| Target tracking and detection both meet the rapid progress because of the application of deep learning.The existing target tracking algorithms show a gratifying trend in speed and accuracy,so that they can maintain good performance in most cases in the face of complex scene environment.Recently,many excellent siamese networks have emerged,which achieve a balance of performance and efficiency.Although target tracking has achieved breakthrough results in various fields,the existing algorithms still face great challenges due to the uncertainty of the target object itself and the variability of the scene.The main contribution of this paper is to propose a siamese network tracking algorithm based on attention mechanism and task-aware.Aiming at the problems of traditional two-stage siamese trackers,this paper mainly puts forward three innovations to optimize it:(1)In view of the neglect of background context information in the feature extraction stage,the image frame feature information has not been fully used.This paper proposes a space-frequency attention module,which aims to fully mine the apparent features of the target image,obtain a more compact feature representation of the intra class distance.Through experiments on major data sets,the improved algorithm effectively improves the robustness in complex scene.(2)In view of the positioning offset of the tracking network detection frame when the target rotates in the tracking process,this paper proposes the rotating target anchor frame learning strategy,dynamically calculates the horizontal anchor frame through the newly defined comprehensive index of matching degree,so as to realize the positioning accuracy of the anchor frame,allocate the anchor frame more effectively,and dynamically select the high-quality anchor frame,achieve accurate target positioning and alleviate the differences between subsequent classification and regression.Experiments show that the proposed anchor box learning strategy significantly improves the comprehensive attributes of the algorithm in the face of large data sets with rotating targets.(3)Aiming at the differences of classification and regression tasks in space tasks,this paper verifies the task-aware perception of target classification and location regression,obtains the accurate candidate boxes suitable for classification and regression,and obtains the identification score and location fine adjustment of candidate targets.In addition,aiming at the conflict of classification regression calculation in training and testing tasks,as well as the quantity and object deviation in the calculation of positive and negative samples in classification regression tasks,this paper uses GFocal loss to optimize the loss function to solve the above problems.Experiments show that the proposed algorithm has achieved great performance improvement in effectiveness,reliability and expected average overlap rate,and meets the requirements of real-time tracking. |
