Research On Target Tracking Algorithm Based On Sharing Backbone Network

Computer vision is an important branch of research in the computer field,and it has always attracted much attention from the scholars and the experts.Among them,the single-target tracking field in machine vision has always been a research hotspot due to its wide application in practice.The target tracking task faces problems such as target occlusion,size change,motion blur,and long-term tracking due to the uncertainty of the target itself and the complexity of the tracking environment.Due to the complexity of its application scenarios,this requires the tracking algorithm to have strong environmental adaptability that good robustness.Aiming at the problems faced by target tracking,this paper proposes a series of solutions to adapt to different target tracking tasks.Through a lot of research work,this paper proposes improved schemes based on sharing backbone network framework.The main research work and achievement of this paper are as follows:(1)In the Siamese network,the template is not updated in the process of online tracking and the edge pixels are filled,which brings about the adaptation of the appearance change of the target.This paper proposes the use of deep features extraction network,and proposes the learning module of the appearance reconstruction of the target and the classification matching loss function which integrates KL loss function and cross entropy loss function.In order to meet the needs of tracking task in complex environment,the Siamese network has been improved from three aspects: feature extraction network,attention mechanism and loss function.(2)The deep and shallow features of deep neural networks have different importance that deep features have strong semantic information,and shallow features have strong spatial location information.This paper proposes a multi-layer features fusion method for these features to apply semantic and spatial information.In addition,the shallow features have a smaller receptive field,and the deep features have a larger receptive field,the information of the tracked small objects is weakened in the deep features,and it is difficult to distinguish small objects from background information.Therefore,this paper proposes to use multi-layer features to determine the final target's location and semantic information.(3)On the basis of the improved object classification module,in order to make the positioning more accurate,this paper proposes to use an accurately positioning loss function to more accurately locate and track the position of the object.In online tracking,when locating the information of the tracking object,the accuracy of the positioning is mainly two points: the center point and the size.Because the traditional IoU(Intersection of Union)cannot reflects the alignment ways and center distance of the rectangular frame,it will cause a positioning offset during positioning.To solve this problem,this paper proposes a distance measurable intersection and union ratio loss function with alignment and center distance to guide the positioning model.In summary,the multi-layer features fusion decision proposed in this paper,the center distance and alignment ways between the integrated rectangular boxes,and the reconstruction of the target through features of the template layer of the Siamese network can automatically adapt to target tracking in different scenarios.In this paper,a large number of experiments have verified the effectiveness of the proposed method.