Research On Target Tracking Based On Siamese Region Proposal Network Framework

Target tracking has a very broad application background in the field of computer vision,whose task is to give the initial state and appearance characteristics of the target in the initial frame of the video sequence,and then predict the state of the target in the subsequent frame images and mark its position and shape.Target tracking is a challenging research direction,some problems such as occlusion,scale change,fast motion can not be solved reliably,and the real-time performance and tracking accuracy of model and algorithm are still difficult to take into account,so further research on tracking model is still imperative.Since the appearance of Fully-Connected Siamese Network(SiamFC),the target tracking model based on Siamese network has entered the climax of research,meeting the real-time requirements that other model structures can not do.Although Siamese Region Proposal Network(SiamRPN)based networks have made great progress in target location and scale prediction,feature fusion mechanism still can not meet the general needs of the model,and online tracking and updating mechanisms are still in absent,which can not show very robust performance undet drastic target appearance changes.In view of the shortcomings of the existing twin network tracker,a new model based on the framework of MobileNet V2 is proposed in this study.The highlights are as follows:(1)The commonly adopted AlexNet feature extraction skeleton is replaced by a deeper and lightweight MobileNetv2 like structure,and depth-wise feature and channel-wise feature attention model are introduced.Deep feature attention fusion model and the RPN structure of channel feature attention are added on this basis.In general,the deeper network and richer depth feature information and more convincing channel correlation are used In addition,in order to increase the receptive field of deep features and prevent the stride and padding from damaging the important information of the image,the skeleton has been rectified based on original MobileNetV2,in which the stride and dilation rate were adjusted.(2)In order to enhance the general appearance of the tracker for specific targets and improve the adaptability to the targets with sharp appearance changes,a Siamese network structure with two templates is proposed,which is based on time in addition to depth and channel,to effectively integrate the tracking results of different templates and different depth features.In addition,this research optimizes the algorithm of long-term tracking and has proposed a way of template updating,sothat the tracker can better adapt to the appearance and state of target changes.In order to verify the performance of the proposed tracking model,a twin network framework based on feature attention and template attention is constructed.The model uses YouTube-BB to pretrain network parameters and keep the parameters fixed during testing processes.OTB100 is selected as the testing dataset.Results show that the proposed model and tracking algorithm are more robust than the depth network and correlation filtering model without reduction of the tracking speed.There are 19 figures,9 tables in this thesis,and 85 references are quoted.