Research On Algorithms Of Visual Target Tracking In Remote Sensing Images

With the advancement of remote sensing UAV(unmanned aerial vehicle)and video satellite technology,the combination of artificial intelligence and remote sensing video has gained extensive application.The target tracking technology based on remote sensing video is one of them.Remote sensing target tracking plays a huge role in military,civilian,scientific research,and other aspects,and the requirements for remote sensing target tracking are becoming increasingly strict.Traditional target tracking algorithms are no longer able to meet the requirements of high accuracy and high tracking speed.Therefore,this article focuses on the research of remote sensing target tracking algorithms,improving single target tracking method based on correlation filtering and multi target tracking method based on deep learning,making them more suitable for remote sensing target tracking.The main research work of this thesis is as follows:(1)In this study,the ECO-HC method,a manual feature version of the classic correlation filtering method ECO,is selected as the baseline method to investigate the application of single target tracking algorithm based on correlation filtering on remote sensing images.To address the limitations of ECO-HC,a joint feature expression method based on the combination of HOG and color features is proposed,along with an adaptive model update strategy to update the tracking model at intervals.The proposed algorithm is evaluated on the UAV-123 aerial remote sensing dataset,and the results show that the improved algorithm has an average success rate and accuracy rate increase of 1.5% and 2.2%,respectively,compared to the baseline method.Notably,the success rate improves by 3.2% when the target moves quickly,the success rate increases by 4% in scenes where the target leaves the field of view,and the accuracy rate increases by 2.4% when the image resolution is low.(2)According to the characteristics of small target size and complex imaging background in satellite remote sensing images,the YOLOv5 detection algorithm combined with the CBAM attention mechanism is designed.For the Io U in the YOLOv5 loss function,an improved GIo U is designed to improve model training efficiency and detection performance.Using YOLOv5 s as the backbone network to design an anchor frame that is conducive to the detection of tiny targets,and training the model on the high-resolution optical satellite remote sensing dataset AIR-MOT,this effectively improves the detection ability of YOLOv5 for small remote sensing targets.The improved YOLOv5 will serve as Deep SORT’s detector.(3)With DeepSORT as the baseline method for multi-target tracking,an improved cascade matching method suitable for remote sensing images is designed to reduce target ID conversion caused by target detection errors and Kalman filter prediction errors,and improve Deep SORT due to detection frame confirmation status.For the problem of no tracking trajectory in the first two frames,the improved YOLOv5 detection algorithm is used to replace Faster R-CNN to improve the target detection ability,and the improved YOLOv5 combined with the attention mechanism is used as the detector for experimental verification.(4)Design a single target tracking experiment based on the aerial remote sensing data set UAV-123,and use the ECO-HC,which applies the adaptive model update strategy based on correlation filtering proposed in this thesis,to compare with the original algorithm and other tracking algorithms of the same type.The results demonstrate the effectiveness of the adaptive model update strategy.A multi-target detection and tracking experiment based on YOLOv5 detection Deep SORT data association was designed on the remote sensing dataset AIR-MOT.Compared with the detection results before and after improvement,the detection accuracy of the algorithm in this thesis increased by 11.2% and 12.5% respectively for aircraft and ships.The average detection accuracy increased by 11.8%,the average m AP@5 increased by12.9%,and the m AP@.5:.95 increased by 5.9%.Compared with the improved detector and the cascaded matching algorithm in this thesis,the multi-target tracking accuracy MOTA in this thesis has increased by 21.6%,and the multi-target tracking accuracy MOTP has increased by 1.4%.