Research On Siamese Network Based Robust Visual Tracking Algorithms

Visual object tracking has always been a basic and active research topic in the field of computer vision.It has a wide range of applications in video surveillance,automatic driving,human-computer interaction and unmanned aerial vehicle.Benefiting from the development of deep learning and big data,as well as the application of convolutional neural networks in the field of object tracking,object tracking algorithms have been significantly developed in the past few years.It’s worth noting that the tracking method based on Siamese network,which usually strikes a good balance between accuracy and speed,has attracted widespread attention.The Siamese tracking methods are first trained offline based on a large dataset.The types of trackers can be divided into online trackers and offline trackers according to whether the model is updated during tracking.The former has better accuracy performance,and the latter is faster for inference.At present,the Siamese tracking method has achieved gratifying achievements in the design of the backbone,matching approach,output representation,etc.However,the efficient solution for dealing with complex scenes such as rapid deformation,distractor interference,and background clutter is still insufficient.In order to resolve the above dilemma,this thesis focuses on two aspects,i.e.,how to effectively integrate online learning and offline learning to improve the robustness of the model,and how to learn better feature representation of the Siamese network.The two works are mainly summarized as follows.(1)A novel object tracking method is proposed by developing an adaptive multi-strategy-based collaborative framework to automatically switch between the online learning tracker and the offline learning tracker during tracking.On the one hand,the online learning tracker and the offline learning Siamese tracker are treated to be the primary and auxiliary trackers,respectively,and combined to track within a frame for improving the accuracy.On the other hand,the tracking speed can be faster by using the offline learning Siamese tracker alone.To make the offline learning Siamese tracker more robust to the target appearance changes,this thesis introduces the memory unit and designs the template update and selection module,which provides the first template,the latest template or the intermediate template(s)for matching.Therefore,considering the tracking speed and robustness,we design a variable action set module,which represents the tracking and template strategies mentioned above with multiple action labels.In addition,an online reliability evaluation module is employed to further evaluate the tracking result,the template and the tracker,so as to update the memory unit and the action set.Finally,a Siamese tracking method based on multi-template matching is implemented,which achieves more robust performance with rich target information.(2)A feature refinement method based on interaction and aggregation is proposed for Siamese tracking,which enhances the salient features of the target and suppresses background noise.This method first analyzes the existing problems of feature representation approaches in the current Siamese framework,and proposes a gated context aggregation scheme,where the local attention and the cross-layer non-local attention are fused adaptively.Especially,the proposed cross-layer non-local attention can integrate the context information of different receptive fields,which plays an important role in the preservation and highlighting of feature details.Since the non-local attention is hard to perceive the target directly,this method uses the refined target features as prior knowledge to guide the self-attention for the search area.In addition,a graph construction method based on correlation is proposed.The refined template feature and search area feature are divided along the channel and space dimensions,and then interacted in the manner of point-to-point.Finally,the spatial neighborhood correlation can be strengthened by fedding the correlation map into the graph convolution network,so that the background noise can be suppressed through fusing the refined graph with the original image.The ablation analysis,quantitative evaluation and qualitative evaluation of the above two methods are conducted on OTB2015,VOT2016,La SOT and UAV20 L.Various experiments demonstrate that the proposed method has superior performance.