of various complex scenarios and challenging factors in the actual tracking process,the object tracking method still has some problems in accuracy and robustness.This dissertation focuses on the basis of the Siamese network framework,and proposes three visual object tracking methods.The main contents of the three methods are described as follows:(1)A object tracking method based on Siamese progressive attention fusion network is proposed.For the most of object tracking algorithms using Siamese networks,the semantic feature derived from the last layer of the backbone network is used to calculate the similarity.However,the use of single deep feature space often leads to partial loss of effective information.To address this issue,the Siamese progressive attention fusion network is proposed.First,the deep and shallow feature information is simultaneously extracted using the backbone network.Second,a top-down strategy is adopted to gradually encode and fuse deep semantic information,as well as shallow spatial structure information is obtained from the progressive feature aggregation module.We then use attention module to reduce feature redundancy that generated by fusion.Last,the optimal solution of object tracking is formed by calculating the similarity between the target and search area.By means of attention module and progressive feature aggregation module,the tracker can enhance the performance of the applications.(2)A Siamese network tracking method with dual-branch synergy is proposed.For the most of object tracking algorithms using Siamese networks,the characteristics of the target template and search image rely on the two branches of the Siamese network to perform independent calculations.However,it is found that the information of each branch of the Siamese network is also more effective for the other branch by research.This dissertation designs a dual-branch synergy mechanism based on the Siamese network architecture,which can effectively share and merge the information between the target template and search image.In addition,a multi-layer aggregation strategy is designed to encode complementary information between high-level semantic features and low-level detailed features to enhance shallow feature representation.And the attention module is introduced to refine the features,which improves the performance of the proposed tracking method significantly.(3)A Siamese network tracking method is proposed which is based on cross-layer non-local fusion.Although the multi-layer fusion strategy can achieve better tracking results,it simply uses the cascading method to aggregate deep and shallow features,which will inevitably lead to feature redundancy in spite of the latter added attention module where features are refined and enhanced.This dissertation proposes a Siamese network tracking method which is based on cross-layer non-local fusion and a more effective feature fusion method is adopted.The cross-layer non-local fusion method can capture remote dependencies from two different levels of feature representations to adaptively aggregate features.In addition,a deformable module is introduced to improve the modeling ability of the Siamese network for target deformation so that the feature representation is improved.