Occlusion And Rotation Robust Target Tracking Algorithm

As an significant branch of computer application,visual target tracking technology has always been a hot research direction.Target tracking technology can continuously track the target in video without human intervention,and it has been widely used in intelligent monitoring,aerial navigation,military security and other fields.With the increasing demand of the market,the application scenarios of the algorithm become more and more complex.Occlusion,disappearance,rotation,deformation,long time tracking becomes the obstacle that the tracking algorithm needs to overcome.This dissertation aims at promoting the real-time target tracking algorithm and enhancing the tracking stability.In view of the deficiency of the traditional Correlation Filter(CF)algorithm in the process of long-term target tracking and target rotation tracking,two improved algorithms are proposed.An experiment is designed to verify the effect of the improved algorithm.The main work contents and innovations are as follows:1.There are two main reasons why it is difficult to solve the problem of long-term target tracking.Firstly,it is difficult to determine the updating time and learning rate of the model.Secondly,after the target disappears,a retracking strategy should be introduced to search the target for the subsequent frames.In order to solve these two problems,an improved long time tracking algorithm is proposed.The algorithm introduces the evaluation index to evaluate the confidence of the tracking results of the algorithm,and dynamically determines the update time,update weight and the subsequent algorithm flow according to the evaluation results.The SVM model which is used to judge whether the inter-frame variation is reasonable is pre-trained.When the evaluation result is not good,the SVM model is used to further judge the target.The evaluation value and classification results of SVM model are used to judge whether the target disappears.Once it is determined that the target has disappeared,the SVM classification model will be enabled to scan the candidate regions to find the target until the target is found again.Finally,the improved algorithm is compared with other algorithms in the data set of VOT and otb-100.The experimental results show that the success rate and accuracy of the modified algorithm are improved.2.In order to cope with the situation that the tracking algorithm based on traditional features is unstable when the target is deformed and rotated,and combining with the deep features with strong expressive ability,this dissertation proposes a CF tracking algorithm that integrates the deep features and the traditional features.The algorithm architecture can be ed into a Total-Sub-Total model.Considering the limited computing power and storage capacity of the embedded platform,the traditional features are preferred to track the target.Once the confidence value is low,the algorithm will stop updating the traditional feature model in time.Then go back to the previous frame or frames,and use the deep feature modeling.After that,the algorithm is divided into the tracking branch based on traditional features and the tracking branch based on deep features,until the tracking results of the two branches are similar for several consecutive frames and the two branch algorithms are combined.Considering the computation amount of deep network,the tracking algorithm based on deep feature can perform frame skipping tracking.The experimental results show that the accuracy and success rate of the new algorithm are improved to some extent.