| As one of the fundamental but challenging tasks in computer vision,visual object tracking has been widely used in various industries in recent years.Given limited information in the initial frame,the tracker aims to predict the position of target object frame by frame.The mainstream tracking methods do not consider the low correlation between classification confidence and localization accuracy as well as the limitation of cross correlation operation itself,which can reduce the localization accuracy of the tracker.To improve the localization accuracy of UAV remote sensing tracker,this paper improves the traditional tracker based on Siamese network and Transformer architecture to make it more suitable for UAV object tracking,the main innovations of this paper are as follows:Firstly,this paper proposes an improved Siamese classification and regression adaptive network for UAV remote sensing object tracking,named ISiam CRAN.This method aims to solve the degradation of localization accuracy caused by the low correlation between classification confidence and localization accuracy.ISiam CRAN applies modified Res Net-50 as the backbone to extract depth features.Then,these extracted features are fed into an improved classification regression adaptive head network for depth feature cross correlation operation.Unlike existing trackers,to remove low-quality prediction bounding boxes,we integrate a quality assessment branch to the improved classification regression adaptive head network and improve localization accuracy.Moreover,ISiam CRAN uses an improved elliptical sample label assignment strategy to replace the traditional method.This way,our tracker can more accurately distinguish the foreground and background and further improves the location accuracy.Finally,ISiam CRAN predicts the position and scale of the target directly in a unified fully convolutional network by an anchor-free manner based on the similarity feature maps.The experimental results show that ISiam CRAN can improve localization accuracy during UAV remote sensing object tracking.Secondly,this paper proposes an improved spatio-temporal hierarchical feature Transformer for UAV remote sensing object tracking,named STHFT.STHFT improves the cross correlation operation used in the feature fusion phase of ISiam CRAN.ISiam CRAN calculates the similarity between template and search region through a simple cross correlation operation and then classifies and regresses the target object based on the similarity feature maps.However,the correlation operation itself is a local linear matching process.Its limitation often causes the loss of semantic information and falls into local optimal solutions easily,which may be the bottleneck of designing highaccuracy tracking algorithms.Inspired by Transformer,this paper uses an improved spatio-temporal hierarchical feature Transformer to achieve the interactive fusion of shallow spatio-temporal and deep semantic information and establishes associations between distant features.In addition,STHFT uses a corner-based bounding box prediction head network instead of the traditional three-layer perceptron structure to estimate the corners of the target and localize it directly,improving the quality of the predicted bounding box.Moreover,the spatio-temporal information of the target is fully exploited by dynamic update templates.Unlike traditional tracking methods,STHFT does not require post-processing steps(such as cosine window and bounding box smoothing)and generates a unique predicted bounding box directly,largely simplifying the tracking process and reducing the computational effort.The verification on the UAV benchmark shows that STHFT can achieve stable tracking of UAV targets. |
