Research On Object Tracking Based On Deep Feature Matching In Complex Scenes

As one of the important research directions in the field of image processing and computer vision,visual object tracking has high application value in the fields such as intelligent transportation,video surveillance,visual navigation,national defense investigation and military observation.Although object tracking has achieved obvious development in the previous decades of research,it still faces many difficulties.On the one hand,in the tracking process,the object often undergoes various complicated changes,such as illumination variation,deformation,occlusion,etc.,which make it particularly difficult to track the object stably and accurately.On the other hand,how to comprehensively and effectively improve the appearance model,the object location method and the template update strategy has always troubled many scholars at home and abroad.In recent years,object tracking based on deep learning shows better performance than other algorithms.Driven by the big data and the end-to-end learning,it can not only train the model efficiently and conveniently,but also greatly improve the robustness and accuracy of the algorithm.As one of the most representative directions,object tracking based on deep feature matching has attracted much attention for its balanced accuracy,robustness and real-time performance.Although this type of algorithm performs well in many aspects,its robustness and accuracy in complex scenes will be significantly reduced due to the relatively simple network structure,the insufficient template update strategy,and the lack of re-detection mechanism.Therefore,based on the research of a large number of object tracking algorithms at home and abroad,the dissertation makes certain improvements to the limitations of the object tracking based on deep feature matching in complex scenes.The dissertation focuses on the research of visible light single object tracking algorithm.The main contributions of the dissertation are listed as follows:1.To improve the accuracy and the robustness of the object tracking based on deep feature matching in complex scenes,a Siamese network combining the re-detection mechanism and the adaptive template updating method is proposed.In the tracking process,when similar interference appears in the background,the response map generated by the Siamese network is prone to appear multiple peaks,which may lead to the tracking failure.In this dissertation,an efficient and accurate Siamese convolutional neural network is utilized as the re-detection mechanism to deal with the multiple peaks.The re-detection network can sample around each peak in the response map,and perform accurate matching calculations.Thus,the elimination of the interference can be realized.In addition,in order to deal with the influence of occlusion,deformation and other complex factors in the tracking process,this dissertation proposes an adaptive template updating method based on the generative model.In this method,the evaluated reliable tracking result is utilized to extract the features and perform the probability statistics.The statistical results are combined with the existing template features adaptively,so as to realize the adaptive updating of the template.The simulation results demonstrate that the proposed algorithm can effectively deal with the challenging problems and improve the tracking performance in complex scenes.2.In most object tracking algorithms based on deep feature matching,the structure of the network is shallow and the convolutional features are not fully utilized.Although most trackers can achieve excellent tracking performance in some simple scenes,they can not perform well in some complex scenes,which have many challenging problems(Occlusion,Deformation,etc.).In this dissertation,an improved Res Net network is used to replace the traditional shallow network,and the deep features with more semantic information can be extracted.At the same time,an adaptive multi-layer feature fusion strategy is adopted to improve the quality of the response map.Thus,the algorithm can significantly reduce the impact of the distractors in complex scenes.In addition,an adaptive feature information fusion is adopted to update the template,so that the algorithm can adapt to various changes of the target appearance.The simulation results show that the proposed algorithm can effectively improve the tracking performance and perform favorably in both robustness and accuracy.3.In most object tracking algorithms based on deep feature matching,the foreground information is not fully utilized and the tracking performance decreases significantly in the semantic backgrounds.For effectively discriminating the object from the semantic backgrounds,a Siamese tracker combining foreground information guidance is proposed in this dissertation.Considering the tracking algorithms based on deep feature matching have limited challenging factors in the positive pairs,an effective sampling strategy is utilized to enrich the factors in the offline training phase.In addition,in order to further improve the discriminating ability of the tracker,the foreground information is highlighted by padding the background.At the same time,a guidance branch is combined to extract the features of foreground and a novel padding loss is adopted to guide the tracker.Furthermore,an improved feature information fusion is used to update the template,so that the algorithm can adapt to various changes of the target appearance.The simulation results demonstrate that the proposed algorithm can effectively deal with the challenging problems and improve the discriminating ability in semantic backgrounds.