Realistic Target Tracking With Multi-Template Evaluation And Selection

Video target tracking based on computer vision is the essential technology in the fields of intelligent video surveillance,human-computer intelligent interaction and robot vision navigation.In recent years,with the development of big data analysis technology,video acquisition technology and high-performance computer,and the increasing need of video intelligence analysis,the research of target tracking algorithms has become one of the hottest spot of computer vision research.The difficulty of target tracking is that we only have true value of the first frame image but needs to achieve accurate tracking result under complex environment and target state changes,while has real-time performance.In this thesis,aiming at study the target tracking problem under complex conditions,apply the feedback mechanism to the tracking model with the advantages of the depth feature,discuss the tracking model structure and the tracking result evaluation method.Here in this thesis proposed a template evaluation and selection tracking algorithm in the Siamese network in order to improve tracking performance.The main research contents and results are as follows:(1)Traditional algorithms often study tracking as a whole.This thesis starts from the single target tracking process,and models the tracking process in complex scenes.The target tracking is divided into four stages as motion model building,image feature extraction appearance model and tracking model update.The main tasks and different methods of each stage are compared and analyzed in this thesis,which lays a foundation for the subsequent research of tracking.(2)Different from other algorithms directly using features,this thesis starts from the feature expression of the target in the tracking task,visualizes the artificial features and depth features,and studies the manual features and different depth convolution features for tracking performance including accuracy and speed.Based on the STRCF and KCF frameworks,qualitative and quantitative comparisons of HOG features and different depth convolution features in complex scenes such as scale and illumination variation are performed on the OTB50 dataset.The higher layer convolution feature is more suitable for target tracking.(3)The current tracking algorithms are based on the forward propagation mechanism.After the tracking drift occurs,the model correction cannot be realized.From the aspect of tracking model updating,study contradiction of the tracking algorithm based on depth features between real-time tracking and computational complexity under complex conditions.Established a multi-template evaluation and selection tracking framework based on Siamese network,through the evaluation of template selection,supervising and correcting the template update process,forming feedback adjustment,using the feedback mechanism in the update,constantly.The changing target feature can be introduced into the tracking model through template update,which replaces the update of the feature extractor.On the OTB-50 dataset,the tracking accuracy is improved by 7.025%and the success rate is improved by 8.061%compared with the Baseline in this thesis.The effectiveness of the multi-template evaluation and selection framework is also presented.This thesis also conducts quantitative comparison experiments and analysis on target scale estimation,model update frequency,template set size and feature usage.Provide experimental basis for the use of the framework of this article The model tracking speed reaches 10-25FPS in i7-4790k,NVIDA GeForce GTX970 configuration.